Advanced Linear Algebra (Graduate Texts in Mathematics)

Graduate Texts in Mathematics 135 Editorial Board S. Axler K.A. Ribet Graduate Texts in Mathematics 1 TAKEUTI/ZARING...

Author: Steven Roman

21 downloads 1129 Views 3MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

Graduate Texts in Mathematics

135

Editorial Board S. Axler K.A. Ribet

Graduate Texts in Mathematics 1 TAKEUTI/ZARING. Introduction to Axiomatic Set Theory. 2nd ed. 2 OXTOBY. Measure and Category. 2nd ed. 3 SCHAEFER. Topological Vector Spaces. 2nd ed. 4 HILTON/STAMMBACH. A Course in Homological Algebra. 2nd ed. 5 MAC LANE. Categories for the Working Mathematician. 2nd ed. 6 HUGHES/PIPER. Projective Planes. 7 J.-P. SERRE. A Course in Arithmetic. 8 TAKEUTI/ZARING. Axiomatic Set Theory. 9 HUMPHREYS. Introduction to Lie Algebras and Representation Theory. 10 COHEN. A Course in Simple Homotopy Theory. 11 CONWAY. Functions of One Complex Variable I. 2nd ed. 12 BEALS. Advanced Mathematical Analysis. 13 ANDERSON/FULLER. Rings and Categories of Modules. 2nd ed. 14 GOLUBITSKY/GUILLEMIN. Stable Mappings and Their Singularities. 15 BERBERIAN. Lectures in Functional Analysis and Operator Theory. 16 WINTER. The Structure of Fields. 17 ROSENBLATT. Random Processes. 2nd ed. 18 HALMOS. Measure Theory. 19 HALMOS. A Hilbert Space Problem Book. 2nd ed. 20 HUSEMOLLER. Fibre Bundles. 3rd ed. 21 HUMPHREYS. Linear Algebraic Groups. 22 BARNES/MACK. An Algebraic Introduction to Mathematical Logic. 23 GREUB. Linear Algebra. 4th ed. 24 HOLMES. Geometric Functional Analysis and Its Applications. 25 HEWITT/STROMBERG. Real and Abstract Analysis. 26 MANES. Algebraic Theories. 27 KELLEY. General Topology. 28 ZARISKI/SAMUEL. Commutative Algebra. Vol. I. 29 ZARISKI/SAMUEL. Commutative Algebra. Vol. II. 30 JACOBSON. Lectures in Abstract Algebra I. Basic Concepts. 31 JACOBSON. Lectures in Abstract Algebra II. Linear Algebra. 32 JACOBSON. Lectures in Abstract Algebra III. Theory of Fields and Galois Theory. 33 HIRSCH. Differential Topology. 34 SPITZER. Principles of Random Walk. 2nd ed. 35 ALEXANDER/WERMER. Several Complex Variables and Banach Algebras. 3rd ed. 36 KELLEY/NAMIOKA et al. Linear Topological Spaces. 37 MONK. Mathematical Logic.

38 GRAUERT/FRITZSCHE. Several Complex Variables. 39 ARVESON. An Invitation to C∗ -Algebras. 40 KEMENY/SNELL/KNAPP. Denumerable Markov Chains. 2nd ed. 41 APOSTOL. Modular Functions and Dirichlet Series in Number Theory. 2nd ed. 42 J.-P. SERRE. Linear Representations of Finite Groups. 43 GILLMAN/JERISON. Rings of Continuous Functions. 44 KENDIG. Elementary Algebraic Geometry. 45 LOÈVE. Probability Theory I. 4th ed. 46 LOÈVE. Probability Theory II. 4th ed. 47 MOISE. Geometric Topology in Dimensions 2 and 3. 48 SACHS/WU. General Relativity for Mathematicians. 49 GRUENBERG/WEIR. Linear Geometry. 2nd ed. 50 EDWARDS. Fermat’s Last Theorem. 51 KLINGENBERG. A Course in Differential Geometry. 52 HARTSHORNE. Algebraic Geometry. 53 MANIN. A Course in Mathematical Logic. 54 GRAVER/WATKINS. Combinatorics with Emphasis on the Theory of Graphs. 55 BROWN/PEARCY. Introduction to Operator Theory I: Elements of Functional Analysis. 56 MASSEY. Algebraic Topology: An Introduction. 57 CROWELL/FOX. Introduction to Knot Theory. 58 KOBLITZ. p-adic Numbers, p-adic Analysis, and Zeta-Functions. 2nd ed. 59 LANG. Cyclotomic Fields. 60 ARNOLD. Mathematical Methods in Classical Mechanics. 2nd ed. 61 WHITEHEAD. Elements of Homotopy Theory. 62 KARGAPOLOV/MERIZJAKOV. Fundamentals of the Theory of Groups. 63 BOLLOBAS. Graph Theory. 64 EDWARDS. Fourier Series. Vol. I. 2nd ed. 65 WELLS. Differential Analysis on Complex Manifolds. 3rd ed. 66 WATERHOUSE. Introduction to Affine Group Schemes. 67 SERRE. Local Fields. 68 WEIDMANN. Linear Operators in Hilbert Spaces. 69 LANG. Cyclotomic Fields II. 70 MASSEY. Singular Homology Theory. 71 FARKAS/KRA. Riemann Surfaces. 2nd ed. 72 STILLWELL. Classical Topology and Combinatorial Group Theory. 2nd ed. 73 HUNGERFORD. Algebra. 74 DAVENPORT. Multiplicative Number Theory. 3rd ed. 75 HOCHSCHILD. Basic Theory of Algebraic Groups and Lie Algebras. (continued after index)

Steven Roman

Advanced Linear Algebra Third Edition

Steven Roman 8 Night Star Irvine, CA 92603 USA [email protected] Editorial Board S. Axler Mathematics Department San Francisco State University San Francisco, CA 94132 USA [email protected]

ISBN-13: 978-0-387-72828-5

K.A. Ribet Mathematics Department University of California at Berkeley Berkeley, CA 94720-3840 USA [email protected]

e-ISBN-13: 978-0-387-72831-5

Library of Congress Control Number: 2007934001 Mathematics Subject Classification (2000): 15-01 c 2008 Springer Science+Business Media, LLC All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed on acid-free paper. 9 8 7 6 5 4 3 2 1 springer.com

To Donna and to Rashelle, Carol and Dan

Preface to the Third Edition

Let me begin by thanking the readers of the second edition for their many helpful comments and suggestions, with special thanks to Joe Kidd and Nam Trang. For the third edition, I have corrected all known errors, polished and refined some arguments (such as the discussion of reflexivity, the rational canonical form, best approximations and the definitions of tensor products) and upgraded some proofs that were originally done only for finite-dimensional/rank cases. I have also moved some of the material on projection operators to an earlier position in the text. A few new theorems have been added in this edition, including the spectral mapping theorem and a theorem to the effect that dim²= ³ dim²= i ³, with equality if and only if = is finite-dimensional. I have also added a new chapter on associative algebras that includes the wellknown characterizations of the finite-dimensional division algebras over the real field (a theorem of Frobenius) and over a finite field (Wedderburn's theorem). The reference section has been enlarged considerably, with over a hundred references to books on linear algebra. Steven Roman

Irvine, California, May 2007

Preface to the Second Edition

Let me begin by thanking the readers of the first edition for their many helpful comments and suggestions. The second edition represents a major change from the first edition. Indeed, one might say that it is a totally new book, with the exception of the general range of topics covered. The text has been completely rewritten. I hope that an additional 12 years and roughly 20 books worth of experience has enabled me to improve the quality of my exposition. Also, the exercise sets have been completely rewritten. The second edition contains two new chapters: a chapter on convexity, separation and positive solutions to linear systems (Chapter 15) and a chapter on the QR decomposition, singular values and pseudoinverses (Chapter 17). The treatments of tensor products and the umbral calculus have been greatly expanded and I have included discussions of determinants (in the chapter on tensor products), the complexification of a real vector space, Schur's theorem and Geršgorin disks. Steven Roman

Irvine, California February 2005

Preface to the First Edition

This book is a thorough introduction to linear algebra, for the graduate or advanced undergraduate student. Prerequisites are limited to a knowledge of the basic properties of matrices and determinants. However, since we cover the basics of vector spaces and linear transformations rather rapidly, a prior course in linear algebra (even at the sophomore level), along with a certain measure of “mathematical maturity,” is highly desirable. Chapter 0 contains a summary of certain topics in modern algebra that are required for the sequel. This chapter should be skimmed quickly and then used primarily as a reference. Chapters 1–3 contain a discussion of the basic properties of vector spaces and linear transformations. Chapter 4 is devoted to a discussion of modules, emphasizing a comparison between the properties of modules and those of vector spaces. Chapter 5 provides more on modules. The main goals of this chapter are to prove that any two bases of a free module have the same cardinality and to introduce Noetherian modules. However, the instructor may simply skim over this chapter, omitting all proofs. Chapter 6 is devoted to the theory of modules over a principal ideal domain, establishing the cyclic decomposition theorem for finitely generated modules. This theorem is the key to the structure theorems for finite-dimensional linear operators, discussed in Chapters 7 and 8. Chapter 9 is devoted to real and complex inner product spaces. The emphasis here is on the finite-dimensional case, in order to arrive as quickly as possible at the finite-dimensional spectral theorem for normal operators, in Chapter 10. However, we have endeavored to state as many results as is convenient for vector spaces of arbitrary dimension. The second part of the book consists of a collection of independent topics, with the one exception that Chapter 13 requires Chapter 12. Chapter 11 is on metric vector spaces, where we describe the structure of symplectic and orthogonal geometries over various base fields. Chapter 12 contains enough material on metric spaces to allow a unified treatment of topological issues for the basic

xii Preface Hilbert space theory of Chapter 13. The rather lengthy proof that every metric space can be embedded in its completion may be omitted. Chapter 14 contains a brief introduction to tensor products. In order to motivate the universal property of tensor products, without getting too involved in categorical terminology, we first treat both free vector spaces and the familiar direct sum, in a universal way. Chapter 15 (Chapter 16 in the second edition) is on affine geometry, emphasizing algebraic, rather than geometric, concepts. The final chapter provides an introduction to a relatively new subject, called the umbral calculus. This is an algebraic theory used to study certain types of polynomial functions that play an important role in applied mathematics. We give only a brief introduction to the subject c emphasizing the algebraic aspects, rather than the applications. This is the first time that this subject has appeared in a true textbook. One final comment. Unless otherwise mentioned, omission of a proof in the text is a tacit suggestion that the reader attempt to supply one. Steven Roman

Irvine, California

Contents

Preface to the Third Edition, vii Preface to the Second Edition, ix Preface to the First Edition, xi

Preliminaries, 1 Part 1: Preliminaries, 1 Part 2: Algebraic Structures, 17

Part I—Basic Linear Algebra, 33 1

Vector Spaces, 35 Vector Spaces, 35 Subspaces, 37 Direct Sums, 40 Spanning Sets and Linear Independence, 44 The Dimension of a Vector Space, 48 Ordered Bases and Coordinate Matrices, 51 The Row and Column Spaces of a Matrix, 52 The Complexification of a Real Vector Space, 53 Exercises, 55

2

Linear Transformations, 59 Linear Transformations, 59 The Kernel and Image of a Linear Transformation, 61 Isomorphisms, 62 The Rank Plus Nullity Theorem, 63 Linear Transformations from - to - , 64 Change of Basis Matrices, 65 The Matrix of a Linear Transformation, 66 Change of Bases for Linear Transformations, 68 Equivalence of Matrices, 68 Similarity of Matrices, 70 Similarity of Operators, 71 Invariant Subspaces and Reducing Pairs, 72 Projection Operators, 73

xiv

Contents Topological Vector Spaces, 79 Linear Operators on = d , 82 Exercises, 83

3

The Isomorphism Theorems, 87 Quotient Spaces, 87 The Universal Property of Quotients and the First Isomorphism Theorem, 90 Quotient Spaces, Complements and Codimension, 92 Additional Isomorphism Theorems, 93 Linear Functionals, 94 Dual Bases, 96 Reflexivity, 100 Annihilators, 101 Operator Adjoints, 104 Exercises, 106

4

Modules I: Basic Properties, 109 Motivation, 109 Modules, 109 Submodules, 111 Spanning Sets, 112 Linear Independence, 114 Torsion Elements, 115 Annihilators, 115 Free Modules, 116 Homomorphisms, 117 Quotient Modules, 117 The Correspondence and Isomorphism Theorems, 118 Direct Sums and Direct Summands, 119 Modules Are Not as Nice as Vector Spaces, 124 Exercises, 125

5

Modules II: Free and Noetherian Modules, 127 The Rank of a Free Module, 127 Free Modules and Epimorphisms, 132 Noetherian Modules, 132 The Hilbert Basis Theorem, 136 Exercises, 137

6

Modules over a Principal Ideal Domain, 139 Annihilators and Orders, 139 Cyclic Modules, 140 Free Modules over a Principal Ideal Domain, 142 Torsion-Free and Free Modules, 145 The Primary Cyclic Decomposition Theorem, 146 The Invariant Factor Decomposition, 156 Characterizing Cyclic Modules, 158

Contents Indecomposable Modules, 158 Exercises, 159

7

The Structure of a Linear Operator, 163 The Module Associated with a Linear Operator, 164 The Primary Cyclic Decomposition of = , 167 The Characteristic Polynomial, 170 Cyclic and Indecomposable Modules, 171 The Big Picture, 174 The Rational Canonical Form, 176 Exercises, 182

8

Eigenvalues and Eigenvectors, 185 Eigenvalues and Eigenvectors, 185 Geometric and Algebraic Multiplicities, 189 The Jordan Canonical Form, 190 Triangularizability and Schur's Theorem, 192 Diagonalizable Operators, 196 Exercises, 198

9

Real and Complex Inner Product Spaces, 205 Norm and Distance, 208 Isometries, 210 Orthogonality, 211 Orthogonal and Orthonormal Sets, 212 The Projection Theorem and Best Approximations, 219 The Riesz Representation Theorem, 221 Exercises, 223

10

Structure Theory for Normal Operators, 227 The Adjoint of a Linear Operator, 227 Orthogonal Projections, 231 Unitary Diagonalizability, 233 Normal Operators, 234 Special Types of Normal Operators, 238 Self-Adjoint Operators, 239 Unitary Operators and Isometries, 240 The Structure of Normal Operators, 245 Functional Calculus, 247 Positive Operators, 250 The Polar Decomposition of an Operator, 252 Exercises, 254

Part II—Topics, 257 11

Metric Vector Spaces: The Theory of Bilinear Forms, 259 Symmetric, Skew-Symmetric and Alternate Forms, 259 The Matrix of a Bilinear Form, 261

xv

xvi

Contents Quadratic Forms, 264 Orthogonality, 265 Linear Functionals, 268 Orthogonal Complements and Orthogonal Direct Sums, 269 Isometries, 271 Hyperbolic Spaces, 272 Nonsingular Completions of a Subspace, 273 The Witt Theorems: A Preview, 275 The Classification Problem for Metric Vector Spaces, 276 Symplectic Geometry, 277 The Structure of Orthogonal Geometries: Orthogonal Bases, 282 The Classification of Orthogonal Geometries: Canonical Forms, 285 The Orthogonal Group, 291 The Witt Theorems for Orthogonal Geometries, 294 Maximal Hyperbolic Subspaces of an Orthogonal Geometry, 295 Exercises, 297

12

Metric Spaces, 301 The Definition, 301 Open and Closed Sets, 304 Convergence in a Metric Space, 305 The Closure of a Set, 306 Dense Subsets, 308 Continuity, 310 Completeness, 311 Isometries, 315 The Completion of a Metric Space, 316 Exercises, 321

13

Hilbert Spaces, 325 A Brief Review, 325 Hilbert Spaces, 326 Infinite Series, 330 An Approximation Problem, 331 Hilbert Bases, 335 Fourier Expansions, 336 A Characterization of Hilbert Bases, 346 Hilbert Dimension, 346 A Characterization of Hilbert Spaces, 347 The Riesz Representation Theorem, 349 Exercises, 352

14

Tensor Products, 355 Universality, 355 Bilinear Maps, 359 Tensor Products, 361

Contents When Is a Tensor Product Zero?, 367 Coordinate Matrices and Rank, 368 Characterizing Vectors in a Tensor Product, 371 Defining Linear Transformations on a Tensor Product, 374 The Tensor Product of Linear Transformations, 375 Change of Base Field, 379 Multilinear Maps and Iterated Tensor Products, 382 Tensor Spaces, 385 Special Multilinear Maps, 390 Graded Algebras, 392 The Symmetric and Antisymmetric Tensor Algebras, 392 The Determinant, 403 Exercises, 406

15

Positive Solutions to Linear Systems: Convexity and Separation, 411 Convex, Closed and Compact Sets, 413 Convex Hulls, 414 Linear and Affine Hyperplanes, 416 Separation, 418 Exercises, 423

16

Affine Geometry, 427 Affine Geometry, 427 Affine Combinations, 428 Affine Hulls, 430 The Lattice of Flats, 431 Affine Independence, 433 Affine Transformations, 435 Projective Geometry, 437 Exercises, 440

17

Singular Values and the Moore–Penrose Inverse, 443 Singular Values, 443 The Moore–Penrose Generalized Inverse, 446 Least Squares Approximation, 448 Exercises, 449

18

An Introduction to Algebras, 451 Motivation, 451 Associative Algebras, 451 Division Algebras, 462 Exercises, 469

19

The Umbral Calculus, 471 Formal Power Series, 471 The Umbral Algebra, 473

xvii

xviii

Contents Formal Power Series as Linear Operators, 477 Sheffer Sequences, 480 Examples of Sheffer Sequences, 488 Umbral Operators and Umbral Shifts, 490 Continuous Operators on the Umbral Algebra, 492 Operator Adjoints, 493 Umbral Operators and Automorphisms of the Umbral Algebra, 494 Umbral Shifts and Derivations of the Umbral Algebra, 499 The Transfer Formulas, 504 A Final Remark, 505 Exercises, 506

References, 507 Index of Symbols, 513 Index, 515

Preliminaries

In this chapter, we briefly discuss some topics that are needed for the sequel. This chapter should be skimmed quickly and used primarily as a reference.

Part 1 Preliminaries Multisets The following simple concept is much more useful than its infrequent appearance would indicate. Definition Let : be a nonempty set. A multiset 4 with underlying set : is a set of ordered pairs 4 ~ ¸² Á ³

:Á {b Á

£

for £ ¹

where {b ~ ¸Á Á Ã ¹. The number is referred to as the multiplicity of the elements in 4 . If the underlying set of a multiset is finite, we say that the multiset is finite. The size of a finite multiset 4 is the sum of the multiplicities of all of its elements. For example, 4 ~ ¸²Á ³Á ²Á ³Á ²Á ³¹ is a multiset with underlying set : ~ ¸Á Á ¹. The element has multiplicity . One often writes out the elements of a multiset according to multiplicities, as in 4 ~ ¸Á Á Á Á Á ¹ . Of course, two mutlisets are equal if their underlying sets are equal and if the multiplicity of each element in the common underlying set is the same in both multisets.

Matrices The set of d matrices with entries in a field - is denoted by CÁ ²- ³ or by CÁ when the field does not require mention. The set CÁ ²< ³ is denoted by C ²- ³ or C À If ( C, the ²Á ³th entry of ( will be denoted by (Á . The identity matrix of size d is denoted by 0 . The elements of the base

2

Advanced Linear Algebra

field - are called scalars. We expect that the reader is familiar with the basic properties of matrices, including matrix addition and multiplication. The main diagonal of an d matrix ( is the sequence of entries (Á Á (Á Á Ã Á (Á where ~ min¸Á ¹. Definition The transpose of ( CÁ is the matrix (! defined by ²(! ³Á ~ (Á A matrix ( is symmetric if ( ~ (! and skew-symmetric if (! ~ c(. Theorem 0.1 (Properties of the transpose) Let (, ) CÁ . Then 1) ²(! ³! ~ ( 2) ²( b )³! ~ (! b ) ! 3) ²(³! ~ (! for all 4) ²()³! ~ ) ! (! provided that the product () is defined 5) det²(! ³ ~ det²(³.

Partitioning and Matrix Multiplication Let 4 be a matrix of size d . If ) ¸Á Ã Á ¹ and * ¸Á Ã Á ¹, then the submatrix 4 ´)Á *µ is the matrix obtained from 4 by keeping only the rows with index in ) and the columns with index in * . Thus, all other rows and columns are discarded and 4 ´)Á *µ has size () ( d (* (. Suppose that 4 CÁ and 5 CÁ . Let 1) F ~ ¸) Á Ã Á ) ¹ be a partition of ¸Á Ã Á ¹ 2) G ~ ¸* Á Ã Á * ¹ be a partition of ¸Á Ã Á ¹ 3) H ~ ¸+ Á Ã Á + ¹ be a partition of ¸Á Ã Á ¹ (Partitions are defined formally later in this chapter.) Then it is a very useful fact that matrix multiplication can be performed at the block level as well as at the entry level. In particular, we have ´4 5 µ´) Á + µ ~ 4 ´) Á * µ5 ´* Á + µ * G

When the partitions in question contain only single-element blocks, this is precisely the usual formula for matrix multiplication

´4 5 µÁ ~ 4Á 5Á ~

Preliminaries

3

Block Matrices It will be convenient to introduce the notational device of a block matrix. If )Á are matrices of the appropriate sizes, then by the block matrix 4~

v )Á Å w )Á

)Á Å )Á

Ä

)Á y Å Ä )Á zblock

we mean the matrix whose upper left submatrix is )Á , and so on. Thus, the )Á 's are submatrices of 4 and not entries. A square matrix of the form v ) x 4 ~x Å w

Ä y Æ Æ Å { { Æ Æ Ä ) zblock

where each ) is square and is a zero submatrix, is said to be a block diagonal matrix.

Elementary Row Operations Recall that there are three types of elementary row operations. Type 1 operations consist of multiplying a row of ( by a nonzero scalar. Type 2 operations consist of interchanging two rows of (. Type 3 operations consist of adding a scalar multiple of one row of ( to another row of (. If we perform an elementary operation of type to an identity matrix 0 , the result is called an elementary matrix of type . It is easy to see that all elementary matrices are invertible. In order to perform an elementary row operation on ( CÁ we can perform that operation on the identity 0 , to obtain an elementary matrix , and then take the product ,(. Note that multiplying on the right by , has the effect of performing column operations. Definition A matrix 9 is said to be in reduced row echelon form if 1) All rows consisting only of 's appear at the bottom of the matrix. 2) In any nonzero row, the first nonzero entry is a . This entry is called a leading entry. 3) For any two consecutive rows, the leading entry of the lower row is to the right of the leading entry of the upper row. 4) Any column that contains a leading entry has 's in all other positions. Here are the basic facts concerning reduced row echelon form.

4

Advanced Linear Algebra

Theorem 0.2 Matrices (Á ) CÁ are row equivalent, denoted by ( ) , if either one can be obtained from the other by a series of elementary row operations. 1) Row equivalence is an equivalence relation. That is, a) ( ( b) ( ) ¬ ) ( c) ( ) , ) * ¬ ( * . 2) A matrix ( is row equivalent to one and only one matrix 9 that is in reduced row echelon form. The matrix 9 is called the reduced row echelon form of (. Furthermore, 9 ~ , Ä, ( where , are the elementary matrices required to reduce ( to reduced row echelon form. 3) ( is invertible if and only if its reduced row echelon form is an identity matrix. Hence, a matrix is invertible if and only if it is the product of elementary matrices. The following definition is probably well known to the reader. Definition A square matrix is upper triangular if all of its entries below the main diagonal are . Similarly, a square matrix is lower triangular if all of its entries above the main diagonal are . A square matrix is diagonal if all of its entries off the main diagonal are .

Determinants We assume that the reader is familiar with the following basic properties of determinants. Theorem 0.3 Let ( CÁ ²- ³. Then det²(³ is an element of - . Furthermore, 1) For any ) C ²- ³, det²()³ ~ det²(³det²)³ 2) ( is nonsingular (invertible) if and only if det²(³ £ . 3) The determinant of an upper triangular or lower triangular matrix is the product of the entries on its main diagonal. 4) If a square matrix 4 has the block diagonal form v ) x 4 ~x Å w then det²4 ³ ~ det²) ³.

Ä y Æ Æ Å { { Æ Æ Ä ) zblock

Preliminaries

5

Polynomials The set of all polynomials in the variable % with coefficients from a field - is denoted by - ´%µ. If ²%³ - ´%µ, we say that ²%³ is a polynomial over - . If ²%³ ~ b % b Ä b % is a polynomial with £ , then is called the leading coefficient of ²%³ and the degree of ²%³ is , written deg ²%³ ~ . For convenience, the degree of the zero polynomial is cB. A polynomial is monic if its leading coefficient is . Theorem 0.4 (Division algorithm) Let ²%³Á ²%³ - ´%µ where deg ²%³ . Then there exist unique polynomials ²%³Á ²%³ - ´%µ for which ²%³ ~ ²%³²%³ b ²%³ where ²%³ ~ or deg ²%³ deg ²%³. If ²%³ divides ²%³, that is, if there exists a polynomial ²%³ for which ²%³ ~ ²%³²%³ then we write ²%³ ²%³. A nonzero polynomial ²%³ - ´%µ is said to split over - if ²%³ can be written as a product of linear factors ²%³ ~ ²% c ³Ä²% c ³ where - . Theorem 0.5 Let ²%³Á ²%³ - ´%µ. The greatest common divisor of ²%³ and ²%³, denoted by gcd² ²%³Á ²%³³, is the unique monic polynomial ²%³ over for which 1) ²%³ ²%³ and ²%³ ²%³ 2) if ²%³ ²%³ and ²%³ ²%³ then ²%³ ²%³. Furthermore, there exist polynomials ²%³ and ²%³ over - for which gcd² ²%³Á ²%³³ ~ ²%³ ²%³ b ²%³²%³

Definition The polynomials ²%³Á ²%³ - ´%µ are relatively prime if gcd² ²%³Á ²%³³ ~ . In particular, ²%³ and ²%³ are relatively prime if and only if there exist polynomials ²%³ and ²%³ over - for which ²%³ ²%³ b ²%³²%³ ~

Definition A nonconstant polynomial ²%³ - ´%µ is irreducible if whenever ²%³ ~ ²%³²%³, then one of ²%³ and ²%³ must be constant. The following two theorems support the view that irreducible polynomials behave like prime numbers.

6

Advanced Linear Algebra

Theorem 0.6 A nonconstant polynomial ²%³ is irreducible if and only if it has the property that whenever ²%³ ²%³²%³, then either ²%³ ²%³ or ²%³ ²%³. Theorem 0.7 Every nonconstant polynomial in - ´%µ can be written as a product of irreducible polynomials. Moreover, this expression is unique up to order of the factors and multiplication by a scalar.

Functions To set our notation, we should make a few comments about functions. Definition Let ¢ : ¦ ; be a function from a set : to a set ; . 1) The domain of is the set : and the range of is ; . 2) The image of is the set im² ³ ~ ¸ ² ³ :¹. 3) is injective (one-to-one), or an injection, if % £ & ¬ ²%³ £ ²&³. 4) is surjective (onto ; ), or a surjection, if im² ³ ~ ; . 5) is bijective, or a bijection, if it is both injective and surjective. 6) Assuming that ; , the support of is supp² ³ ~ ¸ : ² ³ £ ¹

If ¢ : ¦ ; is injective, then its inverse c ¢ im² ³ ¦ : exists and is welldefined as a function on im² ³. It will be convenient to apply to subsets of : and ; . In particular, if ? : and if @ ; , we set ²?³ ~ ¸ ²%³ % ?¹ and c ²@ ³ ~ ¸ : ² ³ @ ¹ Note that the latter is defined even if is not injective. Let ¢ : ¦ ; . If ( : , the restriction of to ( is the function O( ¢ ( ¦ ; defined by O( ²³ ~ ²³ for all (. Clearly, the restriction of an injective map is injective. In the other direction, if ¢ : ¦ ; and if : < , then an extension of to < is a function ¢ < ¦ ; for which O: ~ .

Preliminaries

7

Equivalence Relations The concept of an equivalence relation plays a major role in the study of matrices and linear transformations. Definition Let : be a nonempty set. A binary relation on : is called an equivalence relation on : if it satisfies the following conditions: 1) (Reflexivity) for all : . 2) (Symmetry) ¬ for all Á : . 3) (Transitivity) Á ¬ for all Á Á : . Definition Let be an equivalence relation on : . For : , the set of all elements equivalent to is denoted by ´µ ~ ¸ : ¹ and called the equivalence class of . Theorem 0.8 Let be an equivalence relation on : . Then 1) ´µ ¯ ´µ ¯ ´µ ~ ´µ 2) For any Á : , we have either ´µ ~ ´µ or ´µ q ´µ ~ J. Definition A partition of a nonempty set : is a collection ¸( Á Ã Á ( ¹ of nonempty subsets of : , called the blocks of the partition, for which 1) ( q ( ~ J for all £ 2 ) : ~ ( r Ä r ( . The following theorem sheds considerable light on the concept of an equivalence relation. Theorem 0.9 1) Let be an equivalence relation on : . Then the set of distinct equivalence classes with respect to are the blocks of a partition of : . 2) Conversely, if F is a partition of : , the binary relation defined by if and lie in the same block of F

8

Advanced Linear Algebra

is an equivalence relation on : , whose equivalence classes are the blocks of F . This establishes a one-to-one correspondence between equivalence relations on : and partitions of : . The most important problem related to equivalence relations is that of finding an efficient way to determine when two elements are equivalent. Unfortunately, in most cases, the definition does not provide an efficient test for equivalence and so we are led to the following concepts. Definition Let be an equivalence relation on : . A function ¢ : ¦ ; , where ; is any set, is called an invariant of if it is constant on the equivalence classes of , that is, ¬ ²³ ~ ²³ and a complete invariant if it is constant and distinct on the equivalence classes of , that is, ¯ ²³ ~ ²³ A collection ¸ Á Ã Á ¹ of invariants is called a complete system of invariants if ¯ ²³ ~ ²³ for all ~ Á Ã Á

Definition Let be an equivalence relation on : . A subset * : is said to be a set of canonical forms (or just a canonical form) for if for every : , there is exactly one * such that . Put another way, each equivalence class under contains exactly one member of * . Example 0.1 Define a binary relation on - ´%µ by letting ²%³ ²%³ if and only if ²%³ ~ ²%³ for some nonzero constant - . This is easily seen to be an equivalence relation. The function that assigns to each polynomial its degree is an invariant, since ²%³ ²%³ ¬ deg²²%³³ ~ deg²²%³³ However, it is not a complete invariant, since there are inequivalent polynomials with the same degree. The set of all monic polynomials is a set of canonical forms for this equivalence relation. Example 0.2 We have remarked that row equivalence is an equivalence relation on CÁ ²- ³. Moreover, the subset of reduced row echelon form matrices is a set of canonical forms for row equivalence, since every matrix is row equivalent to a unique matrix in reduced row echelon form.

Preliminaries

9

Example 0.3 Two matrices (, ) C ²- ³ are row equivalent if and only if there is an invertible matrix 7 such that ( ~ 7 ) . Similarly, ( and ) are column equivalent, that is, ( can be reduced to ) using elementary column operations, if and only if there exists an invertible matrix 8 such that ( ~ )8. Two matrices ( and ) are said to be equivalent if there exist invertible matrices 7 and 8 for which ( ~ 7 )8 Put another way, ( and ) are equivalent if ( can be reduced to ) by performing a series of elementary row and/or column operations. (The use of the term equivalent is unfortunate, since it applies to all equivalence relations, not just this one. However, the terminology is standard, so we use it here.) It is not hard to see that an d matrix 9 that is in both reduced row echelon form and reduced column echelon form must have the block form 0 1 ~ > cÁ

Ác cÁc ?block

We leave it to the reader to show that every matrix ( in C is equivalent to exactly one matrix of the form 1 and so the set of these matrices is a set of canonical forms for equivalence. Moreover, the function defined by ²(³ ~ , where ( 1 , is a complete invariant for equivalence. Since the rank of 1 is and since neither row nor column operations affect the rank, we deduce that the rank of ( is . Hence, rank is a complete invariant for equivalence. In other words, two matrices are equivalent if and only if they have the same rank. Example 0.4 Two matrices (, ) C ²- ³ are said to be similar if there exists an invertible matrix 7 such that ( ~ 7 )7 c Similarity is easily seen to be an equivalence relation on C . As we will learn, two matrices are similar if and only if they represent the same linear operators on a given -dimensional vector space = . Hence, similarity is extremely important for studying the structure of linear operators. One of the main goals of this book is to develop canonical forms for similarity. We leave it to the reader to show that the determinant function and the trace function are invariants for similarity. However, these two invariants do not, in general, form a complete system of invariants. Example 0.5 Two matrices (, ) C ²- ³ are said to be congruent if there exists an invertible matrix 7 for which

10

Advanced Linear Algebra

( ~ 7 )7 ! where 7 ! is the transpose of 7 . This relation is easily seen to be an equivalence relation and we will devote some effort to finding canonical forms for congruence. For some base fields - (such as s, d or a finite field), this is relatively easy to do, but for other base fields (such as r), it is extremely difficult.

Zorn's Lemma In order to show that any vector space has a basis, we require a result known as Zorn's lemma. To state this lemma, we need some preliminary definitions. Definition A partially ordered set is a pair ²7 Á ³ where 7 is a nonempty set and is a binary relation called a partial order, read “less than or equal to,” with the following properties: 1) (Reflexivity) For all 7 , 2) (Antisymmetry) For all Á 7 , and implies ~ 3) (Transitivity) For all Á Á 7 , and implies Partially ordered sets are also called posets. It is customary to use a phrase such as “Let 7 be a partially ordered set” when the partial order is understood. Here are some key terms related to partially ordered sets. Definition Let 7 be a partially ordered set. 1) The maximum (largest, top) element of 7 , should it exist, is an element 4 7 with the property that all elements of 7 are less than or equal to 4 , that is, 7 ¬4 Similarly, the mimimum (least, smallest, bottom) element of 7 , should it exist, is an element 5 7 with the property that all elements of 7 are greater than or equal to 5 , that is, 7 ¬5 2) A maximal element is an element 7 with the property that there is no larger element in 7 , that is, 7Á ¬ ~

Preliminaries

11

Similarly, a minimal element is an element 7 with the property that there is no smaller element in 7 , that is, 7Á ¬ ~ 3) Let Á 7 . Then " 7 is an upper bound for and if " and " The unique smallest upper bound for and , if it exists, is called the least upper bound of and and is denoted by lub¸Á ¹. 4) Let Á 7 . Then M 7 is a lower bound for and if M and M The unique largest lower bound for and , if it exists, is called the greatest lower bound of and and is denoted by glb¸Á ¹. Let : be a subset of a partially ordered set 7 . We say that an element " 7 is an upper bound for : if " for all : . Lower bounds are defined similarly. Note that in a partially ordered set, it is possible that not all elements are comparable. In other words, it is possible to have %Á & 7 with the property that % & and & %. Definition A partially ordered set in which every pair of elements is comparable is called a totally ordered set, or a linearly ordered set. Any totally ordered subset of a partially ordered set 7 is called a chain in 7 . Example 0.6 1) The set s of real numbers, with the usual binary relation , is a partially ordered set. It is also a totally ordered set. It has no maximal elements. 2) The set o ~ ¸Á Á Ã ¹ of natural numbers, together with the binary relation of divides, is a partially ordered set. It is customary to write to indicate that divides . The subset : of o consisting of all powers of is a totally ordered subset of o, that is, it is a chain in o. The set 7 ~ ¸Á Á Á Á Á ¹ is a partially ordered set under . It has two maximal elements, namely and . The subset 8 ~ ¸Á Á Á Á ¹ is a partially ordered set in which every element is both maximal and minimal! 3) Let : be any set and let F ²:³ be the power set of : , that is, the set of all subsets of : . Then F ²:³, together with the subset relation , is a partially ordered set. Now we can state Zorn's lemma, which gives a condition under which a partially ordered set has a maximal element.

12

Advanced Linear Algebra

Theorem 0.10 (Zorn's lemma) If 7 is a partially ordered set in which every chain has an upper bound, then 7 has a maximal element. We will use Zorn's lemma to prove that every vector space has a basis. Zorn's lemma is equivalent to the famous axiom of choice. As such, it is not subject to proof from the other axioms of ordinary (ZF) set theory. Zorn's lemma has many important equivalancies, one of which is the well-ordering principle. A well ordering on a nonempty set ? is a total order on ? with the property that every nonempty subset of ? has a least element. Theorem 0.11 (Well-ordering principle) Every nonempty set has a well ordering.

Cardinality Two sets : and ; have the same cardinality, written (: ( ~ (; ( if there is a bijective function (a one-to-one correspondence) between the sets. The reader is probably aware of the fact that ({( ~ (o( and (r( ~ (o( where o denotes the natural numbers, { the integers and r the rational numbers. If : is in one-to-one correspondence with a subset of ; , we write (: ( (; (. If : is in one-to-one correspondence with a proper subset of ; but not all of ; , then we write (: ( (; (. The second condition is necessary, since, for instance, o is in one-to-one correspondence with a proper subset of { and yet o is also in one-to-one correspondence with { itself. Hence, (o( ~ ({(. This is not the place to enter into a detailed discussion of cardinal numbers. The intention here is that the cardinality of a set, whatever that is, represents the “size” of the set. It is actually easier to talk about two sets having the same, or different, size (cardinality) than it is to explicitly define the size (cardinality) of a given set. Be that as it may, we associate to each set : a cardinal number, denoted by (: ( or card²:³, that is intended to measure the size of the set. Actually, cardinal numbers are just very special types of sets. However, we can simply think of them as vague amorphous objects that measure the size of sets. Definition 1) A set is finite if it can be put in one-to-one correspondence with a set of the form { ~ ¸Á Á Ã Á c ¹, for some nonnegative integer . A set that is

Preliminaries

13

not finite is infinite. The cardinal number (or cardinality) of a finite set is just the number of elements in the set. 2) The cardinal number of the set o of natural numbers is L (read “aleph nought”), where L is the first letter of the Hebrew alphabet. Hence, (o( ~ ({( ~ (r( ~ L 3) Any set with cardinality L is called a countably infinite set and any finite or countably infinite set is called a countable set. An infinite set that is not countable is said to be uncountable. Since it can be shown that (s( (o(, the real numbers are uncountable. If : and ; are finite sets, then it is well known that (: ( (; ( and (; ( (: ( ¬ (: ( ~ (; ( The first part of the next theorem tells us that this is also true for infinite sets. The reader will no doubt recall that the power set F²:³ of a set : is the set of all subsets of : . For finite sets, the power set of : is always bigger than the set itself. In fact, (: ( ~ ¬ (F ²:³( ~ The second part of the next theorem says that the power set of any set : is bigger (has larger cardinality) than : itself. On the other hand, the third part of this theorem says that, for infinite sets : , the set of all finite subsets of : is the same size as : . Theorem 0.12 ¨ –Bernstein theorem) For any sets : and ; , 1) (Schroder (: ( (; ( and (; ( (: ( ¬ (: ( ~ (; ( 2) (Cantor's theorem) If F²:³ denotes the power set of : , then (: ( (F ²:³( 3) If F ²:³ denotes the set of all finite subsets of : and if : is an infinite set, then (: ( ~ (F ²:³( Proof. We prove only parts 1) and 2). Let ¢ : ¦ ; be an injective function from : into ; and let ¢ ; ¦ : be an injective function from ; into : . We want to use these functions to create a bijective function from : to ; . For this purpose, we make the following definitions. The descendants of an element : are the elements obtained by repeated alternate applications of the functions and , namely

14

Advanced Linear Algebra

² ³Á ² ² ³³Á ²² ² ³³³Á Ã If ! is a descendant of , then is an ancestor of !. Descendants and ancestors of elements of ; are defined similarly. Now, by tracing an element's ancestry to its beginning, we find that there are three possibilities: the element may originate in : , or in ; , or it may have no point of origin. Accordingly, we can write : as the union of three disjoint sets I: ~ ¸ : originates in :¹ I; ~ ¸ : originates in ; ¹ IB ~ ¸ : has no originator¹ Similarly, ; is the disjoint union of J: , J; and JB . Now, the restriction OI: ¢ I: ¦ J: is a bijection. To see this, note that if ! J: , then ! originated in : and therefore must have the form ² ³ for some : . But ! and its ancestor have the same point of origin and so ! J: implies I: . Thus, OI: is surjective and hence bijective. We leave it to the reader to show that the functions ²OJ; ³c ¢ I; ¦ J; and OIB ¢ IB ¦ JB are also bijections. Putting these three bijections together gives a bijection between : and ; . Hence, (: ( ~ (; (, as desired. We now prove Cantor's theorem. The map ¢ : ¦ F ²:³ defined by ² ³ ~ ¸ ¹ is an injection from : to F ²:³ and so (: ( (F ²:³(. To complete the proof we must show that no injective map ¢ : ¦ F ²:³ can be surjective. To this end, let ?~¸ :

¤ ² ³¹ F ²:³

We claim that ? is not in im² ³. For suppose that ? ~ ²%³ for some % : . Then if % ? , we have by the definition of ? that % ¤ ? . On the other hand, if % ¤ ? , we have again by the definition of ? that % ? . This contradiction implies that ? ¤ im² ³ and so is not surjective. Cardinal Arithmetic Now let us define addition, multiplication and exponentiation of cardinal numbers. If : and ; are sets, the cartesian product : d ; is the set of all ordered pairs : d ; ~ ¸² Á !³

:Á ! ; ¹

The set of all functions from ; to : is denoted by : ; .

Preliminaries

15

Definition Let and denote cardinal numbers. Let : and ; be disjoint sets for which (: ( ~ and (; ( ~ . 1) The sum b is the cardinal number of : r ; . 2) The product is the cardinal number of : d ; . 3) The power is the cardinal number of : ; . We will not go into the details of why these definitions make sense. (For instance, they seem to depend on the sets : and ; , but in fact they do not.) It can be shown, using these definitions, that cardinal addition and multiplication are associative and commutative and that multiplication distributes over addition. Theorem 0.13 Let , and be cardinal numbers. Then the following properties hold: 1) (Associativity) b ² b ³ ~ ² b ³ b and ²³ ~ ²³ 2) (Commutativity) b ~ b and ~ 3) (Distributivity) ² b ³ ~ b 4) (Properties of Exponents) a) b ~ b) ² ³ ~ c) ²³ ~ On the other hand, the arithmetic of cardinal numbers can seem a bit strange, as the next theorem shows. Theorem 0.14 Let and be cardinal numbers, at least one of which is infinite. Then b ~ ~ max¸Á ¹

It is not hard to see that there is a one-to-one correspondence between the power set F²:³ of a set : and the set of all functions from : to ¸Á ¹. This leads to the following theorem. Theorem 0.15 For any cardinal 1) If (: ( ~ , then (F ²:³( ~ 2)

16

Advanced Linear Algebra

We have already observed that (o( ~ L . It can be shown that L is the smallest infinite cardinal, that is, L0 ¬ is a natural number It can also be shown that the set s of real numbers is in one-to-one correspondence with the power set F ²o³ of the natural numbers. Therefore, (s( ~ L The set of all points on the real line is sometimes called the continuum and so L is sometimes called the power of the continuum and denoted by . Theorem 0.14 shows that cardinal addition and multiplication have a kind of “absorption” quality, which makes it hard to produce larger cardinals from smaller ones. The next theorem demonstrates this more dramatically. Theorem 0.16 1) Addition applied a countable number of times or multiplication applied a finite number of times to the cardinal number L , does not yield anything more than L . Specifically, for any nonzero o, we have L h L ~ L and L ~ L 2) Addition and multiplication applied a countable number of times to the cardinal number L does not yield more than L . Specifically, we have L h L ~ L and ²L ³L ~ L

Using this theorem, we can establish other relationships, such as L ²L ³L ²L ³L ~ L which, by the Schro¨der–Bernstein theorem, implies that ²L ³L ~ L We mention that the problem of evaluating in general is a very difficult one and would take us far beyond the scope of this book. We will have use for the following reasonable-sounding result, whose proof is omitted. Theorem 0.17 Let ¸( 2¹ be a collection of sets, indexed by the set 2 , with (2 ( ~ . If (( ( for all 2 , then e ( e 2

Let us conclude by describing the cardinality of some famous sets.

Preliminaries

17

Theorem 0.18 1) The following sets have cardinality L . a) The rational numbers r. b) The set of all finite subsets of o. c) The union of a countable number of countable sets. d) The set { of all ordered -tuples of integers. 2) The following sets have cardinality L . a) The set of all points in s . b) The set of all infinite sequences of natural numbers. c) The set of all infinite sequences of real numbers. d) The set of all finite subsets of s. e) The set of all irrational numbers.

Part 2 Algebraic Structures We now turn to a discussion of some of the many algebraic structures that play a role in the study of linear algebra.

Groups Definition A group is a nonempty set ., together with a binary operation denoted by *, that satisfies the following properties: 1) (Associativity) For all Á Á ., ²i³i ~ i²i³ 2) (Identity) There exists an element . for which i ~ i ~ for all .. 3) (Inverses) For each ., there is an element c . for which ic ~ c i ~

Definition A group . is abelian, or commutative, if i ~ i for all Á . . When a group is abelian, it is customary to denote the operation i by +, thus writing i as b . It is also customary to refer to the identity as the zero element and to denote the inverse c by c, referred to as the negative of . Example 0.7 The set < of all bijective functions from a set : to : is a group under composition of functions. However, in general, it is not abelian. Example 0.8 The set CÁ ²- ³ is an abelian group under addition of matrices. The identity is the zero matrix 0Á of size d . The set C ²- ³ is not a group under multiplication of matrices, since not all matrices have multiplicative

18

Advanced Linear Algebra

inverses. However, the set of invertible matrices of size d is a (nonabelian) group under multiplication. A group . is finite if it contains only a finite number of elements. The cardinality of a finite group . is called its order and is denoted by ².³ or simply (.(. Thus, for example, { ~ ¸Á Á Ã Á c ¹ is a finite group under addition modulo , but CÁ ²s³ is not finite. Definition A subgroup of a group . is a nonempty subset : of . that is a group in its own right, using the same operations as defined on . . Cyclic Groups If is a formal symbol, we can define a group . to be the set of all integral powers of : . ~ ¸ {¹ where the product is defined by the formal rules of exponents: ~ b This group is denoted by º» and called the cyclic group generated by . The identity of º» is ~ . In general, a group . is cyclic if it has the form . ~ º» for some .. We can also create a finite group * ²³ of arbitrary positive order by declaring that ~ . Thus, * ²³ ~ ¸ ~ Á Á Á Ã Á c ¹ where the product is defined by the formal rules of exponents, followed by reduction modulo : ~ ²b³ mod This defines a group of order , called a cyclic group of order . The inverse of is ²c³ mod .

Rings Definition A ring is a nonempty set 9 , together with two binary operations, called addition (denoted by b ) and multiplication (denoted by juxtaposition), for which the following hold: 1) 9 is an abelian group under addition 2) (Associativity) For all Á Á 9 , ²³ ~ ²³

Preliminaries

19

3) (Distributivity) For all Á Á 9 , ² b ³ ~ b and ² b ³ ~ b A ring 9 is said to be commutative if ~ for all Á 9 . If a ring 9 contains an element with the property that ~ ~ for all 9 , we say that 9 is a ring with identity. The identity is usually denoted by . A field - is a commutative ring with identity in which each nonzero element has a multiplicative inverse, that is, if - is nonzero, then there is a for which ~ . Example 0.9 The set { ~ ¸Á Á Ã Á c¹ is a commutative ring under addition and multiplication modulo l ~ ² b ³ mod Á

p ~ mod

The element { is the identity. Example 0.10 The set , of even integers is a commutative ring under the usual operations on {, but it has no identity. Example 0.11 The set C ²- ³ is a noncommutative ring under matrix addition and multiplication. The identity matrix 0 is the identity for C ²- ³. Example 0.12 Let - be a field. The set - ´%µ of all polynomials in a single variable %, with coefficients in - , is a commutative ring under the usual operations of polynomial addition and multiplication. What is the identity for - ´%µ? Similarly, the set - ´% Á Ã Á % µ of polynomials in variables is a commutative ring under the usual addition and multiplication of polynomials. Definition If 9 and : are rings, then a function ¢ 9 ¦ : is a ring homomorphism if ² b ³ ~ b ²³ ~ ²³²³ ~ for all Á 9 . Definition A subring of a ring 9 is a subset : of 9 that is a ring in its own right, using the same operations as defined on 9 and having the same multiplicative identity as 9 .

20

Advanced Linear Algebra

The condition that a subring : have the same multiplicative identity as 9 is required. For example, the set : of all d matrices of the form ( ~ >

?

for - is a ring under addition and multiplication of matrices (isomorphic to - ). The multiplicative identity in : is the matrix ( , which is not the identity 0 of CÁ ²- ³. Hence, : is a ring under the same operations as CÁ ²- ³ but it is not a subring of CÁ ²- ³. Applying the definition is not generally the easiest way to show that a subset of a ring is a subring. The following characterization is usually easier to apply. Theorem 0.19 A nonempty subset : of a ring 9 is a subring if and only if 1) The multiplicative identity 9 of 9 is in : 2) : is closed under subtraction, that is, Á : ¬ c : 3) : is closed under multiplication, that is, Á : ¬ :

Ideals Rings have another important substructure besides subrings. Definition Let 9 be a ring. A nonempty subset ? of 9 is called an ideal if 1) ? is a subgroup of the abelian group 9, that is, ? is closed under subtraction: Á ? ¬ c ? 2) ? is closed under multiplication by any ring element, that is, ? Á 9 ¬ ? and ?

Note that if an ideal ? contains the unit element , then ? ~ 9 . Example 0.13 Let ²%³ be a polynomial in - ´%µ. The set of all multiples of ²%³, º²%³» ~ ¸²%³²%³ ²%³ - ´%µ¹ is an ideal in - ´%µ, called the ideal generated by ²%³. Definition Let : be a subset of a ring 9 with identity. The set º:» ~ ¸

b Ä b

9Á

:Á ¹

Preliminaries

21

of all finite linear combinations of elements of : , with coefficients in 9 , is an ideal in 9 , called the ideal generated by : . It is the smallest (in the sense of set inclusion) ideal of 9 containing : . If : ~ ¸ Á Ã Á ¹ is a finite set, we write º Á Ã Á

»

~ ¸

b Ä b

9Á

:¹

Note that in the previous definition, we require that 9 have an identity. This is to ensure that : º:». Theorem 0.20 Let 9 be a ring. 1) The intersection of any collection ¸? 2¹ of ideals is an ideal. 2) If ? ? Ä is an ascending sequence of ideals, each one contained in the next, then the union ? is also an ideal. 3) More generally, if 9 ~ ¸? 0¹ is a chain of ideals in 9 , then the union @ ~ 0 ? is also an ideal in 9 . Proof. To prove 1), let @ ~ ? . Then if Á @ , we have Á ? for all 2 . Hence, c ? for all 2 and so c @ . Hence, @ is closed under subtraction. Also, if 9 , then ? for all 2 and so @ . Of course, part 2) is a special case of part 3). To prove 3), if Á @ , then ? and ? for some Á 0 . Since one of ? and ? is contained in the other, we may assume that ? ? . It follows that Á ? and so c ? @ and if 9 , then ? @ . Thus @ is an ideal. Note that in general, the union of ideals is not an ideal. However, as we have just proved, the union of any chain of ideals is an ideal. Quotient Rings and Maximal Ideals Let : be a subset of a commutative ring 9 with identity. Let be the binary relation on 9 defined by ¯ c : It is easy to see that is an equivalence relation. When , we say that and are congruent modulo : . The term “mod” is used as a colloquialism for modulo and is often written mod : As shorthand, we write .

22

Advanced Linear Algebra

To see what the equivalence classes look like, observe that ´µ ~ ¸ 9 ¹ ~ ¸ 9 c :¹ ~ ¸ 9 ~ b for some :¹ ~ ¸ b :¹ ~b: The set b : ~ ¸ b

:¹

is called a coset of : in 9 . The element is called a coset representative for b :. Thus, the equivalence classes for congruence mod : are the cosets b : of : in 9 . The set of all cosets is denoted by 9°: ~ ¸ b : 9¹ This is read “9 mod : .” We would like to place a ring structure on 9°: . Indeed, if : is a subgroup of the abelian group 9, then 9°: is easily seen to be an abelian group as well under coset addition defined by ² b :³ b ² b :³ ~ ² b ³ b : In order for the product ² b :³² b :³ ~ b : to be well-defined, we must have b : ~ Z b : ¬ b : ~ Z b : or, equivalently, c Z : ¬ ² c Z ³ : But c Z may be any element of : and may be any element of 9 and so this condition implies that : must be an ideal. Conversely, if : is an ideal, then coset multiplication is well defined. Theorem 0.21 Let 9 be a commutative ring with identity. Then the quotient 9°? is a ring under coset addition and multiplication if and only if ? is an ideal of 9 . In this case, 9°? is called the quotient ring of 9 modulo ? , where addition and multiplication are defined by ² b :³ b ² b :³ ~ ² b ³ b : ² b :³² b :³ ~ b :

Preliminaries

23

Definition An ideal ? in a ring 9 is a maximal ideal if ? £ 9 and if whenever @ is an ideal satisfying ? @ 9 , then either @ ~ ? or @ ~ 9 . Here is one reason why maximal ideals are important. Theorem 0.22 Let 9 be a commutative ring with identity. Then the quotient ring 9°? is a field if and only if ? is a maximal ideal. Proof. First, note that for any ideal ? of 9 , the ideals of 9°? are precisely the quotients @ °? where @ is an ideal for which ? @ 9. It is clear that @ °? is an ideal of 9°? . Conversely, if AZ is an ideal of 9°? , then let A ~ ¸ 9 b ? AZ ¹ It is easy to see that A is an ideal of 9 for which ? A 9 . Next, observe that a commutative ring : with identity is a field if and only if : has no nonzero proper ideals. For if : is a field and ? is an ideal of : containing a nonzero element , then ~ c ? and so ? ~ : . Conversely, if : has no nonzero proper ideals and £ : , then the ideal º » must be : and so there is an : for which ~ . Hence, : is a field. Putting these two facts together proves the theorem. The following result says that maximal ideals always exist. Theorem 0.23 Any nonzero commutative ring 9 with identity contains a maximal ideal. Proof. Since 9 is not the zero ring, the ideal ¸¹ is a proper ideal of 9 . Hence, the set I of all proper ideals of 9 is nonempty. If 9 ~ ¸? 0¹ is a chain of proper ideals in 9 , then the union @ ~ 0 ? is also an ideal. Furthermore, if @ ~ 9 is not proper, then @ and so ? , for some 0 , which implies that ? ~ 9 is not proper. Hence, @ I . Thus, any chain in I has an upper bound in I and so Zorn's lemma implies that I has a maximal element. This shows that 9 has a maximal ideal.

Integral Domains Definition Let 9 be a ring. A nonzero element r 9 is called a zero divisor if there exists a nonzero 9 for which ~ . A commutative ring 9 with identity is called an integral domain if it contains no zero divisors. Example 0.14 If is not a prime number, then the ring { has zero divisors and so is not an integral domain. To see this, observe that if is not prime, then ~ in {, where Á . But in { , we have

24

Advanced Linear Algebra

p ~ mod ~ and so and are both zero divisors. As we will see later, if is a prime, then { is a field (which is an integral domain, of course). Example 0.15 The ring - ´%µ is an integral domain, since ²%³²%³ ~ implies that ²%³ ~ or ²%³ ~ . If 9 is a ring and % ~ & where Á %Á & 9 , then we cannot in general cancel the 's and conclude that % ~ &. For instance, in { , we have h ~ h , but canceling the 's gives ~ . However, it is precisely the integral domains in which we can cancel. The simple proof is left to the reader. Theorem 0.24 Let 9 be a commutative ring with identity. Then 9 is an integral domain if and only if the cancellation law % ~ &Á £ ¬ % ~ & holds.

The Field of Quotients of an Integral Domain Any integral domain 9 can be embedded in a field. The quotient field (or field of quotients) of 9 is a field that is constructed from 9 just as the field of rational numbers is constructed from the ring of integers. In particular, we set 9 b ~ ¸²Á ³ Á 9Á £ ¹ where ²Á ³ ~ ²Z Á Z ³ if and only if Z ~ Z . Addition and multiplication of fractions is defined by ²Á ³ b ²Á ³ ~ ² b Á ³ and ²Á ³ h ²Á ³ ~ ²Á ³ It is customary to write ²Á ³ in the form ° . Note that if 9 has zero divisors, then these definitions do not make sense, because may be even if and are not. This is why we require that 9 be an integral domain.

Principal Ideal Domains Definition Let 9 be a ring with identity and let 9 . The principal ideal generated by is the ideal º» ~ ¸ 9¹ An integral domain 9 in which every ideal is a principal ideal is called a principal ideal domain.

Preliminaries

25

Theorem 0.25 The integers form a principal ideal domain. In fact, any ideal ? in { is generated by the smallest positive integer a that is contained in ? . Theorem 0.26 The ring - ´%µ is a principal ideal domain. In fact, any ideal ? is generated by the unique monic polynomial of smallest degree contained in ? . Moreover, for polynomials ²%³Á Ã Á ²%³, º ²%³Á Ã Á ²%³» ~ ºgcd¸ ²%³Á Ã Á ²%³¹» Proof. Let ? be an ideal in - ´%µ and let ²%³ be a monic polynomial of smallest degree in ? . First, we observe that there is only one such polynomial in ? . For if ²%³ ? is monic and deg²²%³³ ~ deg²²%³³, then ²%³ ~ ²%³ c ²%³ ? and since deg²²%³³ deg²²%³³, we must have ²%³ ~ and so ²%³ ~ ²%³. We show that ? ~ º²%³». Since ²%³ ? , we have º²%³» ? . To establish the reverse inclusion, if ²%³ ? , then dividing ²%³ by ²%³ gives ²%³ ~ ²%³²%³ b ²%³ where ²%³ ~ or deg ²%³ deg ²%³. But since ? is an ideal, ²%³ ~ ²%³ c ²%³²%³ ? and so deg ²%³ deg ²%³ is impossible. Hence, ²%³ ~ and ²%³ ~ ²%³²%³ º²%³» This shows that ? º²%³» and so ? ~ º²%³». To prove the second statement, let ? ~ º ²%³Á Ã Á ²%³». Then, by what we have just shown, ? ~ º ²%³Á Ã Á ²%³» ~ º²%³» where ²%³ is the unique monic polynomial ²%³ in ? of smallest degree. In particular, since ²%³ º²%³», we have ²%³ ²%³ for each ~ Á Ã Á . In other words, ²%³ is a common divisor of the ²%³'s. Moreover, if ²%³ ²%³ for all , then ²%³ º²%³» for all , which implies that ²%³ º²%³» ~ º ²%³Á Ã Á ²%³» º²%³» and so ²%³ ²%³. This shows that ²%³ is the greatest common divisor of the ²%³'s and completes the proof.

26

Advanced Linear Algebra

Example 0.16 The ring 9 ~ - ´%Á &µ of polynomials in two variables % and & is not a principal ideal domain. To see this, observe that the set ? of all polynomials with zero constant term is an ideal in 9 . Now, suppose that ? is the principal ideal ? ~ º²%Á &³». Since %Á & ? , there exist polynomials ²%Á &³ and ²%Á &³ for which % ~ ²%Á &³²%Á &³ and & ~ ²%Á &³²%Á &³

(0.1)

But ²%Á &³ cannot be a constant, for then we would have ? ~ 9 . Hence, deg²²%Á &³³ and so ²%Á &³ and ²%Á &³ must both be constants, which implies that (0.1) cannot hold. Theorem 0.27 Any principal ideal domain 9 satisfies the ascending chain condition, that is, 9 cannot have a strictly increasing sequence of ideals ? ? Ä where each ideal is properly contained in the next one. Proof. Suppose to the contrary that there is such an increasing sequence of ideals. Consider the ideal < ~ ? which must have the form < ~ º» for some < . Since ? for some , we have ? ~ ? for all , contradicting the fact that the inclusions are proper.

Prime and Irreducible Elements We can define the notion of a prime element in any integral domain. For Á 9 , we say that divides (written ) if there exists an % 9 for which ~ %. Definition Let 9 be an integral domain. 1) An invertible element of 9 is called a unit. Thus, " 9 is a unit if "# ~ for some # 9 . 2) Two elements Á 9 are said to be associates if there exists a unit " for which ~ " . We denote this by writing . 3) A nonzero nonunit 9 is said to be prime if ¬ or 4) A nonzero nonunit 9 is said to be irreducible if ~ ¬ or is a unit Note that if is prime or irreducible, then so is " for any unit ". The property of being associate is clearly an equivalence relation.

Preliminaries

27

Definition We will refer to the equivalence classes under the relation of being associate as the associate classes of 9 . Theorem 0.28 Let 9 be a ring. 1) An element " 9 is a unit if and only if º"» ~ 9 . 2) if and only if º» ~ º ». 3) divides if and only if º » º». 4) properly divides , that is, ~ % where % is not a unit, if and only if º » º». In the case of the integers, an integer is prime if and only if it is irreducible. In any integral domain, prime elements are irreducible, but the converse need not hold. (In the ring {´jc µ ~ ¸ b jc Á {¹ the irreducible element divides the product ² b jc ³² c jc ³ ~ but does not divide either factor.) However, in principal ideal domains, the two concepts are equivalent. Theorem 0.29 Let 9 be a principal ideal domain. 1) An 9 is irreducible if and only if the ideal º» is maximal. 2) An element in 9 is prime if and only if it is irreducible. 3) The elements Á 9 are relatively prime, that is, have no common nonunit factors, if and only if there exist Á 9 for which b ~ This is denoted by writing ²Á ³ ~ . Proof. To prove 1), suppose that is irreducible and that º» º» 9 . Then º» and so ~ % for some % 9. The irreducibility of implies that or % is a unit. If is a unit, then º» ~ 9 and if % is a unit, then º» ~ º%» ~ º». This shows that º» is maximal. (We have º» £ 9 , since is not a unit.) Conversely, suppose that is not irreducible, that is, ~ where neither nor is a unit. Then º» º» 9 . But if º» ~ º», then , which implies that is a unit. Hence º» £ º». Also, if º» ~ 9 , then must be a unit. So we conclude that º» is not maximal, as desired. To prove 2), assume first that is prime and ~ . Then or . We may assume that . Therefore, ~ % ~ % . Canceling 's gives ~ % and so is a unit. Hence, is irreducible. (Note that this argument applies in any integral domain.) Conversely, suppose that is irreducible and let . We wish to prove that or . The ideal º» is maximal and so ºÁ » ~ º» or ºÁ » ~ 9 . In the former case, and we are done. In the latter case, we have ~ % b &

28

Advanced Linear Algebra

for some %Á & 9 . Thus, ~ % b & and since divides both terms on the right, we have . To prove 3), it is clear that if b ~ , then and are relatively prime. For the converse, consider the ideal ºÁ », which must be principal, say ºÁ » ~ º%». Then % and % and so % must be a unit, which implies that ºÁ » ~ 9 . Hence, there exist Á 9 for which b ~ .

Unique Factorization Domains Definition An integral domain 9 is said to be a unique factorization domain if it has the following factorization properties: 1) Every nonzero nonunit element 9 can be written as a product of a finite number of irreducible elements ~ Ä . 2) The factorization into irreducible elements is unique in the sense that if ~ Ä and ~ Ä are two such factorizations, then ~ and after a suitable reindexing of the factors, . Unique factorization is clearly a desirable property. Fortunately, principal ideal domains have this property. Theorem 0.30 Every principal ideal domain 9 is a unique factorization domain. Proof. Let 9 be a nonzero nonunit. If is irreducible, then we are done. If not, then ~ , where neither factor is a unit. If and are irreducible, we are done. If not, suppose that is not irreducible. Then ~ , where neither nor is a unit. Continuing in this way, we obtain a factorization of the form (after renumbering if necessary) ~ ~ ² ³ ~ ² ³² ³ ~ ² ³² ³ ~ Ä Each step is a factorization of into a product of nonunits. However, this process must stop after a finite number of steps, for otherwise it will produce an infinite sequence Á Á Ã of nonunits of 9 for which b properly divides . But this gives the ascending chain of ideals º » º » º » º » Ä where the inclusions are proper. But this contradicts the fact that a principal ideal domain satisfies the ascending chain condition. Thus, we conclude that every nonzero nonunit has a factorization into irreducible elements. As to uniqueness, if ~ Ä and ~ Ä are two such factorizations, then because 9 is an integral domain, we may equate them and cancel like factors, so let us assume this has been done. Thus, £ for all Á . If there are no factors on either side, we are done. If exactly one side has no factors left,

Preliminaries

29

then we have expressed as a product of irreducible elements, which is not possible since irreducible elements are nonunits. Suppose that both sides have factors left, that is, Ä ~ Ä where £ . Then Ä , which implies that for some . We can assume by reindexing if necessary that ~ . Since is irreducible must be a unit. Replacing by and canceling gives Äc ~ Äc This process can be repeated until we run out of 's or 's. If we run out of 's first, then we have an equation of the form " Ä ~ where " is a unit, which is not possible since the 's are not units. By the same reasoning, we cannot run out of 's first and so ~ and the 's and 's can be paired off as associates.

Fields For the record, let us give the definition of a field (a concept that we have been using). Definition A field is a set - , containing at least two elements, together with two binary operations, called addition (denoted by b ) and multiplication (denoted by juxtaposition), for which the following hold: 1) - is an abelian group under addition. 2) The set - i of all nonzero elements in - is an abelian group under multiplication. 3) (Distributivity) For all Á Á - , ² b ³ ~ b and ² b ³ ~ b

We require that - have at least two elements to avoid the pathological case in which ~ . Example 0.17 The sets r, s and d, of all rational, real and complex numbers, respectively, are fields, under the usual operations of addition and multiplication of numbers. Example 0.18 The ring { is a field if and only if is a prime number. We have already seen that { is not a field if is not prime, since a field is also an integral domain. Now suppose that ~ is a prime. We have seen that { is an integral domain and so it remains to show that every nonzero element in { has a multiplicative inverse. Let £ { . Since , we know that and are relatively prime. It follows that there exist integers " and # for which

30

Advanced Linear Algebra

" b # ~ Hence, " ² c #³ mod and so " p ~ in { , that is, " is the multiplicative inverse of . The previous example shows that not all fields are infinite sets. In fact, finite fields play an extremely important role in many areas of abstract and applied mathematics. A field - is said to be algebraically closed if every nonconstant polynomial over - has a root in - . This is equivalent to saying that every nonconstant polynomial splits over - . For example, the complex field d is algebraically closed but the real field s is not. We mention without proof that every field - is contained in an algebraically closed field - , called the algebraic closure of - . For example, the algebraic closure of the real field is the complex field.

The Characteristic of a Ring Let 9 be a ring with identity. If is a positive integer, then by h , we simply mean h ~ bÄb terms Now, it may happen that there is a positive integer for which h~ For instance, in { , we have h ~ ~ . On the other hand, in {, the equation h ~ implies ~ and so no such positive integer exists. Notice that in any finite ring, there must exist such a positive integer , since the members of the infinite sequence of numbers h Á h Á h Á Ã cannot be distinct and so h ~ h for some , whence ² c ³ h ~ . Definition Let 9 be a ring with identity. The smallest positive integer for which h ~ is called the characteristic of 9 . If no such number exists, we say that 9 has characteristic . The characteristic of 9 is denoted by char²9³. If char²9³ ~ , then for any 9 , we have h ~ b Ä b ~ ² b Ä b ³ ~ h ~ terms terms

Preliminaries

31

Theorem 0.31 Any finite ring has nonzero characteristic. Any finite integral domain has prime characteristic. Proof. We have already seen that a finite ring has nonzero characteristic. Let be a finite integral domain and suppose that char²- ³ ~ . If ~ , where Á , then h ~ . Hence, ² h ³² h ³ ~ , implying that h ~ or h ~ . In either case, we have a contradiction to the fact that is the smallest positive integer such that h ~ . Hence, must be prime. Notice that in any field - of characteristic , we have ~ for all - . Thus, in - , ~ c for all This property takes a bit of getting used to and makes fields of characteristic quite exceptional. (As it happens, there are many important uses for fields of characteristic .) It can be shown that all finite fields have size equal to a positive integral power of a prime and for each prime power , there is a finite field of size . In fact, up to isomorphism, there is exactly one finite field of size .

Algebras The final algebraic structure of which we will have use is a combination of a vector space and a ring. (We have not yet officially defined vector spaces, but we will do so before needing the following definition, which is placed here for easy reference.) Definition An algebra 7 over a field - is a nonempty set 7, together with three operations, called addition (denoted by b ), multiplication (denoted by juxtaposition) and scalar multiplication (also denoted by juxtaposition), for which the following properties hold: 1) 7 is a vector space over - under addition and scalar multiplication. 2) 7 is a ring under addition and multiplication. 3) If - and Á 7, then ²³ ~ ²³ ~ ²³

Thus, an algebra is a vector space in which we can take the product of vectors, or a ring in which we can multiply each element by a scalar (subject, of course, to additional requirements as given in the definition).

Part I—Basic Linear Algebra

Chapter 1

Vector Spaces

Vector Spaces Let us begin with the definition of one of our principal objects of study. Definition Let - be a field, whose elements are referred to as scalars. A vector space over - is a nonempty set = , whose elements are referred to as vectors, together with two operations. The first operation, called addition and denoted by b , assigns to each pair ²"Á #³ of vectors in = a vector " b # in = . The second operation, called scalar multiplication and denoted by juxtaposition, assigns to each pair ²Á "³ - d = a vector " in = . Furthermore, the following properties must be satisfied: 1) (Associativity of addition) For all vectors "Á #Á $ = , " b ²# b $³ ~ ²" b #³ b $ 2) (Commutativity of addition) For all vectors "Á # = , "b#~#b" 3) (Existence of a zero) There is a vector = with the property that b"~"b~" for all vectors " = . 4) (Existence of additive inverses) For each vector " = , there is a vector in = , denoted by c", with the property that " b ²c"³ ~ ²c"³ b " ~

36

Advanced Linear Algebra

5) (Properties of scalar multiplication) For all scalars Á F and for all vectors "Á # = , ²" b #³ ~ " b # ² b ³" ~ " b " ²³" ~ ²"³ " ~ "

Note that the first four properties in the definition of vector space can be summarized by saying that = is an abelian group under addition. A vector space over a field - is sometimes called an - -space. A vector space over the real field is called a real vector space and a vector space over the complex field is called a complex vector space. Definition Let : be a nonempty subset of a vector space = . A linear combination of vectors in : is an expression of the form # b Ä b # where # Á Ã Á # : and Á Ã Á - . The scalars are called the coefficients of the linear combination. A linear combination is trivial if every coefficient is zero. Otherwise, it is nontrivial.

Examples of Vector Spaces Here are a few examples of vector spaces. Example 1.1 1) Let - be a field. The set - - of all functions from - to - is a vector space over - , under the operations of ordinary addition and scalar multiplication of functions: ² b ³²%³ ~ ²%³ b ²%³ and ² ³²%³ ~ ² ²%³³ 2) The set CÁ ²- ³ of all d matrices with entries in a field - is a vector space over - , under the operations of matrix addition and scalar multiplication. 3) The set - of all ordered -tuples whose components lie in a field - , is a vector space over - , with addition and scalar multiplication defined componentwise: ² Á Ã Á ³ b ² Á Ã Á ³ ~ ² b Á Ã Á b ³ and

Vector Spaces

37

² Á Ã Á ³ ~ ² Á Ã Á ³ When convenient, we will also write the elements of - in column form. When - is a finite field - with elements, we write = ²Á ³ for - . 4) Many sequence spaces are vector spaces. The set Seq²- ³ of all infinite sequences with members from a field - is a vector space under the componentwise operations ² ³ b ²! ³ ~ ²

b ! ³

and ² ³ ~ ² ³ In a similar way, the set of all sequences of complex numbers that converge to is a vector space, as is the set MB of all bounded complex sequences. Also, if is a positive integer, then the set M of all complex sequences ² ³ for which B

( ( B ~

is a vector space under componentwise operations. To see that addition is a binary operation on M , one verifies Minkowski's inequality B

8 ( ~

° b ! ( 9

B

8 ( ( 9 ~

°

B

°

b 8 (! ( 9 ~

which we will not do here.

Subspaces Most algebraic structures contain substructures, and vector spaces are no exception. Definition A subspace of a vector space = is a subset : of = that is a vector space in its own right under the operations obtained by restricting the operations of = to : . We use the notation : = to indicate that : is a subspace of = and : = to indicate that : is a proper subspace of = , that is, : = but : £ = . The zero subspace of = is ¸¹. Since many of the properties of addition and scalar multiplication hold a fortiori in a nonempty subset : , we can establish that : is a subspace merely by checking that : is closed under the operations of = . Theorem 1.1 A nonempty subset : of a vector space = is a subspace of = if and only if : is closed under addition and scalar multiplication or, equivalently,

38

Advanced Linear Algebra

: is closed under linear combinations, that is, Á - Á "Á # : ¬ " b # :

Example 1.2 Consider the vector space = ²Á ³ of all binary -tuples, that is, -tuples of 's and 's. The weight M ²#³ of a vector # = ²Á ³ is the number of nonzero coordinates in #. For instance, M ²³ ~ . Let , be the set of all vectors in = of even weight. Then , is a subspace of = ²Á ³. To see this, note that M ²" b #³ ~ M ²"³ b M ²#³ c M ²" q #³ where " q # is the vector in = ²Á ³ whose th component is the product of the th components of " and #, that is, ²" q #³ ~ " h # Hence, if M ²"³ and M ²#³ are both even, so is M ²" b #³. Finally, scalar multiplication over - is trivial and so , is a subspace of = ²Á ³, known as the even weight subspace of = ²Á ³. Example 1.3 Any subspace of the vector space = ²Á ³ is called a linear code. Linear codes are among the most important and most studied types of codes, because their structure allows for efficient encoding and decoding of information.

The Lattice of Subspaces The set I²= ³ of all subspaces of a vector space = is partially ordered by set inclusion. The zero subspace ¸¹ is the smallest element in I ²= ³ and the entire space = is the largest element. If :Á ; I ²= ³, then : q ; is the largest subspace of = that is contained in both : and ; . In terms of set inclusion, : q ; is the greatest lower bound of : and ; : : q ; ~ glb¸:Á ; ¹ Similarly, if ¸: 2¹ is any collection of subspaces of = , then their intersection is the greatest lower bound of the subspaces:

: ~ glb¸: 2¹ 2

On the other hand, if :Á ; I ²= ³ (and - is infinite), then : r ; I ²= ³ if and only if : ; or ; : . Thus, the union of two subspaces is never a subspace in any “interesting” case. We also have the following.

Vector Spaces

39

Theorem 1.2 A nontrivial vector space = over an infinite field - is not the union of a finite number of proper subspaces. Proof. Suppose that = ~ : r Ä r : , where we may assume that : \ : r Ä r : Let $ : ± ²: r Ä r : ³ and let # ¤ : . Consider the infinite set ( ~ ¸$ b # - ¹ which is the “line” through #, parallel to $. We want to show that each : contains at most one vector from the infinite set (, which is contrary to the fact that = ~ : r Ä r : . This will prove the theorem. If $ b # : for £ , then $ : implies # : , contrary to assumption. Next, suppose that $ b # : and $ b # : , for , where £ . Then : ² $ b #³ c ² $ b #³ ~ ² c ³$ and so $ : , which is also contrary to assumption. To determine the smallest subspace of = containing the subspaces : and ; , we make the following definition. Definition Let : and ; be subspaces of = . The sum : b ; is defined by : b ; ~ ¸" b # " :Á # ; ¹ More generally, the sum of any collection ¸: 2¹ of subspaces is the set of all finite sums of vectors from the union : : : ~ H 2

bÄb

c

: I

2

It is not hard to show that the sum of any collection of subspaces of = is a subspace of = and that the sum is the least upper bound under set inclusion: : b ; ~ lub¸:Á ; ¹ More generally, : ~ lub¸: 2¹ 2

If a partially ordered set 7 has the property that every pair of elements has a least upper bound and greatest lower bound, then 7 is called a lattice. If 7 has a smallest element and a largest element and has the property that every collection of elements has a least upper bound and greatest lower bound, then 7

40

Advanced Linear Algebra

is called a complete lattice. The least upper bound of a collection is also called the join of the collection and the greatest lower bound is called the meet. Theorem 1.3 The set I²= ³ of all subspaces of a vector space = is a complete lattice under set inclusion, with smallest element ¸¹, largest element = , meet glb¸: 2¹ ~ : 2

and join lub¸: 2¹ ~ :

2

Direct Sums As we will see, there are many ways to construct new vector spaces from old ones.

External Direct Sums Definition Let = Á Ã Á = be vector spaces over a field - . The external direct sum of = Á Ã Á = , denoted by = ~ = ^ Ä ^ = is the vector space = whose elements are ordered -tuples: = ~ ¸²# Á Ã Á # ³ # = Á ~ Á Ã Á ¹ with componentwise operations ²" Á Ã Á " ³ b ²# Á Ã Á # ³ ~ ²" b # Á Ã Á " b #³ and ²# Á Ã Á # ³ ~ ²# Á Ã Á # ³ for all - . Example 1.4 The vector space - is the external direct sum of copies of - , that is, - ~ - ^ Ä ^ where there are summands on the right-hand side. This construction can be generalized to any collection of vector spaces by generalizing the idea that an ordered -tuple ²# Á Ã Á # ³ is just a function ¢ ¸Á Ã Á ¹ ¦ = from the index set ¸Á Ã Á ¹ to the union of the spaces with the property that ²³ = .

Vector Spaces

41

Definition Let < ~ ¸= 2¹ be any family of vector spaces over - . The direct product of < is the vector space = ~ H ¢ 2 ¦ = d ²³ = I

2

2

thought of as a subspace of the vector space of all functions from 2 to = . It will prove more useful to restrict the set of functions to those with finite support. Definition Let < ~ ¸= 2¹ be a family of vector spaces over - . The support of a function ¢ 2 ¦ = is the set supp² ³ ~ ¸ 2 ²³ £ ¹ Thus, a function has finite support if ²³ ~ for all but a finite number of 2. The external direct sum of the family < is the vector space ext

= ~ H ¢ 2 ¦ = d ²³ = , has finite supportI 2

2

thought of as a subspace of the vector space of all functions from 2 to = . An important special case occurs when = ~ = for all 2 . If we let = 2 denote the set of all functions from 2 to = and ²= 2 ³ denote the set of all functions in = 2 that have finite support, then ext

= ~ = 2 and = ~ ²= 2 ³ 2

2

Note that the direct product and the external direct sum are the same for a finite family of vector spaces.

Internal Direct Sums An internal version of the direct sum construction is often more relevant. Definition A vector space = is the (internal) direct sum of a family < ~ ¸: 0¹ of subspaces of = , written = ~ <

or = ~ : 0

if the following hold:

42

Advanced Linear Algebra

1) (Join of the family) = is the sum (join) of the family < : = ~ : 0

2) (Independence of the family) For each 0 , : q

p

:

q £

s t

~ ¸¹

In this case, each : is called a direct summand of = . If < ~ ¸: Á Ã Á : ¹ is a finite family, the direct sum is often written = ~ : l Ä l : Finally, if = ~ : l ; , then ; is called a complement of : in = . Note that the condition in part 2) of the previous definition is stronger than saying simply that the members of < are pairwise disjoint: : q : ~ J for all £ 0 . A word of caution is in order here: If : and ; are subspaces of = , then we may always say that the sum : b ; exists. However, to say that the direct sum of : and ; exists or to write : l ; is to imply that : q ; ~ ¸¹. Thus, while the sum of two subspaces always exists, the direct sum of two subspaces does not always exist. Similar statements apply to families of subspaces of = . The reader will be asked in a later chapter to show that the concepts of internal and external direct sum are essentially equivalent (isomorphic). For this reason, the term “direct sum” is often used without qualification. Once we have discussed the concept of a basis, the following theorem can be easily proved. Theorem 1.4 Any subspace of a vector space has a complement, that is, if : is a subspace of = , then there exists a subspace ; for which = ~ : l ; . It should be emphasized that a subspace generally has many complements (although they are isomorphic). The reader can easily find examples of this in s . We can characterize the uniqueness part of the definition of direct sum in other useful ways. First a remark. If : and ; are distinct subspaces of = and if %Á & : q ; , then the sum % b & can be thought of as a sum of vectors from the

Vector Spaces

43

same subspace (say : ) or from different subspaces—one from : and one from ; . When we say that a vector # cannot be written as a sum of vectors from the distinct subspaces : and ; , we mean that # cannot be written as a sum % b & where % and & can be interpreted as coming from different subspaces, even if they can also be interpreted as coming from the same subspace. Thus, if %Á & : q ; , then # ~ % b & does express # as a sum of vectors from distinct subspaces. Theorem 1.5 Let < ~ ¸: 0¹ be a family of distinct subspaces of = . The following are equivalent: 1) (Independence of the family) For each 0 , : q

p

:

q £

s t

~ ¸¹

2) (Uniqueness of expression for ) The zero vector cannot be written as a sum of nonzero vectors from distinct subspaces of < . 3) (Uniqueness of expression) Every nonzero # = has a unique, except for order of terms, expression as a sum #~

bÄb

of nonzero vectors from distinct subspaces in < . Hence, a sum = ~ : 0

is direct if and only if any one of 1)–3) holds. Proof. Suppose that 2) fails, that is, ~ where the nonzero

's

bÄb

are from distinct subspaces : . Then and so c

~

bÄb

which violates 1). Hence, 1) implies 2). If 2) holds and #~

bÄb

where the terms are nonzero and the similarily for the ! 's, then ~

bÄb

and # ~ ! b Ä b ! 's

belong to distinct subspaces in < and c ! c Ä c !

By collecting terms from the same subspaces, we may write ~²

c ! ³ b Ä b ²

c ! ³ b

b

bÄb

c !b c Ä c !

44

Advanced Linear Algebra

Then 2) implies that ~ ~ and implies 3).

"

~ !" for all " ~ Á Ã Á . Hence, 2)

Finally, suppose that 3) holds. If £ # : q then # ~

:

q £

s t

: and

where

p

~

bÄb

: are nonzero. But this violates 3).

Example 1.5 Any matrix ( C can be written in the form (~

²( b (! ³ b ²( c (! ³ ~ ) b *

(1.1)

where (! is the transpose of (. It is easy to verify that ) is symmetric and * is skew-symmetric and so (1.1) is a decomposition of ( as the sum of a symmetric matrix and a skew-symmetric matrix. Since the sets Sym and SkewSym of all symmetric and skew-symmetric matrices in C are subspaces of C , we have C ~ Sym b SkewSym Furthermore, if : b ; ~ : Z b ; Z , where : and : Z are symmetric and ; and ; Z are skew-symmetric, then the matrix < ~ : c :Z ~ ; Z c ; is both symmetric and skew-symmetric. Hence, provided that char²- ³ £ , we must have < ~ and so : ~ : Z and ; ~ ; Z . Thus, C ~ Sym l SkewSym

Spanning Sets and Linear Independence A set of vectors spans a vector space if every vector can be written as a linear combination of some of the vectors in that set. Here is the formal definition. Definition The subspace spanned (or subspace generated) by a nonempty set : of vectors in = is the set of all linear combinations of vectors from : : º:» ~ span²:³ ~ ¸ # b Ä b # - Á # :¹

Vector Spaces

45

When : ~ ¸# Á Ã Á # ¹ is a finite set, we use the notation º# Á Ã Á # » or span²# Á Ã Á # ³. A set : of vectors in = is said to span = , or generate = , if = ~ span²:³. It is clear that any superset of a spanning set is also a spanning set. Note also that all vector spaces have spanning sets, since = spans itself.

Linear Independence Linear independence is a fundamental concept. Definition Let = be a vector space. A nonempty set : of vectors in = is linearly independent if for any distinct vectors Á Ã Á in : ,

b Ä b

~

¬

~

for all

In words, : is linearly independent if the only linear combination of vectors from : that is equal to is the trivial linear combination, all of whose coefficients are . If : is not linearly independent, it is said to be linearly dependent. It is immediate that a linearly independent set of vectors cannot contain the zero vector, since then h ~ violates the condition of linear independence. Another way to phrase the definition of linear independence is to say that : is linearly independent if the zero vector has an “as unique as possible” expression as a linear combination of vectors from : . We can never prevent the zero vector from being written in the form ~ b Ä b , but we can prevent from being written in any other way as a linear combination of the vectors in : . For the introspective reader, the expression ~ b ² c ³ has two interpretations. One is ~ b where ~ and ~ c, but this does not involve distinct vectors so is not relevant to the question of linear independence. The other interpretation is ~ b ! where ! ~ c £ (assuming that £ ). Thus, if : is linearly independent, then : cannot contain both and c . Definition Let : be a nonempty set of vectors in = . To say that a nonzero vector # = is an essentially unique linear combination of the vectors in : is to say that, up to order of terms, there is one and only one way to express # as a linear combination # ~

b Ä b

where the 's are distinct vectors in : and the coefficients are nonzero. More explicitly, # £ is an essentially unique linear combination of the vectors in : if # º:» and if whenever

46

Advanced Linear Algebra

# ~

b Ä b

and # ~ ! b Ä b !

where the 's are distinct, the ! 's are distinct and all coefficients are nonzero, then ~ and after a reindexing of the ! 's if necessary, we have ~ and ~ ! for all ~ Á Ã Á . (Note that this is stronger than saying that ~ ! .) We may characterize linear independence as follows. Theorem 1.6 Let : £ ¸¹ be a nonempty set of vectors in = . The following are equivalent: 1) : is linearly independent. 2) Every nonzero vector # span²:³ is an essentially unique linear combination of the vectors in : . 3) No vector in : is a linear combination of other vectors in : . Proof. Suppose that 1) holds and that £ # ~

b Ä b

~ ! b Ä b !

where the 's are distinct, the ! 's are distinct and the coefficients are nonzero. By subtracting and grouping 's and !'s that are equal, we can write ~ ² c ³ b Ä b ² c ³ b b b b Ä b c b !b c Ä c ! and so 1) implies that ~ ~ and " ~ " and Thus, 1) implies 2). If 2) holds and

"

~ !" for all ~ Á Ã Á .

: can be written as ~

b Ä b

where : are different from , then we may collect like terms on the right and then remove all terms with coefficient. The resulting expression violates 2). Hence, 2) implies 3). If 3) holds and where the

's

b Ä b

~

are distinct and £ , then and we may write

~c

²

b Ä b ³

which violates 3). The following key theorem relates the notions of spanning set and linear independence.

Vector Spaces

47

Theorem 1.7 Let : be a set of vectors in = . The following are equivalent: 1) : is linearly independent and spans = . 2) Every nonzero vector # = is an essentially unique linear combination of vectors in : . 3) : is a minimal spanning set, that is, : spans = but any proper subset of : does not span = . 4) : is a maximal linearly independent set, that is, : is linearly independent, but any proper superset of : is not linearly independent. A set of vectors in = that satisfies any (and hence all) of these conditions is called a basis for = . Proof. We have seen that 1) and 2) are equivalent. Now suppose 1) holds. Then : is a spanning set. If some proper subset : Z of : also spanned = , then any vector in : c : Z would be a linear combination of the vectors in : Z , contradicting the fact that the vectors in : are linearly independent. Hence 1) implies 3). Conversely, if : is a minimal spanning set, then it must be linearly independent. For if not, some vector : would be a linear combination of the other vectors in : and so : c ¸ ¹ would be a proper spanning subset of : , which is not possible. Hence 3) implies 1). Suppose again that 1) holds. If : were not maximal, there would be a vector # = c : for which the set : r ¸#¹ is linearly independent. But then # is not in the span of : , contradicting the fact that : is a spanning set. Hence, : is a maximal linearly independent set and so 1) implies 4). Conversely, if : is a maximal linearly independent set, then : must span = , for if not, we could find a vector # = c : that is not a linear combination of the vectors in : . Hence, : r ¸#¹ would be a linearly independent proper superset of :, which is a contradiction. Thus, 4) implies 1). Theorem 1.8 A finite set : ~ ¸# Á Ã Á # ¹ of vectors in = is a basis for = if and only if = ~ º# » l Ä l º# »

Example 1.6 The th standard vector in - is the vector that has 's in all coordinate positions except the th, where it has a . Thus, ~ ²Á Á Ã Á ³Á

~ ²Á Á Ã Á ³ Á Ã Á

~ ²Á Ã Á Á ³

The set ¸ Á Ã Á ¹ is called the standard basis for - . The proof that every nontrivial vector space has a basis is a classic example of the use of Zorn's lemma.

48

Advanced Linear Algebra

Theorem 1.9 Let = be a nonzero vector space. Let 0 be a linearly independent set in = and let : be a spanning set in = containing 0 . Then there is a basis 8 for = for which 0 8 : . In particular, 1) Any vector space, except the zero space ¸¹, has a basis. 2) Any linearly independent set in = is contained in a basis. 3) Any spanning set in = contains a basis. Proof. Consider the collection 7 of all linearly independent subsets of = containing 0 and contained in : . This collection is not empty, since 0 7 . Now, if 9 ~ ¸0 2¹ is a chain in 7 , then the union < ~ 0 2

is linearly independent and satisfies 0 < : , that is, < 7 . Hence, every chain in 7 has an upper bound in 7 and according to Zorn's lemma, 7 must contain a maximal element 8 , which is linearly independent. Now, 8 is a basis for the vector space º:» ~ = , for if any : is not a linear combination of the elements of 8 , then 8 r ¸ ¹ : is linearly independent, contradicting the maximality of 8 . Hence : º8 » and so = ~ º:» º8 ». The reader can now show, using Theorem 1.9, that any subspace of a vector space has a complement.

The Dimension of a Vector Space The next result, with its classical elegant proof, says that if a vector space = has a finite spanning set : , then the size of any linearly independent set cannot exceed the size of : . Theorem 1.10 Let = be a vector space and assume that the vectors # Á Ã Á # are linearly independent and the vectors Á Ã Á span = . Then . Proof. First, we list the two sets of vectors: the spanning set followed by the linearly independent set: Á Ã Á Â # Á Ã Á # Then we move the first vector # to the front of the first list: # Á Á Ã Á Â # Á Ã Á # Since Á Ã Á span = , # is a linear combination of the 's. This implies that we may remove one of the 's, which by reindexing if necessary can be , from the first list and still have a spanning set # Á Á Ã Á Â # Á Ã Á #

Vector Spaces

49

Note that the first set of vectors still spans = and the second set is still linearly independent. Now we repeat the process, moving # from the second list to the first list # Á # Á Á Ã Á Â # Á Ã Á # As before, the vectors in the first list are linearly dependent, since they spanned = before the inclusion of # . However, since the # 's are linearly independent, any nontrivial linear combination of the vectors in the first list that equals must involve at least one of the 's. Hence, we may remove that vector, which again by reindexing if necessary may be taken to be and still have a spanning set # Á # Á Á Ã Á Â # Á Ã Á # Once again, the first set of vectors spans = and the second set is still linearly independent. Now, if , then this process will eventually exhaust the list # Á # Á Ã Á # Â #b Á Ã Á #

's

and lead to the

where # Á # Á Ã Á # span = , which is clearly not possible since # is not in the span of # Á # Á Ã Á # . Hence, . Corollary 1.11 If = has a finite spanning set, then any two bases of = have the same size. Now let us prove the analogue of Corollary 1.11 for arbitrary vector spaces. Theorem 1.12 If = is a vector space, then any two bases for = have the same cardinality. Proof. We may assume that all bases for = are infinite sets, for if any basis is finite, then = has a finite spanning set and so Corollary 1.11 applies. Let 8 ~ ¸ 0¹ be a basis for = and let 9 be another basis for = . Then any vector 9 can be written as a finite linear combination of the vectors in 8 , where all of the coefficients are nonzero, say ~ <

But because 9 is a basis, we must have < ~ 0 9

50

Advanced Linear Algebra

for if the vectors in 9 can be expressed as finite linear combinations of the vectors in a proper subset 8 Z of 8 , then 8 Z spans = , which is not the case. Since (< ( L for all 9, Theorem 0.17 implies that (8 ( ~ (0 ( L (9( ~ (9( But we may also reverse the roles of 8 and 9, to conclude that (9( (8 ( and so the Schro¨der–Bernstein theorem implies that (8 ( ~ (9(. Theorem 1.12 allows us to make the following definition. Definition A vector space = is finite-dimensional if it is the zero space ¸¹, or if it has a finite basis. All other vector spaces are infinite-dimensional. The dimension of the zero space is and the dimension of any nonzero vector space = is the cardinality of any basis for = . If a vector space = has a basis of cardinality , we say that = is -dimensional and write dim²= ³ ~ . It is easy to see that if : is a subspace of = , then dim²:³ dim²= ³. If in addition, dim²:³ ~ dim²= ³ B, then : ~ = . Theorem 1.13 Let = be a vector space. 1) If 8 is a basis for = and if 8 ~ 8 r 8 and 8 q 8 ~ J, then = ~ º8 » l º8 » 2) Let = ~ : l ; . If 8 is a basis for : and 8 is a basis for ; , then 8 q 8 ~ J and 8 ~ 8 r 8 is a basis for = . Theorem 1.14 Let : and ; be subspaces of a vector space = . Then dim²:³ b dim²; ³ ~ dim²: b ; ³ b dim²: q ; ³ In particular, if ; is any complement of : in = , then dim²:³ b dim²; ³ ~ dim²= ³ that is, dim²: l ; ³ ~ dim²:³ b dim²; ³ Proof. Suppose that 8 ~ ¸ 0¹ is a basis for : q ; . Extend this to a basis 7 r 8 for : where 7 ~ ¸ 1 ¹ is disjoint from 8 . Also, extend 8 to a basis 8 r 9 for ; where 9 ~ ¸ 2¹ is disjoint from 8 . We claim that 7 r 8 r 9 is a basis for : b ; . It is clear that º7 r 8 r 9» ~ : b ; . To see that 7 r 8 r 9 is linearly independent, suppose to the contrary that

Vector Spaces

51

# b Ä b # ~ where # 7 r 8 r 9 and £ for all . There must be vectors # in this expression from both 7 and 9, since 7 r 8 and 8 r 9 are linearly independent. Isolating the terms involving the vectors from 7 on one side of the equality shows that there is a nonzero vector in % º7» q º8 r 9». But then % : q ; and so % º7» q º8 », which implies that % ~ , a contradiction. Hence, 7 r 8 r 9 is linearly independent and a basis for : b ; . Now, dim²:³ b dim²; ³ ~ (7 r 8 ( b (8 r 9( ~ (7( b (8 ( b (8 ( b (9( ~ (7( b (8 ( b (9( b dim²: q ; ³ ~ dim²: b ; ³ b dim²: q ; ³ as desired. It is worth emphasizing that while the equation dim²:³ b dim²; ³ ~ dim²: b ; ³ b dim²: q ; ³ holds for all vector spaces, we cannot write dim²: b ; ³ ~ dim²:³ b dim²; ³ c dim²: q ; ³ unless : b ; is finite-dimensional.

Ordered Bases and Coordinate Matrices It will be convenient to consider bases that have an order imposed on their members. Definition Let = be a vector space of dimension . An ordered basis for = is an ordered -tuple ²# Á Ã Á # ³ of vectors for which the set ¸# Á Ã Á # ¹ is a basis for = . If 8 ~ ²# Á Ã Á # ³ is an ordered basis for = , then for each # = there is a unique ordered -tuple ² Á Ã Á ³ of scalars for which # ~ # b Ä b # Accordingly, we can define the coordinate map 8 ¢ = ¦ - by 8 ²#³ ~ ´#µ8 ~

v y Å w z

(1.3)

52

Advanced Linear Algebra

where the column matrix ´#µ8 is known as the coordinate matrix of # with respect to the ordered basis 8 . Clearly, knowing ´#µ8 is equivalent to knowing # (assuming knowledge of 8 ). Furthermore, it is easy to see that the coordinate map 8 is bijective and preserves the vector space operations, that is, 8 ² # b Ä b # ³ ~ 8 ²# ³ b Ä b 8 ²# ³ or equivalently ´ # b Ä b # µ8 ~ ´# µ8 b Ä b ´# µ8 Functions from one vector space to another that preserve the vector space operations are called linear transformations and form the objects of study in the next chapter.

The Row and Column Spaces of a Matrix Let ( be an d matrix over - . The rows of ( span a subspace of - known as the row space of ( and the columns of ( span a subspace of - known as the column space of (. The dimensions of these spaces are called the row rank and column rank, respectively. We denote the row space and row rank by rs²(³ and rrk²(³ and the column space and column rank by cs²(³ and crk²(³. It is a remarkable and useful fact that the row rank of a matrix is always equal to its column rank, despite the fact that if £ , the row space and column space are not even in the same vector space! Our proof of this fact hinges on the following simple observation about matrices. Lemma 1.15 Let ( be an d matrix. Then elementary column operations do not affect the row rank of (. Similarly, elementary row operations do not affect the column rank of (. Proof. The second statement follows from the first by taking transposes. As to the first, the row space of ( is rs²(³ ~ º (Á Ã Á (» where are the standard basis vectors in - . Performing an elementary column operation on ( is equivalent to multiplying ( on the right by an elementary matrix , . Hence the row space of (, is rs²(,³ ~ º (,Á Ã Á (,» and since , is invertible,

Vector Spaces

53

rrk²(³ ~ dim²rs²(³³ ~ dim²rs²(,³³ ~ rrk²(,³ as desired. Theorem 1.16 If ( CÁ , then rrk²(³ ~ crk²(³. This number is called the rank of ( and is denoted by rk²(³. Proof. According to the previous lemma, we may reduce ( to reduced column echelon form without affecting the row rank. But this reduction does not affect the column rank either. Then we may further reduce ( to reduced row echelon form without affecting either rank. The resulting matrix 4 has the same row and column ranks as (. But 4 is a matrix with 's followed by 's on the main diagonal (entries 4Á Á 4Á Á Ã ) and 's elsewhere. Hence, rrk²(³ ~ rrk²4 ³ ~ crk²4 ³ ~ crk²(³ as desired.

The Complexification of a Real Vector Space If > is a complex vector space (that is, a vector space over d), then we can think of > as a real vector space simply by restricting all scalars to the field s. Let us denote this real vector space by >s and call it the real version of > . On the other hand, to each real vector space = , we can associate a complex vector space = d . This “complexification” process will play a useful role when we discuss the structure of linear operators on a real vector space. (Throughout our discussion = will denote a real vector space.) Definition If = is a real vector space, then the set = d ~ = d = of ordered pairs, with componentwise addition ²"Á #³ b ²%Á &³ ~ ²" b %Á # b &³ and scalar multiplication over d defined by ² b ³²"Á #³ ~ ²" c #Á # b "³ for Á s is a complex vector space, called the complexification of = . It is convenient to introduce a notation for vectors in = d that resembles the notation for complex numbers. In particular, we denote ²"Á #³ = d by " b # and so = d ~ ¸" b # "Á # = ¹ Addition now looks like ordinary addition of complex numbers, ²" b #³ b ²% b &³ ~ ²" b %³ b ²# b &³ and scalar multiplication looks like ordinary multiplication of complex numbers,

54

Advanced Linear Algebra

² b ³²" b #³ ~ ²" c #³ b ²# b "³ Thus, for example, we immediately have for Á s, ²" b #³ ~ " b # ²" b #³ ~ c# b " ² b ³" ~ " b " ² b ³# ~ c# b # The real part of ' ~ " b # is " = and the imaginary part of ' is # = . The essence of the fact that ' ~ " b # = d is really an ordered pair is that ' is if and only if its real and imaginary parts are both . We can define the complexification map cpx¢ = ¦ = d by cpx²#³ ~ # b Let us refer to # b as the complexification, or complex version of # = . Note that this map is a group homomorphism, that is, cpx²³ ~ b

and

cpx²" f #³ ~ cpx²"³ f cpx²#³

and it is injective: cpx²"³ ~ cpx²#³ ¯ " ~ # Also, it preserves multiplication by real scalars: cpx²"³ ~ " b ~ ²" b ³ ~ cpx²"³ for s. However, the complexification map is not surjective, since it gives only “real” vectors in = d . The complexification map is an injective linear transformation (defined in the next chapter) from the real vector space = to the real version ²= d ³s of the complexification = d , that is, to the complex vector space = d provided that scalars are restricted to real numbers. In this way, we see that = d contains an embedded copy of = .

The Dimension of = d The vector-space dimensions of = and = d are the same. This should not necessarily come as a surprise because although = d may seem “bigger” than = , the field of scalars is also “bigger.” Theorem 1.17 If 8 ~ ¸# 0¹ is a basis for = over s, then the complexification of 8 , cpx²8 ³ ~ ¸# b # 8 ¹

Vector Spaces

55

is a basis for the vector space = d over d. Hence, dim²= d ³ ~ dim²= ³ Proof. To see that cpx²8 ³ spans = d over d, let % b & = d . Then %Á & = and so there exist real numbers and (some of which may be ) for which 1

1

% b & ~ # b @ # A ~

~

1

~ ² # b # ³ ~ 1

~ ² b ³²# b ³ ~

To see that cpx²8 ³ is linearly independent, if 1

² b ³²# b ³ ~ b ~

then the previous computations show that 1

1

# ~ and # ~ ~

~

The independence of 8 then implies that ~ and ~ for all . If # = and 8 ~ ¸# 0¹ is a basis for = , then we may write

# ~ # ~

for s. Since the coefficients are real, we have

# b ~ ²# b ³ ~

and so the coordinate matrices are equal: ´# b µcpx²8 ³ ~ ´#µ8

Exercises 1.

Let = be a vector space over - . Prove that # ~ and ~ for all # = and - . Describe the different 's in these equations. Prove that if # ~ , then ~ or # ~ . Prove that # ~ # implies that # ~ or ~ .

56 2. 3.

4.

Advanced Linear Algebra Prove Theorem 1.3. a) Find an abelian group = and a field - for which = is a vector space over - in at least two different ways, that is, there are two different definitions of scalar multiplication making = a vector space over - . b) Find a vector space = over - and a subset : of = that is (1) a subspace of = and (2) a vector space using operations that differ from those of = . Suppose that = is a vector space with basis 8 ~ ¸ 0¹ and : is a subspace of = . Let ¸) Á Ã Á ) ¹ be a partition of 8 . Then is it true that

: ~ ²: q º) »³ ~

5. 6.

What if : q º) » £ ¸¹ for all ? Prove Theorem 1.8. Let :Á ; Á < I ²= ³. Show that if < : , then : q ²; b < ³ ~ ²: q ; ³ b <

7.

This is called the modular law for the lattice I²= ³. For what vector spaces does the distributive law of subspaces : q ²; b < ³ ~ ²: q ; ³ b ²: q < ³

8.

9.

10. 11. 12. 13.

hold? A vector # ~ ² Á Ã Á ³ s is called strongly positive if for all ~ Á Ã Á . a) Suppose that # is strongly positive. Show that any vector that is “close enough” to # is also strongly positive. (Formulate carefully what “close enough” should mean.) b) Prove that if a subspace : of s contains a strongly positive vector, then : has a basis of strongly positive vectors. Let 4 be an d matrix whose rows are linearly independent. Suppose that the columns Á Ã Á of 4 span the column space of 4 . Let * be the matrix obtained from 4 by deleting all columns except Á Ã Á . Show that the rows of * are also linearly independent. Prove that the first two statements in Theorem 1.7 are equivalent. Show that if : is a subspace of a vector space = , then dim²:³ dim²= ³. Furthermore, if dim²:³ ~ dim²= ³ B then : ~ = . Give an example to show that the finiteness is required in the second statement. Let dim²= ³ B and suppose that = ~ < l : ~ < l : . What can you say about the relationship between : and : ? What can you say if : : ? What is the relationship between : l ; and ; l : ? Is the direct sum operation commutative? Formulate and prove a similar statement concerning associativity. Is there an “identity” for direct sum? What about “negatives”?

Vector Spaces

57

14. Let = be a finite-dimensional vector space over an infinite field - . Prove that if : Á Ã Á : are subspaces of = of equal dimension, then there is a subspace ; of = for which = ~ : l ; for all ~ Á Ã Á . In other words, ; is a common complement of the subspaces : . 15. Prove that the vector space 9 of all continuous functions from s to s is infinite-dimensional. 16. Show that Theorem 1.2 need not hold if the base field - is finite. 17. Let : be a subspace of = . The set # b : ~ ¸# b :¹ is called an affine subspace of = . a) Under what conditions is an affine subspace of = a subspace of = ? b) Show that any two affine subspaces of the form # b : and $ b : are either equal or disjoint. 18. If = and > are vector spaces over - for which (= ( ~ (> (, then does it follow that dim²= ³ ~ dim²> ³? 19. Let = be an -dimensional real vector space and suppose that : is a subspace of = with dim²:³ ~ c . Define an equivalence relation on the set = ± : by # $ if the “line segment” 3²#Á $³ ~ ¸# b ² c ³$ ¹ has the property that 3²#Á $³ q : ~ J. Prove that is an equivalence relation and that it has exactly two equivalence classes. 20. Let - be a field. A subfield of - is a subset 2 of - that is a field in its own right using the same operations as defined on - . a) Show that - is a vector space over any subfield 2 of - . b) Suppose that - is an -dimensional vector space over a subfield 2 of - . If = is an -dimensional vector space over - , show that = is also a vector space over 2 . What is the dimension of = as a vector space over 2 ? 21. Let - be a finite field of size and let = be an -dimensional vector space over - . The purpose of this exercise is to show that the number of subspaces of = of dimension is ² c ³Ä² c ³ 4 5 ~ ² c ³Ä² c ³² c c ³Ä² c ³ The expressions ² ³ are called Gaussian coefficients and have properties similar to those of the binomial coefficients. Let :²Á ³ be the number of -dimensional subspaces of = . a) Let 5 ²Á ³ be the number of -tuples of linearly independent vectors ²# Á Ã Á # ³ in = . Show that 5 ²Á ³ ~ ² c ³² c ³Ä² c c ³ b) Now, each of the -tuples in a) can be obtained by first choosing a subspace of = of dimension and then selecting the vectors from this subspace. Show that for any -dimensional subspace of = , the number

58

Advanced Linear Algebra of -tuples of independent vectors in this subspace is ² c ³² c ³Ä² c c ³ c)

Show that 5 ²Á ³ ~ :²Á ³² c ³² c ³Ä² c c ³

How does this complete the proof? 22. Prove that any subspace : of s is a closed set or, equivalently, that its set complement : ~ s ± : is open, that is, for any % : there is an open ball )²%Á ³ centered at % with radius for which )²%Á ³ : . 23. Let 8 ~ ¸ Á Ã Á ¹ and 9 ~ ¸ Á Ã Á ¹ be bases for a vector space = . Let c . Show that there is a permutation of ¸Á Ã Á ¹ such that Á Ã Á Á ²b³ Á Ã Á ²³ and ²³ Á Ã Á ²³ Á b Á Ã Á are both bases for = . Hint: You may use the fact that if 4 is an invertible d matrix and if , then it is possible to reorder the rows so that the upper left d submatrix and the lower right ² c ³ d ² c ³ submatrix are both invertible. (This follows, for example, from the general Laplace expansion theorem for determinants.) 24. Let = be an -dimensional vector space over an infinite field - and suppose that : Á Ã Á : are subspaces of = with dim²: ³ . Prove that there is a subspace ; of = of dimension c for which ; q : ~ ¸¹ for all . 25. What is the dimension of the complexification = d thought of as a real vector space? 26. (When is a subspace of a complex vector space a complexification?) Let = be a real vector space with complexification = d and let < be a subspace of = d . Prove that there is a subspace : of = for which < ~ : d ~ ¸ b ! Á ! :¹ if and only if < is closed under complex conjugation ¢ = d ¦ = d defined by ²" b #³ ~ " c #.

Chapter 2

Linear Transformations

Linear Transformations Loosely speaking, a linear transformation is a function from one vector space to another that preserves the vector space operations. Let us be more precise. Definition Let = and > be vector spaces over a field - . A function ¢ = ¦ > is a linear transformation if ²" b #³ ~ ²"³ b ²#³ for all scalars Á - and vectors ",# = . The set of all linear transformations from = to > is denoted by B²= Á > ³. 1) A linear transformation from = to = is called a linear operator on = . The set of all linear operators on = is denoted by B²= ³. A linear operator on a real vector space is called a real operator and a linear operator on a complex vector space is called a complex operator. 2) A linear transformation from = to the base field - (thought of as a vector space over itself) is called a linear functional on = . The set of all linear functionals on = is denoted by = i and called the dual space of = . We should mention that some authors use the term linear operator for any linear transformation from = to > . Also, the application of a linear transformation on a vector # is denoted by ²#³ or by #, parentheses being used when necessary, as in ²" b #³, or to improve readability, as in ² "³ rather than ² ²"³³. Definition The following terms are also employed: 1) homomorphism for linear transformation 2) endomorphism for linear operator 3) monomorphism (or embedding) for injective linear transformation 4) epimorphism for surjective linear transformation 5) isomorphism for bijective linear transformation.

60

Advanced Linear Algebra

6) automorphism for bijective linear operator. Example 2.1 1) The derivative +¢ = ¦ = is a linear operator on the vector space = of all infinitely differentiable functions on s. 2) The integral operator ¢ - ´%µ ¦ - ´%µ defined by %

~ ²!³!

is a linear operator on - ´%µ. 3) Let ( be an d matrix over - . The function ( ¢ - ¦ - defined by ( # ~ (#, where all vectors are written as column vectors, is a linear transformation from - to - . This function is just multiplication by ( . 4) The coordinate map ¢ = ¦ - of an -dimensional vector space is a linear transformation from = to - . The set B²= Á > ³ is a vector space in its own right and B²= ³ has the structure of an algebra, as defined in Chapter 0. Theorem 2.1 1) The set B²= Á > ³ is a vector space under ordinary addition of functions and scalar multiplication of functions by elements of - . 2) If B²< Á = ³ and B²= Á > ³, then the composition is in B²< Á > ³. 3) If B²= Á > ³ is bijective then c B²> Á = ³. 4) The vector space B²= ³ is an algebra, where multiplication is composition of functions. The identity map B²= ³ is the multiplicative identity and the zero map B²= ³ is the additive identity. Proof. We prove only part 3). Let ¢ = ¦ > be a bijective linear transformation. Then c ¢ > ¦ = is a well-defined function and since any two vectors $ and $ in > have the form $ ~ # and $ ~ # , we have c ²$ b $ ³ ~ c ² # b # ³ ~ c ² ²# b # ³³ ~ # b # ~ c ²$ ³ b c ²$ ³ which shows that c is linear. One of the easiest ways to define a linear transformation is to give its values on a basis. The following theorem says that we may assign these values arbitrarily and obtain a unique linear transformation by linear extension to the entire domain. Theorem 2.2 Let = and > be vector spaces and let 8 ~ ¸# 0¹ be a basis for = . Then we can define a linear transformation B²= Á > ³ by

Linear Transformations

61

specifying the values of # arbitrarily for all # 8 and extending to = by linearity, that is, ² # b Ä b # ³ ~ # b Ä b # This process defines a unique linear transformation, that is, if Á B²= Á > ³ satisfy # ~ # for all # 8 then ~ . Proof. The crucial point is that the extension by linearity is well-defined, since each vector in = has an essentially unique representation as a linear combination of a finite number of vectors in 8 . We leave the details to the reader. Note that if B²= Á > ³ and if : is a subspace of = , then the restriction O: of to : is a linear transformation from : to > .

The Kernel and Image of a Linear Transformation There are two very important vector spaces associated with a linear transformation from = to > . Definition Let B²= Á > ³. The subspace ker² ³ ~ ¸# = # ~ ¹ is called the kernel of and the subspace im² ³ ~ ¸ # # = ¹ is called the image of . The dimension of ker² ³ is called the nullity of and is denoted by null² ³. The dimension of im² ³ is called the rank of and is denoted by rk² ³. It is routine to show that ker² ³ is a subspace of = and im² ³ is a subspace of > . Moreover, we have the following. Theorem 2.3 Let B²= Á > ³. Then 1) is surjective if and only if im² ³ ~ > 2) is injective if and only if ker² ³ ~ ¸¹ Proof. The first statement is merely a restatement of the definition of surjectivity. To see the validity of the second statement, observe that " ~ # ¯ ²" c #³ ~ ¯ " c # ker² ³ Hence, if ker² ³ ~ ¸¹, then " ~ # ¯ " ~ #, which shows that is injective. Conversely, if is injective and " ker² ³, then " ~ and so " ~ . This shows that ker² ³ ~ ¸¹.

62

Advanced Linear Algebra

Isomorphisms Definition A bijective linear transformation ¢ = ¦ > is called an isomorphism from = to > . When an isomorphism from = to > exists, we say that = and > are isomorphic and write = > . Example 2.2 Let dim²= ³ ~ . For any ordered basis 8 of = , the coordinate map 8 ¢ = ¦ - that sends each vector # = to its coordinate matrix ´#µ8 - is an isomorphism. Hence, any -dimensional vector space over - is isomorphic to - . Isomorphic vector spaces share many properties, as the next theorem shows. If B²= Á > ³ and : = we write : ~ ¸

:¹

Theorem 2.4 Let B²= Á > ³ be an isomorphism. Let : = . Then 1) : spans = if and only if : spans > . 2) : is linearly independent in = if and only if : is linearly independent in >. 3) : is a basis for = if and only if : is a basis for > . An isomorphism can be characterized as a linear transformation ¢ = ¦ > that maps a basis for = to a basis for > . Theorem 2.5 A linear transformation B²= Á > ³ is an isomorphism if and only if there is a basis 8 for = for which 8 is a basis for > . In this case, maps any basis of = to a basis of > . The following theorem says that, up to isomorphism, there is only one vector space of any given dimension over a given field. Theorem 2.6 Let = and > be vector spaces over - . Then = > if and only if dim²= ³ ~ dim²> ³. In Example 2.2, we saw that any -dimensional vector space is isomorphic to - . Now suppose that ) is a set of cardinality and let ²- ) ³ be the vector space of all functions from ) to - with finite support. We leave it to the reader to show that the functions ²- ) ³ defined for all ) by ²%³ ~ F

if % ~ if % £

form a basis for ²- ) ³ , called the standard basis. Hence, dim²²- ) ³ ³ ~ () (. It follows that for any cardinal number , there is a vector space of dimension . Also, any vector space of dimension is isomorphic to ²- ) ³ .

Linear Transformations

63

Theorem 2.7 If is a natural number, then any -dimensional vector space over - is isomorphic to - . If is any cardinal number and if ) is a set of cardinality , then any -dimensional vector space over - is isomorphic to the vector space ²- ) ³ of all functions from ) to - with finite support.

The Rank Plus Nullity Theorem Let B²= Á > ³. Since any subspace of = has a complement, we can write = ~ ker² ³ l ker² ³ where ker² ³ is a complement of ker² ³ in = . It follows that dim²= ³ ~ dim²ker² ³³ b dim²ker² ³ ³ Now, the restriction of to ker² ³ , ¢ ker² ³ ¦ > is injective, since ker² ³ ~ ker² ³ q ker² ³ ~ ¸¹ Also, im² ³ im² ³. For the reverse inclusion, if # im² ³, then since # ~ " b $ for " ker² ³ and $ ker² ³ , we have # ~ " b $ ~ $ ~ $ im² ³ Thus im² ³ ~ im² ³. It follows that ker² ³ im² ³ From this, we deduce the following theorem. Theorem 2.8 Let B²= Á > ³. 1) Any complement of ker² ³ is isomorphic to im² ³ 2) (The rank plus nullity theorem) dim²ker² ³³ b dim²im² ³³ ~ dim²= ³ or, in other notation, rk² ³ b null² ³ ~ dim²= ³

Theorem 2.8 has an important corollary. Corollary 2.9 Let B²= Á > ³, where dim²= ³ ~ dim²> ³ B. Then is injective if and only if it is surjective. Note that this result fails if the vector spaces are not finite-dimensional. The reader is encouraged to find an example to support this statement.

64

Advanced Linear Algebra

Linear Transformations from - to - Recall that for any d matrix ( over - the multiplication map ( ²#³ ~ (# is a linear transformation. In fact, any linear transformation B²- Á - ³ has this form, that is, is just multiplication by a matrix, for we have 2 Ä 3 ~ 2 Ä 3²³ ~ and so ~ ( , where ( ~ 2 Ä 3 Theorem 2.10 1) If ( is an d matrix over - then ( B²- Á - ³. 2) If B²- Á - ³ then ~ ( , where ( ~ ² Ä ³ The matrix ( is called the matrix of . Example 2.3 Consider the linear transformation ¢ - ¦ - defined by ²%Á &Á '³ ~ ²% c &Á 'Á % b & b '³ Then we have, in column form,

v % y v %c& y v & ~ ' ~ w ' z w %b&b' z w

c

y v%y & z w' z

and so the standard matrix of is (~

v w

c

y z

If ( CÁ , then since the image of ( is the column space of (, we have dim²ker²( ³³ b rk²(³ ~ dim²- ³ This gives the following useful result. Theorem 2.11 Let ( be an d matrix over - . 1) ( ¢ - ¦ - is injective if and only if rk²(³ ~ n. 2) ( ¢ - ¦ - is surjective if and only if rk²(³ ~ m.

Linear Transformations

65

Change of Basis Matrices Suppose that 8 ~ ² Á Ã Á ³ and 9 ~ ² Á Ã Á ³ are ordered bases for a vector space = . It is natural to ask how the coordinate matrices ´#µ8 and ´#µ9 are related. Referring to Figure 2.1,

Fn

IB IC(IB)-1

V

IC

Fn

Figure 2.1 the map that takes ´#µ8 to ´#µ9 is 8Á9 ~ 9 8c and is called the change of basis operator (or change of coordinates operator). Since 8Á9 is an operator on - , it has the form ( , where ( ~ ²8Á9 ² ³ Ä 8Á9 ² ³³ ~ ²9 8c ²´ µ8 ³ Ä 9 8c ²´ µ8 ³³ ~ ²´ µ9 Ä ´ µ9 ³³ We denote ( by 48,9 and call it the change of basis matrix from 8 to 9. Theorem 2.12 Let 8 ~ ² Á Ã Á ³ and 9 be ordered bases for a vector space = . Then the change of basis operator 8Á9 ~ 9 8c is an automorphism of - , whose standard matrix is 48,9 ~ ²´ µ9 Ä ´ µ9 ³³ Hence ´#µ9 ~ 48Á9 ´#µ8 and 49Á8 ~ 48c ,9 . Consider the equation ( ~ 48 Á 9 or equivalently, ( ~ ²´ µ9 Ä ´ µ9 ³³ Then given any two of ( (an invertible d matrix)Á 8 (an ordered basis for - ) and 9 (an ordered basis for - ), the third component is uniquely determined by this equation. This is clear if 8 and 9 are given or if ( and 9 are

66

Advanced Linear Algebra

given. If ( and 8 are given, then there is a unique 9 for which (c ~ 49Á8 and so there is a unique 9 for which ( ~ 48Á9 . Theorem 2.13 If we are given any two of the following: 1) an invertible d matrix ( 2) an ordered basis 8 for - 3) an ordered basis 9 for - . then the third is uniquely determined by the equation ( ~ 48 Á 9

The Matrix of a Linear Transformation Let ¢ = ¦ > be a linear transformation, where dim²= ³ ~ and dim²> ³ ~ and let 8 ~ ² Á Ã Á ³ be an ordered basis for = and 9 an ordered basis for > . Then the map ¢ ´#µ8 ¦ ´ #µ9 is a representation of as a linear transformation from - to - , in the sense that knowing (along with 8 and 9, of course) is equivalent to knowing . Of course, this representation depends on the choice of ordered bases 8 and 9. Since is a linear transformation from - to - , it is just multiplication by an d matrix (, that is, ´ #µ9 ~ (´#µ8 Indeed, since ´ µ8 ~ , we get the columns of ( as follows: (²³ ~ ( ~ (´# µ8 ~ ´ µ9 Theorem 2.14 Let B²= Á > ³ and let 8 ~ ² Á Ã Á ³ and 9 be ordered bases for = and > , respectively. Then can be represented with respect to 8 and 9 as matrix multiplication, that is, ´ #µ9 ~ ´ µ8,9 ´#µ8 where ´ µ8,9 ~ ²´ µ9 Ä ´ µ9 ³ is called the matrix of with respect to the bases 8 and 9. When = ~ > and 8 ~ 9, we denote ´ µ8,8 by ´ µ8 and so ´ #µ8 ~ ´ µ8 ´#µ8

Example 2.4 Let +¢ F ¦ F be the derivative operator, defined on the vector space of all polynomials of degree at most . Let 8 ~ 9 ~ ²Á %Á % ³. Then

Linear Transformations

´+²³µ9 ~ ´µ9 ~

67

vy vy vy , ´+²%³µ9 ~ ´µ9 ~ Á ´+²% ³µ9 ~ ´%µ9 ~ wz wz wz

and so ´+µ8 ~

v w

y z

Hence, for example, if ²%³ ~ b % b % , then ´+²%³µ9 ~ ´+µ8 ´²%³µ8 ~

v w

yv y v y ~ zw z w z

and so +²%³ ~ b %. The following result shows that we may work equally well with linear transformations or with the matrices that represent them (with respect to fixed ordered bases 8 and 9). This applies not only to addition and scalar multiplication, but also to matrix multiplication. Theorem 2.15 Let = and > be finite-dimensional vector spaces over - , with ordered bases 8 ~ ² Á Ã Á ³ and 9 ~ ² Á Ã Á ³, respectively. 1) The map ¢ B²= Á > ³ ¦ CÁ ²- ³ defined by ² ³ ~ ´ µ8,9 is an isomorphism and so B²= Á > ³ CÁ ²- ³. Hence, dim²B²= Á > ³³ ~ dim²CÁ ²- ³³ ~ d 2) If B²< Á = ³ and B²= Á > ³ and if 8 , 9 and : are ordered bases for < , = and > , respectively, then ´µ8,: ~ ´ µ9,: ´µ8,9 Thus, the matrix of the product (composition) is the product of the matrices of and . In fact, this is the primary motivation for the definition of matrix multiplication. Proof. To see that is linear, observe that for all , ´ b ! µ8Á9 ´ µ8 ~ ´² b ! ³² ³µ9 ~ ´ ² ³ b ! ² ³µ9 ~ ´² ³µ9 b !´ ² ³µ9 ~ ´µ8Á9 ´ µ8 b !´ µ8Á9 ´ µ8 ~ ² ´µ8Á9 b !´ µ8Á9 ³´ µ8

68

Advanced Linear Algebra

and since ´ µ8 ~ is a standard basis vector, we conclude that ´ b ! µ8Á9 ~ ´µ8Á9 b !´ µ8Á9 and so is linear. If ( CÁ , we define by the condition ´ µ9 ~ (²³ , whence ² ³ ~ ( and is surjective. Also, ker²³ ~ ¸¹ since ´ µ8 ~ implies that ~ . Hence, the map is an isomorphism. To prove part 2), we have ´µ8Á: ´#µ8 ~ ´ ²#³µ: ~ ´ µ9,: ´#µ9 ~ ´ µ9Á: ´µ8Á9 ´#µ8

Change of Bases for Linear Transformations Since the matrix ´ µ8,9 that represents depends on the ordered bases 8 and 9 , it is natural to wonder how to choose these bases in order to make this matrix as simple as possible. For instance, can we always choose the bases so that is represented by a diagonal matrix? As we will see in Chapter 7, the answer to this question is no. In that chapter, we will take up the general question of how best to represent a linear operator by a matrix. For now, let us take the first step and describe the relationship between the matrices ´ µ8Á9 and ´ µ8Z Á9Z of with respect to two different pairs ²8 Á 9³ and ²8 Z Á 9Z ³ of ordered bases. Multiplication by ´ µ8Z Á9Z sends ´#µ8Z to ´ #µ9Z . This can be reproduced by first switching from 8 Z to 8 , then applying ´ µ8Á9 and finally switching from 9 to 9Z , that is, ´ µ8Z ,9Z ~ 49Á9Z ´ µ8,9 48Z Á8 ~ 49Á9Z ´ µ8,9 48c Á8 Z Theorem 2.16 Let B²= ,> ³ and let ²8 Á 9³ and ²8 Z Á 9Z ³ be pairs of ordered bases of = and > , respectively. Then ´ µ8Z Á9Z ~ 49Á9Z ´ µ8Á9 48Z Á8

(2.1)

When B²= ³ is a linear operator on = , it is generally more convenient to represent by matrices of the form ´ µ8 , where the ordered bases used to represent vectors in the domain and image are the same. When 8 ~ 9, Theorem 2.16 takes the following important form. Corollary 2.17 Let B²= ³ and let 8 and 9 be ordered bases for = . Then the matrix of with respect to 9 can be expressed in terms of the matrix of with respect to 8 as follows: ´ µ9 ~ 48Á9 ´ µ8 48c Á9

(2.2)

Equivalence of Matrices Since the change of basis matrices are precisely the invertible matrices, (2.1) has the form

Linear Transformations

69

´ µ8Z Á9Z ~ 7 ´ µ8Á9 8c where 7 and 8 are invertible matrices. This motivates the following definition. Definition Two matrices ( and ) are equivalent if there exist invertible matrices 7 and 8 for which ) ~ 7 (8c

We have remarked that ) is equivalent to ( if and only if ) can be obtained from ( by a series of elementary row and column operations. Performing the row operations is equivalent to multiplying the matrix ( on the left by 7 and performing the column operations is equivalent to multiplying ( on the right by 8c . In terms of (2.1), we see that performing row operations (premultiplying by 7 ) is equivalent to changing the basis used to represent vectors in the image and performing column operations (postmultiplying by 8c ) is equivalent to changing the basis used to represent vectors in the domain. According to Theorem 2.16, if ( and ) are matrices that represent with respect to possibly different ordered bases, then ( and ) are equivalent. The converse of this also holds. Theorem 2.18 Let = and > be vector spaces with dim²= ³ ~ and dim²> ³ ~ . Then two d matrices ( and ) are equivalent if and only if they represent the same linear transformation B²= Á > ³, but possibly with respect to different ordered bases. In this case, ( and ) represent exactly the same set of linear transformations in B²= Á > ³. Proof. If ( and ) represent , that is, if ( ~ ´ µ8,9

and ) ~ ´ µ8Z ,9Z

for ordered bases 8 Á 9Á 8 Z and 9Z , then Theorem 2.16 shows that ( and ) are equivalent. Now suppose that ( and ) are equivalent, say ) ~ 7 (8c where 7 and 8 are invertible. Suppose also that ( represents a linear transformation B²= Á > ³ for some ordered bases 8 and 9, that is, ( ~ ´ µ8Á9 Theorem 2.9 implies that there is a unique ordered basis 8 Z for = for which 8 ~ 48Á8Z and a unique ordered basis 9Z for > for which 7 ~ 49Á9Z . Hence ) ~ 49Á9Z ´ µ8Á9 48Z Á8 ~ ´ µ8Z Á9Z

70

Advanced Linear Algebra

Hence, ) also represents . By symmetry, we see that ( and ) represent the same set of linear transformations. This completes the proof. We remarked in Example 0.3 that every matrix is equivalent to exactly one matrix of the block form 0 1 ~ > cÁ

Ác cÁc ?block

Hence, the set of these matrices is a set of canonical forms for equivalence. Moreover, the rank is a complete invariant for equivalence. In other words, two matrices are equivalent if and only if they have the same rank.

Similarity of Matrices When a linear operator B²= ³ is represented by a matrix of the form ´ µ8 , equation (2.2) has the form ´ µ8Z ~ 7 ´ µ8 7 c where 7 is an invertible matrix. This motivates the following definition. Definition Two matrices ( and ) are similar, denoted by ( ) , if there exists an invertible matrix 7 for which ) ~ 7 (7 c The equivalence classes associated with similarity are called similarity classes. The analog of Theorem 2.18 for square matrices is the following. Theorem 2.19 Let = be a vector space of dimension . Then two d matrices ( and ) are similar if and only if they represent the same linear operator B²= ³, but possibly with respect to different ordered bases. In this case, ( and ) represent exactly the same set of linear operators in B²= ³. Proof. If ( and ) represent B²= ³, that is, if ( ~ ´ µ8

and ) ~ ´ µ9

for ordered bases 8 and 9, then Corollary 2.17 shows that ( and ) are similar. Now suppose that ( and ) are similar, say ) ~ 7 (7 c Suppose also that ( represents a linear operator B²= ³ for some ordered basis 8 , that is, ( ~ ´ µ8 Theorem 2.9 implies that there is a unique ordered basis 9 for = for which

Linear Transformations

71

7 ~ 48Á9 . Hence ) ~ 48Á9 ´ µ8 48c Á 9 ~ ´ µ9 Hence, ) also represents . By symmetry, we see that ( and ) represent the same set of linear operators. This completes the proof. We will devote much effort in Chapter 7 to finding a canonical form for similarity.

Similarity of Operators We can also define similarity of operators. Definition Two linear operators Á B²= ³ are similar, denoted by , if there exists an automorphism B²= ³ for which ~ c The equivalence classes associated with similarity are called similarity classes. Note that if 8 ~ ² Á Ã Á ³ and 9 ~ ² Á Ã Á ³ are ordered bases for = , then 49Á8 ~ ²´ µ8 Ä ´ µ8 ³ Now, the map defined by ² ³ ~ is an automorphism of = and 49Á8 ~ ²´² ³µ8 Ä ´² ³µ8 ³ ~ ´µ8 Conversely, if ¢ = ¦ = is an automorphism and 8 ~ ² Á Ã Á ³ is an ordered basis for = , then 9 ~ ² ~ ² ³Á Ã Á ~ ² ³³ is also a basis: ´µ8 ~ ²´² ³µ8 Ä ´² ³µ8 ³ ~ 49Á8 The analog of Theorem 2.19 for linear operators is the following. Theorem 2.20 Let = be a vector space of dimension . Then two linear operators and on = are similar if and only if there is a matrix ( C that represents both operators, but with respect to possibly different ordered bases. In this case, and are represented by exactly the same set of matrices in C . Proof. If and are represented by ( C , that is, if ´ µ8 ~ ( ~ ´µ9 for ordered bases 8 and 9, then ´µ9 ~ ´ µ8 ~ 49Á8 ´ µ9 48Á9 As remarked above, if ¢ = ¦ = is defined by ² ³ ~ , then

72

Advanced Linear Algebra

´µ9 ~ 48Á9 and so c ´µ9 ~ ´µc 9 ´ µ9 ´µ9 ~ ´ µ9

from which it follows that and are similar. Conversely, suppose that and are similar, say ~ c where is an automorphism of = . Suppose also that is represented by the matrix ( C , that is, ( ~ ´ µ8 for some ordered basis 8 . Then ´µ8 ~ 49Á8 and so c ´µ8 ~ ´c µ8 ~ ´µ8 ´ µ8 ´µc 8 ~ 4 9 Á8 ´ µ 8 4 9 Á 8

It follows that ( ~ ´ µ8 ~ 48Á9 ´µ8 48c Á9 ~ ´ µ9 and so ( also represents . By symmetry, we see that and are represented by the same set of matrices. This completes the proof. We can summarize the sitiation with respect to similarity in Figure 2.2. Each similarity class I in B²= ³ corresponds to a similarity class J in C ²- ³: J is the set of all matrices that represent any I and I is the set of all operators in B²= ³ that are represented by any 4 J .

V

similarity classes of L(V)

[W]B [V]B J [W]C [V]C

Similarity classes of matrices

W I

V W

Figure 2.2

Invariant Subspaces and Reducing Pairs The restriction of a linear operator B²= ³ to a subspace : of = is not necessarily a linear operator on : . This prompts the following definition.

Linear Transformations

73

Definition Let B²= ³. A subspace : of = is said to be invariant under or -invariant if : : , that is, if : for all : . Put another way, : is invariant under if the restriction O: is a linear operator on : . If = ~: l; then the fact that : is -invariant does not imply that the complement ; is also -invariant. (The reader may wish to supply a simple example with = ~ s .) Definition Let B²= ³. If = ~ : l ; and if both : and ; are -invariant, we say that the pair ²:Á ; ³ reduces . A reducing pair can be used to decompose a linear operator into a direct sum as follows. Definition Let B²= ³. If ²:Á ; ³ reduces we write ~ O: l O; and call the direct sum of O: and O; . Thus, the expression ~l means that there exist subspaces : and ; of = for which ²:Á ; ³ reduces and ~ O: and ~ O;

The concept of the direct sum of linear operators will play a key role in the study of the structure of a linear operator.

Projection Operators We will have several uses for a special type of linear operator that is related to direct sums. Definition Let = ~ : l ; . The linear operator :Á; ¢ = ¦ = defined by :Á; ² b !³ ~ where : and ! ; is called projection onto : along ; . Whenever we say that the operator :Á; is a projection, it is with the understanding that = ~ : l ; . The following theorem describes a few basic properties of projection operators. We leave proof as an exercise. Theorem 2.21 Let = be a vector space and let B²= ³.

74

Advanced Linear Algebra

1) If = ~ : l ; then :Á; b ; Á: ~ 2) If ~ :Á; then im²³ ~ :

and

ker²³ ~ ;

and so = ~ im²³ l ker²³ In other words, is projection onto its image along its kernel. Moreover, # im²³

¯

# ~ #

3) If B²= ³ has the property that = ~ im²³ l ker²³

and

Oim²³ ~

then is projection onto im²³ along ker²³. Projection operators are easy to characterize. Definition A linear operator B²= ³ is idempotent if ~ . Theorem 2.22 A linear operator B²= ³ is a projection if and only if it is idempotent. Proof. If ~ :Á; , then for any : and ! ; , ² b !³ ~ ~

~ ² b !³

and so ~ . Conversely, suppose that is idempotent. If # im²³ q ker²³, then # ~ % and so ~ # ~ % ~ % ~ # Hence im²³ q ker²³ ~ ¸¹. Also, if # = , then # ~ ²# c #³ b # ker²³ l im²³ and so = ~ ker²³ l im²³. Finally, ²%³ ~ % ~ % and so Oim²³ ~ . Hence, is projection onto im²³ along ker²³.

Projections and Invariance Projections can be used to characterize invariant subspaces. Let B²= ³ and let : be a subspace of = . Let ~ :Á; for any complement ; of : . The key is that the elements of : can be characterized as those vectors fixed by , that is,

Linear Transformations

75

: if and only if ~ . Hence, the following are equivalent: : : : for all : ² ³ ~ for all : ² ³ ~ for all : Thus, : is -invariant if and only if ~ for all vectors : . But this is also true for all vectors in ; , since both sides are equal to on ; . This proves the following theorem. Theorem 2.23 Let B²= ³. Then a subspace : of = is -invariant if and only if there is a projection ~ :Á; for which ~ in which case this holds for all projections of the form ~ :Á; . We also have the following relationship between projections and reducing pairs. Theorem 2.24 Let = ~ : l ; . Then ²:Á ; ³ reduces B²= ³ if and only if commutes with :Á; . Proof. Theorem 2.23 implies that : and ; are -invariant if and only if and

:Á; :Á; ~ :Á;

² c :Á; ³ ² c :Á; ³ ~ ² c :Á; ³

and a little algebra shows that this is equivalent to and

:Á; :Á; ~ :Á;

:Á; ~ :Á;

which is equivalent to :Á; ~ :Á; .

Orthogonal Projections and Resolutions of the Identity Observe that if is a projection, then ² c ³ ~ ² c ³ ~ Definition Two projections Á B²= ³ are orthogonal, written , if ~ ~

Note that if and only if im²³ ker²³

and

im²³ ker²³

The following example shows that it is not enough to have ~ in the definition of orthogonality. In fact, it is possible for ~ and yet is not even a projection.

76

Advanced Linear Algebra

Example 2.5 Let = ~ - and consider the ? - and @ -axes and the diagonal: ? ~ ¸²%Á ³ % - ¹ @ ~ ¸²Á &³ & - ¹ + ~ ¸²%Á %³ % - ¹ Then +Á? +Á@ ~ +Á@ £ +Á? ~ +Á@ +Á? From this we deduce that if and are projections, it may happen that both products and are projections, but that they are not equal. We leave it to the reader to show that @ Á? ?Á+ ~ (which is a projection), but that ?Á+ @ Á? is not a projection. Since a projection is idempotent, we can write the identity operator as s sum of two orthogonal projections: b ² c ³ ~ Á

² c ³

Let us generalize this to more than two projections. Definition A resolution of the identity on = is a sum of the form b Ä b ~ where the 's are pairwise orthogonal projections, that is, for £ . There is a connection between the resolutions of the identity on = and direct sum decompositions of = . In general terms, if b Ä b ~ for any linear operators B²= ³, then for all # = , # ~ # b Ä b # im² ³ b Ä b im² ³ and so = ~ im² ³ b Ä b im² ³ However, the sum need not be direct. Theorem 2.25 Let = be a vector space. Resolutions of the identity on = correspond to direct sum decompositions of = as follows: 1) If b Ä b ~ is a resolution of the identity, then = ~ im² ³ l Ä l im² ³

Linear Transformations

77

and is projection onto im² ³ along ker² ³ ~ im² ³ £

2) Conversely, if = ~ : l Ä l : and if is projection onto : along the direct sum £ : ,, then b Ä b ~ is a resolution of the identity. Proof. To prove 1), if b Ä b ~ is a resolution of the identity, then = ~ im² ³ b Ä b im² ³ Moreover, if % b Ä b % ~ then applying gives % ~ and so the sum is direct. As to the kernel of , we have im² ³ l ker² ³ ~ = ~ im² ³ l

p

im² ³

q £

s t

and since ~ , it follows that im² ³ ker² ³ £

and so equality must hold. For part 2), suppose that = ~ : l Ä l : and is projection onto : along £ : . If £ , then im² ³ ~ : ker² ³ and so . Also, if # ~ #~

bÄb

bÄb

for

: , then

~ # b Ä b # ~ ² b Ä b ³#

and so ~ b Ä b is a resolution of the identity.

The Algebra of Projections If and are projections, it does not necessarily follow that b , c or is a projection. For example, the sum b is a projection if and only if ² b ³ ~ b

78

Advanced Linear Algebra

which is equivalent to ~ c Of course, this holds if ~ ~ , that is, if . But the converse is also true, provided that char²- ³ £ . To see this, we simply evaluate in two ways: ²³ ~ c²³ ~ c and ²³ ~ c²³ ~ c Hence, ~ ~ c and so ~ . It follows that ~ c ~ and so . Thus, for char²- ³ £ , we have b is a projection if and only if . Now suppose that b is a projection. For the kernel of b , note that ² b ³# ~

¬

² b ³# ~

¬

# ~

and similarly, # ~ . Hence, ker² b ³ ker²³ q ker²³. But the reverse inclusion is obvious and so ker² b ³ ~ ker²³ q ker²³ As to the image of b , we have # im² b ³

¬

# ~ ² b ³# ~ # b # im²³ b im²³

and so im² b ³ im²³ b im²³. For the reverse inclusion, if # ~ % b &, then ² b ³# ~ ² b ³²% b &³ ~ % b & ~ # and so # im² b ³. Thus, im² b ³ ~ im²³ b im²³. Finally, ~ implies that im²³ ker²³ and so the sum is direct and im² b ³ ~ im²³ l im²³ The following theorem also describes the situation for the difference and product. Proof in these cases is left for the exercises. Theorem 2.26 Let = be a vector space over a field - of characteristic £ and let and be projections. 1) The sum b is a projection if and only if , in which case im² b ³ ~ im²³ l im²³

and

ker² b ³ ~ ker²³ q ker²³

2) The difference c is a projection if and only if ~ ~

Linear Transformations

79

in which case im² c ³ ~ im²³ q ker²³

and

ker² c ³ ~ ker²³ l im²³

3) If and commute, then is a projection, in which case im²³ ~ im²³ q im²³

and

ker²³ ~ ker²³ b ker²³

(Example 2.5 shows that the converse may be false.)

Topological Vector Spaces This section is for readers with some familiarity with point-set topology.

The Definition A pair ²= Á J ³ where = is a real vector space = and J is a topology on the set = is called a topological vector space if the operations of addition 7¢ = d = ¦ = Á

7²#Á $³ ~ # b $

and scalar multiplication C¢ s d = ¦ = Á

C²Á #³ ~ #

are continuous functions.

The Standard Topology on s The vector space s is a topological vector space under the standard topology, which is the topology for which the set of open rectangles 8 ~ ¸0 d Ä d 0 0 's are open intervals in s¹ is a base, that is, a subset of s is open if and only if it is a union of open rectangles. The standard topology is also the topology induced by the Euclidean metric on s , since an open rectangle is the union of Euclidean open balls and an open ball is the union of open rectangles. The standard topology on s has the property that the addition function 7¢ s d s ¦ s ¢ ²#Á $³ ¦ # b $ and the scalar multiplication function C¢ s d s ¦ s ¢ ²Á #³ ¦ # are continuous and so s is a topological vector space under this topology. Also, the linear functionals ¢ s ¦ s are continuous maps. For example, to see that addition is continuous, if ²" Á Ã Á " ³ b ²# Á Ã Á # ³ ² Á ³ d Ä d ²Á ³ 8

80

Advanced Linear Algebra

then " b # ² Á ³ and so there is an for which ²" c Á " b ³ b ²# c Á # b ³ ² Á ³ for all . It follows that if ²" Á Ã Á " ³ 0 ²" c Á " b ³ d Ä d ²" c Á " b ³ 8 and ²# Á Ã Á # ³ 1 ²# c Á # b ³ d Ä d ²# c Á # b ³ 8 then ²" Á Ã Á " ³ b ²# Á Ã Á # ³ 7²0Á 1 ³ ² Á ³ d Ä d ²Á ³

The Natural Topology on = Now let = be a real vector space of dimension and fix an ordered basis 8 ~ ²# Á Ã Á # ³ for = . We wish to show that there is precisely one topology J on = for which ²= Á J ³ is a topological vector space and all linear functionals are continuous. This topology is called the natural topology on = . Our plan is to show that if ²= Á J ³ is a topological vector space and if all linear functionals on = are continuous, then the coordinate map 8 ¢ = s is a homeomorphism. This implies that if J does exist, it must be unique. Then we use ~ 8c to move the standard topology from s to = , thus giving = a topology J for which 8 is a homeomorphism. Finally, we show that ²= Á J ³ is a topological vector space and that all linear functionals on = are continuous. The first step is to show that if ²= Á J ³ is a topological vector space, then is continuous. Since ~ where ¢ s ¦ = is defined by ² Á Ã Á ³ ~ # it is sufficient to show that these maps are continuous. (The sum of continuous maps is continuous.) Let 6 be an open set in J . Then Cc ²6³ ~ ¸²Á %³ s d = % 6¹ is open in s d = . This implies that if % 6, then there is an open interval 0 s containing for which 0% ~ ¸ %

0¹ 6

We need to show that the set c ²6³ is open. But c ²6³ ~ ¸² Á Ã Á ³ s # 6¹ ~ s d Ä d s d ¸ s # 6¹ d s d Ä d s

In words, an -tuple ² Á Ã Á ³ is in c ²6³ if the th coordinate times # is

Linear Transformations

81

in 6. But if # 6, then there is an open interval 0 s for which 0 and 0# 6. Hence, the entire open set < ~sdÄdsd0 dsdÄds where the factor 0 is in the th position is in c ²6³, that is, ² Á Ã Á ³ < c ²6³ Thus, c ²6³ is open and , and therefore also , is continuous. Next we show that if every linear functional on = is continuous under a topology J on = , then the coordinate map is continuous. If # = denote by ´#µ8Á the th coordinate of ´#µ8 . The map ¢ = ¦ s defined by # ~ ´#µ8Á is a linear functional and so is continuous by assumption. Hence, for any open interval 0 s the set ( ~ ¸# = ´#µ8Á 0 ¹ is open. Now, if 0 are open intervals in s, then c ²0 d Ä d 0 ³ ~ ¸# = ´#µ8 0 d Ä d 0 ¹ ~ ( is open. Thus, is continuous. We have shown that if a topology J has the property that ²= Á J ³ is a topological vector space under which every linear functional is continuous, then and ~ c are homeomorphisms. This means that if J exists, its open sets must be the images under of the open sets in the standard topology of s . It remains to prove that the topology J on = that makes a homeomorphism makes ²= Á J ³ a topological vector space for which any linear functional on = is continuous. The addition map on = is a composition 7 ~ c k 7Z k ² d ³ where 7Z ¢ s d s ¦ s is addition in s and since each of the maps on the right is continuous, so is 7. Similarly, scalar multiplication in = is C ~ c k CZ k ² d ³ where CZ ¢ s d s ¦ s is scalar multiplication in s . Hence, C is continuous. Now let be a linear functional. Since is continuous if and only if k c is continuous, we can confine attention to = ~ s . In this case, if Á Ã Á is the standard basis for s and ( ² ³( 4 for all , then for any

82

Advanced Linear Algebra

% ~ ² Á Ã Á ³ s , we have ( ²%³( ~ c ² ³c ( (( ² ³( 4 ( ( Now, if (%( °4 , then ( ( °4 and so ( ²%³( , which implies that is continuous at % ~ . According to the Riesz representation theorem (Theorem 9.18) and the Cauchy– Schwarz inequality, we have ) ²%³) )H ))%) where 9 s . Hence, % ¦ implies ²% ³ ¦ and so by linearity, % ¦ % implies ²% ³ ¦ % and so is continuous at all %. Theorem 2.27 Let = be a real vector space of dimension . There is a unique topology on = , called the natural topology, for which = is a topological vector space and for which all linear functionals on = are continuous. This topology is determined by the fact that the coordinate map ¢ = ¦ s is a homeomorphism, where s has the standard topology induced by the Euclidean metric.

Linear Operators on = d A linear operator on a real vector space = can be extended to a linear operator d on the complexification = d by defining d ²" b #³ ~ ²"³ b ²#³ Here are the basic properties of this complexification of . Theorem 2.28 If Á B²= ³, then 1) ² ³d ~ d , s 2) ² b ³d ~ d b d 3) ²³d ~ d d 4) ´ #µd ~ d ²#d ³. Let us recall that for any ordered basis 8 for = and any vector # = we have ´# b µcpx²8 ³ ~ ´#µ8 Now, if 8 is an ordered basis for = , then the th column of ´ µ8 is ´ µ8 ~ ´ b µcpx²8³ ~ ´ d ² b ³µcpx²8³ which is the th column of the coordinate matrix of d with respect to the basis cpx²8 ³. Thus we have the following theorem.

Linear Transformations

83

Theorem 2.29 Let B²= ³ where = is a real vector space. The matrix of d with respect to the ordered basis cpx²8 ³ is equal to the matrix of with respect to the ordered basis 8 : ´ d µcpx²8³ ~ ´ µ8 Hence, if a real matrix ( represents a linear operator on = , then ( also represents the complexification d of on = d .

Exercises 1.

2. 3. 4.

Let ( CÁ have rank . Prove that there are matrices ? CÁ and @ CÁ , both of rank , for which ( ~ ?@ . Prove that ( has rank if and only if it has the form ( ~ %! & where % and & are row matrices. Prove Corollary 2.9 and find an example to show that the corollary does not hold without the finiteness condition. Let B²= Á > ³. Prove that is an isomorphism if and only if it carries a basis for = to a basis for > . If B²= Á > ³ and B²= Á > ³ we define the external direct sum ^ B²= ^ = Á > ^ > ³ by ² ^ ³²²# Á # ³³ ~ ² # Á # ³

Show that ^ is a linear transformation. Let = ~ : l ; . Prove that : l ; : ^ ; . Thus, internal and external direct sums are equivalent up to isomorphism. 6. Let = ~ ( b ) and consider the external direct sum , ~ ( ^ ) . Define a map ¢ ( ^ ) ¦ = by ²#Á $³ ~ # b $. Show that is linear. What is the kernel of ? When is an isomorphism? 7. Let B- ²= ³ where dim²= ³ ~ B. Let ( C ²- ³. Suppose that there is an isomorphism ¢ = - with the property that ² #³ ~ (²#³. Prove that there is an ordered basis 8 for which ( ~ ´ µ8 . 8. Let J be a subset of B²= ³. A subspace : of = is J -invariant if : is invariant for every J . Also, = is J -irreducible if the only J -invariant subspaces of = are ¸¹ and = . Prove the following form of Schur's lemma. Suppose that J= B²= ³ and J> B²> ³ and = is J= -irreducible and > is J> -irreducible. Let B²= Á > ³ satisfy J= ~ J> , that is, for any J= there is a J> such that ~ and for any J> there is a J= such that ~ . Prove that ~ or is an isomorphism. 9. Let B²= ³ where dim²= ³ B. If rk² ³ ~ rk² ³ show that im² ³ q ker² ³ ~ ¸¹. 10. Let B²< ,= ³ and B²= Á > ³. Show that 5.

rk² ³ min¸rk² ³Á rk²³¹ 11. Let B²< Á = ³ and B²= Á > ³. Show that null² ³ null² ³ b null²³

84

Advanced Linear Algebra

12. Let Á B²= ³ where is invertible. Show that rk²³ ~ rk² ³ ~ rk²³ 13. Let Á B²= Á > ³. Show that rk² b ³ rk² ³ b rk²³ 14. Let : be a subspace of = . Show that there is a B²= ³ for which ker² ³ ~ : . Show also that there exists a B²= ³ for which im²³ ~ : . 15. Suppose that Á B²= ³. a) Show that ~ for some B²= ³ if and only if im²³ im² ³. b) Show that ~ for some B²= ³ if and only if ker² ³ ker²³. 16. Let dim²= ³ B and suppose that B²= ³ satisfies ~ . Show that rk² ³ dim²= ³. 17. Let ( be an d matrix over - . What is the relationship between the linear transformation ( ¢ - ¦ - and the system of equations (? ~ ) ? Use your knowledge of linear transformations to state and prove various results concerning the system (? ~ ) , especially when ) ~ . 18. Let = have basis 8 ~ ¸# Á Ã Á # ¹ and assume that the base field - for = has characteristic . Suppose that for each Á we define Á B²= ³ by Á ²# ³ ~ F

# # b #

if £ if ~

Prove that the Á are invertible and form a basis for B²= ³. 19. Let B²= ³. If : is a -invariant subspace of = must there be a subspace ; of = for which ²:Á ; ³ reduces ? 20. Find an example of a vector space = and a proper subspace : of = for which = : . 21. Let dim²= ³ B. If , B²= ³ prove that ~ implies that and are invertible and that ~ ² ³ for some polynomial ²%³ - ´%µ. 22. Let B²= ³. If ~ for all B²= ³ show that ~ , for some - , where is the identity map. 23. Let = be a vector space over a field - of characteristic £ and let and be projections. Prove the following: a) The difference c is a projection if and only if ~ ~ in which case im² c ³ ~ im²³ q ker²³

and

ker² c ³ ~ ker²³ l im²³

Hint: is a projection if and only if c is a projection and so c is a projection if and only if

Linear Transformations

85

~ c ² c ³ ~ ² c ³ b is a projection. b) If and commute, then is a projection, in which case im²³ ~ im²³ q im²³

and

ker²³ ~ ker²³ b ker²³

24. Let ¢ s ¦ s be a continuous function with the property that ²% b &³ ~ ²%³ b ²&³ 25. 26. 27. 28. 29.

30.

31. 32.

33.

Prove that is a linear functional on s . Prove that any linear functional ¢ s ¦ s is a continuous map. Prove that any subspace : of s is a closed set or, equivalently, that : ~ s ± : is open, that is, for any % : there is an open ball )²%Á ³ centered at % with radius for which )²%Á ³ : . Prove that any linear transformation ¢ = ¦ > is continuous under the natural topologies of = and > . Prove that any surjective linear transformation from = to > (both finitedimensional topological vector spaces under the natural topology) is an open map, that is, maps open sets to open sets. Prove that any subspace : of a finite-dimensional vector space = is a closed set or, equivalently, that : is open, that is, for any % : there is an open ball )²%Á ³ centered at % with radius for which )²%Á ³ : . Let : be a subspace of = with dim²= ³ B. a) Show that the subspace topology on : inherited from = is the natural topology. b) Show that the natural topology on = °: is the topology for which the natural projection map ¢ = ¦ = °: continuous and open. If = is a real vector space, then = d is a complex vector space. Thinking of = d as a vector space ²= d ³s over s, show that ²= d ³s is isomorphic to the external direct product = ^ = . (When is a complex linear map a complexification?) Let = be a real vector space with complexification = d and let B²= d ³. Prove that is a complexification, that is, has the form d for some B²= ³ if and only if commutes with the conjugate map ¢ = d ¦ = d defined by ²" b #³ ~ " c #. Let > be a complex vector space. a) Consider replacing the scalar multiplication on > by the operation ²'Á $³ ¦ '$ where ' d and $ > . Show that the resulting set with the addition defined for the vector space > and with this scalar multiplication is a complex vector space, which we denote by > . b) Show, without using dimension arguments, that ²>s ³d > ^ > .

Chapter 3

The Isomorphism Theorems

Quotient Spaces Let : be a subspace of a vector space = . It is easy to see that the binary relation on = defined by "#

¯

"c#:

is an equivalence relation. When " #, we say that " and # are congruent modulo : . The term mod is used as a colloquialism for modulo and " # is often written " # mod : When the subspace in question is clear, we will simply write " #. To see what the equivalence classes look like, observe that ´#µ ~ ¸" = " #¹ ~ ¸" = " c # :¹ ~ ¸" = " ~ # b for some :¹ ~ ¸# b :¹ ~#b: The set ´#µ ~ # b : ~ ¸# b

:¹

is called a coset of : in = and # is called a coset representative for # b : . (Thus, any member of a coset is a coset representative.) The set of all cosets of : in = is denoted by = °: ~ ¸# b : # = ¹ This is read “= mod : ” and is called the quotient space of = modulo : . Of

88

Advanced Linear Algebra

course, the term space is a hint that we intend to define vector space operations on = °: . The natural choice for these vector space operations is ²" b :³ b ²# b :³ ~ ²" b #³ b : and ²" b :³ ~ ²"³ b : but we must check that these operations are well-defined, that is, 1) " b : ~ " b :Á # b : ~ # b : ¬ ²" b # ³ b : ~ ²" b # ³ b : 2) " b : ~ " b : ¬ " b : ~ " b : Equivalently, the equivalence relation must be consistent with the vector space operations on = , that is, 3) " " Á # # ¬ ²" b # ³ ²" b # ³ 4) " " ¬ " " This senario is a recurring one in algebra. An equivalence relation on an algebraic structure, such as a group, ring, module or vector space is called a congruence relation if it preserves the algebraic operations. In the case of a vector space, these are conditions 3) and 4) above. These conditions follow easily from the fact that : is a subspace, for if " " and # # , then " c " :Á # c # : ¬ ²" c " ³ b ²# c # ³ : ¬ ²" b # ³ c ²" b # ³ : ¬ " b # " b # which verifies both conditions at once. We leave it to the reader to verify that = °: is indeed a vector space over - under these well-defined operations. Actually, we are lucky here: For any subspace : of = , the quotient = °: is a vector space under the natural operations. In the case of groups, not all subgroups have this property. Indeed, it is precisely the normal subgroups 5 of . that have the property that the quotient .°5 is a group. Also, for rings, it is precisely the ideals (not the subrings) that have the property that the quotient is a ring. Let us summarize.

The Isomorphism Theorems

89

Theorem 3.1 Let : be a subspace of = . The binary relation "#

¯

"c#:

is an equivalence relation on = , whose equivalence classes are the cosets # b : ~ ¸# b

:¹

of : in = . The set = °: of all cosets of : in = , called the quotient space of = modulo : , is a vector space under the well-defined operations ²" b :³ ~ " b : ²" b :³ b ²# b :³ ~ ²" b #³ b : The zero vector in = °: is the coset b : ~ : .

The Natural Projection and the Correspondence Theorem If : is a subspace of = , then we can define a map : ¢ = ¦ = °: by sending each vector to the coset containing it: : ²#³ ~ # b : This map is called the canonical projection or natural projection of = onto = °: , or simply projection modulo : . (Not to be confused with the projection operators :Á; .) It is easily seen to be linear, for we have (writing for : ) ²" b #³ ~ ²" b #³ b : ~ ²" b :³ b ²# b :³ ~ ²"³ b ²#³ The canonical projection is clearly surjective. To determine the kernel of , note that # ker²³ ¯ ²#³ ~ ¯ # b : ~ : ¯ # : and so ker²³ ~ : Theorem 3.2 The canonical projection : ¢ = ¦ = °: defined by : ²#³ ~ # b : is a surjective linear transformation with ker²: ³ ~ : . If : is a subspace of = , then the subspaces of the quotient space = °: have the form ; °: for some intermediate subspace ; satisfying : ; = . In fact, as shown in Figure 3.1, the projection map : provides a one-to-one correspondence between intermediate subspaces : ; = and subspaces of the quotient space = °: . The proof of the following theorem is left as an exercise.

90

Advanced Linear Algebra

V V/S

T S

T/S

{0}

{0}

Figure 3.1: The correspondence theorem Theorem 3.3 (The correspondence theorem) Let : be a subspace of = . Then the function that assigns to each intermediate subspace : ; = the subspace ; °: of = °: is an order-preserving (with respect to set inclusion) one-to-one correspondence between the set of all subspaces of = containing : and the set of all subspaces of = °: . Proof. We prove only that the correspondence is surjective. Let ? ~ ¸" b : " < ¹ be a subspace of = °: and let ; be the union of all cosets in ? : ; ~ ²" b :³ "<

We show that : ; = and that ; °: ~ ? . If %Á & ; , then % b : and & b : are in ? and since ? = °: , we have % b :Á ²% b &³ b : ? which implies that %Á % b & ; . Hence, ; is a subspace of = containing : . Moreover, if ! b : ; °: , then ! ; and so ! b : ? . Conversely, if " b : ? , then " ; and therefore " b : ; °: . Thus, ? ~ ; °: .

The Universal Property Isomorphism Theorem

of

Quotients

and

the

First

Let : be a subspace of = . The pair ²= °:Á : ³ has a very special property, known as the universal property—a term that comes from the world of category theory. Figure 3.2 shows a linear transformation B²= Á > ³, along with the canonical projection : from = to the quotient space = °: .

The Isomorphism Theorems

W

V

91

W

Ss W' V/S Figure 3.2: The universal property The universal property states that if ker² ³ : , then there is a unique Z ¢ = °: ¦ > for which Z k : ~ Another way to say this is that any such B²= Á > ³ can be factored through the canonical projection : . Theorem 3.4 Let : be a subspace of = and let B²= Á > ³ satisfy : ker² ³. Then, as pictured in Figure 3.2, there is a unique linear transformation Z ¢ = °: ¦ > with the property that Z k : ~ Moreover, ker² Z ³ ~ ker² ³°: and im² Z ³ ~ im² ³. Proof. We have no other choice but to define Z by the condition Z k : ~ , that is, Z ²# b :³ ~ # This function is well-defined if and only if # b : ~ " b : ¬ Z ²# b :³ ~ Z ²" b :³ which is equivalent to each of the following statements: #b: ~"b: #c": %: :

¬ # ~ " ¬ ²# c "³ ~ ¬ % ~ ker² ³

Thus, Z ¢ = °: ¦ > is well-defined. Also, im² Z ³ ~ ¸ Z ²# b :³ # = ¹ ~ ¸ # # = ¹ ~ im² ³ and

92

Advanced Linear Algebra

ker² Z ³ ~ ¸# b : Z ²# b :³ ~ ¹ ~ ¸# b : # ~ ¹ ~ ¸# b : # ker² ³¹ ~ ker² ³°: The uniqueness of Z is evident. Theorem 3.4 has a very important corollary, which is often called the first isomorphism theorem and is obtained by taking : ~ ker² ³. Theorem 3.5 (The first isomorphism theorem) Let ¢ = ¦ > be a linear transformation. Then the linear transformation Z ¢ = °ker² ³ ¦ > defined by Z ²# b ker² ³³ ~ # is injective and = im² ³ ker² ³

According to Theorem 3.5, the image of any linear transformation on = is isomorphic to a quotient space of = . Conversely, any quotient space = °: of = is the image of a linear transformation on = : the canonical projection : . Thus, up to isomorphism, quotient spaces are equivalent to homomorphic images.

Quotient Spaces, Complements and Codimension The first isomorphism theorem gives some insight into the relationship between complements and quotient spaces. Let : be a subspace of = and let ; be a complement of : , that is, = ~: l; Applying the first isomorphism theorem to the projection operator ; Á: ¢ = ¦ ; gives ; = °: Theorem 3.6 Let : be a subspace of = . All complements of : in = are isomorphic to = °: and hence to each other. The previous theorem can be rephrased by writing (l) ~(l* ¬) * On the other hand, quotients and complements do not behave as nicely with respect to isomorphisms as one might casually think. We leave it to the reader to show the following:

The Isomorphism Theorems

93

1) It is possible that (l) ~* l+ ° +. Hence, ( * does not imply that a complement with ( * but ) of ( is isomorphic to a complement of * . 2) It is possible that = > and = ~ : l ) and > ~ : l + ° +. Hence, = > does not imply that = °: > °: . (However, but ) according to the previous theorem, if = equals > then ) +.) Corollary 3.7 Let : be a subspace of a vector space = . Then dim²= ³ ~ dim²:³ b dim²= °:³

Definition If : is a subspace of = , then dim²= °:³ is called the codimension of : in = and is denoted by codim²:³ or codim= ²:³. Thus, the codimension of : in = is the dimension of any complement of : in = and when = is finite-dimensional, we have codim= ²:³ ~ dim²= ³ c dim²:³ (This makes no sense, in general, if = is not finite-dimensional, since infinite cardinal numbers cannot be subtracted.)

Additional Isomorphism Theorems There are other isomorphism theorems that are direct consequences of the first isomorphism theorem. As we have seen, if = ~ : l ; then = °; : . This can be written : l; : ; : q; This applies to nondirect sums as well. Theorem 3.7 (The second isomorphism theorem) Let = be a vector space and let : and ; be subspaces of = . Then : b; : ; : q; Proof. Let ¢ ²: b ; ³ ¦ :°²: q ; ³ be defined by ² b !³ ~ b ²: q ; ³ We leave it to the reader to show that is a well-defined surjective linear transformation, with kernel ; . An application of the first isomorphism theorem then completes the proof.

94

Advanced Linear Algebra

The following theorem demonstrates one way in which the expression = °: behaves like a fraction. Theorem 3.8 (The third isomorphism theorem) Let = be a vector space and suppose that : ; = are subspaces of = . Then = °: = ; °: ; Proof. Let ¢ = °: ¦ = °; be defined by ²# b :³ ~ # b ; . We leave it to the reader to show that is a well-defined surjective linear transformation whose kernel is ; °: . The rest follows from the first isomorphism theorem. The following theorem demonstrates one way in which the expression = °: does not behave like a fraction. Theorem 3.9 Let = be a vector space and let : be a subspace of = . Suppose that = ~ = l = and : ~ : l : with : = . Then = = l = = = ^ ~ : : l : : : Proof. Let ¢ = ¦ ²= °: ³ ^ ²= °: ³ be defined by ²# b # ³ ~ ²# b : Á # b : ³ This map is well-defined, since the sum = ~ = l = is direct. We leave it to the reader to show that is a surjective linear transformation, whose kernel is : l : . The rest follows from the first isomorphism theorem.

Linear Functionals Linear transformations from = to the base field - (thought of as a vector space over itself) are extremely important. Definition Let = be a vector space over - . A linear transformation B²= Á - ³ whose values lie in the base field - is called a linear functional (or simply functional) on = . (Some authors use the term linear function.) The vector space of all linear functionals on = is denoted by = * and is called the algebraic dual space of = . The adjective algebraic is needed here, since there is another type of dual space that is defined on general normed vector spaces, where continuity of linear transformations makes sense. We will discuss the so-called continuous dual space briefly in Chapter 13. However, until then, the term “dual space” will refer to the algebraic dual space.

The Isomorphism Theorems

95

To help distinguish linear functionals from other types of linear transformations, we will usually denote linear functionals by lowercase italic letters, such as , and . Example 3.1 The map ¢ - ´%µ ¦ - defined by ²²%³³ ~ ²³ is a linear functional, known as evaluation at . Example 3.2 Let 9´Á µ denote the vector space of all continuous functions on ´Á µ s. Let ¢ 9´Á µ ¦ s be defined by

²²%³³ ~ ²%³ %

i

Then 9´Á µ . For any = * , the rank plus nullity theorem is dim²ker² ³³ b dim²im² ³³ ~ dim²= ³ But since im² ³ - , we have either im² ³ ~ ¸¹, in which case is the zero linear functional, or im² ³ ~ - , in which case is surjective. In other words, a nonzero linear functional is surjective. Moreover, if £ , then codim²ker² ³³ ~ dim6

= 7~ ker² ³

and if dim²= ³ B, then dim²ker² ³³ ~ dim²= ³ c Thus, in dimensional terms, the kernel of a linear functional is a very “large” subspace of the domain = . The following theorem will prove very useful. Theorem 3.10 1) For any nonzero vector # = , there exists a linear functional = * for which ²#³ £ . 2) A vector # = is zero if and only if ²#³ ~ for all = * . 3) Let = i . If ²%³ £ , then = ~ º%» l ker² ³ 4) Two nonzero linear functionals Á = i have the same kernel if and only if there is a nonzero scalar such that ~ . Proof. For part 3), if £ # º%» q ker² ³, then ²#³ ~ and # ~ % for £ - , whence ²%³ ~ , which is false. Hence, º%» q ker² ³ ~ ¸¹ and the direct sum : ~ º%» l ker² ³ exists. Also, for any # = we have

96

Advanced Linear Algebra

²#³ ²#³ % b 6# c %7 º%» b ker² ³ ²%³ ²%³

#~

and so = ~ º%» l ker² ³. For part 4), if ~ for £ , then ker² ³ ~ ker²³. Conversely, if 2 ~ ker² ³ ~ ker²³, then for % ¤ 2 we have by part 3), = ~ º%» l 2 Of course, O2 ~ O2 for any . Therefore, if ~ ²%³°²%³, it follows that ²%³ ~ ²%³ and hence ~ .

Dual Bases Let = be a vector space with basis 8 ~ ¸# 0¹. For each 0 , we can define a linear functional #i = * by the orthogonality condition #i ²# ³ ~ Á where Á is the Kronecker delta function, defined by if ~ if £

Á ~ F

Then the set 8 i ~ ¸#i 0¹ is linearly independent, since applying the equation ~ #i b Ä b #i to the basis vector # gives

~ #i ²# ³ ~ Á ~ ~

~

for all . Theorem 3.11 Let = be a vector space with basis 8 ~ ¸# 0¹. 1) The set 8 i ~ ¸#i 0¹ is linearly independent. 2) If = is finite-dimensional, then 8 i is a basis for = i , called the dual basis of 8. Proof. For part 2), for any = i , we have ²# ³#i ²# ³ ~ ²# ³Á ~ ²# ³

and so ~ ²# ³#i is in the span of 8 i . Hence, 8 * is a basis for = i .

The Isomorphism Theorems

97

It follows from the previous theorem that if dim²= ³ B, then dim²= i ³ ~ dim²= ³ since the dual vectors also form a basis for = i . Our goal now is to show that the converse of this also holds. But first, let us consider an example. Example 3.3 Let = be an infinite-dimensional vector space over the field - ~ { ~ ¸Á ¹, with basis 8 . Since the only coefficients in - are and , a finite linear combination over - is just a finite sum. Hence, = is the set of all finite sums of vectors in 8 and so according to Theorem 0.12, (= ( (F ²8 ³( ~ (8 ( On the other hand, each linear functional = i is uniquely defined by specifying its values on the basis 8 . Since these values must be either or , specifying a linear functional is equivalent to specifying the subset of 8 on which takes the value . In other words, there is a one-to-one correspondence between linear functionals on = and all subsets of 8 . Hence, (= i ( ~ (F ²8 ³( (8 ( (= ( This shows that = i cannot be isomorphic to = , nor to any proper subset of = . Hence, dim²= i ³ dim²= ³. We wish to show that the behavior in the previous example is typical, in particular, that dim²= ³ dim²= i ³ with equality if and only if = is finite-dimensional. The proof uses the concept of the prime subfield of a field 2 , which is defined as the smallest subfield of the field 2 . Since Á 2 , it follows that 2 contains a copy of the integers Á Á ~ b Á ~ b b Á Ã If 2 has prime characteristic , then ~ and so 2 contains the elements { ~ ¸Á Á Á Á Ã Á c ¹ which form a subfield of 2 . Since any subfield - of 2 contains and , we see that { - and so { is the prime subfield of 2 . On the other hand, if 2 has characteristic , then 2 contains a “copy” of the integers { and therefore also the rational numbers r, which is the prime subfield of 2 . Our main interest in the prime subfield is that in either case, the prime subfield is countable. Theorem 3.12 Let = be a vector space. Then dim²= ³ dim²= i ³ with equality if and only if = is finite-dimensional.

98

Advanced Linear Algebra

Proof. For any vector space = , we have dim²= ³ dim²= i ³ since the dual vectors to a basis 8 for = are linearly independent in = i . We have already seen that if = is finite-dimensional, then dim²= ³ ~ dim²= i ³. We wish to show that if = is infinite-dimensional, then dim²= ³ dim²= i ³. (The author is indebted to Professor Richard Foote for suggesting this line of proof.) If 8 is a basis for = and if 2 is the base field for = , then Theorem 2.7 implies that = ²2 8 ³ where ²2 8 ³ is the set of all functions with finite support from 8 to 2 and = i 28 where 2 8 is the set of all functions from 8 to 2 . Thus, we can work with the vector spaces ²2 8 ³ and 2 8 . The plan is to show that if - is a countable subfield of 2 and if 8 is infinite, then dim2 2²2 8 ³ 3 ~ dim- 2²- 8 ³ 3 dim- ²- 8 ³ dim2 22 8 3 Since we may take - to be the prime subfield of 2 , this will prove the theorem. The first equality follows from the fact that the 2 -space ²2 8 ³ and the - -space ²- 8 ³ each have a basis consisting of the “standard” linear functionals ¸ 8 ¹ defined by # ~ Á for all # 8 , where Á is the Kronecker delta function. For the final inequality, suppose that ¸ ¹ - 8 is linearly independent over and that ~

where 2 . If ¸ ¹ is a basis for 2 over - , then ~ Á for Á and so ~ ~ Á

Evaluating at any # 8 gives

The Isomorphism Theorems

99

~ Á ²#³ ~ 8 Á ²#³9

and since the inner sums are in - and ¸ ¹ is - -independent, the inner sums must be zero: Á ²#³ ~

Since this holds for all # 8 , we have Á ~

which implies that Á ~ for all Á . Hence, ¸ ¹ is linearly independent over 2 . This proves that dim- ²- 8 ³ dim2 22 8 3. For the center inequality, it is clear that dim- 2²- 8 ³ 3 dim- ²- 8 ³ We will show that the inequality must be strict by showing that the cardinality of ²- 8 ³ is (8 ( whereas the cardinality of - 8 is greater than (8 (. To this end, the set ²- 8 ³ can be partitioned into blocks based on the support of the function. In particular, for each finite subset : of 8 , if we let (: ~ ¸ ²- 8 ³ supp² ³ ~ :¹ then ²- 8 ³ ~ (: :8 : finite

where the union is disjoint. Moreover, if (: ( ~ , then ((: ( (- ( L and so b²- 8 ³ b ~ ((: ( (8 ( h L ~ max²(8 (Á L ³ ~ (8 ( :8 : finite

But since the reverse inequality is easy to establish, we have b²- 8 ³ b ~ (8 ( As to the cardinality of - 8 , for each subset ; of 8 , there is a function ; - 8 that sends every element of ; to and every element of 8 ± ; to . Clearly, each distinct subset ; gives rise to a distinct function ; and so Cantor's

100

Advanced Linear Algebra

theorem implies that b- 8 b b8 b (8 ( ~ b²- 8 ³ b This shows that dim- 2²- 8 ³ 3 dim- ²- 8 ³ and completes the proof.

Reflexivity If = is a vector space, then so is the dual space = i and so we may form the double (algebraic) dual space = ii , which consists of all linear functionals ¢ = i ¦ - . In other words, an element of = ** is a linear functional that assigns a scalar to each linear functional on = . With this firmly in mind, there is one rather obvious way to obtain an element of = ii . Namely, if # = , consider the map #¢ = i ¦ - defined by #² ³ ~ ²#³ which sends the linear functional to the scalar ²#³. The map # is called evaluation at #. To see that # = ii , if Á = i and Á - , then #² b ³ ~ ² b ³²#³ ~ ²#³ b ²#³ ~ #² ³ b #²³ and so # is indeed linear. We can now define a map ¢ = ¦ = ii by # ~ # This is called the canonical map (or the natural map) from = to = ii . This map is injective and hence in the finite-dimensional case, it is also surjective. Theorem 3.13 The canonical map ¢ = ¦ = ii defined by # ~ #, where # is evaluation at #, is a monomorphism. If = is finite-dimensional, then is an isomorphism. Proof. The map is linear since " b #² ³ ~ ²" b #³ ~ ²"³ b ²#³ ~ ²" b #³² ³ for all = i . To determine the kernel of , observe that # ~ ¬ # ~ ¬ #² ³ ~ for all = i ¬ ²#³ ~ for all = i ¬ #~ by Theorem 3.10 and so ker² ³ ~ ¸¹. In the finite-dimensional case, since

The Isomorphism Theorems

101

dim²= ii ³ ~ dim²= i ³ ~ dim²= ³ it follows that is also surjective, hence an isomorphism. Note that if dim²= ³ B, then since the dimensions of = and = ii are the same, we deduce immediately that = = ii . This is not the point of Theorem 3.13. The point is that the natural map # ¦ # is an isomorphism. Because of this, = is said to be algebraically reflexive. Theorem 3.13 and Theorem 3.12 together imply that a vector space is algebraically reflexive if and only if it is finitedimensional. If = is finite-dimensional, it is customary to identify the double dual space = ii with = and to think of the elements of = ii simply as vectors in = . Let us consider a specific example to show how algebraic reflexivity fails in the infinite-dimensional case. Example 3.4 Let = be the vector space over { with basis ~ ²Á Ã Á Á Á Á Ã ³ where the is in the th position. Thus, = is the set of all infinite binary sequences with a finite number of 's. Define the order ²#³ of any # = to be the largest coordinate of # with value . Then ²#³ B for all # = . Consider the dual vectors i , defined (as usual) by i ² ³ ~ Á For any # = , the evaluation functional # has the property that #²i ³ ~ i ²#³ ~ if ²#³ However, since the dual vectors i are linearly independent, there is a linear functional = ii for which ²i ³ ~ for all . Hence, does not have the form # for any # = . This shows that the canonical map is not surjective and so = is not algebraically reflexive.

Annihilators The functions = i are defined on vectors in = , but we may also define on subsets 4 of = by letting ²4 ³ ~ ¸ ²#³ # 4 ¹

102

Advanced Linear Algebra

Definition Let 4 be a nonempty subset of a vector space = . The annihilator 4 of 4 is 4 ~ ¸ = i ²4 ³ ~ ¸¹¹

The term annihilator is quite descriptive, since 4 consists of all linear functionals that annihilate (send to ) every vector in 4 . It is not hard to see that 4 is a subspace of = i , even when 4 is not a subspace of = . The basic properties of annihilators are contained in the following theorem. Theorem 3.14 1) (Order-reversing) If 4 and 5 are nonempty subsets of = , then 4 5 ¬ 5 4 2) If dim²= ³ B, then for any nonempty subset 4 of = the natural map ¢ span²4 ³ 4 is an isomorphism from span²4 ³ onto 4 . In particular, if : is a subspace of = , then : : . 3) If : and ; are subspaces of = , then ²: q ; ³ ~ : b ; and ²: b ; ³ ~ : q ; Proof. We leave proof of part 1) for the reader. For part 2), since 4 ~ ²span²4 ³³ it is sufficient to prove that ¢ : : is an isomorphism, where : is a subspace of = . Now, we know that is a monomorphism, so it remains to prove that : ~ : . If : , then ~ has the property that for all : , ² ³ ~ ~ and so ~ : , which implies that : : . Moreover, if # : , then for all : we have ²#³ ~ #² ³ ~ and so every linear functional that annihilates : also annihilates #. But if # ¤ : , then there is a linear functional = i for which ²:³ ~ ¸¹ and ²#³ £ . (We leave proof of this as an exercise.) Hence, # : and so # ~ # : and so : : . For part 3), it is clear that annihilates : b ; if and only if annihilates both : and ; . Hence, ²: b ; ³ ~ : q ; . Also, if ~ b : b ; where : and ; , then Á ²: q ; ³ and so ²: q ; ³ . Thus,

The Isomorphism Theorems

103

: b ; ²: q ; ³ For the reverse inclusion, suppose that ²: q ; ³ . Write = ~ : Z l ²: q ; ³ l ; Z l < where : ~ : Z l ²: q ; ³ and ; ~ ²: q ; ³ l ; Z . Define = i by O: Z ~ Á

O:q; ~ O:q; ~ Á

O; Z ~ Á

O< ~

O:q; ~ O:q; ~ Á

O; Z ~ Á

O< ~

and define = i by O: Z ~ Á

It follows that ; , : and b ~ .

Annihilators and Direct Sums Consider a direct sum decomposition = ~: l; Then any linear functional ; i can be extended to a linear functional on = by setting ²:³ ~ . Let us call this extension by . Clearly, : and it is easy to see that the extension by map ¦ is an isomorphism from ; i to : , whose inverse is the restriction to ; . Theorem 3.15 Let = ~ : l ; . a) The extension by map is an isomorphism from ; i to : and so ; i : b) If = is finite-dimensional, then dim²: ³ ~ codim= ²:³ ~ dim²= ³ c dim²:³

Example 3.5 Part b) of Theorem 3.15 may fail in the infinite-dimensional case, since it may easily happen that : = i . As an example, let = be the vector space over { with a countably infinite ordered basis 8 ~ ² Á Á Ã ³. Let : ~ º » and ; ~ º Á Á Ã ». It is easy to see that : ; i = i and that dim²= i ³ dim²= ³. The annihilator provides a way to describe the dual space of a direct sum. Theorem 3.16 A linear functional on the direct sum = ~ : l ; can be written as a sum of a linear functional that annihilates : and a linear functional that annihilates ; , that is, ²: l ; ³i ~ : l ;

104

Advanced Linear Algebra

Proof. Clearly : q ; ~ ¸¹, since any functional that annihilates both : and ; must annihilate : l ; ~ = . Hence, the sum : b ; is direct. The rest follows from Theorem 3.14, since = i ~ ¸¹ ~ ²: q ; ³ ~ : b ; ~ : l ; Alternatively, since ; b : ~ is the identity map, if = i , then we can write ~ k ²; b : ³ ~ ² k ; ³ b ² k : ³ : l ; and so = i ~ : l ; .

Operator Adjoints If B²= Á > ³, then we may define a map d ¢ > * ¦ = i by d ² ³ ~ k ~ for > * . (We will write composition as juxtaposition.) Thus, for any # = , ´ d ² ³µ²#³ ~ ² #³ The map d is called the operator adjoint of and can be described by the phrase “apply first.” Theorem 3.17 (Properties of the Operator Adjoint) 1) For Á B²= Á > ³ and Á - , ² b ³d ~ d b d 2) For B²= Á > ³ and B²> Á < ³, ²³d ~ d d 3) For any invertible B²= ³, ² c ³d ~ ² d ³c Proof. Proof of part 1) is left for the reader. For part 2), we have for all ³, then d B(> i Á = i ) and so dd B²= ii Á > ii ³. Of course, dd is not equal to . However, in the finite-dimensional case, if we use the natural maps to identify = ii with = and > ii with > , then we can think of dd

The Isomorphism Theorems

105

as being in B²= Á > ³. Using these identifications, we do have equality in the finite-dimensional case. Theorem 3.18 Let = and > be finite-dimensional and let B²= Á > ³. If we identify = ii with = and > ii with > using the natural maps, then dd is identified with . Proof. For any % = let the corresponding element of = ii be denoted by % and similarly for > . Then before making any identifications, we have for # = , dd ²#³² ³ ~ #´ d ² ³µ ~ #² ³ ~ ² #³ ~ #² ³ for all > * and so dd ²#³ ~ # > ii Therefore, using the canonical identifications for both = ii and > ii we have dd ²#³ ~ # for all # = . The next result describes the kernel and image of the operator adjoint. Theorem 3.19 Let B²= Á > ³. Then 1) ker² d ³ ~ im² ³ 2) im² d ³ ~ ker² ³ Proof. For part 1), ker² d ³ ~ ¸ > i d ² ³ ~ ¹ ~ ¸ > i ² = ³ ~ ¸¹¹ ~ ¸ > i ²im² ³³ ~ ¸¹¹ ~ im² ³ For part 2), if ~ ~ d im² d ³, then ker² ³ ker² ³ and so ker² ³ . For the reverse inclusion, let ker² ³ = i . We wish to show that ~ d ~ for some > i . On 2 ~ ker² ³, there is no problem since and d ~ agree on 2 for any > i . Let : be a complement of ker² ³. Then maps a basis 8 ~ ¸ 0¹ for : to a linearly independent set 8 ~ ¸ 0¹ in > and so we can define > i on 8 by setting ² ³ ~ and extending to all of > . Then ~ ~ d on 8 and therefore on : . Thus, ~ d im² d ³.

106

Advanced Linear Algebra

Corollary 3.20 Let B²= Á > ³, where = and > are finite-dimensional. Then rk² ³ ~ rk² d ³. In the finite-dimensional case, and d can both be represented by matrices. Let 8 ~ ² Á Ã Á ³ and 9 ~ ² Á Ã Á ³ be ordered bases for = and > , respectively, and let i 8 i ~ ²i Á Ã Á i ³ and 9i ~ ²i Á Ã Á ³

be the corresponding dual bases. Then ²´ µ8Á9 ³Á ~ ²´ µ9 ³ ~ i ´ µ and ²´ d µ9i Á8i ³Á ~ ²´ d ²i ³µ8i ³ ~ ii ´ d ²i ³µ ~ d ²i ³² ³ ~ i ² ³ Comparing the last two expressions we see that they are the same except that the roles of and are reversed. Hence, the matrices in question are transposes. Theorem 3.21 Let B²= Á > ³, where = and > are finite-dimensional. If 8 and 9 are ordered bases for = and > , respectively, and 8 i and 9i are the corresponding dual bases, then ´ d µ9i Á8i ~ ²´ µ8Á9 ³! In words, the matrices of and its operator adjoint d are transposes of one another.

Exercises 1. 2. 3. 4. 5. 6.

If = is infinite-dimensional and : is an infinite-dimensional subspace, must the dimension of = °: be finite? Explain. Prove the correspondence theorem. Prove the first isomorphism theorem. Complete the proof of Theorem 3.9. Let : be a subspace of = . Starting with a basis ¸ Á Ã Á ¹ for :Á how would you find a basis for = °: ? Use the first isomorphism theorem to prove the rank-plus-nullity theorem rk² ³ b null² ³ ~ dim²= ³

7.

for B²= Á > ³ and dim²= ³ B. Let B²= ³ and suppose that : is a subspace of = . Define a map Z ¢ = °: ¦ = °: by

The Isomorphism Theorems

107

Z ²# b :³ ~ # b : When is Z well-defined? If Z is well-defined, is it a linear transformation? What are im² Z ³ and ker² Z ³? 8. Show that for any nonzero vector # = , there exists a linear functional = i for which ²#³ £ . 9. Show that a vector # = is zero if and only if ²#³ ~ for all = i . 10. Let : be a proper subspace of a finite-dimensional vector space = and let # = ± : . Show that there is a linear functional = i for which ²#³ ~ and ² ³ ~ for all : . 11. Find a vector space = and decompositions = ~(l) ~* l+ ° +. Hence, ( * does not imply that ( * . with ( * but ) 12. Find isomorphic vectors spaces = and > with = ~ : l ) and > ~ : l + ° +. Hence, = > does not imply that = °: > °: . but ) 13. Let = be a vector space with = ~ : l ; ~ : l ; Prove that if : and : have finite codimension in = , then so does : q : and codim²: q : ³ dim²; ³ b dim²; ³ 14. Let = be a vector space with = ~ : l ; ~ : l ;

15.

16. 17. 18. 19. 20.

Suppose that : and : have finite codimension. Hence, by the previous exercise, so does : q : . Find a direct sum decomposition = ~ > l ? for which (1) > has finite codimension, (2) > : q : and (3) ? ; b ; . Let 8 be a basis for an infinite-dimensional vector space = and define, for all 8 , the map Z = i by Z ²³ ~ if ~ and otherwise, for all 8 . Does ¸ Z 8 ¹ form a basis for = i ? What do you conclude about the concept of a dual basis? Prove that if : and ; are subspaces of = , then ²: l ; ³i : i ^ ; i . Prove that d ~ and d ~ where is the zero linear operator and is the identity. Let : be a subspace of = . Prove that ²= °:³i : . Verify that a) ² b ³d ~ d b d for Á B²= Á > ³. b) ² ³d ~ d for any - and B²= Á > ³ Let B²= Á > ³, where = and > are finite-dimensional. Prove that rk² ³ ~ rk² d ³.

Chapter 4

Modules I: Basic Properties

Motivation Let = be a vector space over a field - and let B²= ³. Then for any polynomial ²%³ - ´%µ, the operator ² ³ is well-defined. For instance, if ²%³ ~ b % b % , then ² ³ ~ b b where is the identity operator and is the threefold composition k k . Thus, using the operator we can define the product of a polynomial ²%³ - ´%µ and a vector # = by ²%³# ~ ² ³²#³

(4.1)

This product satisfies the usual properties of scalar multiplication, namely, for all ²%³Á ²%³ - ´%µ and "Á # = , ²%³²" b #³ ~ ²%³" b ²%³# ²²%³ b ²%³³" ~ ²%³" b ²%³" ´²%³ ²%³µ" ~ ²%³´ ²%³"µ " ~ " Thus, for a fixed B²= ³, we can think of = as being endowed with the operations of addition and multiplication of an element of = by a polynomial in - ´%µ. However, since - ´%µ is not a field, these two operations do not make = into a vector space. Nevertheless, the situation in which the scalars form a ring but not a field is extremely important, not only in this context but in many others.

Modules Definition Let 9 be a commutative ring with identity, whose elements are called scalars. An 9 -module (or a module over 9 ) is a nonempty set 4 ,

110

Advanced Linear Algebra

together with two operations. The first operation, called addition and denoted by b , assigns to each pair ²"Á #³ 4 d 4 , an element " b # 4 . The second operation, denoted by juxtaposition, assigns to each pair ²Á #³ 9 d 4 , an element # 4 . Furthermore, the following properties must hold: 1) 4 is an abelian group under addition. 2) For all Á 9 and "Á # 4 ²" b #³ ~ " b # ² b ³" ~ " b " ² ³" ~ ² "³ " ~ " The ring 9 is called the base ring of 4 . Note that vector spaces are just special types of modules: a vector space is a module over a field. When we turn in a later chapter to the study of the structure of a linear transformation B²= ³, we will think of = as having the structure of a vector space over - as well as a module over - ´%µ and we will use the notation = . Put another way, = is an abelian group under addition, with two scalar multiplications—one whose scalars are elements of - and one whose scalars are polynomials over - . This viewpoint will be of tremendous benefit for the study of . For now, we concentrate only on modules. Example 4.1 1) If 9 is a ring, the set 9 of all ordered -tuples whose components lie in 9 is an 9 -module, with addition and scalar multiplication defined componentwise (just as in - ), ² Á Ã Á ³ b ² Á Ã Á ³ ~ ² b Á Ã Á b ³ and ² Á Ã Á ³ ~ ² Á Ã Á ³ for , Á 9 . For example, { is the {-module of all ordered -tuples of integers. 2) If 9 is a ring, the set CÁ ²9³ of all matrices of size d is an 9 module, under the usual operations of matrix addition and scalar multiplication over 9 . Since 9 is a ring, we can also take the product of matrices in CÁ ²9³. One important example is 9 ~ - ´%µ, whence CÁ ²- ´%µ³ is the - ´%µ-module of all d matrices whose entries are polynomials. 3) Any commutative ring 9 with identity is a module over itself, that is, 9 is an 9 -module. In this case, scalar multiplication is just multiplication by

Modules I: Basic Properties

111

elements of 9, that is, scalar multiplication is the ring multiplication. The defining properties of a ring imply that the defining properties of the 9 module 9 are satisfied. We shall use this example many times in the sequel.

Importance of the Base Ring Our definition of a module requires that the ring 9 of scalars be commutative. Modules over noncommutative rings can exhibit quite a bit more unusual behavior than modules over commutative rings. Indeed, as one would expect, the general behavior of 9 -modules improves as we impose more structure on the base ring 9. If we impose the very strict structure of a field, the result is the very well behaved vector space. To illustrate, we will give an example of a module over a noncommutative ring that has a basis of size for every integer ! As another example, if the base ring is an integral domain, then whenever # Á Ã Á # are linearly independent over 9 so are # Á Ã Á # for any nonzero 9 . This can fail when 9 is not an integral domain. We will also consider the property on the base ring 9 that all of its ideals are finitely generated. In this case, any finitely generated 9 -module 4 has the property that all of its submodules are also finitely generated. This property of 9 -modules fails if 9 does not have the stated property. When 9 is a principal ideal domain (such as { or - ´%µ), each of its ideals is generated by a single element. In this case, the 9 -modules are “reasonably” well behaved. For instance, in general, a module may have a basis and yet possess a submodule that has no basis. However, if 9 is a principal ideal domain, this cannot happen. Nevertheless, even when 9 is a principal ideal domain, 9 -modules are less well behaved than vector spaces. For example, there are modules over a principal ideal domain that do not have any linearly independent elements. Of course, such modules cannot have a basis.

Submodules Many of the basic concepts that we defined for vector spaces can also be defined for modules, although their properties are often quite different. We begin with submodules. Definition A submodule of an 9 -module 4 is a nonempty subset : of 4 that is an 9 -module in its own right, under the operations obtained by restricting the operations of 4 to : . We write : 4 to denote the fact that : is a submodule of 4 .

112

Advanced Linear Algebra

Theorem 4.1 A nonempty subset : of an 9 -module 4 is a submodule if and only if it is closed under the taking of linear combinations, that is, Á 9Á "Á # : ¬ " b # :

Theorem 4.2 If : and ; are submodules of 4 , then : q ; and : b ; are also submodules of 4 . We have remarked that a commutative ring 9 with identity is a module over itself. As we will see, this type of module provides some good examples of nonvector-space-like behavior. When we think of a ring 9 as an 9-module rather than as a ring, multiplication is treated as scalar multiplication. This has some important implications. In particular, if : is a submodule of 9, then it is closed under scalar multiplication, which means that it is closed under multiplication by all elements of the ring 9 . In other words, : is an ideal of the ring 9. Conversely, if ? is an ideal of the ring 9 , then ? is also a submodule of the module 9 . Hence, the submodules of the 9 -module 9 are precisely the ideals of the ring 9 .

Spanning Sets The concept of spanning set carries over to modules as well. Definition The submodule spanned (or generated) by a subset : of a module 4 is the set of all linear combinations of elements of : : ºº:»» ~ ¸ # b Ä b # 9Á # :Á ¹ A subset : 4 is said to span 4 or generate 4 if 4 ~ ºº:»». We use a double angle bracket notation for the submodule generated by a set because when we study the - -vector space/- ´%µ-module = , we will need to make a distinction between the subspace º#» ~ - # generated by # = and the submodule ºº#»» ~ - ´%µ# generated by #. One very important point to note is that if a nontrivial linear combination of the elements # Á Ã Á # in an 9 -module 4 is , # b Ä b # ~ where not all of the coefficients are , then we cannot conclude, as we could in a vector space, that one of the elements # is a linear combination of the others. After all, this involves dividing by one of the coefficients, which may not be possible in a ring. For instance, for the {-module { d { we have ²Á ³ c ²Á ³ ~ ²Á ³ but neither ²Á ³ nor ²Á ³ is an integer multiple of the other.

Modules I: Basic Properties

113

The following simple submodules play a special role in the theory. Definition Let 4 be an 9 -module. A submodule of the form ºº#»» ~ 9# ~ ¸# 9¹ for # 4 is called the cyclic submodule generated by #. Of course, any finite-dimensional vector space is the direct sum of cyclic submodules, that is, one-dimensional subspaces. One of our main goals is to show that a finitely generated module over a principal ideal domain has this property as well. Definition An 9 -module 4 is said to be finitely generated if it contains a finite set that generates 4 . More specifically, 4 is -generated if it has a generating set of size (although it may have a smaller generating set as well). Of course, a vector space is finitely generated if and only if it has a finite basis, that is, if and only if it is finite-dimensional. For modules, life is more complicated. The following is an example of a finitely generated module that has a submodule that is not finitely generated. Example 4.2 Let 9 be the ring - ´% Á % Á Ã µ of all polynomials in infinitely many variables over a field - . It will be convenient to use ? to denote % Á % Á Ã and write a polynomial in 9 in the form ²?³. (Each polynomial in 9 , being a finite sum, involves only finitely many variables, however.) Then 9 is an 9 -module and as such, is finitely generated by the identity element ²?³ ~ . Now consider the submodule : of all polynomials with zero constant term. This module is generated by the variables themselves, : ~ ºº% Á % Á Ã »» However, : is not finitely generated. To see this, suppose that . ~ ¸ Á Ã Á ¹ is a finite generating set for : . Choose a variable % that does not appear in any of the polynomials in .. Then no linear combination of the polynomials in . can be equal to % . For if

% ~ ²?³ ²?³ ~

114

Advanced Linear Algebra

then let ²?³ ~ % ²?³ b ²?³ where ²?³ does not involve % . This gives

% ~ ´% ²?³ b ²?³µ ²?³ ~

~ % ²?³ ²?³ b ²?³ ²?³ ~

~

The last sum does not involve % and so it must equal . Hence, the first sum must equal , which is not possible since ²?³ has no constant term.

Linear Independence The concept of linear independence also carries over to modules. Definition A subset : of an 9-module 4 is linearly independent if for any distinct # Á Ã Á # : and Á Ã Á 9 , we have # b Ä b # ~ ¬ ~ for all A set : that is not linearly independent is linearly dependent. It is clear from the definition that any subset of a linearly independent set is linearly independent. Recall that in a vector space, a set : of vectors is linearly dependent if and only if some vector in : is a linear combination of the other vectors in : . For arbitrary modules, this is not true. Example 4.3 Consider { as a {-module. The elements Á { are linearly dependent, since ²³ c ²³ ~ but neither one is a linear combination (i.e., integer multiple) of the other. The problem in the previous example (as noted earlier) is that # b Ä b # ~ implies that # ~ c # c Ä c # but in general, we cannot divide both sides by , since it may not have a multiplicative inverse in the ring 9 .

Modules I: Basic Properties

115

Torsion Elements In a vector space = over a field - , singleton sets ¸#¹ where # £ are linearly independent. Put another way, £ and # £ imply # £ . However, in a module, this need not be the case. Example 4.4 The abelian group { ~ ¸Á Á Ã Á c¹ is a {-module, with scalar multiplication defined by ' ~ ²' h ³ mod , for all ' { and { . However, since ~ for all { , no singleton set ¸¹ is linearly independent. Indeed, { has no linearly independent sets. This example motivates the following definition. Definition Let 4 be an 9 -module. A nonzero element # 4 for which # ~ for some nonzero 9 is called a torsion element of 4 . A module that has no nonzero torsion elements is said to be torsion-free. If all elements of 4 are torsion elements, then 4 is a torsion module. The set of all torsion elements of 4 , together with the zero element, is denoted by 4tor . If 4 is a module over an integral domain, it is not hard to see that 4tor is a submodule of 4 and that 4 °4tor is torsion-free. (We will define quotient modules shortly: they are defined in the same way as for vector spaces.)

Annihilators Closely associated with the notion of a torsion element is that of an annihilator. Definition Let 4 be an 9 -module. The annihilator of an element # 4 is ann²#³ ~ ¸ 9 # ~ ¹ and the annihilator of a submodule 5 of 4 is ann²5 ³ ~ ¸ 9 5 ~ ¸¹¹ where 5 ~ ¸# # 5 ¹. Annihilators are also called order ideals. It is easy to see that ann²#³ and ann²5 ³ are ideals of 9 . Clearly, # 4 is a torsion element if and only if ann²#³ £ ¸¹. Also, if ( and ) are submodules of 4 , then ()

¬

ann²)³ ann²(³

(note the reversal of order). Let 4 ~ ºº" Á Ã Á " »» be a finitely generated module over an integral domain 9 and assume that each of the generators " is torsion, that is, for each , there is a nonzero ann²" ³. Then, the nonzero product ~ Ä annihilates each generator of 4 and therefore every element of 4 , that is, ann²4 ³. This

116

Advanced Linear Algebra

shows that ann²4 ³ £ ¸¹. On the other hand, this may fail if 9 is not an integral domain. Also, there are torsion modules whose annihilators are trivial. (We leave verification of these statements as an exercise.)

Free Modules The definition of a basis for a module parallels that of a basis for a vector space. Definition Let 4 be an 9 -module. A subset 8 of 4 is a basis if 8 is linearly independent and spans 4 . An 9-module 4 is said to be free if 4 ~ ¸¹ or if 4 has a basis. If 8 is a basis for 4 , we say that 4 is free on 8 . We have the following analog of part of Theorem 1.7. Theorem 4.3 A subset 8 of a module 4 is a basis if and only if every nonzero # 4 is an essentially unique linear combination of the vectors in 8 . In a vector space, a set of vectors is a basis if and only if it is a minimal spanning set, or equivalently, a maximal linearly independent set. For modules, the following is the best we can do in general. We leave proof to the reader. Theorem 4.4 Let 8 be a basis for an 9 -module 4 . Then 1) 8 is a minimal spanning set. 2) 8 is a maximal linearly independent set. The {-module { has no basis since it has no linearly independent sets. But since the entire module is a spanning set, we deduce that a minimal spanning set need not be a basis. In the exercises, the reader is asked to give an example of a module 4 that has a finite basis, but with the property that not every spanning set in 4 contains a basis and not every linearly independent set in 4 is contained in a basis. It follows in this case that a maximal linearly independent set need not be a basis. The next example shows that even free modules are not very much like vector spaces. It is an example of a free module that has a submodule that is not free. Example 4.5 The set { d { is a free module over itself, using componentwise scalar multiplication ²Á ³²Á ³ ~ ²Á ³ with basis ¸²Á ³¹. But the submodule { d ¸¹ is not free since it has no linearly independent elements and hence no basis. Theorem 2.2 says that a linear transformation can be defined by specifying its values arbitrarily on a basis. The same is true for free modules.

Modules I: Basic Properties

117

Theorem 4.5 Let 4 and 5 be 9-modules where 4 is free with basis 8 ~ ¸ 0¹. Then we can define a unique 9 -map ¢ 4 ¦ 5 by specifying the values of arbitrarily for all 8 and then extending to 4 by linearity, that is, ² # b Ä b # ³ ~ # b Ä b #

Homomorphisms The term linear transformation is special to vector spaces. However, the concept applies to most algebraic structures. Definition Let 4 and 5 be 9 -modules. A function ¢ 4 ¦ 5 is an 9 homomorphism or 9 -map if it preserves the module operations, that is, ²" b #³ ~ ²"³ b ²#³ for all Á 9 and "Á # 4 . The set of all 9 -homomorphisms from 4 to 5 is denoted by hom9 ²4 Á 5 ³. The following terms are also employed: 1) An 9 -endomorphism is an 9 -homomorphism from 4 to itself. 2) An 9 -monomorphism or 9 -embedding is an injective 9 -homomorphism. 3) An 9 -epimorphism is a surjective 9 -homomorphism. 4) An 9 -isomorphism is a bijective 9 -homomorphism. It is easy to see that hom9 ²4 Á 5 ³ is itself an 9 -module under addition of functions and scalar multiplication defined by ² ³²#³ ~ ² #³ ~ ²#³ Theorem 4.6 Let hom9 ²4 Á 5 ³. The kernel and image of , defined as for linear transformations by ker² ³ ~ ¸# 4 # ~ ¹ and im² ³ ~ ¸ # # 4 ¹ are submodules of 4 and 5 , respectively. Moreover, is a monomorphism if and only if ker² ³ ~ ¸¹. If 5 is a submodule of the 9 -module 4 , then the map ¢ 5 ¦ 4 defined by ²#³ ~ # is evidently an 9 -monomorphism, called injection of 5 into 4 .

Quotient Modules The procedure for defining quotient modules is the same as that for defining quotient vector spaces. We summarize in the following theorem.

118

Advanced Linear Algebra

Theorem 4.7 Let : be a submodule of an 9 -module 4 . The binary relation "#¯"c#: is an equivalence relation on 4 , whose equivalence classes are the cosets # b : ~ ¸# b

:¹

of : in 4 . The set 4 °: of all cosets of : in 4 , called the quotient module of 4 modulo : , is an 9 -module under the well-defined operations ²" b :³ b ²# b :³ ~ ²" b #³ b : ²" b :³ ~ " b : The zero element in 4 °: is the coset b : ~ : . One question that immediately comes to mind is whether a quotient module of a free module must be free. As the next example shows, the answer is no. Example 4.6 As a module over itself, { is free on the set ¸¹. For any , the set { ~ ¸' ' {¹ is a free cyclic submodule of {, but the quotient {module {°{ is isomorphic to { via the map ²" b {³ ~ " mod and since { is not free as a {-module, neither is {°{.

The Correspondence and Isomorphism Theorems The correspondence and isomorphism theorems for vector spaces have analogs for modules. Theorem 4.8 (The correspondence theorem) Let : be a submodule of 4 . Then the function that assigns to each intermediate submodule : ; 4 the quotient submodule ; °: of 4 °: is an order-preserving (with respect to set inclusion) one-to-one correspondence between submodules of 4 containing : and all submodules of 4 °: . Theorem 4.9 (The first isomorphism theorem) Let ¢ 4 ¦ 5 be an 9 homomorphism. Then the map Z ¢ 4 °ker² ³ ¦ 5 defined by Z ²# b ker² ³³ ~ # is an 9 -embedding and so 4 im² ³ ker² ³

Modules I: Basic Properties

119

Theorem 4.10 (The second isomorphism theorem) Let 4 be an 9 -module and let : and ; be submodules of 4 . Then : b; : ; : q;

Theorem 4.11 (The third isomorphism theorem) Let 4 be an 9 -module and suppose that : ; are submodules of 4 . Then 4 °: 4 ; °: ;

Direct Sums and Direct Summands The definition of direct sum of a family of submodules is a direct analog of the definition for vector spaces. Definition The external direct sum of 9 -modules 4 Á Ã Á 4 , denoted by 4 ~ 4 ^ Ä ^ 4 is the -module whose elements are ordered -tuples 4 ~ ¸²# Á Ã Á # ³ # 4 Á ~ Á Ã Á ¹ with componentwise operations ²" Á Ã Á " ³ b ²# Á Ã Á # ³ ~ ²" b # Á Ã Á " b #³ and ²# Á Ã Á # ³ ~ ²# Á Ã Á # ³ for 9 . We leave it to the reader to formulate the definition of external direct sums and products for arbitrary families of modules, in direct analogy with the case of vector spaces. Definition An 9 -module 4 is the (internal) direct sum of a family < ~ ¸: 0¹ of submodules of 4 , written 4 ~ <

or 4 ~ : 0

if the following hold: 1) (Join of the family) 4 is the sum (join) of the family < : = ~ : 0

120

Advanced Linear Algebra

2) (Independence of the family) For each 0 , : q

p

:

q £

s t

~ ¸¹

In this case, each : is called a direct summand of 4 . If < ~ ¸: Á Ã Á : ¹ is a finite family, the direct sum is often written 4 ~ : l Ä l : Finally, if 4 ~ : l ; , then : is said to be complemented and ; is called a complement of : in 4 . As with vector spaces, we have the following useful characterization of direct sums. Theorem 4.12 Let < ~ ¸: 0¹ be a family of distinct submodules of an 9 module 4 . The following are equivalent: 1) (Independence of the family) For each 0 , : q

p

:

q £

s t

~ ¸¹

2) (Uniqueness of expression for ) The zero element cannot be written as a sum of nonzero elements from distinct submodules in < . 3) (Uniqueness of expression) Every nonzero # 4 has a unique, except for order of terms, expression as a sum #~

bÄb

of nonzero elements from distinct submodules in < . Hence, a sum 4 ~ : 0

is direct if and only if any one of 1)–3) holds. In the case of vector spaces, every subspace is a direct summand, that is, every subspace has a complement. However, as the next example shows, this is not true for modules. Example 4.7 The set { of integers is a {-module. Since the submodules of { are precisely the ideals of the ring { and since { is a principal ideal domain, the submodules of { are the sets ºº»» ~ { ~ ¸' ' {¹

Modules I: Basic Properties

121

Hence, any two nonzero proper submodules of { have nonzero intersection, for if £ , then { q { ~ { where ~ lcm¸Á ¹. It follows that the only complemented submodules of { are { and ¸¹. In the case of vector spaces, there is an intimate connection between subspaces and quotient spaces, as we saw in Theorem 3.6. The problem we face in generalizing this to modules is that not all submodules are complemented. However, this is the only problem. Theorem 4.13 Let : be a complemented submodule of 4 . All complements of : are isomorphic to 4 °: and hence to each other. Proof. For any complement ; of : , the first isomorphism theorem applied to the projection ; Á: ¢ 4 ¦ ; gives ; 4 °: .

Direct Summands and Extensions of Isomorphisms Direct summands play a role in questions relating to whether certain module homomorphisms ¢ 5 ¦ 4 can be extended from a submodule 5 4 to the full module 4 . The discussion will be a bit simpler if we restrict attention to epimorphisms. If 4 ~ 5 l / , then a module epimorphism ¢ 5 ¦ 4 can be extended to an epimorphism ¢ 4 ¦ 4 simply by sending the elements of / to zero, that is, by setting ² b ³ ~ This is easily seen to be an 9 -map with ker²³ ~ ker²³ l / Moreover, if is another extension of with the same kernel as , then and agree on / as well as on 5 , whence ~ . Thus, there is a unique extension of with kernel ker²³ l / . Now suppose that ¢ 5 4 is an isomorphism. If 5 is complemented, that is, if . ~5 l/ then we have seen that there is a unique extension of for which ker²³ ~ / . Thus, the correspondence / ª ,

where ker²³ ~ /

from complements of 5 to extensions of is an injection. To see that this correspondence is a bijection, if ¢ 4 ¦ 4 is an extension of , then

122

Advanced Linear Algebra

4 ~ 5 l ker²³ To see this, we have 5 q ker²³ ~ ker²³ ~ ¸¹ and if 4 , then there is a 5 for which ~ and so ² c ³ ~ c ~ Thus, ~ b ² c ³ 5 b ker²³ which shows that ker²³ is a complement of 5 . Theorem 4.14 Let 4 and 4 be 9 -modules and let 5 4 . 1) If 4 ~ 5 l / , then any 9 -epimorphism ¢ 5 ¦ 4 has a unique extension ¢ 4 ¦ 4 to an epimorphism with ker²³ ~ ker²³ l / 2) Let ¢ 5 4 be an 9 -isomorphism. Then the correspondence / ª ,

where ker²³ ~ /

is a bijection from complements of 5 onto the extensions of . Thus, an isomorphism ¢ 5 4 has an extension to 4 if and only if 5 is complemented. Definition Let 5 4 . When the identity map ¢ 5 5 has an extension to ¢ 4 ¦ 5 , the submodule 5 is called a retract of 4 and is called the retraction map. Corollary 4.15 A submodule 5 4 is a retract of 4 if and only if 5 has a complement in 4 .

Direct Summands and One-Sided Invertibility Direct summands are also related to one-sided invertibility of 9 -maps. Definition Let ¢ ( ¦ ) be a module homomorphism. 1) A left inverse of is a module homomorphism 3 ¢ ) ¦ ( for which 3 k ~ . 2) A right inverse of is a module homomorphism 9 ¢ ) ¦ ( for which k 9 ~ . Left and right inverses are called one-sided inverses. An ordinary inverse is called a two-sided inverse. Unlike a two-sided inverse, one-sided inverses need not be unique.

Modules I: Basic Properties

123

A left-invertible homomorphism must be injective, since ~ ¬ 3 k ~ 3 k ¬ ~ Also, a right-invertible homomorphism ¢ ( ¦ ) must be surjective, since if ), then ~ ´9 ²³µ im²³ For set functions, the converses of these statements hold: is left-invertible if and only if it is injective and is right-invertible if and only if it is surjective. However, this is not the case for 9 -maps. Let ¢ 4 ¦ 4 be an injective 9 -map. Referring to Figure 4.1,

H V_im(V)

im(V)

)

im(V) -1

V_

M

M1

Figure 4.1 the map Oim²³ ¢ 4 im²³ obtained from by restricting its range to im²³ is an isomorphism and the left inverses 3 of are precisely the extensions of ²Oim²³ ³c ¢ im²³ 4 to 4 . Hence, Theorem 4.14 says that the correspondence / ª extension of ²Oim²³ ³c with kernel / is a bijection from the complements / of im²³ onto the left inverses of . Now let ¢ 4 ¦ 4 be a surjective 9 -map. Referring to Figure 4.2,

ker(V) H

V|H

M

VR=(V|H)-1 Figure 4.2

if ker²³ is complemented, that is, if 4 ~ ker²³ l /

M1

124

Advanced Linear Algebra

then O/ ¢ / 4 is an isomorphism. Thus, a map ¢ 4 ¦ 4 is a right inverse of if and only if is a range-extension of ²O/ ³c ¢ 4 / , the only difference being in the ranges of the two functions. Hence, ²O/ ³c ¢ 4 ¦ 4 is the only right inverse of with image / . It follows that the correspondence / ª ²O/ ³c ¢ 4 ¦ 4 is an injection from the complements / of ker²³ to the right inverses of . Moreover, this map is a bijection, since if 9 ¢ 4 ¦ 4 is a right inverse of , c ¢ im²9 ³ 4 , which then 9 ¢ 4 im²9 ³ and is an extension of 9 implies that 4 ~ im²9 ³ l ker²³ Theorem 4.16 Let 4 and 4 be 9 -modules and let ¢ 4 ¦ 4 be an 9 -map. 1) Let ¢ 4 Æ 4 be injective. The map / ª extension of ²Oim²³ ³c with kernel / is a bijection from the complements / of im²³ onto the left inverses of . Thus, there is exactly one left inverse of for each complement of im²³ and that complement is the kernel of the left inverse. 2) Let ¢ 4 ¦ 4 be surjective. The map / ª ²O/ ³c ¢ 4 ¦ 4 is a bijection from the complements / of ker²³ to the right inverses of . Thus, there is exactly one right inverse of for each complement / of ker²³ and that complement is the image of the right inverse. Thus, 4 ~ ker²³ l / ker²³ ^ im²³

The last part of the previous theorem is worth further comment. Recall that if ¢ = ¦ > is a linear transformation on vector spaces, then = ker² ³ ^ im² ³ This holds for modules as well provided that ker² ³ is a direct summand.

Modules Are Not as Nice as Vector Spaces Here is a list of some of the properties of modules (over commutative rings with identity) that emphasize the differences between modules and vector spaces. 1) A submodule of a module need not have a complement. 2) A submodule of a finitely generated module need not be finitely generated. 3) There exist modules with no linearly independent elements and hence with no basis. 4) A minimal spanning set or maximal linearly independent set is not necessarily a basis.

Modules I: Basic Properties

125

5) There exist free modules with submodules that are not free. 6) There exist free modules with linearly independent sets that are not contained in a basis and spanning sets that do not contain a basis. Recall also that a module over a noncommutative ring may have bases of different sizes. However, all bases for a free module over a commutative ring with identity have the same size, as we will prove in the next chapter.

Exercises 1. 2. 3.

Give the details to show that any commutative ring with identity is a module over itself. Let : ~ ¸# Á Ã Á # ¹ be a subset of a module 4 . Prove that 5 ~ ºº:»» is the smallest submodule of 4 containing : . First you will need to formulate precisely what it means to be the smallest submodule of 4 containing : . Let 4 be an 9 -module and let 0 be an ideal in 9 . Let 04 be the set of all finite sums of the form # b Ä b #

4.

where 0 and # 4 . Is 04 a submodule of 4 ? Show that if : and ; are submodules of 4 , then (with respect to set inclusion) : q ; ~ glb¸:Á ; ¹ and : b ; ~ lub¸:Á ; ¹

5. 6. 7. 8. 9.

Let : : Ä be an ascending sequence of submodules of an 9 module 4 . Prove that the union : is a submodule of 4 . Give an example of a module 4 that has a finite basis but with the property that not every spanning set in 4 contains a basis and not every linearly independent set in 4 is contained in a basis. Show that, just as in the case of vector spaces, an 9 -homomorphism can be defined by assigning arbitrary values on the elements of a basis and extending by linearity. Let hom9 ²4 Á 5 ³ be an 9 -isomorphism. If 8 is a basis for 4 , prove that 8 ~ ¸ 8 ¹ is a basis for 5 . Let 4 be an 9 -module and let hom9 ²4 Á 4 ³ be an 9 -endomorphism. If is idempotent, that is, if ~ , show that 4 ~ ker² ³ l im² ³

Does the converse hold? 10. Consider the ring 9 ~ - ´%Á &µ of polynomials in two variables. Show that the set 4 consisting of all polynomials in 9 that have zero constant term is an 9 -module. Show that 4 is not a free 9 -module. 11. Prove that if 9 is an integral domain, then all 9 -modules 4 have the following property: If # Á Ã Á # is linearly independent over 9 , then so is # Á Ã Á # for any nonzero 9 .

126

Advanced Linear Algebra

12. Prove that if a nonzero commutative ring 9 with identity has the property that every finitely generated 9 -module is free then 9 is a field. 13. Let 4 and 5 be 9 -modules. If : is a submodule of 4 and ; is a submodule of 5 show that 4 l5 4 5 ^ : l; : ; 14. If 9 is a commutative ring with identity and ? is an ideal of 9 , then ? is an 9-module. What is the maximum size of a linearly independent set in ? ? Under what conditions is ? free? 15. a) Show that for any module 4 over an integral domain the set 4tor of all torsion elements in a module 4 is a submodule of 4 . b) Find an example of a ring 9 with the property that for some 9 -module 4 the set 4tor is not a submodule. c) Show that for any module 4 over an integral domain, the quotient module 4 °4tor is torsion-free. 16. a) Find a module 4 that is finitely generated by torsion elements but for which ann²4 ³ ~ ¸¹. b) Find a torsion module 4 for which ann²4 ³ ~ ¸¹. 17. Let 5 be an abelian group together with a scalar multiplication over a ring 9 that satisfies all of the properties of an 9 -module except that # does not necessarily equal # for all # 5 . Show that 5 can be written as a direct sum of an 9 -module 5 and another “pseudo 9 -module” 5 . 18. Prove that hom9 ²4 Á 5 ³ is an 9 -module under addition of functions and scalar multiplication defined by ² ³²#³ ~ ² #³ ~ ²#³ 19. Prove that any 9 -module 4 is isomorphic to the 9 -module hom9 ²9Á 4 ³. 20. Let 9 and : be commutative rings with identity and let ¢ 9 ¦ : be a ring homomorphism. Show that any : -module is also an 9 -module under the scalar multiplication # ~ ²³# 21. Prove that hom{ ²{ Á { ³ { where ~ gcd²Á ³. 22. Suppose that 9 is a commutative ring with identity. If ? and @ are ideals of 9 for which 9°? 9°@ as 9 -modules, then prove that ? ~ @ . Is the result true if 9°? 9°@ as rings?

Chapter 5

Modules II: Free and Noetherian Modules

The Rank of a Free Module Since all bases for a vector space = have the same cardinality, the concept of vector space dimension is well-defined. A similar statement holds for free 9 modules when the base ring is commutative (but not otherwise). Theorem 5.1 Let 4 be a free module over a commutative ring 9 with identity. 1) Then any two bases of 4 have the same cardinality. 2) The cardinality of a spanning set is greater than or equal to that of a basis. Proof. The plan is to find a vector space = with the property that, for any basis for 4 , there is a basis of the same cardinality for = . Then we can appeal to the corresponding result for vector spaces. Let ? be a maximal ideal of 9 , which exists by Theorem 0.23. Then 9°? is a field. Our first thought might be that 4 is a vector space over 9°? , but that is not the case. In fact, scalar multiplication using the field 9°? , ² b ? ³# ~ # is not even well-defined, since this would require that ? 4 ~ ¸¹. On the other hand, we can fix precisely this problem by factoring out the submodule ? 4 ~ ¸ # b Ä b # ? Á # 4 ¹ Indeed, 4 °? 4 is a vector space over 9°? , with scalar multiplication defined by ² b ? ³²" b ? 4 ³ ~ " b ? 4 To see that this is well-defined, we must show that the conditions b ? ~ Z b ? " b ? 4 ~ "Z b ? 4 imply

128

Advanced Linear Algebra

" b ? 4 ~ Z "Z b ? 4 But this follows from the fact that " c Z "Z ~ ²" c "Z ³ b ² c Z ³"Z ? 4 Hence, scalar multiplication is well-defined. We leave it to the reader to show that 4 °? 4 is a vector space over 9°? . Consider now a set 8 ~ ¸ 0¹ 4 and the corresponding set 8 b ? 4 ~ ¸ b ? 4 0¹

4 ?4

If 8 spans 4 over 9, then 8 b ? 4 spans 4 °? 4 over 9°? . To see this, note that any # 4 has the form # ~ ' for 9 and so # b ? 4 ~ 8 9 b ? 4

~ ² b ? 4 ³

~ ² b ? ³² b ? 4 ³

which shows that 8 b ? 4 spans 4 °? 4 . Now suppose that 8 ~ ¸ 0¹ is a basis for 4 over 9 . We show that 8 b ? 4 is a basis for 4 °? 4 over 9°? . We have seen that 8 b ? 4 spans 4 °? 4 . Also, if ² b ? ³² b ? 4 ³ ~ ? 4

then ? 4 and so ~

where ? . From the linear independence of 8 we deduce that ? for all and so b ? ~ ? . Hence 8 b ? 4 is linearly independent and therefore a basis, as desired. To see that (8 ( ~ (8 b ? 4 (, note that if b ? 4 ~ b ? 4 , then c ~

where ? . If £ , then the coefficient of on the right must be equal to

Modules II: Free and Noetherian Modules

129

and so ? , which is not possible since ? is a maximal ideal. Hence, ~ . Thus, if 8 is a basis for 4 over 9 , then (8 ( ~ (8 b ? 4 ( ~ dim9°? ²4 °? 4 ³ and so all bases for 4 over 9 have the same cardinality, which proves part 1). Finally, if 8 spans 4 over 9 , then 8 b ? 4 spans 4 °? 4 and so dim9°? ²4 °? 4 ³ (8 b ? 4 ( (8 ( Thus, 8 has cardinality at least as great as that of any basis for 4 over 9 . The previous theorem allows us to define the rank of a free module. (The term dimension is not used for modules in general.) Definition Let 9 be a commutative ring with identity. The rank rk²4 ³ of a nonzero free 9 -module 4 is the cardinality of any basis for 4 . The rank of the trivial module ¸¹ is . Theorem 5.1 fails if the underlying ring of scalars is not commutative. The next example describes a module over a noncommutative ring that has the remarkable property of possessing a basis of size for any positive integer . Example 5.1 Let = be a vector space over - with a countably infinite basis 8 ~ ¸ Á Á Ã ¹. Let B²= ³ be the ring of linear operators on = . Observe that B²= ³ is not commutative, since composition of functions is not commutative. The ring B²= ³ is an B²= ³-module and as such, the identity map forms a basis for B²= ³. However, we can also construct a basis for B²= ³ of any desired finite size . To understand the idea, consider the case ~ and define the operators and by ² ³ ~ Á ²b ³ ~ and ² ³ ~ Á ²b ³ ~ These operators are linearly independent essentially because they are surjective and their supports are disjoint. In particular, if b ~ then ~ ² b ³² ³ ~ ² ³

130

Advanced Linear Algebra

and ~ ² b ³²b ³ ~ ² ³ which shows that ~ and ~ . Moreover, if B²= ³, then we define and by ² ³ ~ ² ³ ² ³ ~ ²b ³ from which it follows easily that ~ b which shows that ¸ Á ¹ is a basis for B²= ³. More generally, we begin by partitioning 8 into blocks. For each ~ Á Ã Á c , let 8 ~ ¸ mod ¹ Now we define elements B²= ³ by ²b! ³ ~ !Á where ! and where !Á is the Kronecker delta function. These functions are surjective and have disjoint support. It follows that 9 ~ ¸0 Á Ã Á c ¹ is linearly independent. For if ~ b Ä b c c where B²= ³, then, applying this to b! gives ~ ! ! ²b! ³ ~ ! ² ³ for all . Hence, ! ~ . Also, 9 spans B²= ³, for if B²= ³, we define B²= ³ by ² ³ ~ ²+ ³ to get ² b Ä b c c ³²+! ³ ~ ! ! ²+! ³ ~ ! ² ³ ~ ²+! ³ and so ~ b Ä b c c Thus, 9 ~ ¸0 Á Ã Á c ¹ is a basis for B²= ³ of size . Recall that if ) is a basis for a vector space = over - , then = is isomorphic to the vector space ²- ) ³ of all functions from ) to - that have finite support. A

Modules II: Free and Noetherian Modules

131

similar result holds for free 9 -modules. We begin with the fact that ²9 ) ³ is a free 9 -module. The simple proof is left to the reader. Theorem 5.2 Let ) be any set and let 9 be a commutative ring with identity. The set ²9 ) ³ of all functions from ) to 9 that have finite support is a free 9 module of rank () ( with basis 8 ~ ¸ ¹ where ²%³ ~ F

if % ~ if % £

This basis is referred to as the standard basis for ²9 ) ³ . Theorem 5.3 Let 4 be an 9 -module. If ) is a basis for 4 , then 4 is isomorphic to ²9 ) ³ . Proof. Consider the map ¢ 4 ¦ ²9 ) ³ defined by setting ~ where is defined in Theorem 5.2 and extending to 4 by linearity. Since maps a basis for 4 to a basis 8 ~ ¸ ¹ for ²9 ) ³ , it follows that is an isomorphism from 4 to ²9 ) ³ . Theorem 5.4 Two free 9 -modules (over a commutative ring) are isomorphic if and only if they have the same rank. Proof. If 4 5 , then any isomorphism from 4 to 5 maps a basis for 4 to a basis for 5 . Since is a bijection, we have rk²4 ³ ~ rk²5 ³. Conversely, suppose that rk²4 ³ ~ rk²5 ³. Let 8 be a basis for 4 and let 9 be a basis for 5 . Since (8 ( ~ (9(, there is a bijective map ¢ 8 ¦ 9 . This map can be extended by linearity to an isomorphism of 4 onto 5 and so 4 5 . We have seen that the cardinality of a (minimal) spanning set for a free module 4 is at least equal to rk²4 ³. Let us now speak about the cardinality of maximal linearly independent sets. Theorem 5.5 Let 9 be an integral domain and let 4 be a free 9 -module. Then all linearly independent sets have cardinality at most rk²4 ³. Proof. Since 4 ²9 ³ we need only prove the result for ²9 ³ . Let 8 be the field of quotients of 9 . Then ²8 ³ is a vector space. Now, if 8 ~ ¸# 0¹ ²9 ³ ²8 ³ is linearly independent over 8 as a subset of ²8 ³ , then 8 is clearly linearly independent over 9 as a subset of ²9 ³ . Conversely, suppose that 8 is linearly independent over 9 and # b Ä b # ~

132

Advanced Linear Algebra

where £ for all and £ for some . Multiplying by produces a nontrivial linear dependency over 9 ,

# b Ä b

~

Ä

£

# ~

which implies that ~ for all . Thus 8 is linearly dependent over 9 if and only if it is linearly dependent over 8. But in the vector space ²8 ³ , all sets of cardinality greater than are linearly dependent over 8 and hence all subsets of ²9 ³ of cardinality greater than are linearly dependent over 9 .

Free Modules and Epimorphisms If ¢ 4 ¦ - is a module epimorphism where - is free on 8, then it is easy to define a right inverse for , since we can define an 9 -map 9 ¢ - ¦ 4 by specifying its values arbitrarily on 8 and extending by linearity. Thus, we take 9 ²³ to be any member of c ²³. Then Theorem 4.16 implies that ker²³ is a direct summand of 4 and 4 ker²³ ^ This discussion applies to the canonical projection ¢ 4 ¦ 4 °: provided that the quotient 4 °: is free. Theorem 5.6 Let 9 be a commutative ring with identity. 1) If ¢ 4 ¦ - is an 9 -epimorphism and - is free, then ker²³ is complemented and 4 ~ ker²³ l 5 ker²³ ^ where 5 - . 2) If : is a submodule of 4 and if 4 °: is free, then : is complemented and 4 :^

4 :

If 4 Á : and 4 °: are free, then rk²4 ³ ~ rk²:³ b rk6

4 7 :

and if the ranks are all finite, then rk6

4 7 ~ rk²4 ³ c rk²:³ :

Noetherian Modules One of the most desirable properties of a finitely generated 9 -module 4 is that all of its submodules be finitely generated:

Modules II: Free and Noetherian Modules

4 finitely generated, : 4

¬

133

: finitely generated

Example 4.2 shows that this is not always the case and leads us to search for conditions on the ring 9 that will guarantee this property for 9 -modules. Definition An 9 -module 4 is said to satisfy the ascending chain condition (abbreviated ACC) on submodules if every ascending sequence of submodules : : : Ä of 4 is eventually constant, that is, there exists an index for which : ~ :b ~ :kb ~ Ä Modules with the ascending chain condition on submodules are also called Noetherian modules (after Emmy Noether, one of the pioneers of module theory). Since a ring 9 is a module over itself and since the submodules of the module 9 are precisely the ideals of the ring 9 , the preceding definition can be formulated for rings as follows. Definition A ring 9 is said to satisfy the ascending chain condition (abbreviated ACC) on ideals if any ascending sequence ? ? ? Ä of ideals of 9 is eventually constant, that is, there exists an index for which ? ~ ?b ~ ?b2 ~ Ä A ring that satisfies the ascending chain condition on ideals is called a Noetherian ring. The following theorem describes the relevance of this to the present discussion. Theorem 5.7 1) An 9 -module 4 is Noetherian if and only if every submodule of 4 is finitely generated. 2) In particular, a ring 9 is Noetherian if and only if every ideal of 9 is finitely generated. Proof. Suppose that all submodules of 4 are finitely generated and that 4 contains an infinite ascending sequence : : :3 Ä of submodules. Then the union

(5.1)

134

Advanced Linear Algebra

: ~ :

is easily seen to be a submodule of 4 . Hence, : is finitely generated, say : ~ ºº" Á Ã Á " »». Since " : , there exists an index such that " : . Therefore, if ~ max¸ Á Ã Á ¹, we have ¸" Á Ã Á " ¹ : and so : ~ ºº" Á Ã Á " »» : :b :b Ä : which shows that the chain (5.1) is eventually constant. For the converse, suppose that 4 satisfies the ACC on submodules and let : be a submodule of 4 . Pick " : and consider the submodule : ~ ºº" »» : generated by " . If : ~ : , then : is finitely generated. If : £ : , then there is a " : c : . Now let : ~ ºº" ," »». If : ~ : , then : is finitely generated. If : £ : , then pick "3 : c : and consider the submodule :3 ~ ºº" ," ,"3 »». Continuing in this way, we get an ascending chain of submodules ºº" »» ºº" Á " »» ºº" Á " Á " »» Ä : If none of these submodules were equal to : , we would have an infinite ascending chain of submodules, each properly contained in the next, which contradicts the fact that 4 satisfies the ACC on submodules. Hence, : ~ ºº" Á Ã Á " »» for some and so : is finitely generated. Our goal is to find conditions under which all finitely generated 9 -modules are Noetherian. The very pleasing answer is that all finitely generated 9 -modules are Noetherian if and only if 9 is Noetherian as an 9 -module, or equivalently, as a ring. Theorem 5.8 Let 9 be a commutative ring with identity. 1) 9 is Noetherian if and only if every finitely generated 9 -module is Noetherian. 2) Let 9 be a principal ideal domain. If an 9 -module 4 is -generated, then any submodule of 4 is also -generated. Proof. For part 1), one direction is evident. Assume that 9 is Noetherian and let 4 ~ ºº" Á Ã Á " »» be a finitely generated 9 -module. Consider the epimorphism ¢ 9 ¦ 4 defined by ² Á Ã Á ³ ~ " b Ä b " Let : be a submodule of 4 . Then

Modules II: Free and Noetherian Modules

135

c ²:³ ~ ¸" 9 " :¹ is a submodule of 9 and ² c :³ ~ : . If every submodule of 9 is finitely generated, then c ²:³ is finitely generated and so c ²:³ ~ ºº# Á Ã Á # »». Then : is finitely generated by ¸ # Á Ã Á # ¹. Thus, it is sufficient to prove the theorem for 9 , which we do by induction on . If ~ , any submodule of 9 is an ideal of 9 , which is finitely generated by assumption. Assume that every submodule of 9 is finitely generated for all and let : be a submodule of 9 . If , we can extract from : something that is isomorphic to an ideal of 9 and so will be finitely generated. In particular, let : be the “last coordinates” in :, specifically, let : ~ ¸²Á Ã Á Á ³ ² Á Ã Á c Á ³ : for some Á Ã Á c 9¹ The set : is isomorphic to an ideal of 9 and is therefore finitely generated, say : ~ ºº= »», where = ~ ¸ Á Ã Á ¹ is a finite subset of : . Also, let : ~ ¸# : # ~ ² Á Ã Á c Á ³ for some Á Ã Á c 9¹ be the set of all elements of : that have last coordinate equal to . Note that : is a submodule of 9 and is isomorphic to a submodule of 9 c . Hence, the inductive hypothesis implies that : is finitely generated, say : ~ ºº= »», where = is a finite subset of : . By definition of : , each = has the form ~ ²Á Ã Á Á Á ³ for Á 9 where there is a : of the form ~ ²Á Á Ã Á Ác Á Á ³ Let = ~ ¸ Á Ã Á ¹. We claim that : is generated by the finite set = r = . To see this, let # ~ ² Á Ã Á ³ : . Then ²Á Ã Á Á ³ : and so

²Á Ã Á Á ³ ~ ~

for 9. Consider now the sum

136

Advanced Linear Algebra

$ ~ ºº= »» ~

The last coordinate of this sum is

Á ~ ~

and so the difference # c $ has last coordinate and is thus in : ~ ºº= »». Hence # ~ ²# c $³ b $ ºº= »» b ºº= »» ~ ºº= r = »» as desired. For part 2), we leave it to the reader to review the proof and make the necessary changes. The key fact is that : is isomorphic to an ideal of 9 , which is principal. Hence, : is generated by a single element of 4 .

The Hilbert Basis Theorem Theorem 5.8 naturally leads us to ask which familiar rings are Noetherian. The following famous theorem describes one very important case. Theorem 5.9 (Hilbert basis theorem) If a ring 9 is Noetherian, then so is the polynomial ring 9´%µ. Proof. We wish to show that any ideal ? in 9´%µ is finitely generated. Let 3 denote the set of all leading coefficients of polynomials in ? , together with the element of 9 . Then 3 is an ideal of 9 . To see this, observe that if 3 is the leading coefficient of ²%³ ? and if 9 , then either ~ or else is the leading coefficient of ²%³ ? . In either case, 3. Similarly, suppose that 3 is the leading coefficient of ²%³ ? . We may assume that deg ²%³ ~ and deg ²%³ ~ , with . Then ²%³ ~ %c ²%³ is in ? , has leading coefficient and has the same degree as ²%³. Hence, either c is or c is the leading coefficient of ²%³ c ²%³ ? . In either case c 3. Since 3 is an ideal of the Noetherian ring 9 , it must be finitely generated, say 3 ~ º Á Ã Á ». Since 3, there exist polynomials ²%³ ? with leading coefficient . By multiplying each ²%³ by a suitable power of %, we may assume that deg ²%³ ~ ~ max¸deg ²%³¹ for all ~ Á Ã Á .

Modules II: Free and Noetherian Modules

137

Now for ~ Á Ã Á c let 3 be the set of all leading coefficients of polynomials in ? of degree , together with the element of 9 . A similar argument shows that 3 is an ideal of 9 and so 3 is also finitely generated. Hence, we can find polynomials 7 ~ ¸Á ²%³Á Ã Á Á ²%³¹ in ? whose leading coefficients constitute a generating set for 3 . Consider now the finite set c

7 ~ 8 7 9 r ¸ ²%³Á Ã Á ²%³¹ ~

If @ is the ideal generated by 7 , then @ ? . An induction argument can be used to show that @ ~ ? . If ²%³ ? has degree , then it is a linear combination of the elements of 7 (which are constants) and is thus in @ . Assume that any polynomial in ? of degree less than is in @ and let ²%³ ? have degree . If , then some linear combination ²%³ over 9 of the polynomials in 7 has the same leading coefficient as ²%³ and if , then some linear combination ²%³ of the polynomials B%c ²%³Á Ã Á %c ²%³C @ has the same leading coefficient as ²%³. In either case, there is a polynomial ²%³ @ that has the same leading coefficient as ²%³. Since ²%³ c ²%³ ? has degree strictly smaller than that of ²%³ the induction hypothesis implies that ²%³ c ²%³ @ and so ²%³ ~ ´²%³ c ²%³µ b ²%³ @ This completes the induction and shows that ? ~ @ is finitely generated.

Exercises 1. 2. 3. 4. 5. 6.

If 4 is a free 9 -module and ¢ 4 ¦ 5 is an epimorphism, then must 5 also be free? Let ? be an ideal of 9 . Prove that if 9°? is a free 9 -module, then ? is the zero ideal. Prove that the union of an ascending chain of submodules is a submodule. Let : be a submodule of an 9-module 4 . Show that if 4 is finitely generated, so is the quotient module 4 °: . Let : be a submodule of an 9 -module. Show that if both : and 4 °: are finitely generated, then so is 4 . Show that an 9-module 4 satisfies the ACC for submodules if and only if the following condition holds. Every nonempty collection I of submodules

138

7.

8. 9. 10. 11.

12. 13.

14. 15.

Advanced Linear Algebra of 4 has a maximal element. That is, for every nonempty collection I of submodules of 4 there is an : I with the property that ; I ¬ ; :. Let ¢ 4 ¦ 5 be an 9 -homomorphism. a) Show that if 4 is finitely generated, then so is im² ³. b) Show that if ker² ³ and im² ³ are finitely generated, then 4 ~ ker² ³ b : where : is a finitely generated submodule of 4 . Hence, 4 is finitely generated. If 9 is Noetherian and ? is an ideal of 9 show that 9°? is also Noetherian. Prove that if 9 is Noetherian, then so is 9´% Á Ã Á % µ. Find an example of a commutative ring with identity that does not satisfy the ascending chain condition. a) Prove that an 9-module 4 is cyclic if and only if it is isomorphic to 9°? where ? is an ideal of 9 . b) Prove that an 9 -module 4 is simple (4 £ ¸¹ and 4 has no proper nonzero submodules) if and only if it is isomorphic to 9°? where ? is a maximal ideal of 9 . c) Prove that for any nonzero commutative ring 9 with identity, a simple 9-module exists. Prove that the condition that 9 be a principal ideal domain in part 2) of Theorem 5.8 is required. Prove Theorem 5.8 in the following way. a) Show that if ; : are submodules of 4 and if ; and :°; are finitely generated, then so is : . b) The proof is again by induction. Assuming it is true for any module generated by elements, let 4 ~ ºº# Á Ã Á #b »» and let 4 Z ~ ºº# Á Ã Á # »». Then let ; ~ : q 4 Z in part a). Prove that any 9 -module 4 is isomorphic to the quotient of a free module - . If 4 is finitely generated, then - can also be taken to be finitely generated. Prove that if : and ; are isomorphic submodules of a module 4 it does not necessarily follow that the quotient modules 4 °: and 4 °; are isomorphic. Prove also that if : l ; : l ; as modules it does not necessarily follow that ; ; . Prove that these statements do hold if all modules are free and have finite rank.

Chapter 6

Modules over a Principal Ideal Domain

We remind the reader of a few of the basic properties of principal ideal domains. Theorem 6.1 Let 9 be a principal ideal domain. 1) An element 9 is irreducible if and only if the ideal º» is maximal. 2) An element in 9 is prime if and only if it is irreducible. 3) 9 is a unique factorization domain. 4) 9 satisfies the ascending chain condition on ideals. Hence, so does any finitely generated 9 -module 4 . Moreover, if 4 is -generated, then any submodule of 4 is -generated.

Annihilators and Orders When 9 is a principal ideal domain, all annihilators are generated by a single element. This permits the following definition. Definition Let 9 be a principal ideal domain and let 4 be an 9 -module. 1) If 5 is a submodule of 4 , then any generator of ann²5 ³ is called an order of 5 . 2) An order of an element # 4 is an order of the submodule ºº#»». For readers acquainted with group theory, we mention that the order of a module corresponds to the smallest exponent of a group, not to the order of the group. Theorem 6.2 Let 9 be a principal ideal domain and let 4 be an 9 -module. 1) If is an order of 5 4 , then the orders of 5 are precisely the associates of . We denote any order of 5 by ²5 ³ and, as is customary, refer to ²5 ³ as “the” order of 5 . 2) If 4 ~ ( l ) , then ²4 ³ ~ lcm²²(³Á ²)³³

140

Advanced Linear Algebra

that is, the orders of 4 are precisely the least common multiples of the orders of ( and ) . Proof. We leave proof of part 1) for the reader. For part 2), suppose that ²4 ³ ~ Á

²(³ ~ Á

²)³ ~ Á

~ lcm²Á ³

Then ( ~ ¸¹ and ) ~ ¸¹ imply that and and so . On the other hand, annihilates both ( and ) and therefore also 4 ~ ( l ) . Hence, and so is an order of 4 .

Cyclic Modules The simplest type of nonzero module is clearly a cyclic module. Despite their simplicity, cyclic modules will play a very important role in our study of linear operators on a finite-dimensional vector space and so we want to explore some of their basic properties, including their composition and decomposition. Theorem 6.3 Let 9 be a principal ideal domain. 1) If ºº#»» is a cyclic 9 -module with annihilator º», then the multiplication map ¢ 9 ¦ ºº#»» defined by ~ # is an 9 -epimorphism with kernel º». Hence the induced map ¢

9 ¦ ºº#»» º »

defined by ² b º»³ ~ # is an isomorphism. In other words, cyclic 9 -modules are isomorphic to quotient modules of the base ring 9 . 2) Any submodule of a cyclic 9 -module is cyclic. 3) If ºº#»» is a cyclic submodule of 4 of order , then for 9 , ²ºº #»»³ ~ gcd² Á ³ Also, ºº #»» ~ ºº#»»

¯

²²#³Á ³ ~

¯

²#³ ~ ²#³

Proof. We leave proof of part 1) as an exercise. For part 2), let : ºº#»». Then 0 ~ ¸ 9 # :¹ is an ideal of 9 and so 0 ~ º » for some 9 . Thus, : ~ 0# ~ 9 # ~ ºº #»» For part 3), we have ² #³ ~ if and only if ² ³# ~ , that is, if and only if , which is equivalent to

Modules Over a Principal Ideal Domain

141

d gcd²Á ³

Thus, ann² #³ if and only if º » and so ann² #³ ~ º ». For the second statement, if ²Á ³ ~ then there exist Á 9 for which b ~ and so # ~ ² b ³# ~ # ºº #»» ºº#»» and so ºº #»» ~ ºº#»». Of course, if ºº #»» ~ ºº#»» then ²#³ ~ . Finally, if ² #³ ~ , then ~ ² #³ ~ gcd²Á ³ and so ²Á ³ ~ .

The Decomposition of Cyclic Modules The following theorem shows how cyclic modules can be composed and decomposed. Theorem 6.4 Let 4 be an 9 -module. 1) (Composing cyclic modules) If " Á Ã Á " 4 have relatively prime orders, then ²" b Ä b " ³ ~ ²" ³Ä²" ³ and ºº" »» l Ä l ºº" »» ~ ºº" b Ä b " »» Consequently, if 4 ~ ( b Ä b ( where the submodules ( have relatively prime orders, then the sum is direct. 2) (Decomposing cyclic modules) If ²#³ ~ Ä where the 's are pairwise relatively prime, then # has the form # ~ " b Ä b " where ²" ³ ~ and so ºº#»» ~ ºº" b Ä b " »» ~ ºº" »» l Ä l ºº"»» Proof. For part 1), let ~ ²" ³, Ä and # " b Ä b " . Then since annihilates #, the order of # divides . If ²#³ is a proper divisor of , then for some index , there is a prime for which ° annihilates #. But ° annihilates each " for £ . Thus,

142

Advanced Linear Algebra

~

# ~ " ~ 6 7"

Since ²" ³ and ° are relatively prime, the order of ²° ³" is equal to ²" ³ ~ , which contradicts the equation above. Hence, ²#³ ~ . It is clear that ºº" b Ä b " »» ºº" »» l Ä l ºº" »». For the reverse inclusion, since and ° are relatively prime, there exist Á 9 for which b ~ Hence " ~ 6 b

" ~ ²" b Ä b " ³ ºº" b Ä b " »» 7" ~

Similarly, " ºº" b Ä b " »» for all and so we get the reverse inclusion. Finally, to see that the sum above is direct, note that if # b Ä b # ~ where # ( , then each # must be , for otherwise the order of the sum on the left would be different from . For part 2), the scalars ~ ° are relatively prime and so there exist 9 for which b Ä b ~ Hence, # ~ ² b Ä b ³# ~ # b Ä b # Since ² #³ ~ °gcd²Á ³ ~ and since and are relatively prime, we have ² #³ ~ . The second statement follows from part 1).

Free Modules over a Principal Ideal Domain We have seen that a submodule of a free module need not be free: The submodule { d ¸¹ of the module { d { over itself is not free. However, if 9 is a principal ideal domain this cannot happen. Theorem 6.5 Let 4 be a free module over a principal ideal domain 9 . Then any submodule : of 4 is also free and rk²:³ rk²4 ³. Proof. We will give the proof first for modules of finite rank and then generalize to modules of arbitrary rank. Since 4 9 where ~ rk²4 ³ is finite, we may in fact assume that 4 ~ 9 . For each , let

Modules Over a Principal Ideal Domain

143

0 ~ ¸ 9 ² Á Ã Á c Á Á Á Ã Á ³ : for some Á Ã Á c 9¹ Then it is easy to see that 0 is an ideal of 9 and so 0 ~ º » for some 9 . Let " ~ ² Á Ã Á c Á Á Á Ã Á ³ : We claim that 8 ~ ¸" ~ Á Ã Á and £ ¹ is a basis for : . As to linear independence, suppose that 8 ~ ¸" Á Ã Á " ¹ and that " b Ä b " ~ Then comparing the th coordinates gives ~ and since £ , it follows that ~ . In a similar way, all coefficients are and so 8 is linearly independent. To see that 8 spans : , we partition the elements % : according to the largest coordinate index ²%³ with nonzero entry and induct on ²%³. If ²%³ ~ , then % ~ , which is in the span of 8 . Suppose that all % : with ²%³ are in the span of 8 and let ²%³ ~ , that is, % ~ ² Á Ã Á Á Á Ã Á ³ where £ . Then 0 and so £ and ~ for some 9 . Hence, ²% c " ³ and so & ~ % c " ºº8 »» and therefore % ºº8»». Thus, 8 is a basis for : . The previous proof can be generalized in a more or less direct way to modules of arbitrary rank. In this case, we may assume that 4 ~ ²9 ³ is the 9 -module of functions with finite support from to 9 , where is a cardinal number. We use the fact that is a well-ordered set, that is, is a totally ordered set in which any nonempty subset has a smallest element. If , the closed interval ´Á µ is ´Á µ ~ ¸% % ¹ Let : 4 . For each , let 4 ~ ¸ : supp² ³ ´Á µ¹ Then the set 0 ~ ¸ ²³ 4 ¹

144

Advanced Linear Algebra

is an ideal of 9 and so 0 ~ º ²³» for some : . We show that 8 ~ ¸ Á ²³ £ ¹ is a basis for : . First, suppose that b Ä b ~ where for . Applying this to gives ² ³ ~ and since 9 is an integral domain, ~ . Similarly, ~ for all and so 8 is linearly independent. To show that 8 spans : , since any : has finite support, there is a largest index ~ ² ³ for which ² ³ £ . Now, if ºº8 »» : , then since is wellordered, we may choose a : ± ºº8 »» for which ~ ~ ²³ is as small as possible. Then 4 . Moreover, since £ ²³ 0 , it follows that ²³ £ and ²³ ~ ²³ for some 9 . Then supp² c ³ ´Á µ and ² c ³²³ ~ ²³ c ²³ ~ and so ² c ³ , which implies that c ºº8 »». But then ~ ² c ³ b ºº8 »» a contradiction. Thus, 8 is a basis for : . In a vector space of dimension , any set of linearly independent vectors is a basis. This fails for modules. For example, { is a {-module of rank but the independent set ¸¹ is not a basis. On the other hand, the fact that a spanning set of size is a basis does hold for modules over a principal ideal domain, as we now show. Theorem 6.6 Let 4 be a free 9 -module of finite rank , where 9 is a principal ideal domain. Let : ~ ¸ Á Ã Á ¹ be a spanning set for 4 . Then : is a basis for 4 . Proof. Let 8 ~ ¸ Á Ã Á ¹ be a basis for 4 and define the map ¢ 4 ¦ 4 by ~ and extending to a surjective 9 -homomorphism. Since 4 is free, Theorem 5.6 implies that 4 ker² ³ ^ im² ³ ~ ker² ³ ^ 4 Since ker² ³ is a submodule of the free module and since 9 is a principal ideal domain, we know that ker² ³ is free of rank at most . It follows that

Modules Over a Principal Ideal Domain

145

rk²4 ³ ~ rk²ker² ³³ b rk²4 ³ and so rk²ker² ³³ ~ , that is, ker² ³ ~ ¸¹, which implies that is an 9 isomorphism and so : is a basis. In general, a basis for a submodule of a free module over a principal ideal domain cannot be extended to a basis for the entire module. For example, the set ¸¹ is a basis for the submodule { of the {-module {, but this set cannot be extended to a basis for { itself. We state without proof the following result along these lines. Theorem 6.7 Let 4 be a free 9 -module of rank , where 9 is a principal ideal domain. Let 5 be a submodule of 4 that is free of rank . Then there is a basis 8 for 4 that contains a subset : ~ ¸# Á Ã Á # ¹ for which ¸ # Á Ã Á # ¹ is a basis for 5 , for some nonzero elements Á Ã Á of 9 .

Torsion-Free and Free Modules Let us explore the relationship between the concepts of torsion-free and free. It is not hard to see that any free module over an integral domain is torsion-free. The converse does not hold, unless we strengthen the hypotheses by requiring that the module be finitely generated. Theorem 6.8 A finitely generated module over a principal ideal domain is free if and only if it is torsion-free. Proof. We leave proof that a free module over an integral domain is torsion-free to the reader. Let . ~ ¸# Á Ã Á # ¹ be a generating set for 4 . Consider first the case ~ , whence . ~ ¸#¹. Then . is a basis for 4 since singleton sets are linearly independent in a torsion-free module. Hence, 4 is free. Now suppose that . ~ ¸"Á #¹ is a generating set with "Á # £ . If . is linearly independent, we are done. If not, then there exist nonzero Á 9 for which " ~ #. It follows that 4 ~ ºº"Á #»» ºº"»» and so 4 is a submodule of a free module and is therefore free by Theorem 6.5. But the map ¢ 4 ¦ 4 defined by # ~ # is an isomorphism because 4 is torsion-free. Thus 4 is also free. Now we can do the general case. Write . ~ ¸" Á Ã Á " Á # Á Ã Á #c ¹ where : ~ ¸" Á Ã Á " ¹ is a maximal linearly independent subset of . . (Note that : is nonempty because singleton sets are linearly independent.) For each # , the set ¸" Á Ã Á " Á # ¹ is linearly dependent and so there exist 9 and Á Ã Á 9 for which

146

Advanced Linear Algebra

# b " b Ä b " ~ If ~ Äc , then 4 ~ ºº" Á Ã Á " Á # Á Ã Á #c »» ºº" Á Ã Á " »» and since the latter is a free module, so is 4 , and therefore so is 4 .

The Primary Cyclic Decomposition Theorem The first step in the decomposition of a finitely generated module 4 over a principal ideal domain 9 is an easy one. Theorem 6.9 Any finitely generated module 4 over a principal ideal domain 9 is the direct sum of a finitely generated free 9 -module and a finitely generated torsion 9 -module 4 ~ 4free l 4tor The torsion part 4tor is unique, since it must be the set of all torsion elements of 4 , whereas the free part 4free is unique only up to isomorphism, that is, the rank of the free part is unique. Proof. It is easy to see that the set 4tor of all torsion elements is a submodule of 4 and the quotient 4 °4tor is torsion-free. Moreover, since 4 is finitely generated, so is 4 °4tor . Hence, Theorem 6.8 implies that 4 °4tor is free. Hence, Theorem 5.6 implies that 4 ~ 4tor l where - 4 °4tor is free. As to the uniqueness of the torsion part, suppose that 4 ~ ; l . where ; is torsion and . is free. Then ; 4tor . But if # ~ ! b 4tor for ! ; and ., then ~ # c ! 4tor and so ~ and # ; . Thus, ; ~ 4tor . For the free part, since 4 ~ 4tor l - ~ 4tor l . , the submodules - and . are both complements of 4tor and hence are isomorphic. Note that if ¸$ Á Ã Á $ ¹ is a basis for 4free we can write 4 ~ ºº$ »» l Ä l ºº$ »» l 4tor where each cyclic submodule ºº$ »» has zero annihilator. This is a partial decomposition of 4 into a direct sum of cyclic submodules.

The Primary Decomposition In view of Theorem 6.9, we turn our attention to the decomposition of finitely generated torsion modules 4 over a principal ideal domain. The first step is to decompose 4 into a direct sum of primary submodules, defined as follows.

Modules Over a Principal Ideal Domain

147

Definition Let be a prime in 9 . A -primary (or just primary) module is a module whose order is a power of . Theorem 6.10 (The primary decomposition theorem) Let 4 be a torsion module over a principal ideal domain 9 , with order ~ Ä where the 's are distinct nonassociate primes in 9 . 1) 4 is the direct sum 4 ~ 4 l Ä l 4 where 4 ~

4 ~ ¸# 4 # ~ ¹

is a primary submodule of order . This decomposition of 4 into primary submodules is called the primary decomposition of 4 . 2) The primary decomposition of 4 is unique up to order of the summands. That is, if 4 ~ 5 l Ä l 5 where 5 is primary of order and Á Ã Á are distinct nonassociate primes, then ~ and, after a possible reindexing, 5 ~ 4 . Hence, ~ and , for ~ Á Ã Á . 3) Two 9 -modules 4 and 5 are isomorphic if and only if the summands in their primary decompositions are pairwise isomorphic, that is, if 4 ~ 4 l Ä l 4 and 5 ~ 5 l Ä l 5 are primary decompositions, then ~ and, after a possible reindexing, 4 5 for ~ Á Ã Á . Proof. Let us write ~ ° and show first that 4 ~ 4 ~ ¸ # # 4 ¹ Since ² 4 ³ ~ 4 ~ ¸¹, we have 4 4 . On the other hand, since and are relatively prime, there exist Á 9 for which b ~ and so if % 4 then

148

Advanced Linear Algebra

% ~ ² b ³% ~ % 4 Hence 4 ~ 4 . For part 1), since gcd² Á Ã Á ³ ~ , there exist scalars for which b Ä b ~ and so for any % 4 ,

% ~ ² b Ä b ³% 4 ~

Moreover, since the ² 4 ³ and the 's are pairwise relatively prime, it follows that the sum of the submodules 4 is direct, that is, 4 ~ 4 l Ä l 4 ~ 4 l Ä l 4 As to the annihilators, it is clear that º » ann² 4 ³. For the reverse inclusion, if ann² 4 ³, then ann²4 ³ and so , that is, and so º ». Thus ann² 4 ³ ~ º ». As to uniqueness, we claim that ~ Ä is an order of 4 . It is clear that annihilates 4 and so . On the other hand, 5 contains an element " of order and so the sum # ~ " b Ä b " has order , which implies that . Hence, and are associates.

Unique factorization in 9 now implies that ~ and, after a suitable reindexing, that ~ and and are associates. Hence, 5 is primary of order . For convenience, we can write 5 as 5 . Hence, 5 ¸# 4 # ~ ¹ ~ 4 But if 5 l Ä l 5 ~ 4 l Ä l 4 and 5 4 for all , we must have 5 ~ 4 for all . For part 3), if ~ and ¢ 4 5 , then the map ¢ 4 ¦ 5 defined by ² b Ä b ³ ~ ² ³ b Ä b ² ³ is an isomorphism and so 4 5 . Conversely, suppose that ¢ 4 5 . Then 4 and 5 have the same annihilators and therefore the same order ~ Ä Hence, part 1) and part 2) imply that ~ and after a suitable reindexing,

Modules Over a Principal Ideal Domain

149

~ . Moreover, since 4 ¯ ~ ¯ ² ³ ~ ¯ ~ ¯ 5 it follows that ¢ 4 5 .

The Cyclic Decomposition of a Primary Module The next step in the decomposition process is to show that a primary module can be decomposed into a direct sum of cyclic submodules. While this decomposition is not unique (see the exercises), the set of annihilators is unique, as we will see. To establish this uniqueness, we use the following result. Lemma 6.11 Let 4 be a module over a principal ideal domain 9 and let 9 be a prime. 1) If 4 ~ ¸¹, then 4 is a vector space over the field 9°º» with scalar multiplication defined by ² b º»³# ~ # for all # 4 . 2) For any submodule : of 4 the set : ²³ ~ ¸# : # ~ ¹ is also a submodule of 4 and if 4 ~ : l ; , then 4 ²³ ~ : ²³ l ; ²³ Proof. For part 1), since is prime, the ideal º» is maximal and so 9°º» is a field. We leave the proof that 4 is a vector space over 9°º» to the reader. For part 2), it is straightforward to show that : ²³ is a submodule of 4 . Since : ²³ : and ; ²³ ; we see that : ²³ q ; ²³ ~ ¸¹. Also, if # 4 ²³ , then # ~ . But # ~ b ! for some : and ! ; and so ~ # ~ b !. Since : and ! ; we deduce that ~ ! ~ , whence # : ²³ l ; ²³ . Thus, 4 ²³ : ²³ l ; ²³ . But the reverse inequality is manifest. Theorem 6.12 (The cyclic decomposition theorem of a primary module) Let 4 be a primary finitely generated torsion module over a principal ideal domain 9 , with order . 1) 4 is a direct sum 4 ~ ºº# »» l Ä l ºº# »»

(6.1)

of cyclic submodules with annihilators ann²ºº# »»³ ~ º », which can be arranged in ascending order ann²ºº# »»³ Ä ann²ºº# »»³

150

Advanced Linear Algebra or equivalently, ~ Ä

2) As to uniqueness, suppose that 4 is also the direct sum 4 ~ ºº" »» l Ä l ºº" »» of cyclic submodules with annihilators ann²ºº" »»³ ~ º », arranged in ascending order ann²ºº" »»³ Ä ann²ºº" »»³ or equivalently Ä Then the two chains of annihilators are identical, that is, ~ and ann²ºº" »»³ ~ ann²ºº# »»³ for all . Thus, and ~ for all . 3) Two -primary 9 -modules 4 ~ ºº# »» l Ä l ºº# »» and 5 ~ ºº" »» l Ä l ºº" »» are isomorphic if and only if they have the same annihilator chains, that is, if and only if ~ and, after a possible reindexing, ann²ºº" »»³ ~ ann²ºº# »»³ Proof. Let # 4 have order equal to the order of 4 , that is, ann²# ³ ~ ann²4 ³ ~ º » Such an element must exist since ²# ³ for all # 4 and if this inequality is strict, then c will annihilate 4 . If we show that ºº# »» is complemented, that is, 4 ~ ºº# »» l : for some submodule : , then since : is also a finitely generated primary torsion module over 9, we can repeat the process to get 4 ~ ºº# »» l ºº# »» l : where ann²# ³ ~ º ». We can continue this decomposition: 4 ~ ºº# »» l ºº# »» l Ä l ºº# »» l : as long as : £ ¸¹. But the ascending sequence of submodules

Modules Over a Principal Ideal Domain

151

ºº# »» ºº# »» l ºº# »» Ä must terminate since 4 is Noetherian and so there is an integer for which eventually : ~ ¸¹, giving (6.1). Let # ~ # . The direct sum 4 ~ ºº#»» l ¸¹ clearly exists. Suppose that the direct sum 4 ~ ºº#»» l : exists. We claim that if 4 4 , then it is possible to find a submodule :b for which : :b and for which the direct sum 4b ~ ºº#»» l :b also exists. This process must also stop after a finite number of steps, giving 4 ~ ºº#»» l : as desired. If 4 4 and " 4 ± 4 let :b ~ ºº: Á " c #»» for 9. Then : :b since " ¤ 4 . We wish to show that for some 9, the direct sum ºº#»» l :b exists, that is, % ºº#»» q ºº: Á " c #»» ¬ % ~ Now, there exist scalars and for which % ~ # ~ b ²" c #³ for : and so if we find a scalar for which ²" c #³ :

(6.2)

then ºº#»» q : ~ ¸¹ implies that % ~ and the proof of existence will be complete. Solving for " gives " ~ ² b ³# c ºº#»» l : ~ 4 so let us consider the ideal of all such scalars: ? ~ ¸ 9 " 4 ¹ Since ? and ? is principal, we have ? ~ º » for some . Also, since " ¤ 4 implies that ¤ ? .

152

Advanced Linear Algebra

Since ? , we have ~ and there exist 9 and ! : for which " ~ # b ! Hence, " ~ " ~ ²# b !³ ~ # b ! Now we need more information about . Multiplying the expression for " by c gives ~ " ~ c ² "³ ~ c # b c ! and since ºº#»» q : ~ ¸¹, it follows that c # ~ . Hence, c , that is, and so ~ for some 9 . Now we can write " ~ # b ! and so ²" c #³ ~ ! : Thus, we take ~ to get (6.2) and that completes the proof of existence. For uniqueness, note first that 4 has orders and and so and are associates and ~ . Next we show that ~ . According to part 2) of Lemma 6.10, 4 ²³ ~ ºº# »»²³ l Ä l ºº# »»²³ and 4 ²³ ~ ºº" »»²³ l Ä l ºº" »»²³ where all summands are nonzero. Since 4 ²³ ~ ¸¹, it follows from Lemma 6.10 that 4 ²³ is a vector space over 9°º» and so each of the preceding decompositions expresses 4 ²³ as a direct sum of one-dimensional vector subspaces. Hence, ~ dim²4 ²³ ³ ~ . Finally, we show that the exponents and are equal using induction on . If ~ , then ~ for all and since ~ , we also have ~ for all . Suppose the result is true whenever c and let ~ . Write ² Á Ã Á ³ ~ ² Á Ã Á Á Á Ã Á ³Á and ² Á Ã Á ³ ~ ² Á Ã Á ! Á Á Ã Á ³Á !

Modules Over a Principal Ideal Domain

153

Then 4 ~ ºº# »» l Ä l ºº# »» and 4 ~ ºº" »» l Ä l ºº"! »» But ºº# »» ~ ºº# »» is a cyclic submodule of 4 with annihilator º c » and so by the induction hypothesis ~ ! and ~ Á Ã Á ~ which concludes the proof of uniqueness. For part 3), suppose that ¢ 4 5 and 4 has annihilator chain ann²ºº# »»³ Ä ann²ºº# »»³ and 5 has annihilator chain ann²ºº" »»³ Ä ann²ºº" »»³ Then 5 ~ 4 ~ ºº# »» l Ä l ºº# »» and so ~ and after a suitable reindexing, ann²ºº# »»³ ~ ann²ºº# »»³ ~ ann²ºº" »»³ Conversely, suppose that 4 ~ ºº# »» l Ä l ºº# »» and 5 ~ ºº" »» l Ä l ºº" »» have the same annihilator chains, that is, ~ and ann²ºº" »»³ ~ ann²ºº# »»³ Then ºº" »»

9 9 ~ ºº# »» ann²ºº" »»³ ann²ºº# »»³

The Primary Cyclic Decomposition Now we can combine the various decompositions. Theorem 6.13 (The primary cyclic decomposition theorem) Let 4 be a finitely generated torsion module over a principal ideal domain 9 .

154

Advanced Linear Algebra

1) If 4 has order ~ Ä where the 's are distinct nonassociate primes in 9 , then 4 can be uniquely decomposed (up to the order of the summands) into the direct sum 4 ~ 4 l Ä l 4 where 4 ~

4 ~ ¸# 4 # ~ ¹

is a primary submodule with annihilator º ». Finally, each primary submodule 4 can be written as a direct sum of cyclic submodules, so that 4 ~ ´ ºº#Á »» l Ä l ºº#Á »» µ l Ä l ´ºº# Á »» l Ä l ºº#Á »»µ 4

4

where ann²ºº#Á »»³ ~ º Á » and the terms in each cyclic decomposition can be arranged so that, for each , ann²ºº#Á »»³ Ä ann²ºº#Á »»³ or, equivalently, ~ Á Á Ä Á 2) As for uniqueness, suppose that 4 ~ ´ ºº"Á »» l Ä l ºº"Á »» µ l Ä l ´ºº" Á »» l Ä l ºº"Á »»µ 5

5

is also a primary cyclic decomposition of 4 . Then, a) The number of summands is the same in both decompositions; in fact, ~ and after possible reindexing, " ~ " for all ". b) The primary submodules are the same; that is, after possible reindexing, and 5 ~ 4 c) For each primary submodule pair 5 ~ 4 , the cyclic submodules have the same annihilator chains; that is, after possible reindexing, ann²ºº"Á »»³ ~ ann²ºº#Á »»³ for all Á . In summary, the primary submodules and annihilator chains are uniquely determined by the module 4 . 3) Two 9 -modules 4 and 5 are isomorphic if and only if they have the same annihilator chains.

Modules Over a Principal Ideal Domain

155

Elementary Divisors Since the chain of annihilators

ann²ºº#Á »»³ ~ º Á »

is unique except for order, the multiset ¸ Á ¹ of generators is uniquely determined up to associate. The generators Á are called the elementary divisors of 4 . Note that for each prime , the elementary divisor Á of largest exponent is precisely the factor of ²4 ³ associated to . Let us write ElemDiv²4 ³ to denote the multiset of all elementary divisors of 4 . Thus, if ElemDiv²4 ³, then any associate of is also in ElemDiv²4 ³. We can now say that ElemDiv²4 ³ is a complete invariant for isomorphism. Technically, the function 4 ª ElemDiv²4 ³ is the complete invariant, but this hair is not worth splitting. Also, we could work with a system of distinct representatives for the associate classes of the elementary divisors, but in general, there is no way to single out a special representative. Theorem 6.14 Let 9 be a principal ideal domain. The multiset ElemDiv²4 ³ is a complete invariant for isomorphism of finitely generated torsion 9 -modules, that is, 4 5

¯

ElemDiv²4 ³ ~ ElemDiv²5 ³

We have seen (Theorem 6.2) that if 4 ~(l) then ²4 ³ ~ lcm²²(³Á ²)³³ Let us now compare the elementary divisors of 4 to those of ( and ) . Theorem 6.15 Let 4 be a finitely generated torsion module over a principal ideal domain and suppose that 4 ~(l) 1) The primary cyclic decomposition of 4 is the direct sum of the primary cyclic decompositons of ( and ) ; that is, if ( ~ ººÁ »»

and ) ~ ººÁ »»

are the primary cyclic decompositions of ( and ) , respectively, then M ~ 4ººÁ »»5 l 4ººÁ »»5 is the primary cyclic decomposition of M.

156

Advanced Linear Algebra

2) The elementary divisors of 4 are ElemDiv²4 ³ ~ ElemDiv²(³ r ElemDiv²)³ where the union is a multiset union; that is, we keep all duplicate members.

The Invariant Factor Decomposition According to Theorem 6.4, if : and ; are cyclic submodules with relatively prime orders, then : l ; is a cyclic submodule whose order is the product of the orders of : and ; . Accordingly, in the primary cyclic decomposition of 4 , 4 ~ ´ ºº#Á »» l Ä l ºº#Á »» µ l Ä l ´ºº# Á »» l Ä l ºº#Á »»µ 4

4

with elementary divisors Á satisfying (6.3)

~ Á Á Ä Á

we can combine cyclic summands with relatively prime orders. One judicious way to do this is to take the leftmost (highest-order) cyclic submodules from each group to get + ~ ºº#Á »» l Ä l ºº#Á »» and repeat the process + ~ ºº#Á »» l Ä l ºº#Á »» + ~ ºº#Á »» l Ä l ºº#Á »» Å Of course, some summands may be missing here since different primary modules 4 do not necessarily have the same number of summands. In any case, the result of this regrouping and combining is a decomposition of the form 4 ~ + l Ä l + which is called an invariant factor decomposition of 4 . For example, suppose that 4 ~ ´ºº#Á »» l ºº#Á »»µ l ´ºº#Á »»µ l ´ºº#Á »» l ºº#Á»» l ºº#Á»»µ Then the resulting regrouping and combining gives 4 ~ ´ ºº#Á »» l ºº#Á »» l ºº#Á »» µ l ´ ºº#Á »» l ºº#Á »» µ l ´ ºº#Á»» µ +

+

+

As to the orders of the summands, referring to (6.3), if + has order , then since the highest powers of each prime are taken for , the second–highest for and so on, we conclude that

Modules Over a Principal Ideal Domain

157

(6.4)

c Ä or equivalently, ann²+ ³ ann²+ ³ Ä The numbers are called invariant factors of the decomposition. For instance, in the example above suppose that the elementary divisors are Á Á Á Á Á Then the invariant factors are ~ ~ ~

The process described above that passes from a sequence Á of elementary divisors in order (6.3) to a sequence of invariant factors in order (6.4) is reversible. The inverse process takes a sequence Á Ã Á satisfying (6.4), factors each into a product of distinct nonassociate prime powers with the primes in the same order and then “peels off” like prime powers from the left. (The reader may wish to try it on the example above.) This fact, together with Theorem 6.4, implies that primary cyclic decompositions and invariant factor decompositions are essentially equivalent. Therefore, since the multiset of elementary divisors of 4 is unique up to associate, the multiset of invariant factors of 4 is also unique up to associate. Furthermore, the multiset of invariant factors is a complete invariant for isomorphism. Theorem 6.16 (The invariant factor decomposition theorem) Let 4 be a finitely generated torsion module over a principal ideal domain 9 . Then 4 ~ + l Ä l + where D is a cyclic submodule of 4 , with order , where c Ä This decomposition is called an invariant factor decomposition of 4 and the scalars are called the invariant factors of 4 . 1) The multiset of invariant factors is uniquely determined up to associate by the module 4 . 2) The multiset of invariant factors is a complete invariant for isomorphism. The annihilators of an invariant factor decomposition are called the invariant ideals of 4 . The chain of invariant ideals is unique, as is the chain of

158

Advanced Linear Algebra

annihilators in the primary cyclic decomposition. Note that is an order of 4 , that is, ann²4 ³ ~ º » Note also that the product ~ Ä of the invariant factors of 4 has some nice properties. For example, is the product of all the elementary divisors of 4 . We will see in a later chapter that in the context of a linear operator on a vector space, is the characteristic polynomial of .

Characterizing Cyclic Modules The primary cyclic decomposition can be used to characterize cyclic modules via their elementary divisors. Theorem 6.17 Let 4 be a finitely generated torsion module over a principal ideal domain, with order ~ Ä The following are equivalent: 1) 4 is cyclic. 2) 4 is the direct sum 4 ~ ºº# »» l Ä l ºº# »» of primary cyclic submodules ºº# »» of order . 3) The elementary divisors of 4 are precisely the prime power factors of : ElemDiv²4 ³ ~ ¸ Á Ã Á ¹ Proof. Suppose that 4 is cyclic. Then the primary decomposition of 4 is a primary cyclic decomposition, since any submodule of a cyclic module is cyclic. Hence, 1) implies 2). Conversely, if 2) holds, then since the orders are relatively prime, Theorem 6.4 implies that 4 is cyclic. We leave the rest of the proof to the reader.

Indecomposable Modules The primary cyclic decomposition of 4 is a decomposition of 4 into a direct sum of submodules that cannot be further decomposed. In fact, this characterizes the primary cyclic decomposition of 4 . Before justifying these statements, we make the following definition. Definition A module 4 is indecomposable if it cannot be written as a direct sum of proper submodules.

Modules Over a Principal Ideal Domain

159

We leave proof of the following as an exercise. Theorem 6.18 Let 4 be a finitely generated torsion module over a principal ideal domain. The following are equivalent: 1) 4 is indecomposable 2) 4 is primary cyclic 3) 4 has only one elementary divisor: ElemDiv²4 ³ ~ ¸ ¹

Thus, the primary cyclic decomposition of 4 is a decomposition of 4 into a direct sum of indecomposable modules. Conversely, if 4 ~ ( l Ä l ( is a decomposition of 4 into a direct sum of indecomposable submodules, then each submodule ( is primary cyclic and so this is the primary cyclic decomposition of 4 .

Indecomposable Submodules of Prime Order Readers acquainted with group theory know that any group of prime order is cyclic. However, as mentioned earlier, the order of a module corresponds to the smallest exponent of a group, not to the order of a group. Indeed, there are modules of prime order that are not cyclic. Nevertheless, cyclic modules of prime order are important. Indeed, if 4 is a finitely generated torsion module over a principal ideal domain, with order , then each prime factor of gives rise to a cyclic submodule > of 4 whose order is and so > is also indecomposable. Unfortunately, > need not be complemented and so we cannot use it to decompose 4 . Nevertheless, the theorem is still useful, as we will see in a later chapter. Theorem 6.19 Let 4 be a finitely generated torsion module over a principal ideal domain, with order . If is a prime divisor of , then 4 has a cyclic (equivalently, indecomposable) submodule > of prime order . Proof. If ~ , then there is a # 4 for which $ ~ # £ but $ ~ . Then > ~ ºº$»» is annihilated by and so ²$³ . But is prime and ²$³ £ and so ²$³ ~ . Since > has prime order, Theorem 6.18 implies that > is cyclic if and only if it is indecomposable.

Exercises 1. 2.

Show that any free module over an integral domain is torsion-free. Let 4 be a finitely generated torsion module over a principal ideal domain. Prove that the following are equivalent: a) 4 is indecomposable b) 4 has only one elementary divisor (including multiplicity)

160

3. 4.

5. 6. 7.

Advanced Linear Algebra c) 4 is cyclic of prime power order. Let 9 be a principal ideal domain and 9 b the field of quotients. Then 9 b is an 9 -module. Prove that any nonzero finitely generated submodule of 9 b is a free module of rank . Let 9 be a principal ideal domain. Let 4 be a finitely generated torsionfree 9 -module. Suppose that 5 is a submodule of 4 for which 5 is a free 9 -module of rank and 4 °5 is a torsion module. Prove that 4 is a free 9 -module of rank . Show that the primary cyclic decomposition of a torsion module over a principal ideal domain is not unique (even though the elementary divisors are). Show that if 4 is a finitely generated 9 -module where 9 is a principal ideal domain, then the free summand in the decomposition 4 ~ - l 4tor need not be unique. If ºº#»» is a cyclic 9 -module of order show that the map ¢ 9 ¦ ºº#»» defined by ~ # is a surjective 9 -homomorphism with kernel º» and so ºº#»»

8. 9.

10.

11. 12.

13.

9 º»

If 9 is an integral domain with the property that all submodules of cyclic 9 -modules are cyclic, show that 9 is a principal ideal domain. Suppose that - is a finite field and let - i be the set of all nonzero elements of - . a) Show that if ²%³ - ´%µ is a nonconstant polynomial over - and if - is a root of ²%³, then % c is a factor of ²%³. b) Prove that a nonconstant polynomial ²%³ - ´%µ of degree can have at most distinct roots in - . c) Use the invariant factor or primary cyclic decomposition of a finite {module to prove that - i is cyclic. Let 9 be a principal ideal domain. Let 4 ~ ºº#»» be a cyclic 9 -module with order . We have seen that any submodule of 4 is cyclic. Prove that for each 9 such that there is a unique submodule of 4 of order . Suppose that 4 is a free module of finite rank over a principal ideal domain 9 . Let 5 be a submodule of 4 . If 4 °5 is torsion, prove that rk²5 ³ ~ rk²4 ³. Let - ´%µ be the ring of polynomials over a field - and let - Z ´%µ be the ring of all polynomials in - ´%µ that have coefficient of % equal to . Then - ´%µ is an - Z ´%µ-module. Show that - ´%µ is finitely generated and torsion-free but not free. Is - Z ´%µ a principal ideal domain? Show that the rational numbers r form a torsion-free {-module that is not free.

More on Complemented Submodules 14. Let 9 be a principal ideal domain and let 4 be a free 9 -module.

Modules Over a Principal Ideal Domain

161

a)

Prove that a submodule 5 of 4 is complemented if and only if 4 °5 is free. b) If 4 is also finitely generated, prove that 5 is complemented if and only if 4 °5 is torsion-free. 15. Let 4 be a free module of finite rank over a principal ideal domain 9 . a) Prove that if 5 is a complemented submodule of 4 , then rk²5 ³ ~ rk²4 ³ if and only if 5 ~ 4 . b) Show that this need not hold if 5 is not complemented. c) Prove that 5 is complemented if and only if any basis for 5 can be extended to a basis for 4 . 16. Let 4 and 5 be free modules of finite rank over a principal ideal domain 9 . Let ¢ 4 ¦ 5 be an 9 -homomorphism. a) Prove that ker² ³ is complemented. b) What about im² ³? c) Prove that rk²4 ³ ~ rk²ker² ³³ b rk²im² ³³ ~ rk²ker² ³³ b rk6

4 7 ker² ³

d) If is surjective, then is an isomorphism if and only if rk²4 ³ ~ rk²5 ³. e) If 3 is a submodule of 4 and if 4 °3 is free, then rk6

4 7 ~ rk²4 ³ c rk²3³ 3

17. A submodule 5 of a module 4 is said to be pure in 4 if whenever # ¤ 4 ± 5 , then # ¤ 5 for all nonzero 9 . a) Show that 5 is pure if and only if # 5 and # ~ $ for 9 implies $ 5. b) Show that 5 is pure if and only if 4 °5 is torsion-free. c) If 9 is a principal ideal domain and 4 is finitely generated, prove that 5 is pure if and only if 4 °5 is free. d) If 3 and 5 are pure submodules of 4 , then so are 3 q 5 and 3 r 5 . What about 3 b 5 ? e) If 5 is pure in 4 , then show that 3 q 5 is pure in 3 for any submodule 3 of 4 . 18. Let 4 be a free module of finite rank over a principal ideal domain 9 . Let 3 and 5 be submodules of 4 with 3 complemented in 4 . Prove that rk²3 b 5 ³ b rk²3 q 5 ³ ~ rk²3³ b rk²5 ³

Chapter 7

The Structure of a Linear Operator

In this chapter, we study the structure of a linear operator on a finitedimensional vector space, using the powerful module decomposition theorems of the previous chapter. Unless otherwise noted, all vector spaces will be assumed to be finite-dimensional. Let = be a finite-dimensional vector space. Let us recall two earler theorems (Theorem 2.19 and Theorem 2.20). Theorem 7.1 Let = be a vector space of dimension . 1) Two d matrices ( and ) are similar (written ( ) ) if and only if they represent the same linear operator B²= ³, but possibly with respect to different ordered bases. In this case, the matrices ( and ) represent exactly the same set of linear operators in B²= ³. 2) Then two linear operators and on = are similar (written ) if and only if there is a matrix ( C that represents both operators, but with respect to possibly different ordered bases. In this case, and are represented by exactly the same set of matrices in C . Theorem 7.1 implies that the matrices that represent a given linear operator are precisely the matrices that lie in one similarity class. Hence, in order to uniquely represent all linear operators on = , we would like to find a set consisting of one simple representative of each similarity class, that is, a set of simple canonical forms for similarity. One of the simplest types of matrix is the diagonal matrix. However, these are too simple, since some operators cannot be represented by a diagonal matrix. A less simple type of matrix is the upper triangular matrix. However, these are not simple enough: Every operator (over an algebraically closed field) can be represented by an upper triangular matrix but some operators can be represented by more than one upper triangular matrix.

164

Advanced Linear Algebra

This gives rise to two different directions for further study. First, we can search for a characterization of those linear operators that can be represented by diagonal matrices. Such operators are called diagonalizable. Second, we can search for a different type of “simple” matrix that does provide a set of canonical forms for similarity. We will pursue both of these directions.

The Module Associated with a Linear Operator If B²= ³, we will think of = not only as a vector space over a field - but also as a module over - ´%µ, with scalar multiplication defined by ²%³# ~ ² ³²#³ We will write = to indicate the dependence on . Thus, = and = are modules with the same ring of scalars - ´%µ, although with different scalar multiplication if £ . Our plan is to interpret the concepts of the previous chapter for the module = . First, if dim²= ³ ~ , then dim²B²= ³³ ~ . This implies that = is a torsion module. In fact, the b vectors

Á Á Á Ã Á

are linearly dependent in B²= ³, which implies that ² ³ ~ for some nonzero polynomial ²%³ - ´%µ. Hence, ²%³ ann²= ³ and so ann²= ³ is a nonzero principal ideal of - ´%µ. Also, since = is finitely generated as a vector space, it is, a fortiori, finitely generated as an - ´%µ-module. Thus, = is a finitely generated torsion module over a principal ideal domain - ´%µ and so we may apply the decomposition theorems of the previous chapter. In the first part of this chapter, we embark on a “translation project” to translate the powerful results of the previous chapter into the language of the modules = . Let us first characterize when two modules = and = are isomorphic. Theorem 7.2 If Á B²= ³, then = =

¯

In particular, ¢ = ¦ = is a module isomorphism if and only if is a vector space automorphism of = satisfying ~ c Proof. Suppose that ¢ = ¦ = is a module isomorphism. Then for # = , ²%#³ ~ %²#³ which is equivalent to

The Structure of a Linear Operator

165

² #³ ~ ²#³ and since is bijective, this is equivalent to ²c ³# ~ # that is, ~ c . Since a module isomorphism from = to = is a vector space isomorphism as well, the result follows. For the converse, suppose that is a vector space automorphism of = and ~ c , that is, ~ . Then ²% #³ ~ ² #³ ~ ²#³ ~ % ²#³ and the - -linearity of implies that for any polynomial ²%³ - ´%µ, ²² ³#³ ~ ²³# Hence, is a module isomorphism from = to = .

Submodules and Invariant Subspaces There is a simple connection between the submodules of the - ´%µ-module = and the subspaces of the vector space = . Recall that a subspace : of = is invariant if : : . Theorem 7.3 A subset : = is a submodule of = if and only if : is a invariant subspace of = .

Orders and the Minimal Polynomial We have seen that the annihilator of = , ann²= ³ ~ ¸²%³ - ´%µ ²%³= ~ ¸¹¹ is a nonzero principal ideal of - ´%µ, say ann²= ³ ~ º²%³» Since the elements of the base ring - ´%µ of = are polynomials, for the first time in our study of modules there is a logical choice among all scalars in a given associate class: Each associate class contains exactly one monic polynomial. Definition Let B²= ³. The unique monic order of = is called the minimal polynomial for and is denoted by ²%³ or min² ³. Thus, ann²= ³ ~ º ²%³»

In treatments of linear algebra that do not emphasize the role of the module = , the minimal polynomial of a linear operator is simply defined as the unique

166

Advanced Linear Algebra

monic polynomial ²%³ of smallest degree for which ² ³ ~ . This definition is equivalent to our definition. The concept of minimal polynomial is also defined for matrices. The minimal polynomial A ²%³ of matrix ( C ²- ³ is defined as the minimal polynomial of the multiplication operator ( . Equivalently, ( ²%³ is the unique monic polynomial ²%³ - ´%µ of smallest degree for which ²(³ ~ . Theorem 7.4 1) If are similar linear operators on = , then ²%³ ~ ²%³. Thus, the minimal polynomial is an invariant under similarity of operators. 2) If ( ) are similar matrices, then ( ²%³ ~ ) ²%³. Thus, the minimal polynomial is an invariant under similarity of matrices. 3) The minimal polynomial of B²= ³ is the same as the minimal polynomial of any matrix that represents .

Cyclic Submodules and Cyclic Subspaces Let us now look at the cyclic submodules of = : ºº#»» ~ - ´%µ# ~ ¸² ³²#³ ²%³ - ´%µ¹ which are -invariant subspaces of = . Let ²%³ be the minimal polynomial of Oºº#»» and suppose that deg²²%³³ ~ . If ²%³# ºº#»», then writing ²%³ ~ ²%³²%³ b ²%³ where deg ²%³ deg ²%³ gives ²%³# ~ ´²%³²%³ b ²%³µ# ~ ²%³# and so ºº#»» ~ ¸²%³# deg ²%³ ¹ Hence, the set 8 ~ ¸#Á %#Á Ã Á %c #¹ ~ ¸#Á #Á Ã Á c #¹ spans the vector space ºº#»». To see that 8 is a basis for ºº#»», note that any linear combination of the vectors in 8 has the form ²%³# for deg²²%³³ and so is equal to if and only if ²%³ ~ . Thus, 8 is an ordered basis for ºº#»». Definition Let B²= ³. A -invariant subspace : of = is -cyclic if : has a basis of the form 8 ~ ¸#Á #Á Ã Á c #¹ for some # = and . The basis 8 is called a -cyclic basis for = .

The Structure of a Linear Operator

167

Thus, a cyclic submodule ºº#»» of = with order ²%³ of degree is a -cyclic subspace of = of dimension . The converse is also true, for if 8 ~ ¸#Á #Á Ã Á c #¹ is a basis for a -invariant subspace : of = , then : is a submodule of = . Moreover, the minimal polynomial of O: has degree , since if # ~ c # c # c Ä c c c # then O: satisfies the polynomial ²%³ ~ b % b Ä b c %c b % but none of smaller degree since 8 is linearly independent. Theorem 7.5 Let = be a finite-dimenional vector space and let : = . The following are equivalent: 1) : is a cyclic submodule of = with order ²%³ of degree 2) : is a -cyclic subspace of = of dimension . We will have more to say about cyclic modules a bit later in the chapter.

Summary The following table summarizes the connection between the module concepts and the vector space concepts that we have discussed so far. - ´%µ-Module = Scalar multiplication: ²%³# Submodule of = Annihilator: ann²= ³ ~ ¸²%³ ²%³= ~ ¸¹¹ Monic order ²%³ of = : ann²= ³ ~ º²%³» Cyclic submodule of = : ºº#»» ~ ¸²%³# deg ²%³ deg ²%³¹

- -Vector Space = Action of ² ³: ² ³²#³ -Invariant subspace of = Annihilator: ann²= ³ ~ ¸²%³ ² ³²= ³ ~ ¸¹¹ Minimal polynomial of : ²%³ has smallest deg with ² ³ ~ -cyclic subspace of = : º#Á #Á Ã Á c²#³»Á ~ deg²²%³³

The Primary Cyclic Decomposition of = We are now ready to translate the cyclic decomposition theorem into the language of = . Definition Let B²= ³. 1) The elementary divisors and invariant factors of are the monic elementary divisors and invariant factors, respectively, of the module = . We denote the multiset of elementary divisors of by ElemDiv² ³ and the multiset of invariant factors of by InvFact² ³.

168

Advanced Linear Algebra

2) The elementary divisors and invariant factors of a matrix ( are the elementary divisors and invariant factors, respectively, of the multiplication operator ( : ElemDiv²(³ ~ ElemDiv²( ³

and

InvFact²(³ ~ InvFact²( ³

We emphasize that the elementary divisors and invariant factors of an operator or matrix are monic by definition. Thus, we no longer need to worry about uniqueness up to associate. Theorem 7.6 (The primary cyclic decomposition theorem for = ³ Let = be finite-dimensional and let B²= ³ have minimal polynomial ²%³ ~ ²%³Ä ²%³ where the polynomials ²%³ are distinct monic primes. 1) (Primary decomposition) The - ´%µ-module = is the direct sum = ~ = l Ä l = where = ~

²%³ = ~ ¸# = ² ³²#³ ~ ¹ ²%³

is a primary submodule of = of order ²%³. In vector space terms, = is a -invariant subspace of = and the minimal polynomial of O= is min² O= ³ ~ ²%³ 2) (Cyclic decomposition) Each primary summand = can be decomposed into a direct sum = ~ ºº#Á »» l Ä l ºº#Á »»

of -cyclic submodules ºº#Á »» of order Á ²%³ with ~ Á Á Ä Á In vector space terms, ºº#Á »» is a -cyclic subspace of = and the minimal polynomial of Oºº#Á »» is

min² Oºº#Á »» ³ ~ Á ²%³ 3) (The complete decomposition) This yields the decomposition of = into a direct sum of -cyclic subspaces = ~ ²ºº#Á »» l Ä l ºº#Á »»³ l Ä l ²ºº#Á »» l Ä l ºº#Á »»³ 4) (Elementary divisors and dimensions) The multiset of elementary divisors ¸ Á ²%³¹ is uniquely determined by . If deg² Á ²%³³ ~ Á , then the -

The Structure of a Linear Operator

169

cyclic subspace ºº#Á »» has -cyclic basis 8Á ~ 2#Á Á #Á Á Ã Á Á c #Á 3

and dim²ºº#Á »»³ ~ deg² Á ³. Hence,

dim²= ³ ~ deg² Á ³ ~

We will call the basis H ~ 8Á Á

for = the elementary divisor basis for = . Recall that if = ~ ( l ) and if both ( and ) are -invariant subspaces of = , the pair ²(Á )³ is said to reduce . In module language, the pair ²(Á )³ reduces if ( and ) are submodules of = and = ~ ( l ) We can now translate Theorem 6.15 into the current context. Theorem 7.7 Let B²= ³ and let = ~ ( l ) 1) The minimal polynomial of is ²%³ ~ lcm² O( ²%³Á O) ²%³³ 2) The primary cyclic decomposition of = is the direct sum of the primary cyclic decompositons of ( and ) ; that is, if ( ~ ººÁ »»

and ) ~ ººÁ »»

are the primary cyclic decompositions of ( and ) , respectively, then = ~ 4ººÁ »»5 l 4ººÁ »»5 is the primary cyclic decomposition of = . 3) The elementary divisors of are ElemDiv² ³ ~ ElemDiv² O( ³ r ElemDiv² O) ³ where the union is a multiset union; that is, we keep all duplicate members.

170

Advanced Linear Algebra

The Characteristic Polynomial To continue our translation project, we need a definition. Recall that in the characterization of cyclic modules in Theorem 6.17, we made reference to the product of the elementary divisors, one from each associate class. Now that we have singled out a special representative from each associate class, we can make a useful definition. Definition Let B²= ³. The characteristic polynomial ²%³ of is the product of all of the elementary divisors of :

²%³ ~ Á ²%³ Á

Hence, deg² ²%³³ ~ dim²= ³ Similarly, the characteristic polynomial 4 ²%³ of a matrix 4 is the product of the elementary divisors of 4 . The following theorem describes the relationship between the minimal and characteristic polynomials. Theorem 7.8 Let B²= ³. 1) (The Cayley–Hamilton theorem) The minimal polynomial of divides the characteristic polynomial of : ²%³ ²%³ Equivalently, satisfies its own characteristic polynomial, that is, ² ³ ~ 2) The minimal polynomial

²%³ ~ Á ²%³ÄÁ ²%³ and characteristic polynomial

²%³ ~ Á ²%³ Á

of have the same set of prime factors ²%³ and hence the same set of roots (not counting multiplicity). We have seen that the multiset of elementary divisors forms a complete invariant for similarity. The reader should construct an example to show that the pair ² ²%³Á ²%³³ is not a complete invariant for similarity, that is, this pair of

The Structure of a Linear Operator

171

polynomials does not uniquely determine the multiset of elementary divisors of the operator . In general, the minimal polynomial of a linear operator is hard to find. One of the virtues of the characteristic polynomial is that it is comparatively easy to find and we will discuss this in detail a bit later in the chapter. Note that since ²%³ ²%³ and both polynomials are monic, it follows that ²%³ ~ ²%³

¯

deg² ²%³³ ~ deg² ²%³³

Definition A linear operator B²= ³ is nonderogatory if its minimal polynomial is equal to its characteristic polynomial: ²%³ ~ ²%³ or equivalently, if deg² ²%³³ ~ deg² ²%³³ or if deg² ²%³³ ~ dim²= ³ Similar statements hold for matrices.

Cyclic and Indecomposable Modules We have seen (Theorem 6.17) that cyclic submodules can be characterized by their elementary divisors. Let us translate this theorem into the language of = (and add one more equivalence related to the characteristic polynomial). Theorem 7.9 Let B²= ³ have minimal polynomial ²%³ ~ ²%³Ä ²%³ where ²%³ are distinct monic primes. The following are equivalent: 1) = is cyclic. 2) = is the direct sum = ~ ºº# »» l Ä l ºº# »» of -cyclic submodules ºº# »» of order ²%³. 3) The elementary divisors of are ElemDiv² ³ ~ ¸ ²%³Á Ã Á ²%³¹ 4) is nonderogatory, that is, ²%³ ~ ²%³

172

Advanced Linear Algebra

Indecomposable Modules We have also seen (Theorem 6.19) that, in the language of = , each prime factor ²%³ of the minimal polynomial ²%³ gives rise to a cyclic submodule > of = of prime order ²%³. Theorem 7.10 Let B²= ³ and let ²%³ be a prime factor of ²%³. Then = has a cyclic submodule > of prime order ²%³. For a module of prime order, we have the following. Theorem 7.11 For a module > of prime order ²%³, the following are equivalent: 1) > is cyclic 2) > is indecomposable 3) ²%³ is irreducible 4) is nonderogatory, that is, ²%³ ~ ²%³ 5) dim²> ³ ~ deg²²%³³. Our translation project is now complete and we can begin to look at issues that are specific to the modules = .

Companion Matrices We can also characterize the cyclic modules = via the matrix representations of the operator , which is obviously something that we could not do for arbitrary modules. Let = ~ ºº#»» be a cyclic module, with order ²%³ ~ b % b Ä b c %c b % and ordered -cyclic basis 8 ~ 2#Á #Á Ã Á c #3 Then ² #³ ~ b # for c and ² c #³ ~ # ~ c² b b Ä b c c ³# ~ c # c # c Ä c c c # and so

The Structure of a Linear Operator

v x x ´ µ8 ~ x x Å w

Ä Ä Æ Å Æ Ä

173

c y c { { Å { { cc2 cc z

This matrix is known as the companion matrix for the polynomial ²%³. Definition The companion matrix of a monic polyomial ²%³ ~ b % b Ä b c %c b % is the matrix v x x *´²%³µ ~ x x Å w

Ä Ä Æ Å Æ Ä

c y c { { Å { { cc2 cc z

Note that companion matrices are defined only for monic polynomials. Companion matrices are nonderogatory. Also, companion matrices are precisely the matrices that represent operators on -cyclic subspaces. Theorem 7.12 Let ²%³ - ´%µ. 1) A companion matrix ( ~ *´²%³µ is nonderogatory; in fact, ( ²%³ ~ ( ²%³ ~ ²%³ 2) = is cyclic if and only if can be represented by a companion matrix, in which case the representing basis is -cyclic. Proof. For part 1), let ; ~ ² Á Ã Á ³ be the standard basis for - . Since ~ (c for , it follows that for any polynomial ²%³, ²(³ ~

¯

²(³ ~ for all

¯

²(³ ~

If ²%³ ~ b % b Ä b c %c b % , then c

c

c

²(³ ~ ( b ( ~ b c b ~ ~

~

~

and so ²(³ ~ , whence ²(³ ~ . Also, if ²%³ ~ b % b Ä b c %c b % is nonzero and has degree , then ²(³ ~ b b Ä b c b b £

174

Advanced Linear Algebra

since ; is linearly independent. Hence, ²%³ has smallest degree among all polynomials satisfied by ( and so ²%³ ~ ( ²%³. Finally, deg²( ²%³³ ~ deg²²%³³ ~ deg²( ²%³³ For part 2), we have already proved that if = is cyclic with -cyclic basis 8 , then ´ µ8 ~ *´²%³µ. For the converse, if ´ µ8 ~ *´²%³µ, then part 1) implies that is nonderogatory. Hence, Theorem 7.11 implies that = is cyclic. It is clear from the form of *´²%³µ that 8 is a -cyclic basis for = .

The Big Picture If Á B²= ³, then Theorem 7.2 and the fact that the elementary divisors form a complete invariant for isomorphism imply that

¯

= =

¯

ElemDiv² ³ ~ ElemDiv²³

Hence, the multiset of elementary divisors is a complete invariant for similarity of operators. Of course, the same is true for matrices: ()

¯

-( -)

¯

ElemDiv²(³ ~ ElemDiv²)³

where we write -( in place of -( . The connection between the elementary divisors of an operator and the elementary divisors of the matrix representations of is described as follows. If ( ~ ´ µ8 , then the coordinate map 8 ¢ = - is also a module isomorphism 8 ¢ = ¦ -( . Specifically, we have 8 ²² ³#³ ~ ´² ³#µ8 ~ ²´ µ8 ³´#µ8 ~ ²(³8 ²#³ and so 8 preserves - ´%µ-scalar multiplication. Hence, ( ~ ´ µ8 for some 8

¬

= -(

For the converse, suppose that ¢ = -( . If we define = by ~ , where is the th standard basis vector, then 8 ~ ² Á Ã Á ³ is an ordered basis for = and ~ 8 is the coordinate map for 8 . Hence, 8 is a module isomorphism and so 8 ² #³ ~ ( ²8 #³ for all # = , that is, ´ #µ8 ~ ( ²´#µ8 ³ which shows that ( ~ ´ µ8 . Theorem 7.13 Let = be a finite-dimensional vector space over - . Let Á B²= ³ and let (Á ) C ²- ³.

The Structure of a Linear Operator

175

1) The multiset of elementary divisors (or invariant factors) is a complete invariant for similarity of operators, that is, ¯ = = ¯ ElemDiv² ³ ~ ElemDiv²³ ¯ InvFact² ³ ~ InvFact²³ A similar statement holds for matrices: ( ) ¯ -( -) ¯ ElemDiv²(³ ~ ElemDiv²)³ ¯ InvFact²(³ ~ InvFact²)³ 2) The connection between operators and their representing matrices is ( ~ ´ µ8 for some 8 ¯ = -( ¯ ElemDiv² ³ ~ ElemDiv²(³ ¯ InvFact² ³ ~ InvFact²(³

Theorem 7.13 can be summarized in Figure 7.1, which shows the big picture.

W

V

similarity classes of L(V)

VW

VV

isomorphism classes of F[x]-modules

{ED1} {ED2}

Multisets of elementary divisors

[W]B [V]B [W]R [V]R

Similarity classes of matrices Figure 7.1

Figure 7.1 shows that the similarity classes of B²= ³ are in one-to-one correspondence with the isomorphism classes of - ´%µ-modules = and that these are in one-to-one correspondence with the multisets of elementary divisors, which, in turn, are in one-to-one correspondence with the similarity classes of matrices. We will see shortly that any multiset of prime power polynomials is the multiset of elementary divisors for some operator (or matrix) and so the third family in

176

Advanced Linear Algebra

the figure could be replaced by the family of all multisets of prime power polynomials.

The Rational Canonical Form We are now ready to determine a set of canonical forms for similarity. Let B²= ³. The elementary divisor basis H for = that gives the primary cyclic decomposition of = , = ~ ²ºº#Á »» l Ä l ºº#Á »»³ l Ä l ²ºº#Á »» l Ä l ºº#Á »»³ is the union of the bases 8Á ~ ²#Á Á #Á Á Ã Á Á c #Á ³ and so the matrix of with respect to H is the block diagonal matrix

Á

´ µH ~ diag²*´Á ²%³µÁ Ã Á *´ ²%³µÁ Ã Á *´Á ²%³µÁ Ã Á *´Á ²%³µ³ with companion matrices on the block diagonal. This matrix has the following form. Definition A matrix ( is in the elementary divisor form of rational canonical form if ( ~ diag4*´ ²%³µÁ Ã Á *´ ²%³µ5 where the ²%³ are monic prime polynomials. Thus, as shown in Figure 7.1, each similarity class I contains at least one matrix in the elementary divisor form of rational canonical form. On the other hand, suppose that 4 is a rational canonical matrix

Á

4 ~ diag²*´Á ²%³µÁ Ã Á *´ ²%³µÁ Ã Á *´Á ²%³µÁ Ã Á *´ Á ²%³µ³ of size d . Then 4 represents the matrix multiplication operator 4 under the standard basis ; on - . The basis ; can be partitioned into blocks ;Á corresponding to the position of each of the companion matrices on the block diagonal of 4 . Since

´4 Oº;Á » µ;Á ~ *´ Á ²%³µ it follows from Theorem 7.12 that each subspace º;Á » is 4 -cyclic with monic order Á ²%³ and so Theorem 7.9 implies that the multiset of elementary divisors of 4 is ¸ Á ²%³¹. This shows two important things. First, any multiset of prime power polynomials is the multiset of elementary divisors for some matrix. Second, 4

The Structure of a Linear Operator

177

lies in the similarity class that is associated with the elementary divisors ¸ Á ²%³¹. Hence, two matrices in the elementary divisor form of rational canonical form lie in the same similarity class if and only if they have the same multiset of elementary divisors. In other words, the elementary divisor form of rational canonical form is a set of canonical forms for similarity, up to order of blocks on the block diagonal. Theorem 7.14 (The rational canonical form: elementary divisor version) Let = be a finite-dimensional vector space and let B²= ³ have minimal polynomial ²%³ ~ ²%³Ä ²%³ where the ²%³'s are distinct monic prime polynomials. 1) If H is an elementary divisor basis for = , then ´ µH is in the elementary divisor form of rational canonical form: Á

´ µH ~ diag4*´Á ²%³µÁ Ã Á *´ ²%³µÁ Ã Á *´Á ²%³µÁ Ã Á *´ Á ²%³µ 5

where Á ²%³ are the elementary divisors of . This block diagonal matrix is called an elementary divisor version of a rational canonical form of . 2) Each similarity class I of matrices contains a matrix 9 in the elementary divisor form of rational canonical form. Moreover, the set of matrices in I that have this form is the set of matrices obtained from 4 by reordering the block diagonal matrices. Any such matrix is called an elementary divisor verison of a rational canonical form of (. 3) The dimension of = is the sum of the degrees of the elementary divisors of , that is,

dim²= ³ ~ deg² Á ³

~ ~

Example 7.1 Let be a linear operator on the vector space s7 and suppose that has minimal polynomial ²%³ ~ ²% c ³²% b ³ Noting that % c and ²% b ³ are elementary divisors and that the sum of the degrees of all elementary divisors must equal , we have two possibilities: 1) % c Á ²% b 1³ Á % b 2) % c Á % c Á % c Á ²% b 1³ These correspond to the following rational canonical forms:

178

Advanced Linear Algebra

v x x x x 1) x x x x w

c c

v x x x x 2) x x x x w

y { { { { { { { { c z y { { { { c { { { { c z

The rational canonical form may be far from the ideal of simplicity that we had in mind for a set of simple canonical forms. Indeed, the rational canonical form can be important as a theoretical tool, more so than a practical one.

The Invariant Factor Version There is also an invariant factor version of the rational canonical form. We begin with the following simple result. Theorem 7.15 If ²%³Á ²%³ - ´%µ are relatively prime polynomials, then *´²%³²%³µ 6

*´²%³µ

*´²%³µ 7block

Proof. Speaking in general terms, if an d matrix ( has minimal polynomial ²%³ ~ ²%³Ä ²%³ of degree equal to the size of the matrix, then Theorem 7.14 implies that the elementary divisors of ( are precisely ²%³Á Ã Á ²%³ Since the matrices *´²%³²%³µ and diag²*´²%³µÁ *´²%³µ³ have the same size d and the same minimal polynomial ²%³²%³ of degree , it follows that they have the same multiset of elementary divisors and so are similar. Definition A matrix ( is in the invariant factor form of rational canonical form if

The Structure of a Linear Operator

179

( ~ diag4*´ ²%³µÁ Ã Á *´ ²%³µ5 where

b ²%³

²%³

for ~ Á Ã Á c .

Theorem 7.15 can be used to rearrange and combine the companion matrices in an elementary divisor version of a rational canonical form 9 to produce an invariant factor version of rational canonical form that is similar to 9 . Also, this process is reversible. Theorem 7.16 (The rational canonical form: invariant factor version) Let dim²= ³ B and suppose that B²= ³ has minimal polynomial ²%³ ~ ²%³Ä ²%³ where the monic polynomials ²%³ are distinct prime (irreducible) polynomials 1) = has an invariant factor basis 8 , that is, a basis for which ´ µ8 ~ diag4*´ ²%³µÁ Ã Á *´ ²%³µ5 where the polynomials ²%³ are the invariant factors of and b ²%³ ²%³. This block diagonal matrix is called an invariant factor version of a rational canonical form of . 2) Each similarity class I of matrices contains a matrix 9 in the invariant factor form of rational canonical form. Moreover, the set of matrices in I that have this form is the set of matrices obtained from 4 by reordering the block diagonal matrices. Any such matrix is called an invariant factor verison of a rational canonical form of (. 3) The dimension of = is the sum of the degrees of the invariant factors of , that is,

dim²= ³ ~ deg² ³

~

The Determinant Form of the Characteristic Polynomial In general, the minimal polynomial of an operator is hard to find. One of the virtues of the characteristic polynomial is that it is comparatively easy to find. This also provides a nice example of the theoretical value of the rational canonical form. Let us first take the case of a companion matrix. If ( ~ *´ ²%³µ is the companion matrix of a monic polynomial ²%Â Á Ã Á c ³ ~ b % b Ä b c %c b % then how can we recover ²%³ ~ ( ²%³ from *´²%³µ by arithmetic operations?

180

Advanced Linear Algebra

When ~ , we can write ²%³ as ²%Â Á ³ ~ b % b % ~ %²% b ³ b which looks suspiciously like a determinant: % b ? c ~ det6%0 c > c ?7 ~ det²%0 c *´ ²%³µ³

²%Â Á ³ ~ det>

% c

So, let us define (²%Â Á Ã Á c ³ ~ %0 c *´ ²%³µ Ä v % x c % Ä x ~ x c Æ x Å Å Æ % w Ä c

y { { Å { { c % b c z

where % is an independent variable. The determinant of this matrix is a polynomial in % whose degree equals the number of parameters Á Ã Á c . We have just seen that det²(²%Â Á ³³ ~ ²%Â Á ³ and this is also true for ~ . As a basis for induction, if det²(²%Â Á Ã Á c ³³ ~ ²%Â Á Ã Á c ³ then expanding along the first row gives det²(²%Á Á Ã Á ³³ v c x ~ % det²(²%Á Á Ã Á ³³ b ²c³ detx Å w ~ % det²(²%Á Á Ã Á ³³ b ~ % ²%Â Á Ã Á ³ b ~ % b % b Ä b % b %b b ~ b ²%Â Á Ã Á ³ We have proved the following.

% c Å

Ä y Æ { { Æ % Ä c zd

The Structure of a Linear Operator

181

Lemma 7.17 For any ²%³ - ´%µ, det²%0 c *´²%³µ³ ~ ²%³

Now suppose that 9 is a matrix in the elementary divisor form of rational canonical form. Since the determinant of a block diagonal matrix is the product of the determinants of the blocks on the diagonal, it follows that

det²%0 c 9³ ~ Á ²%³ ~ 9 ²%³ Á

Moreover, if ( 9, say ( ~ 7 97 c , then det²%0 c (³ ~ det²%0 c 7 97 c ³ ~ det ´7 ²%0 c 9³7 c µ ~ det²7 ³det²%0 c 9³det²7 c ³ ~ det²%0 c 9³ and so det²%0 c (³ ~ det²%0 c 9³ ~ 9 ²%³ ~ ( ²%³ Hence, the fact that all matrices have a rational canonical form allows us to deduce the following theorem. Theorem 7.18 Let B²= ³. If ( is any matrix that represents , then ²%³ ~ ( ²%³ ~ det²%0 c (³

Changing the Base Field A change in the base field will generally change the primeness of polynomials and therefore has an effect on the multiset of elementary divisors. It is perhaps a surprising fact that a change of base field has no effect on the invariant factors— hence the adjective invariant. Theorem 7.19 Let - and 2 be fields with - 2 . Suppose that the elementary divisors of a matrix ( C ²- ³ are Á

7 ~ ¸Á Á Ã Á Á Ã Á Á Á Ã Á Á ¹ Suppose also that the polynomials can be further factored over 2 , say

Á

~ ÁÁ ÄÁ where Á is prime over 2 . Then the prime powers Á

Á 8 ~ ¸Á

Á Á

Á Ã Á Á

Á Á Ã Á Ã Á Á

are the elementary divisors of ( over 2 .

Á

Á Á Ã Á Á

Á

¹

182

Advanced Linear Algebra

Proof. Consider the companion matrix *´ Á ²%³µ in the rational canonical form of ( over - . This is a matrix over 2 as well and Theorem 7.15 implies that

Á Á

*´ Á ²%³µ diag²*´ÁÁ Á µÁ Ã Á *´Á µ³ Hence, 8 is an elementary divisor basis for ( over 2 . As mentioned, unlike the elementary divisors, the invariant factors are field independent. This is equivalent to saying that the invariant factors of a matrix ( 4 ²- ³ are polynomials over the smallest subfield of - that contains the entries of (À Theorem 7.20 Let ( C ²- ³ and let , - be the smallest subfield of that contains the entries of (. 1) The invariant factors of ( are polynomials over , . 2) Two matrices (Á ) C ²- ³ are similar over - if and only if they are similar over , . Proof. Part 1) follows immediately from Theorem 7.19, since using either 7 or 8 to compute invariant factors gives the same result. Part 2) follows from the fact that two matrices are similar over a given field if and only if they have the same multiset of invariant factors over that field. Example 7.2 Over the real field, the matrix (~6

c 7

is the companion matrix for the polynomial % b , and so ElemDivs ²(³ ~ ¸% b ¹ ~ InvFacts ²(³ However, as a complex matrix, the rational canonical form for ( is (~6

c 7

and so ElemDivd ²(³ ~ ¸% c Á % b ¹

and

InvFactd ²(³ ~ ¸% b ¹

Exercises 1. 2.

We have seen that any B²= ³ can be used to make = into an - ´%µmodule. Does every module = over - ´%µ come from some B²= ³? Explain. Let B²= ³ have minimal polynomial ²%³ ~ ²%³Ä ²%³

The Structure of a Linear Operator

183

where ²%³ are distinct monic primes. Prove that the following are equivalent: a) = is -cyclic. b) deg² ²%³³ ~ dim²= ³. c) The elementary divisors of are the prime power factors ²%³ and so = ~ ºº# »» l Ä l ºº# »» is a direct sum of -cyclic submodules ºº# »» of order ²%³. 3. Prove that a matrix ( C ²- ³ is nonderogatory if and only if it is similar to a companion matrix. 4. Show that if ( and ) are block diagonal matrices with the same blocks, but in possibly different order, then ( and ) are similar. 5. Let ( C ²- ³. Justify the statement that the entries of any invariant factor version of a rational canonical form for ( are “rational” expressions in the coefficients of (, hence the origin of the term rational canonical form. Is the same true for the elementary divisor version? 6. Let B²= ³ where = is finite-dimensional. If ²%³ - ´%µ is irreducible and if ² ³ is not one-to-one, prove that ²%³ divides the minimal polynomial of . 7. Prove that the minimal polynomial of B²= ³ is the least common multiple of its elementary divisors. 8. Let B²= ³ where = is finite-dimensional. Describe conditions on the minimal polynomial of that are equivalent to the fact that the elementary divisor version of the rational canonical form of is diagonal. What can you say about the elementary divisors? 9. Verify the statement that the multiset of elementary divisors (or invariant factors) is a complete invariant for similarity of matrices. 10. Prove that given any multiset of monic prime power polynomials

Á

4 ~ ¸Á ²%³Á Ã Á ²%³Á Ã Á Ã Á Á ²%³Á Ã Á Á ²%³¹ and given any vector space = of dimension equal to the sum of the degrees of these polynomials, there is an operator B²= ³ whose multiset of elementary divisors is 4 . 11. Find all rational canonical forms ²up to the order of the blocks on the diagonal) for a linear operator on s6 having minimal polynomial ²% c 1³ ²% b 1³ . 12. How many possible rational canonical forms (up to order of blocks) are there for linear operators on s6 with minimal polynomial ²% c 1³²% b 1³ ? 13. a) Show that if ( and ) are d matrices, at least one of which is invertible, then () and )( are similar.

184

Advanced Linear Algebra b) What do the matrices (~>

c)

and ?

)~>

?

have to do with this issue? Show that even without the assumption on invertibility the matrices () and )( have the same characteristic polynomial. Hint: Write ( ~ 7 0Á 8

where 7 and 8 are invertible and 0Á is an d matrix that has the d identity in the upper left-hand corner and 's elsewhere. Write ) Z ~ 8)7 . Compute () and )( and find their characteristic polynomials. 14. Let be a linear operator on - with minimal polynomial ²%³ ~ ²% b 1³²% c 2³. Find the rational canonical form for if - ~ r, - ~ s or - ~ d. 15. Suppose that the minimal polynomial of B²= ³ is irreducible. What can you say about the dimension of = ? 16. Let B²= ³ where = is finite-dimensional. Suppose that ²%³ is an irreducible factor of the minimal polynomial ²%³ of . Suppose further that "Á # = have the property that ²"³ ~ ²#³ ~ ²%³. Prove that " ~ ² ³# for some polyjomial ²%³ if and only if # ~ ² ³" for some polynomial ²%³.

Chapter 8

Eigenvalues and Eigenvectors

Unless otherwise noted, we will assume throughout this chapter that all vector spaces are finite-dimensional.

Eigenvalues and Eigenvectors We have seen that for any B²= ³, the minimal and characteristic polynomials have the same set of roots (but not generally the same multiset of roots). These roots are of vital importance. Let ( ~ ´ µ8 be a matrix that represents . A scalar - is a root of the characteristic polynomial ²%³ ~ ( ²%³ ~ det²%0 c (³ if and only if det²0 c (³ ~

(8.1)

that is, if and only if the matrix 0 c ( is singular. In particular, if dim²= ³ ~ , then (8.1) holds if and only if there exists a nonzero vector % - for which ²0 c (³% ~ or equivalently, ( % ~ % If ´#µ8 ~ %, then this is equivalent to ´ µ8 ´#µ8 ~ ´#µ8 or in operator language, # ~ # This prompts the following definition. Definition Let = be a vector space over a field - and let B²= ³. 1) A scalar - is an eigenvalue (or characteristic value) of if there exists a nonzero vector # = for which

186

Advanced Linear Algebra

# ~ # In this case, # is called an eigenvector (or characteristic vector) of associated with . 2) A scalar - is an eigenvalue for a matrix ( if there exists a nonzero column vector % for which (% ~ % In this case, % is called an eigenvector (or characteristic vector) for ( associated with . 3) The set of all eigenvectors associated with a given eigenvalue , together with the zero vector, forms a subspace of = , called the eigenspace of and denoted by ; . This applies to both linear operators and matrices. 4) The set of all eigenvalues of an operator or matrix is called the spectrum of the operator or matrix. We denote the spectrum of by Spec² ³. Theorem 8.1 Let B²= ³ have minimal polynomial ²%³ and characteristic polynomial ²%³. 1) The spectrum of is the set of all roots of ²%³ or of ²%³, not counting multiplicity. 2) The eigenvalues of a matrix are invariants under similarity. 3) The eigenspace ; of the matrix ( is the solution space to the homogeneous system of equations ²0 c (³²%³ ~

One way to compute the eigenvalues of a linear operator is to first represent by a matrix ( and then solve the characteristic equation det²%0 c (³ ~ Unfortunately, it is quite likely that this equation cannot be solved when dim²= ³ . As a result, the art of approximating the eigenvalues of a matrix is a very important area of applied linear algebra. The following theorem describes the relationship between eigenspaces and eigenvectors of distinct eigenvalues. Theorem 8.2 Suppose that Á Ã Á are distinct eigenvalues of a linear operator B²= ³. 1) Eigenvectors associated with distinct eigenvalues are linearly independent; that is, if # ; , then the set ¸# Á Ã Á # ¹ is linearly independent. 2) The sum ; b Ä b ; is direct; that is, ; l Ä l ; exists. Proof. For part 1), if ¸# Á Ã Á # ¹ is linearly dependent, then by renumbering if necessary, we may assume that among all nontrivial linear combinations of

Eigenvalues and Eigenvectors

187

these vectors that equal , the equation # b Ä b # ~

(8.2)

has the fewest number of terms. Applying gives # b Ä b # ~

(8.3)

Multiplying (8.2) by and subtracting from (8.3) gives ² c ³# b Ä b ² c ³# ~ But this equation has fewer terms than (8.2) and so all of its coefficients must equal . Since the 's are distinct, ~ for and so ~ as well. This contradiction implies that the # 's are linearly independent. The next theorem describes the spectrum of a polynomial ² ³ in . Theorem 8.3 (The spectral mapping theorem) Let = be a vector space over an algebraically closed field - . Let B²= ³ and let ²%³ - ´%µ. Then Spec²² ³³ ~ ²Spec² ³³ ~ ¸²³ Spec² ³¹ Proof. We leave it as an exercise to show that if is an eigenvalue of , then ²³ is an eigenvalue of ² ³. Hence, ²Spec² ³³ Spec²² ³³. For the reverse inclusion, let Spec²² ³³, that is, ²² ³ c ³# ~ for # £ . If ²%³ c ~ ²% c ³ Ä²% c ³ where - , then writing this as a product of (not necessarily distinct) linear factors, we have ² c ³Ä² c ³Ä² c ³Ä² c ³# ~ (The operator is written for convenience.) We can remove factors from the left end of this equation one by one until we arrive at an operator (perhaps the identity) for which # £ but ² c ³# ~ . Then # is an eigenvector for with eigenvalue . But since ² ³ c ~ , it follows that ~ ² ³ ²Spec² ³³. Hence, Spec²² ³³ ²Spec² ³³.

The Trace and the Determinant Let - be algebraically closed and let ( C ²- ³ have characteristic polynomial ( ²%³ ~ % b c %c b Ä b % b ~ ²% c ³Ä²% c ³

188

Advanced Linear Algebra

where Á Ã Á are the eigenvalues of (. Then ( ²%³ ~ det²%0 c (³ and setting % ~ gives det²(³ ~ c ~ ²c³c Ä Hence, if - is algebraically closed then, up to sign, det²(³ is the constant term of ( ²%³ and the product of the eigenvalues of (, including multiplicity. The sum of the eigenvalues of a matrix over an algebraically closed field is also an interesting quantity. Like the determinant, this quantity is one of the coefficients of the characteristic polynomial (up to sign) and can also be computed directly from the entries of the matrix, without knowing the eigenvalues explicitly. Definition The trace of a matrix ( C ²- ³, denoted by tr²(³, is the sum of the elements on the main diagonal of (. Here are the basic propeties of the trace. Proof is left as an exercise. Theorem 8.4 Let (Á ) C ²- ³. 1) tr²A³ ~ tr²(³, for - . 2) tr²( b )³ ~ tr²(³ b tr²)³. 3) tr²()³ ~ tr²)(³. 4) tr²()*³ ~ tr²*()³ ~ tr²)*(³. However, tr²()*³ may not equal tr²(*)³. 5) The trace is an invariant under similarity. 6) If - is algebraically closed, then tr²(³ is the sum of the eigenvalues of (, including multiplicity, and so tr²(³ ~ cc where ( ²%³ ~ % b c %c b Ä b % b . Since the trace is invariant under similarity, we can make the following definition. Definition The trace of a linear operator B²= ³ is the trace of any matrix that represents . As an aside, the reader who is familar with symmetric polynomials knows that the coefficients of any polynomial ²%³ ~ % b c %c b Ä b % b ~ ²% c ³Ä²% c ³

Eigenvalues and Eigenvectors

189

are the elementary symmetric functions of the roots: c ~ ²c³ c ~ ²c³

c ~ ²c³

Å

~ ²c³ ~

The most important elementary symmetric functions of the eigenvalues are the first and last ones: c ~ c b Ä b ~ tr²(³

and ~ ²c³ Ä ~ det²(³

Geometric and Algebraic Multiplicities Eigenvalues actually have two forms of multiplicity, as described in the next definition. Definition Let be an eigenvalue of a linear operator B²= ³. 1) The algebraic multiplicity of is the multiplicity of as a root of the characteristic polynomial ²%³. 2) The geometric multiplicity of is the dimension of the eigenspace ; . Theorem 8.5 The geometric multiplicity of an eigenvalue of B²= ³ is less than or equal to its algebraic multiplicity. Proof. We can extend any basis 8 ~ ¸# Á Ã Á # ¹ of ; to a basis 8 for = . Since ; is invariant under , the matrix of with respect to 8 has the block form 0 ´ µ8 ~ 6

( ) 7block

where ( and ) are matrices of the appropriate sizes and so ²%³ ~ det²%0 c ´ µ8 ³ ~ det²%0 c 0 ³det²%0c c )³ ~ ²% c ³ det²%0c c )³ (Here is the dimension of = .) Hence, the algebraic multiplicity of is at least equal to the the geometric multiplicity of .

190

Advanced Linear Algebra

The Jordan Canonical Form One of the virtues of the rational canonical form is that every linear operator on a finite-dimensional vector space has a rational canonical form. However, as mentioned earlier, the rational canonical form may be far from the ideal of simplicity that we had in mind for a set of simple canonical forms and is really more of a theoretical tool than a practical tool. When the minimal polynomial ²%³ of splits over - , ²%³ ~ ²% c ³ Ä²% c ³ there is another set of canoncial forms that is arguably simpler than the set of rational canonical forms. In some sense, the complexity of the rational canonical form comes from the choice of basis for the cyclic submodules ºº#Á »». Recall that the -cyclic bases have the form 8Á ~ 2#Á Á #Á Á Ã Á Á c #Á 3

where Á ~ deg² Á ³. With this basis, all of the complexity comes at the end, so to speak, when we attempt to express ² Á c ²#Á ³³ ~ Á ²#Á ³ as a linear combination of the basis vectors. However, since 8Á has the form 2#Á #Á #Á Ã Á c #3 any ordered set of the form ² ³#Á ² ³#Á Ã Á c ² ³# where deg² ²%³³ ~ will also be a basis for ºº#Á »». In particular, when ²%³ splits over - , the elementary divisors are

Á ²%³ ~ ²% c ³Á and so the set 9Á ~ 2#Á Á ² c ³#Á Á Ã Á ² c ³Á c #Á 3 is also a basis for ºº#Á »». If we temporarily denote the th basis vector in 9Á by , then for ~ Á Ã Á Á c ,

Eigenvalues and Eigenvectors

191

~ ´² c ³ ²#Á ³µ ~ ² c b ³´² c ³ ²#Á ³µ ~ ² c ³b ²#Á ³ b ² c ³ ²#Á ³ ~ b b For ~ Á c , a similar computation, using the fact that ² c ³b ²#Á ³ ~ ² c ³Á ²#Á ³ ~ gives ²Á c ³ ~ Á c Thus, for this basis, the complexity is more or less spread out evenly, and the matrix of Oºº#Á »» with respect to 9Á is the Á d Á matrix v x x @ ² Á Á ³ ~ x x Å w

Æ Ä

Ä Ä y Æ Å { { Æ Æ Å { { Æ Æ z

which is called a Jordan block associated with the scalar . Note that a Jordan block has 's on the main diagonal, 's on the subdiagonal and 's elsewhere. Let us refer to the basis 9 ~ 9Á as a Jordan basis for . Theorem 8.6 (The Jordan canonical form) Suppose that the minimal polynomial of B²= ³ splits over the base field - , that is, ²%³ ~ ²% c ³ Ä²% c ³ where - . 1) The matrix of with respect to a Jordan basis 9 is diag@ ² Á Á ³Á Ã Á @ ² Á Á ³Á Ã Á @ ² Á Á ³Á Ã Á @ ² Á Á ³ where the polynomials ²% c ³Á are the elementary divisors of . This block diagonal matrix is said to be in Jordan canonical form and is called the Jordan canonical form of . 2) If - is algebraically closed, then up to order of the block diagonal matrices, the set of matrices in Jordan canonical form constitutes a set of canonical forms for similarity. Proof. For part 2), the companion matrix and corresponding Jordan block are similar:

192

Advanced Linear Algebra

*´²% c ³Á µ @ ² Á Á ³ since they both represent the same operator on the subspace ºº#Á »». It follows that the rational canonical matrix and the Jordan canonical matrix for are similar. Note that the diagonal elements of the Jordan canonical form @ of are precisely the eigenvalues of , each appearing a number of times equal to its algebraic multiplicity. In general, the rational canonical form does not “expose” the eigenvalues of the matrix, even when these eigenvalues lie in the base field.

Triangularizability and Schur's Lemma We have discussed two different canonical forms for similarity: the rational canonical form, which applies in all cases and the Jordan canonical form, which applies only when the base field is algebraically closed. Moreover, there is an annoying sense in which these sets of canoncial forms leave something to be desired: One is too complex and the other does not always exist. Let us now drop the rather strict requirements of canonical forms and look at two classes of matrices that are too large to be canonical forms (the upper triangular matrices and the almost upper triangular matrices) and one class of matrices that is too small to be a canonical form (the diagonal matrices). The upper triangular matrices (or lower triangular matrices) have some nice algebraic properties and it is of interest to know when an arbitrary matrix is similar to a triangular matrix. We confine our attention to upper triangular matrices, since there are direct analogs for lower triangular matrices as well. Definition A linear operator B²= ³ is upper triangularizable if there is an ordered basis 8 ~ ²# Á Ã Á # ³ of = for which the matrix ´ µ8 is upper triangular, or equivalently, if # º# Á Ã Á # » for all ~ Á Ã Á . As we will see next, when the base field is algebraically closed, all operators are upper triangularizable. However, since two distinct upper triangular matrices can be similar, the class of upper triangular matrices is not a canonical form for similarity. Simply put, there are just too many upper triangular matrices. Theorem 8.7 (Schur's theorem) Let = be a finite-dimensional vector space over a field - . 1) If the characteristic polynomial (or minimal polynomial) of B²= ³ splits over - , then is upper triangularizable. 2) If - is algebraically closed, then all operators are upper triangularizable.

Eigenvalues and Eigenvectors

193

Proof. Part 2) follows from part 1). The proof of part 1) is most easily accomplished by matrix means, namely, we prove that every square matrix ( 4 ²- ³ whose characteristic polynomial splits over - is similar to an upper triangular matrix. If ~ there is nothing to prove, since all d matrices are upper triangular. Assume the result is true for c and let ( 4 ²- ³. Let # be an eigenvector associated with the eigenvalue - of ( and extend ¸# ¹ to an ordered basis 8 ~ ²# Á Ã Á # ³ for s . The matrix of ( with respect to 8 has the form ´( µ8 ~ >

i ( ?block

for some ( 4c ²- ³. Since ´( µ8 and ( are similar, we have det ²%0 c (³ ~ det ²%0 c ´( µ8 ³ ~ ²% c ³ det ²%0 c ( ³ Hence, the characteristic polynomial of ( also splits over - and the induction hypothesis implies that there is an invertible matrix 7 4c ²- ³ for which < ~ 7 ( 7 c is upper triangular. Hence, if 8~>

7 ?block

then 8 is invertible and 8´(µ8 8c ~ >

7 ?>

i ( ?>

~ 7 c ? >

i
is upper triangular.

The Real Case When the base field is - ~ s, an operator is upper triangularizable if and only if its characteristic polynomial splits over s. (Why?) We can, however, always achieve a form that is close to triangular by permitting values on the first subdiagonal. Before proceeding, let us recall Theorem 7.11, which says that for a module > of prime order ²%³, the following are equivalent: 1) 2) 3) 4) 5)

> is cyclic > is indecomposable ²%³ is irreducible is nonderogatory, that is, ²%³ ~ ²%³ dim²> ³ ~ deg²²%³³.

194

Advanced Linear Algebra

Now suppose that - ~ s and ²%³ ~ % b % b ! is an irreducible quadratic. If 8 is a -cyclic basis for > , then ´ µ8 ~ >

c! c ?

However, there is a more appealing matrix representation of . To this end, let ( be the matrix above. As a complex matrix, ( has two distinct eigenvalues: ~c f

j! c

Now, a matrix of the form )~>

c ?

has characteristic polynomial ²%³ ~ ²% c ³ b and eigenvalues f . So if we set ~c

and ~ c

j! c

then ) has the same two distinct eigenvalues as ( and so ( and ) have the same Jordan canonical form over d. It follows that ( and ) are similar over d and therefore also over s, by Theorem 7.20. Thus, there is an ordered basis 9 for which ´ µ9 ~ ) . Theorem 8.8 If - ~ s and > is cyclic and deg² ²%³³ ~ , then there is an ordered basis 9 for which ´ µ9 ~ >

c ?

Now we can proceed with the real version of Schur's theorem. For the sake of the exposition, we make the following definition. Definition A matrix ( 4 ²- ³ is almost upper triangular if it has the form v ( x (~x w where

i ( Æ

y { { ( zblock

Eigenvalues and Eigenvectors

( ~ ´µ or ( ~ >

195

c ?

for Á - . A linear operator B²= ³ is almost upper triangularizable if there is an ordered basis 8 for which ´ µ8 is almost upper triangular. To see that every real linear operator is almost upper triangularizable, we use Theorem 7.19, which states that if ²%³ is a prime factor of ²%³, then = has a cyclic submodule > of order ²%³. Hence, > is a -cyclic subspace of dimension deg²²%³³ and O> has characteristic polynomial ²%³. Now, the minimal polynomial of a real operator B²= ³ factors into a product of linear and irreducible quadratic factors. If ²%³ has a linear factor over - , then = has a one-dimensional -invariant subspace > . If ²%³ has an irreducible quadratic factor ²%³, then = has a cyclic submodule > of order ²%³ and so a matrix representation of on > is given by the matrix (~>

c ?

This is the basis for an inductive proof, as in the complex case. Theorem 8.9 (Schur's theorem: real case) If = is a real vector space, then every linear operator on = is almost upper triangularizable. Proof. As with the complex case, it is simpler to proceed using matrices, by showing that any d real matrix ( is similar to an almost upper triangular matrix. The result is clear if ~ . Assume for the purposes of induction that any square matrix of size less than d is almost upper triangularizable. We have just seen that - has a one-dimensional ( -invariant subspace > or a two-dimensional ( -cyclic subspace > , where ( has irreducible characteristic polynomial on > . Hence, we may choose a basis 8 for - for which the first one or first two vectors are a basis for > . Then ´( µ8 ~ >

(

i ( ?block

where ( ~ ´µ or ( ~ >

c ?

and ( has size d . The induction hypothesis applied to ( gives an invertible matrix 7 4 for which < ~ 7 ( 7 c

196

Advanced Linear Algebra

is almost upper triangular. Hence, if 8~>

0c

7 ?block

then 8 is invertible and 8´(µ8 8c ~ >

0c

( 7 ?>

0c i ( ?>

( ~ 7 c ? >

i
is almost upper triangular.

Unitary Triangularizability Although we have not yet discussed inner product spaces and orthonormal bases, the reader may very well be familiar with these concepts. For those who are, we mention that when = is a real or complex inner product space, then if an operator on = can be triangularized (or almost triangularized) using an ordered basis 8 , it can also be triangularized (or almost triangularized) using an orthonormal ordered basis E . To see this, suppose we apply the Gram–Schmidt orthogonalization process to a basis 8 ~ ²# Á Ã Á # ³ that triangularizes (or almost triangularizes) . The resulting ordered orthonormal basis E ~ ²" Á Ã Á " ³ has the property that º# Á Ã Á # » ~ º" Á Ã Á " » for all . Since ´ µ8 is (almost) upper triangular, that is, # º# Á Ã Á # » for all , it follows that " º # Á Ã Á # » º# Á Ã Á # » ~ º" Á Ã Á "» and so the matrix ´ µE is also (almost) upper triangular. A linear operator is unitarily upper triangularizable if there is an ordered orthonormal basis with respect to which is upper triangular. Accordingly, when = is an inner product space, we can replace the term “upper triangularizable” with “unitarily upper triangularizable” in Schur's theorem. (A similar statement holds for almost upper triangular matrices.)

Diagonalizable Operators Definition A linear operator B²= ³ is diagonalizable if there is an ordered basis 8 ~ ²# Á Ã Á # ³ of = for which the matrix ´ µ8 is diagonal, or equivalently, if

Eigenvalues and Eigenvectors

197

# ~ # for all ~ Á Ã Á . The previous definition leads immediately to the following simple characterization of diagonalizable operators. Theorem 8.10 Let B²= ³. The following are equivalent: 1) is diagonalizable. 2) = has a basis consisting entirely of eigenvectors of . 3) = has the form = ~ ; l Ä l ; where Á Ã Á are the distinct eigenvalues of . Diagonalizable operators can also be characterized in a simple way via their minimal polynomials. Theorem 8.11 A linear operator B²= ³ on a finite-dimensional vector space is diagonalizable if and only if its minimal polynomial is the product of distinct linear factors. Proof. If is diagonalizable, then = ~ ; l Ä l ; and Theorem 7.7 implies that ²%³ is the least common multiple of the minimal polynomials % c of restricted to ; . Hence, ²%³ is a product of distinct linear factors. Conversely, if ²%³ is a product of distinct linear factors, then the primary decomposition of = has the form = ~ = l Ä l = where = ~ ¸# = ² c ³# ~ ¹ ~ ; and so is diagonalizable.

Spectral Resolutions We have seen (Theorem 2.25) that resolutions of the identity on a vector space = correspond to direct sum decompositions of = . We can do something similar for any diagonalizable linear operator on = (not just the identity operator). Suppose that has the form ~ b Ä b where b Ä b ~ is a resolution of the identity and the - are distinct. This is referred to as a spectral resolution of .

198

Advanced Linear Algebra

We claim that the 's are the eigenvalues of and im² ³ ~ ; . Theorem 2.25 implies that = ~ im² ³ l Ä l im² ³ If # im² ³, then ² #³ ~ ² b Ä b ³ # ~ ² #³ and so # ; . Hence, im² ³ ; and so = ~ im² ³ l Ä l im² ³ ; l Ä l ; = which implies that im² ³ ~ ; and = ~ ; l Ä l ; The converse also holds, for if = ~ ; l Ä l ; and if is projection onto ; along the direct sum of the other eigenspaces, then b Ä b ~ and since ~ , it follows that ~ ² b Ä b ³ ~ b Ä b Theorem 8.12 A linear operator B²= ³ is diagonalizable if and only if it has a spectral resolution ~ b Ä b In this case, ¸ Á Ã Á ¹ is the spectrum of and im² ³ ~ ;

and

ker² ³ ~ ;

£

Exercises 1. 2. 3. 4. 5. 6.

Let 1 be the d matrix all of whose entries are equal to . Find the minimal polynomial and characteristic polynomial of 1 and the eigenvalues. Prove that the eigenvalues of a matrix do not form a complete set of invariants under similarity. Show that B²= ³ is invertible if and only if is not an eigenvalue of . Let ( be an d matrix over a field - that contains all roots of the characteristic polynomial of (. Prove that det²(³ is the product of the eigenvalues of (, counting multiplicity. Show that if is an eigenvalue of , then ²³ is an eigenvalue of ² ³, for any polynomial ²%³. Also, if £ , then c is an eigenvalue for c . An operator B²= ³ is nilpotent if ~ for some positive o.

Eigenvalues and Eigenvectors

199

a) Show that if is nilpotent, then the spectrum of is ¸¹. b) Find a nonnilpotent operator with spectrum ¸¹. 7. Show that if Á B²= ³ and one of and is invertible, then and so and have the same eigenvalues, counting multiplicty. 8. (Halmos) a) Find a linear operator that is not idempotent but for which ² c ³ ~ . b) Find a linear operator that is not idempotent but for which ² c ³ ~ . c) Prove that if ² c ³ ~ ² c ³ ~ , then is idempotent. 9. An involution is a linear operator for which ~ . If is idempotent what can you say about c ? Construct a one-to-one correspondence between the set of idempotents on = and the set of involutions. 10. Let (Á ) 4 ²d³ and suppose that ( ~ ) ~ 0Á ()( ~ ) c but ( £ 0 and ) £ 0 . Show that if * 4 ²d³ commutes with both ( and ) , then * ~ 0 for some scalar d. 11. Let B²= ³ and let : ~ º#Á #Á Ã Á c #» be a -cyclic submodule of = with minimal polynomial ²%³ where ²%³ is prime of degree . Let ~ ² ³ restricted to º#». Show that : is the direct sum of -cyclic submodules each of dimension , that is, : ~ ; l Ä l ; Hint: For each , consider the set 8 ~ ¸ #Á ² ³ #Á Ã Á ² ³c #» 12. Fix . Show that any complex matrix is similar to a matrix that looks just like a Jordan matrix except that the entries that are equal to are replaced by entries with value , where is any complex number. Thus, any complex matrix is similar to a matrix that is “almost” diagonal. Hint: consider the fact that v w

yv zw

yv zw

c

y v ~ c z w

y z

13. Show that the Jordan canonical form is not very robust in the sense that a small change in the entries of a matrix ( may result in a large jump in the entries of the Jordan form 1 . Hint: consider the matrix ( ~ >

?

What happens to the Jordan form of ( as ¦ ?

200

Advanced Linear Algebra

14. Give an example of a complex nonreal matrix all of whose eigenvalues are real. Show that any such matrix is similar to a real matrix. What about the type of the invertible matrices that are used to bring the matrix to Jordan form? 15. Let 1 ~ ´ µ8 be the Jordan form of a linear operator B²= ³. For a given Jordan block of 1 ²Á ³ let < be the subspace of = spanned by the basis vectors of 8 associated with that block. a) Show that O< has a single eigenvalue with geometric multiplicity . In other words, there is essentially only one eigenvector (up to scalar multiple) associated with each Jordan block. Hence, the geometric multiplicity of for is the number of Jordan blocks for . Show that the algebraic multiplicity is the sum of the dimensions of the Jordan blocks associated with . b) Show that the number of Jordan blocks in 1 is the maximum number of linearly independent eigenvectors of . c) What can you say about the Jordan blocks if the algebraic multiplicity of every eigenvalue is equal to its geometric multiplicity? 16. Assume that the base field - is algebraically closed. Then assuming that the eigenvalues of a matrix ( are known, it is possible to determine the Jordan form 1 of ( by looking at the rank of various matrix powers. A matrix ) is nilpotent if ) ~ for some . The smallest such exponent is called the index of nilpotence. a) Let 1 ~ 1 ²Á ³ be a single Jordan block of size d . Show that 1 c 0 is nilpotent of index . Thus, is the smallest integer for which rk²1 c 0³ ~ . Now let 1 be a matrix in Jordan form but possessing only one eigenvalue . b) Show that 1 c 0 is nilpotent. Let be its index of nilpotence. Show that is the maximum size of the Jordan blocks of 1 and that rk²1 c 0³c is the number of Jordan blocks in 1 of maximum size. c) Show that rk²1 c 0³c is equal to times the number of Jordan blocks of maximum size plus the number of Jordan blocks of size one less than the maximum. d) Show that the sequence rk²1 c 0³ for ~ Á Ã Á uniquely determines the number and size of all of the Jordan blocks in 1 , that is, it uniquely determines 1 up to the order of the blocks. e) Now let 1 be an arbitrary Jordan matrix. If is an eigenvalue for 1 show that the sequence rk²1 c 0³ for ~ Á Ã Á where is the first integer for which rk²1 c 0³ ~ rk²1 c 0³b uniquely determines 1 up to the order of the blocks. f) Prove that for any matrix ( with spectrum ¸ Á Ã Á ¹ the sequence rk²( c 0³ for ~ Á Ã Á and ~ Á Ã Á where is the first integer for which rk²( c 0³ ~ rk²( c 0³b uniquely determines the Jordan matrix 1 for ( up to the order of the blocks. 17. Let ( C ²- ³.

Eigenvalues and Eigenvectors a)

201

If all the roots of the characteristic polynomial of ( lie in - prove that ( is similar to its transpose (! . Hint: Let ) be the matrix v xÅ )~x w

Ä y Ç { { Ç Ç Å Ä z

with 's on the diagonal that moves up from left to right and 's elsewhere. Let 1 be a Jordan block of the same size as ) . Show that )1 ) c ~ 1 ! . b) Let (Á ) C ²- ³. Let 2 be a field containing - . Show that if ( and ) are similar over 2 , that is, if ) ~ 7 (7 c where 7 C ²2³, then ( and ) are also similar over - , that is, there exists 8 C ²- ³ for which ) ~ 8(8c . c) Show that any matrix is similar to its transpose.

The Trace of a Matrix 18. Let ( C ²- ³. Verify the following statements. a) tr²A³ ~ tr²(³, for - . b) tr²( b )³ ~ tr²(³ b tr²)³. c) tr²()³ ~ tr²)(³. d) tr²()*³ ~ tr²*()³ ~ tr²)*(³. Find an example to show that tr²()*³ may not equal tr²(*)³. e) The trace is an invariant under similarity. f) If - is algebraically closed, then the trace of ( is the sum of the eigenvalues of (. 19. Use the concept of the trace of a matrix, as defined in the previous exercise, to prove that there are no matrices (, ) C ²d³ for which () c )( ~ 0 20. Let ; ¢ C ²- ³ ¦ - be a function with the following properties. For all matrices (Á ) C ²- ³ and - , 1) ; ²A³ ~ ; ²(³ 2) ; ²( b )³ ~ ; ²(³ b ; ²)³ 3) ; ²()³ ~ ; ²)(³ Show that there exists - for which ; ²(³ ~ tr²(³, for all ( C ²- ³.

Commuting Operators Let < ~ ¸ B²= ³ ? ¹ be a family of operators on a vector space = . Then < is a commuting family if every pair of operators commutes, that is, ~ for all Á < . A subspace

202

Advanced Linear Algebra

< of = is < -invariant if it is -invariant for every < . It is often of interest to know whether a family < of linear operators on = has a common eigenvector, that is, a single vector # = that is an eigenvector for every < (the corresponding eigenvalues may be different for each operator, however). 21. A pair of linear operators Á B²= ³ is simultaneously diagonalizable if there is an ordered basis 8 for = for which ´ µ8 and ´µ8 are both diagonal, that is, 8 is an ordered basis of eigenvectors for both and . Prove that two diagonalizable operators and are simultaneously diagonalizable if and only if they commute, that is, ~ . Hint: If ~ , then the eigenspaces of are invariant under . 22. Let Á B²= ³. Prove that if and commute, then every eigenspace of is -invariant. Thus, if < is a commuting family, then every eigenspace of any member of < is < -invariant. 23. Let < be a family of operators in B²= ³ with the property that each operator in < has a full set of eigenvalues in the base field - , that is, the characteristic polynomial splits over - . Prove that if < is a commuting family, then < has a common eigenvector # = . 24. What do the real matrices (~>

c

and ) ~ > ? c

?

have to do with the issue of common eigenvectors?

Geršgorin Disks It is generally impossible to determine precisely the eigenvalues of a given complex operator or matrix ( C ²d³, for if , then the characteristic equation has degree and cannot in general be solved. As a result, the approximation of eigenvalues is big business. Here we consider one aspect of this approximation problem, which also has some interesting theoretical consequences. Let ( C ²d³ and suppose that (# ~ # where # ~ ² Á Ã Á ³! . Comparing th rows gives

( ~ ~

which can also be written in the form

² c ( ³ ~ ( ~ £

If has the property that ( ( ( ( for all , we have

Eigenvalues and Eigenvectors

203

( (( c ( ( (( (( ( ( ((( ( ~ £

~ £

and thus

( c ( ( (( (

(8.7)

~ £

The right-hand side is the sum of the absolute values of all entries in the th row of ( except the diagonal entry ( . This sum 9 ²(³ is the th deleted absolute row sum of (. The inequality (8.7) says that, in the complex plane, the eigenvalue lies in the disk centered at the diagonal entry ( with radius equal to 9 ²(³. This disk GR ²(³ ~ ¸' d (' c ( ( 9 ²(³¹ is called the Geršgorin row disk for the th row of (. The union of all of the Geršgorin row disks is called the Geršgorin row region for (. Since there is no way to know in general which is the index for which ( ( ( (, the best we can say in general is that the eigenvalues of ( lie in the union of all Geršgorin row disks, that is, in the Geršgorin row region of (. Similar definitions can be made for columns and since a matrix has the same eigenvalues as its transpose, we can say that the eigenvalues of ( lie in the Geršgorin column region of (. The Geršgorin region .²(³ of a matrix ( 4 ²- ³ is the intersection of the Geršgorin row region and the Geršgorin column region and we can say that all eigenvalues of ( lie in the Geršgorin region of (. In symbols, ( .(. 25. Find and sketch the Geršgorin region and the eigenvalues for the matrix (~

v w

y

z

26. A matrix ( 4 ²d³ is diagonally dominant if for each ~ Á Ã Á , (( ( 9 ²(³ and it is strictly diagonally dominant if strict inequality holds. Prove that if ( is strictly diagonally dominant, then it is invertible. 27. Find a matrix ( 4 ²d³ that is diagonally dominant but not invertible. 28. Find a matrix ( 4 ²d³ that is invertible but not strictly diagonally dominant.

Chapter 9

Real and Complex Inner Product Spaces

We now turn to a discussion of real and complex vector spaces that have an additional function defined on them, called an inner product, as described in the following definition. In this chapter, - will denote either the real or complex field. Also, the complex conjugate of d is denoted by . Definition Let = be a vector space over - ~ s or - ~ d. An inner product on = is a function ºÁ »¢ = d = ¦ - with the following properties: 1) (Positive definiteness) For all # = , º#Á #»

and º#Á #» ~ ¯ # ~

2) For - ~ d: (Conjugate symmetry) º"Á #» ~ º#Á "» For - ~ s: (Symmetry) º"Á #» ~ º#Á "» 3) (Linearity in the first coordinate) For all "Á # = and Á º" b #Á $» ~ º"Á $» b º#Á $» A real (or complex) vector space = , together with an inner product, is called a real (or complex) inner product space. If ?Á @ = , then we let º?Á @ » ~ ¸º%Á &» % ?Á & @ ¹ and º#Á ?» ~ ¸º#Á %» % ?¹ Note that a vector subspace : of an inner product space = is also an inner product space under the restriction of the inner product of = to : .

206

Advanced Linear Algebra

We will study bilinear forms (also called inner products) on vector spaces over fields other than s or d in Chapter 11. Note that property 1) implies that º#Á #» is always real, even if = is a complex vector space. If - ~ s, then properties 2) and 3) imply that the inner product is linear in both coordinates, that is, the inner product is bilinear. However, if - ~ d, then º$Á " b #» ~ º" b #Á $» ~ º"Á $» b º#Á $» ~ º$Á "» b º$Á #» This is referred to as conjugate linearity in the second coordinate. Specifically, a function ¢ = ¦ > between complex vector spaces is conjugate linear if ²" b #³ ~ ²"³ b ²#³ and ²"³ ~ ²"³ for all "Á # = and d. Thus, a complex inner product is linear in its first coordinate and conjugate linear in its second coordinate. This is often described by saying that a complex inner product is sesquilinear. (Sesqui means “one and a half times.”) Example 9.1 1) The vector space s is an inner product space under the standard inner product, or dot product, defined by º² Á Ã Á ³Á ² Á Ã Á

³»

~

b Ä b

The inner product space s is often called -dimensional Euclidean space. 2) The vector space d is an inner product space under the standard inner product defined by º² Á Ã Á ³Á ² Á Ã Á

³»

~

b Ä b

This inner product space is often called -dimensional unitary space. 3) The vector space *´Á µ of all continuous complex-valued functions on the closed interval ´Á µ is a complex inner product space under the inner product

º Á » ~ ²%³²%³ %

Example 9.2 One of the most important inner product spaces is the vector space M of all real (or complex) sequences ² ³ with the property that ( ( B

Real and Complex Inner Product Spaces

207

under the inner product B

º² ³Á ²! ³» ~ ! ~

Such sequences are called square summable. Of course, for this inner product to make sense, the sum on the right must converge. To see this, note that if ² ³Á ²! ³ M , then ²( ( c (! (³ ~ ( ( c ( ((! ( b (! ( and so ( ! ( ( ( b (! ( which implies that ² ! ³ M . We leave it to the reader to verify that M is an inner product space. The following simple result is quite useful. Lemma 9.1 If = is an inner product space and º"Á %» ~ º#Á %» for all % = , then " ~ #. The next result points out one of the main differences between real and complex inner product spaces and will play a key role in later work. Theorem 9.2 Let = be an inner product space and let B²= ³. 1) º #Á $» ~ for all #Á $ =

¬

~

2) If = is a complex inner product space, then º #Á #» ~ for all # =

¬

~

but this does not hold in general for real inner product spaces. Proof. Part 1) follows directly from Lemma 9.1. As for part 2), let # ~ % b &, for %Á & = and - . Then ~ º ²% b &³Á % b &» ~ (( º %Á %» b º &Á &» b º %Á &» b º &Á %» ~ º %Á &» b º &Á %» Setting ~ gives º %Á &» b º &Á %» ~ and setting ~ gives

208

Advanced Linear Algebra

º %Á &» c º &Á %» ~ These two equations imply that º %Á &» ~ for all %Á & = and so part 1) implies that ~ . For the last statement, rotation by degrees in the real plane s has the property that º #Á #» ~ for all #.

Norm and Distance If = is an inner product space, the norm, or length of # = is defined by )#) ~ jº#Á #»

(9.1)

A vector # is a unit vector if )#) ~ . Here are the basic properties of the norm. Theorem 9.3 1) )#) and )#) ~ if and only if # ~ . 2) For all - and # = , )#) ~ (()#) 3) (The Cauchy–Schwarz inequality) For all "Á # = , (º"Á #»( )"))#) with equality if and only if one of " and # is a scalar multiple of the other. 4) (The triangle inequality) For all "Á # = , )" b #) )") b )#) with equality if and only if one of " and # is a scalar multiple of the other. 5) For all "Á #Á % = , )" c #) )" c %) b )% c #) 6) For all "Á # = , ()") c )#)( )" c #) 7) (The parallelogram law) For all "Á # = , )" b #) b )" c #) ~ )") b )#) Proof. We prove only Cauchy–Schwarz and the triangle inequality. For Cauchy–Schwarz, if either " or # is zero the result follows, so assume that "Á # £ . Then, for any scalar - , )" c #) ~ º" c #Á " c #» ~ º"Á "» c º"Á #» c ´º#Á "» c º#Á #»µ Choosing ~ º#Á "»°º#Á #» makes the value in the square brackets equal to

Real and Complex Inner Product Spaces

209

and so º"Á "» c

º#Á "»º"Á #» (º"Á #»( ~ )") c º#Á #» )#)

which is equivalent to the Cauchy–Schwarz inequality. Furthermore, equality holds if and only if )" c #) ~ , that is, if and only if " c # ~ , which is equivalent to " and # being scalar multiples of one another. To prove the triangle inequality, the Cauchy–Schwarz inequality gives )" b #) ~ º" b #Á " b #» ~ º"Á "» b º"Á #» b º#Á "» b º#Á #» )") b )"))#) b )#) ~ ²)") b )#)³ from which the triangle inequality follows. The proof of the statement concerning equality is left to the reader. Any vector space = , together with a function ) h )¢ = ¦ s that satisfies properties 1), 2) and 4) of Theorem 9.3, is called a normed linear space and the function ) h ) is called a norm. Thus, any inner product space is a normed linear space, under the norm given by (9.1). It is interesting to observe that the inner product on = can be recovered from the norm. Thus, knowing the length of all vectors in = is equivalent to knowing all inner products of vectors in = . Theorem 9.4 (The polarization identities) 1) If = is a real inner product space, then º"Á #» ~

²)" b #) c )" c #) ³

2) If = is a complex inner product space, then º"Á #» ~

²)" b #) c )" c #) ³ b ²)" b #) c )" c #) ³

The norm can be used to define the distance between any two vectors in an inner product space. Definition Let = be an inner product space. The distance ²"Á #³ between any two vectors " and # in = is ²"Á #³ ~ )" c #) Here are the basic properties of distance.

(9.2)

210

Advanced Linear Algebra

Theorem 9.5 1) ²"Á #³ and ²"Á #³ ~ if and only if " ~ # 2) (Symmetry) ²"Á #³ ~ ²#Á "³ 3) (The triangle inequality) ²"Á #³ ²"Á $³ b ²$Á #³

Any nonempty set = , together with a function ¢ = d = ¦ s that satisfies the properties of Theorem 9.5, is called a metric space and the function is called a metric on = . Thus, any inner product space is a metric space under the metric (9.2). Before continuing, we should make a few remarks about our goals in this and the next chapter. The presence of an inner product, and hence a metric, permits the definition of a topology on = , and in particular, convergence of infinite sequences. A sequence ²# ³ of vectors in = converges to # = if lim )# c #) ~

¦B

Some of the more important concepts related to convergence are closedness and closures, completeness and the continuity of linear operators and linear functionals. In the finite-dimensional case, the situation is very straightforward: All subspaces are closed, all inner product spaces are complete and all linear operators and functionals are continuous. However, in the infinite-dimensional case, things are not as simple. Our goals in this chapter and the next are to describe some of the basic properties of inner product spaces—both finite and infinite-dimensional—and then discuss certain special types of operators (normal, unitary and self-adjoint) in the finite-dimensional case only. To achieve the latter goal as rapidly as possible, we will postpone a discussion of convergence-related properties until Chapter 12. This means that we must state some results only for the finitedimensional case in this chapter.

Isometries An isomorphism of vector spaces preserves the vector space operations. The corresponding concept for inner product spaces is the isometry. Definition Let = and > be inner product spaces and let B²= Á > ³.

Real and Complex Inner Product Spaces

211

1) is an isometry if it preserves the inner product, that is, if º "Á #» ~ º"Á #» for all "Á # = . 2) A bijective isometry is called an isometric isomorphism. When ¢ = ¦ > is an isometric isomorphism, we say that = and > are isometrically isomorphic. It is clear that an isometry is injective and so it is an isometric isomorphism provided it is surjective. Moreover, if dim²= ³ ~ dim²> ³ B injectivity implies surjectivity and is an isometry if and only if is an isometric isomorphism. On the other hand, the following simple example shows that this is not the case for infinite-dimensional inner product spaces. Example 9.3 The map ¢ M ¦ M defined by ²% Á % Á % Á Ã ³ ~ ²Á % Á % Á Ã ³ is an isometry, but it is clearly not surjective. Since the norm determines the inner product, the following should not come as a surprise. Theorem 9.6 A linear transformation B²= Á > ³ is an isometry if and only if it preserves the norm, that is, if and only if )#) ~ )#) for all # = . Proof. Clearly, an isometry preserves the norm. The converse follows from the polarization identities. In the real case, we have ² ) " b # ) c ) " c # ) ³ ~ ²) ²" b #³) c ) ²" c #³) ³ ~ ²)" b #) c )" c #) ³ ~ º"Á #»

º "Á #» ~

and so is an isometry. The complex case is similar.

Orthogonality The presence of an inner product allows us to define the concept of orthogonality.

212

Advanced Linear Algebra

Definition Let = be an inner product space. 1) Two vectors "Á # = are orthogonal, written " #, if º"Á #» ~ 2) Two subsets ?Á @ = are orthogonal, written ? @ , if º?Á @ » ~ ¸¹, that is, if % & for all % ? and & @ . We write # ? in place of ¸#¹ ?. 3) The orthogonal complement of a subset ? = is the set ? ~ ¸# = # ?¹

The following result is easily proved. Theorem 9.7 Let = be an inner product space. 1) The orthogonal complement ? of any subset ? = is a subspace of = . 2) For any subspace : of = , : q : ~ ¸¹

Definition An inner product space = is the orthogonal direct sum of subspaces : and ; if = ~ : l ;Á

:;

In this case, we write : p; More generally, = is the orthogonal direct sum of the subspaces : Á Ã Á : , written : ~ : p Ä p : if = ~ : l Ä l :

and

: : for £

Theorem 9.8 Let = be an inner product space. The following are equivalent. 1) = ~ : p ; 2) = ~ : l ; and ; ~ : Proof. If = ~ : p ; , then by definition, ; : . However, if # : , then # ~ b ! where : and ! ; . Then is orthogonal to both ! and # and so is orthogonal to itself, which implies that ~ and so # ; . Hence, ; ~ : . The converse is clear.

Orthogonal and Orthonormal Sets Definition A nonempty set E ~ ¸" 2¹ of vectors in an inner product space is said to be an orthogonal set if " " for all £ 2 . If, in addition, each vector " is a unit vector, then E is an orthonormal set. Thus, a

Real and Complex Inner Product Spaces

213

set is orthonormal if º" Á " » ~ Á for all Á 2 , where Á is the Kronecker delta function. Of course, given any nonzero vector # = , we may obtain a unit vector " by multiplying # by the reciprocal of its norm: "~

# )#)

This process is referred to as normalizing the vector #. Thus, it is a simple matter to construct an orthonormal set from an orthogonal set of nonzero vectors. Note that if " #, then )" b #) ~ )") b )#) and the converse holds if - ~ s. Orthogonality is stronger than linear independence. Theorem 9.9 Any orthogonal set of nonzero vectors in = is linearly independent. Proof. If E ~ ¸" 2¹ is an orthogonal set of nonzero vectors and " b Ä b " ~ then ~ º " b Ä b " Á " » ~ º" Á " » and so ~ , for all . Hence, E is linearly independent.

Gram–Schmidt Orthogonalization The Gram–Schmidt process can be used to transform a sequence of vectors into an orthogonal sequence. We begin with the following. Theorem 9.10 (Gram–Schmidt augmentation) Let = be an inner product space and let E ~ ¸" Á Ã Á " ¹ be an orthogonal set of vectors in = . If # ¤ º" Á Ã Á " », then there is a nonzero " = for which ¸" Á Ã Á " Á "¹ is orthogonal and º" Á Ã Á " Á "» ~ º" Á Ã Á " Á #» In particular,

214

Advanced Linear Algebra

" ~ # c " ~

where ~ H

if " ~ if " £

º#Á" » º" Á" »

Proof. We simply set " ~ # c " c Ä c " and force " " for all , that is, ~ º"Á " » ~ º# c " c Ä c " Á " » ~ º#Á " » c º"Á "» Thus, if " ~ , take ~ and if " £ , take ~

º#Á " » º" Á " »

The Gram–Schmidt augmentation is traditionally applied to a sequence of linearly independent vectors, but it also applies to any sequence of vectors. Theorem 9.11 (The Gram–Schmidt orthogonalization process) Let 8 ~ ²# Á # Á Ã ³ be a sequence of vectors in an inner product space = . Define a sequence E ~ ²" Á " Á Ã ³ by repeated Gram–Schmidt augmentation, that is, c

" ~ # c Á " ~

where " ~ # and Á ~ H º# Á" » º" Á" »

if " ~ if " £

Then E is an orthogonal sequence in = with the property that º" Á Ã Á " » ~ º# Á Ã Á # » for all . Also, " ~ if and only if # º# Á Ã Á #c ». Proof. The result holds for ~ . Assume it holds for c . If # º# Á Ã Á #c », then # º# Á Ã Á #c » ~ º" Á Ã Á "c » Writing

Real and Complex Inner Product Spaces

215

c

# ~ " ~

we have º# Á " » ~ F

º" Á " »

if " ~ if " £

Therefore, ~ Á when " £ and so " ~ . Hence, º" Á Ã Á " » ~ º" Á Ã Á "c Á » ~ º# Á Ã Á #c » ~ º# Á Ã Á # » If # ¤ º# Á Ã Á #c » then º" Á Ã Á " » ~ º# Á Ã Á #c Á " » ~ º# Á Ã Á #c Á # »

Example 9.4 Consider the inner product space s´%µ of real polynomials, with inner product defined by

º²%³Á ²%³» ~ ²%³²%³% c

Applying the Gram–Schmidt process to the sequence 8 ~ ²Á %Á % Á % Á Ã ³ gives " ²%³ ~ " ²%³ ~ % c

% % c h~% % c

% % % % "3 ²%³ ~ % c c h c c h % ~ % c % % % c c % ²% c ³% % % % % 3 "4 ²%³ ~ % c c h c c h % c c h >% c ? % % % ²% c ³ % c c c 3 ~ % c % and so on. The polynomials in this sequence are (at least up to multiplicative constants) the Legendre polynomials.

The QR Factorization The Gram–Schmidt process can be used to factor any real or complex matrix into a product of a matrix with orthogonal columns and an upper triangular matrix. Suppose that ( ~ ²# # Ä # ³ is an d matrix with columns # , where . The Gram–Schmidt process applied to these columns gives orthogonal vectors 6 ~ ²" " Ä " ³ for which

216

Advanced Linear Algebra

º" Á Ã Á " » ~ º# Á Ã Á # » for all . In particular, c

# ~ " b Á " ~

where Á ~ H º# Á" » º" Á" »

if " ~ if " £

In matrix terms, v x ²# # Ä # ³ ~ ²" " Ä " ³x

Á

w

Ä Á y Ä Á { { Æ z

that is, ( ~ 6) where 6 has orthogonal columns and ) is upper triangular. We may normalize the nonzero columns " of 6 and move the positive constants to ) . In particular, if ~ )" ) for " £ and ~ for " ~ , then v " " " x ²# # Ä # ³ ~ 6 c c Ä c 7x w

Á

Ä Á y Ä Á { { Æ z

and so ( ~ 89 where the columns of 8 are orthogonal and each column is either a unit vector or the zero vector and 9 is upper triangular with positive entries on the main diagonal. Moreover, if the vectors # Á Ã Á # are linearly independent, then the columns of 8 are nonzero. Also, if ~ and ( is nonsingular, then 8 is unitary/orthogonal. If the columns of ( are not linearly independent, we can make one final adjustment to this matrix factorization. If a column " ° is zero, then we may replace this column by any vector as long as we replace the ²Á ³th entry in 9 by . Therefore, we can take nonzero columns of 8, extend to an orthonormal basis for the span of the columns of 8 and replace the zero columns of 8 by the additional members of this orthonormal basis. In this way, 8 is replaced by a unitary/orthogonal matrix 8Z and 9 is replaced by an upper triangular matrix 9 Z that has nonnegative entries on the main diagonal.

Real and Complex Inner Product Spaces

217

Theorem 9.12 Let ( CÁ ²- ³, where - ~ d or - ~ s. There exists a matrix 8 CÁ ²- ³ with orthonormal columns and an upper triangular matrix 9 C ²- ³ with nonnegative real entries on the main diagonal for which ( ~ 89 Moreover, if ~ , then 8 is unitary/orthogonal. If ( is nonsingular, then 9 can be chosen to have positive entries on the main diagonal, in which case the factors 8 and 9 are unique. The factorization ( ~ 89 is called the 89 factorization of the matrix (. If ( is real, then 8 and 9 may be taken to be real. Proof. As to uniqueness, if ( is nonsingular and 89 ~ 8 9 then c 8c 8 ~ 9 9

and the right side is upper triangular with nonzero entries on the main diagonal and the left side is unitary. But an upper triangular matrix with positive entries on the main diagonal is unitary if and only if it is the identity and so 8 ~ 8 and 9 ~ 9. Finally, if ( is real, then all computations take place in the real field and so 8 and 9 are real. The 89 decomposition has important applications. For example, a system of linear equations (% ~ " can be written in the form 89% ~ " and since 8c ~ 8i , we have 9% ~ 8i " This is an upper triangular system, which is easily solved by back substitution; that is, starting from the bottom and working up. We mention also that the 89 factorization is associated with an algorithm for approximating the eigenvalues of a matrix, called the 89 algorithm. Specifically, if ( ~ ( is an d matrix, define a sequence of matrices as follows: 1) Let ( ~ 8 9 be the 89 factorization of ( and let ( ~ 9 8 . 2) Once ( has been defined, let ( ~ 8 9 be the 89 factorization of ( and let (b ~ 9 8 . Then ( is unitarily/orthogonally similar to (, since i 8c ( 8ic ~ 8c ²9c 8c ³8c ~ 8c 9c ~ (c

For complex matrices, it can be shown that under certain circumstances, such as when the eigenvalues of ( have distinct norms, the sequence ( converges

218

Advanced Linear Algebra

(entrywise) to an upper triangular matrix < , which therefore has the eigenvalues of ( on its main diagonal. Results can be obtained in the real case as well. For more details, we refer the reader to [48], page 115.

Hilbert and Hamel Bases Definition A maximal orthonormal set in an inner product space = is called a Hilbert basis for = . Zorn's lemma can be used to show that any nontrivial inner product space has a Hilbert basis. We leave the details to the reader. Some care must be taken not to confuse the concepts of a basis for a vector space and a Hilbert basis for an inner product space. To avoid confusion, a vector space basis, that is, a maximal linearly independent set of vectors, is referred to as a Hamel basis. We will refer to an orthonormal Hamel basis as an orthonormal basis. To be perfectly clear, there are maximal linearly independent sets called (Hamel) bases and maximal orthonormal sets (called Hilbert bases). If a maximal linearly independent set (basis) is orthonormal, it is called an orthonormal basis. Moreover, since every orthonormal set is linearly independent, it follows that an orthonormal basis is a Hilbert basis, since it cannot be properly contained in an orthonormal set. For finite-dimensional inner product spaces, the two types of bases are the same. Theorem 9.13 Let = be an inner product space. A finite subset E ~ ¸" Á Ã Á " ¹ of = is an orthonormal (Hamel) basis for = if and only if it is a Hilbert basis for = . Proof. We have seen that any orthonormal basis is a Hilbert basis. Conversely, if E is a finite maximal orthonormal set and E F , where F is linearly independent, then we may apply part 1) to extend E to a strictly larger orthonormal set, in contradiction to the maximality of E . Hence, E is maximal linearly independent. The following example shows that the previous theorem fails for infinitedimensional inner product spaces. Example 9.5 Let = ~ M and let 4 be the set of all vectors of the form ~ ²Á Ã Á Á Á Á Ã ³ where has a in the th coordinate and 's elsewhere. Clearly, 4 is an orthonormal set. Moreover, it is maximal. For if # ~ ²% ³ M has the property that # 4 , then

Real and Complex Inner Product Spaces

219

% ~ º#Á » ~ for all and so # ~ . Hence, no nonzero vector # ¤ 4 is orthogonal to 4 . This shows that 4 is a Hilbert basis for the inner product space M . On the other hand, the vector space span of 4 is the subspace : of all sequences in M that have finite support, that is, have only a finite number of nonzero terms and since span²4 ³ ~ : £ M , we see that 4 is not a Hamel basis for the vector space M .

The Projection Theorem and Best Approximations Orthonormal bases have a great practical advantage over arbitrary bases. From a computational point of view, if 8 ~ ¸# Á Ã Á # ¹ is a basis for = , then each # = has the form # ~ # b Ä b # In general, determining the coordinates requires solving a system of linear equations of size d . On the other hand, if E ~ ¸" Á Ã Á " ¹ is an orthonormal basis for = and # ~ " b Ä b " then the coefficients are quite easily computed: º#Á " » ~ º " b Ä b " Á " » ~ º" Á " » ~ Even if E ~ ¸" Á Ã Á " ¹ is not a basis (but just an orthonormal set), we can still consider the expansion V# ~ º#Á " »" b Ä b º#Á " »" Theorem 9.14 Let E ~ ¸" Á Ã Á " ¹ be an orthonormal subset of an inner product space = and let : ~ ºE». The Fourier expansion with respect to E of a vector # = is V# ~ º#Á " »" b Ä b º#Á " »" Each coefficient º#Á " » is called a Fourier coefficient of # with respect to E . The vector V# can be characterized as follows: 1) V# is the unique vector : for which ²# c ³ : . 2) V# is the best approximation to # from within : , that is, V# is the unique vector : that is closest to #, in the sense that )# c V#) )# c ) for all : ± ¸#¹ V .

220

Advanced Linear Algebra

3) Bessel's inequality holds for all # = , that is )V#) )#) Proof. For part 1), since º# c V#Á " » ~ º#Á " » c º#Á V " » ~ it follows that # c V# : . Also, if # c : for : , then c V# : and c V# ~ ²# c V#³ c ²# c ³ : and so ~ V#. For part 2), if ²# c V#³ ²# V c ³ and so

: , then # c V# : implies that

)# c ) ~ )# c V# b V# c ) ~ )# c V#) b )V# c ) Hence, )# c ) is smallest if and only if ~ V# and the smallest value is )# c V#). We leave proof of Bessel's inequality as an exercise. Theorem 9.15 (The projection theorem) If : is a finite-dimensional subspace of an inner product space = , then : ~ : p : In particular, if # = , then # ~ V# b ²# c V#³ : p : It follows that dim²= ³ ~ dim²:³ b dim²: ³ Proof. We have seen that # c V# : and so = ~ : b : . But : q : ~ ¸¹ and so = ~ : p : . The following example shows that the projection theorem may fail if : is not finite-dimensional. Indeed, in the infinite-dimensional case, : must be a complete subspace, but we postpone a discussion of this case until Chapter 13. Example 9.6 As in Example 9.5, let = ~ M and let : be the subspace of all sequences with finite support, that is, : is spanned by the vectors ~ ²Á Ã Á Á Á Á Ã ³ where has a in the th coordinate and 's elsewhere. If % ~ ²% ³ : , then % ~ º%Á » ~ for all and so % ~ . Therefore, : ~ ¸¹. However, : p : ~ : £ M The projection theorem has a variety of uses.

Real and Complex Inner Product Spaces

221

Theorem 9.16 Let = be an inner product space and let : be a finitedimensional subspace of = . 1) : ~ : 2) If ? = and dim²º?»³ B, then ? ~ º?» Proof. For part 1), it is clear that : : . On the other hand, if # : , then the projection theorem implies that # ~ b Z where : and Z : . Then Z is orthogonal to both and # and so Z is orthogonal to itself. Hence, Z ~ and # ~ : and so : ~ : . We leave the proof of part 2) as an exercise.

Characterizing Orthonormal Bases We can characterize orthonormal bases using Fourier expansions. Theorem 9.17 Let E ~ ¸" Á Ã Á " ¹ be an orthonormal subset of an inner product space = and let : ~ ºE». The following are equivalent: 1) E is an orthonormal basis for = . 2) ºE» ~ ¸¹ 3) Every vector is equal to its Fourier expansion, that is, for all # = , V# ~ # 4) Bessel's identity holds for all # = , that is, )V#) ~ )#) 5) Parseval's identity holds for all #Á $ = , that is, º#Á $» ~ ´#µ V E h ´$µ VE where ´#µ V E h ´$µ V E ~ º#Á " »º$Á " » b Ä b º#Á " »º$Á " » is the standard dot product in - . Proof. To see that 1) implies 2), if # ºE» is nonzero, then E r ¸#°)#)¹ is orthonormal and so E is not maximal. Conversely, if E is not maximal, there is an orthonormal set F for which E F . Then any nonzero # F ± E is in ºE» . Hence, 2) implies 1). We leave the rest of the proof as an exercise.

The Riesz Representation Theorem We have been dealing with linear maps for some time. We now have a need for conjugate linear maps. Definition A function ¢ = ¦ > on complex vector spaces is conjugate linear if it is additive, ²# b # ³ ~ # b #

222

Advanced Linear Algebra

and ²#³ ~ # for all d. A conjugate isomorphism is a bijective conjugate linear map. If % = , then the inner product function º h Á %»¢ = ¦ - defined by º h Á %»# ~ º#Á %» is a linear functional on = . Thus, the linear map ¢ = ¦ = i defined by % ~ º h Á %» is conjugate linear. Moreover, since º h Á %» ~ º h Á &» implies % ~ &, it follows that is injective and therefore a conjugate isomorphism (since = is finitedimensional). Theorem 9.18 (The Riesz representation theorem) Let = be a finitedimensional inner product space. 1) The map ¢ = ¦ = i defined by % ~ º h Á %» is a conjugate isomorphism. In particular, for each = i , there exists a unique vector % = for which ~ º h Á %», that is, # ~ º#Á %» for all # = . We call % the Riesz vector for and denote it by 9 . 2) The map 9¢ = i ¦ = defined by 9 ~ 9 is also a conjugate isomorphism, being the inverse of . We will call this map the Riesz map. Proof. Here is the usual proof that is surjective. If ~ , then 9 ~ , so let us assume that £ . Then 2 ~ ker² ³ has codimension and so = ~ º$» p 2 for $ 2 . Letting % ~ $ for - , we require that ²#³ ~ º#Á $» and since this clearly holds for any # 2 , it is sufficient to show that it holds for # ~ $, that is, ²$³ ~ º$Á $» ~ º$Á $» Thus, ~ ²$³°)$) and

Real and Complex Inner Product Spaces

9 ~

²$³ ) $)

223

$

For part 2), we have º#Á 9 b » ~ ² b ³²#³ ~ ²#³ b ²#³ ~ º#Á 9 » b º#Á 9 » ~ º#Á 9 b 9 » for all # = and so 9 b ~ 9 b 9

Note that if = ~ s , then 9 ~ ² ² ³Á Ã Á ² ³³, where ² Á Ã Á ³ is the standard basis for s .

Exercises 1. 2. 3. 4. 5.

Prove that if a matrix 4 is unitary, upper triangular and has positive entries on the main diagonal, must be the identity matrix. Use the QR factorization to show that any triangularizable matrix is unitarily (orthogonally) triangularizable. Verify the statement concerning equality in the triangle inequality. Prove the parallelogram law. Prove the Apollonius identity ) $ c " ) b )$ c # ) ~

6. 7.

)" c #) b h$ c ²" b #³h

Let = be an inner product space with basis 8 . Show that the inner product is uniquely defined by the values º"Á #», for all "Á # 8 . Prove that two vectors " and # in a real inner product space = are orthogonal if and only if )" b #) ~ )") b )#)

8. 9.

Show that an isometry is injective. Use Zorn's lemma to show that any nontrivial inner product space has a Hilbert basis. 10. Prove Bessel's inequality. 11. Prove that an orthonormal set E is a Hilbert basis for a finite-dimensional vector space = if and only if V# ~ #, for all # = . 12. Prove that an orthonormal set E is a Hilbert basis for a finite-dimensional vector space = if and only if Bessel's identity holds for all # = , that is, if and only if

224

Advanced Linear Algebra

)V#) ~ )#) for all # = . 13. Prove that an orthonormal set E is a Hilbert basis for a finite-dimensional vector space = if and only if Parseval's identity holds for all #Á $ = , that is, if and only if º#Á $» ~ ´#µ V E h ´$µ VE for all #Á $ = . 14. Let " ~ ² Á Ã Á ³ and # ~ ² Á Ã Á inequality states that (

³

be in s . The Cauchy–Schwarz

b Ä b ( ² b Ä b ³²

bÄb

³

Prove that we can do better: ²( ( b Ä b ( (³ ² b Ä b ³²

bÄb

³

15. Let = be a finite-dimensional inner product space. Prove that for any subset ? of = , we have ? ~ span²?³. 16. Let F3 be the inner product space of all polynomials of degree at most 3, under the inner product B

º²%³Á ²%³» ~

²%³²%³ c% %

cB

Apply the Gram–Schmidt process to the basis ¸Á %Á % Á % ¹, thereby computing the first four Hermite polynomials (at least up to a multiplicative constant). 17. Verify uniqueness in the Riesz representation theorem. 18. Let = be a complex inner product space and let : be a subspace of = . Suppose that # = is a vector for which º#Á » b º Á #» º Á » for all : . Prove that # : . 19. If = and > are inner product spaces, consider the function on = ^ > defined by º²# Á $ ³Á ²# Á $ ³» ~ º# Á # » b º$ Á $ » Is this an inner product on = ^ > ? 20. A normed vector space over s or d is a vector space (over s or d) together with a function ))¢ = ¦ s for which for all "Á # = and scalars we have a) )#) ~ (()#) b) )" b #) )") b )#) c) )#) ~ if and only if # ~ If = is a real normed space (over s) and if the norm satisfies the parallelogram law

Real and Complex Inner Product Spaces

225

)" b #) b )" c #) ~ )") b )#) prove that the polarization identity º"Á #» ~

²)" b #) c )" c #) ³

defines an inner product on = . Hint: Evaluate º"Á %» b º#Á %» to show that º"Á %» ~ º"Á %» and º"Á %» b º#Á %» ~ º" b #Á %». Then complete the proof that º"Á %» ~ º"Á %». 21. Let : be a subspace of a finite-dimensional inner product space = . Prove that each coset in = °: contains exactly one vector that is orthogonal to : .

Extensions of Linear Functionals 22. Let be a linear functional on a subspace : of a finite-dimensional inner product space = . Let ²#³ ~ º#Á 9 ». Suppose that = i is an extension of , that is, O: ~ . What is the relationship between the Riesz vectors 9 and 9 ? 23. Let be a nonzero linear functional on a subspace : of a finite-dimensional inner product space = and let 2 ~ ker² ³. Show that if = i is an extension of , then 9 2 ± : . Moreover, for each vector " 2 ± : there is exactly one scalar for which the linear functional ²?³ ~ º?Á "» is an extension of .

Positive Linear Functionals on s A vector # ~ ² Á Ã Á ³ in s is nonnegative (also called positive), written # , if for all . The vector # is strictly positive, written # , if # is nonnegative but not . The set sb of all strictly positive vectors in s is called the nonnegative orthant in s À The vector # is strongly positive, written # , if for all . The set sbb , of all strongly positive vectors in s is the strongly positive orthant in s À Let ¢ : ¦ s be a linear functional on a subspace : of s . Then is nonnegative (also called positive), written , if # ¬ ²#³ for all # : and is strictly positive, written , if # ¬ ²#³ for all # :À 24. Prove that a linear functional on s is positive if and only if 9 and strictly positive if and only if 9 . If : is a subspace of s is it true that a linear functional on : is nonnegative if and only if 9 ?

226

Advanced Linear Algebra

25. Let ¢ : ¦ s be a strictly positive linear functional on a subspace : of s . Prove that has a strictly positive extension to s . Use the fact that if < q s b ~ ¸¹, where sb ~ ¸² Á Ã Á ³ all ¹ and < is a subspace of s , then < contains a strongly positive vector. 26. If = is a real inner product space, then we can define an inner product on its complexification = d as follows (this is the same formula as for the ordinary inner product on a complex vector space): º" b #Á % b &» ~ º"Á %» b º#Á &» b ²º#Á %» c º"Á &»³ Show that )²" b #³) ~ )") b )#) where the norm on the left is induced by the inner product on = d and the norm on the right is induced by the inner product on = .

Chapter 10

Structure Theory for Normal Operators

Throughout this chapter, all vector spaces are assumed to be finite-dimensional unless otherwise noted. Also, the field - is either s or d.

The Adjoint of a Linear Operator The purpose of this chapter is to study the structure of certain special types of linear operators on finite-dimensional real and complex inner product spaces. In order to define these operators, we introduce another type of adjoint (different from the operator adjoint of Chapter 3). Theorem 10.1 Let = and > be finite-dimensional inner product spaces over and let B²= Á > ³. Then there is a unique function i ¢ > ¦ = , defined by the condition º #Á $» ~ º#Á i $» for all # = and $ > . This function is in B²> Á = ³ and is called the adjoint of . Proof. If i exists, then it is unique, for if º #Á $» ~ º#Á $» then º#Á $» ~ º#Á i $» for all # and $ and so ~ i . We seek a linear map i ¢ > ¦ = for which º#Á i $» ~ º #Á $» By way of motivation, the vector i $, if it exists, looks very much like a linear map sending # to º #Á $». The only problem is that i # is supposed to be a vector, not a linear map. But the Riesz representation theorem tells us that linear maps can be represented by vectors.

228

Advanced Linear Algebra

Specifically, for each $ > , the linear functional $ = i defined by $ # ~ º #Á $» has the form $ # ~ º#Á 9$ » where 9$ = is the Riesz vector for $ . If i ¢ > ¦ = is defined by i $ ~ 9$ ~ 9²$ ³ where 9 is the Riesz map, then º#Á i $» ~ º#Á 9$ » ~ $ # ~ º #Á $» Finally, since i ~ 9 k is the composition of the Riesz map 9 and the map ¢ $ ª $ and since both of these maps are conjugate linear, their composition is linear. Here are some of the basic properties of the adjoint. Theorem 10.2 Let = and > be finite-dimensional inner product spaces. For every Á B²= Á > ³ and - , 1) ² b ³i ~ i b i 2) ² ³i ~ i 3) ii ~ and so º i #Á $» ~ º#Á $» 4) If = ~ > , then ² ³i ~ i i 5) If is invertible, then ² c ³i ~ ² i ³c 6) If = ~ > and ²%³ s´%µ, then ² ³i ~ ² i ³. Moreover, if B²= ³ and : is a subspace of = , then 7) : is -invariant if and only if : is i -invariant. 8) ²:Á : ³ reduces if and only if : is both -invariant and i -invariant, in which case ² O: ³i ~ ² i ³O: Proof. For part 7), let : and ' : and write º i 'Á » ~ º'Á » Now, if : is -invariant, then º i 'Á » ~ for all : and so i ' : and : is i -invariant. Conversely, if : is i -invariant, then º'Á » ~ for all ' : and so : ~ : , whence : is -invariant. The first statement in part 8) follows from part 7) applied to both : and : . For the second statement, since : is both -invariant and i -invariant, if Á ! : ,

Structure Theory for Normal Operators

229

then º Á ² i ³O: ²!³» ~ º Á i !» ~ º Á !» ~ º O: ² ³Á !» Hence, by definition of adjoint, ² i ³O: ~ ² O: ³i . Now let us relate the kernel and image of a linear transformation to those of its adjoint. Theorem 10.3 Let B²= Á > ³, where = and > are finite-dimensional inner product spaces. 1) ker² i ³ ~ im² ³

and

im² i ³ ~ ker² ³

surjective injective

¯ ¯

i injective i surjective

ker² i ³ ~ ker² ³

and

ker² i ³ ~ ker² i ³

and so

2)

3) im² i ³ ~ im² i ³

and

im² i ³ ~ im² ³

4) ²:Á; ³i ~ ; Á: Proof. For part 1), " ker² i ³ ¯ i " ~ ¯ º i "Á = » ~ ¸¹ ¯ º"Á = » ~ ¸¹ ¯ " im² ³ and so ker² i ³ ~ im² ³ . The second equation in part 1) follows by replacing by i and taking complements. For part 2), it is clear that ker² ³ ker² i ³. For the reverse inclusion, we have i " ~

¬

º i "Á "» ~

¬

º "Á "» ~

¬

" ~

and so ker² i ³ ker² ³. The second equation follows from the first by replacing with i . We leave the rest of the proof for the reader.

230

Advanced Linear Algebra

The Operator Adjoint and the Hilbert Space Adjoint We should make some remarks about the relationship between the operator adjoint d of , as defined in Chapter 3 and the adjoint i that we have just defined, which is sometimes called the Hilbert space adjoint. In the first place, if ¢ = ¦ > , then d and i have different domains and ranges: d¢ > i ¦ = i

and

i¢ > ¦ =

The two maps are shown in Figure 10.1, along with the conjugate Riesz isomorphisms 9 = ¢ = i ¦ = and 9 > ¢ > i ¦ > .

V*

Wx

W*

RV

RW

V

W* W

W

Figure 10.1 i

i

The composite map ¢ > ¦ = defined by ~ ²9 = ³c k i k 9 > is linear. Moreover, for all > i and # = , ² d ² ³³# ~ ² #³ ~ º #Á 9 > ² ³» ~ º#Á i 9 > ² ³» ~ ´²9 = ³c ² i 9 > ² ³³µ²#³ ~ ² ³# and so ~ d . Hence, the relationship between d and i is d ~ ²9 = ³c k i k 9 > Loosely speaking, the Riesz functions are like “change of variables” functions from linear functionals to vectors, and we can say that i does to Riesz vectors what d does to the corresponding linear functionals. Put another way (and just as loosely), and i are the same, up to conjugate Riesz isomorphism. In Chapter 3, we showed that the matrix of the operator adjoint d is the transpose of the matrix of the map . For Hilbert space adjoints, the situation is slightly different (due to the conjugate linearity of the inner product). Suppose that 8 ~ ² Á Ã Á ³ and 9 ~ ² Á Ã Á ³ are ordered orthonormal bases for = and > , respectively. Then

Structure Theory for Normal Operators

231

²´ i µ9Á8 ³Á ~ º i Á » ~ º Á » ~ º Á » ~ ²´ µ8Á9 ³Á and so ´ i µ9Á8 and ´ µ8Á9 are conjugate transposes. The conjugate transpose of a matrix ( ~ ²Á ³ is (i ~ ²Á ³! and is called the adjoint of (. Theorem 10.4 Let B²= Á > ³, where = and > are finite-dimensional inner product spaces. 1) The operator adjoint d and the Hilbert space adjoint i are related by d ~ ²9 = ³c k i k 9 > where 9 = and 9 > are the conjugate Riesz isomorphisms on = and > , respectively. 2) If 8 and 9 are ordered orthonormal bases for = and > , respectively, then ´ i µ9Á8 ~ ²´ µ8Á9 ³i In words, the matrix of the adjoint i is the adjoint (conjugate transpose) of the matrix of .

Orthogonal Projections In an inner product space, we can single out some special projection operators. Definition A projection of the form :Á: is said to be orthogonal. Equivalently, a projection is orthogonal if ker²³ im²³. Some care must be taken to avoid confusion between orthogonal projections and two projections that are orthogonal to each other, that is, for which ~ ~ . We have seen that an operator is a projection operator if and only if it is idempotent. Here is the analogous characterization of orthogonal projections. Theorem 10.5 Let = be a finite-dimensional inner product space. The following are equivalent for an operator on = : 1) is an orthogonal projection 2) is idempotent and self-adjoint 3) is idempotent and does not expand lengths, that is )#) )#) for all # = .

232

Advanced Linear Algebra

Proof. Since ²:Á; ³i ~ ; Á: it follows that ~ i if and only if : ~ ; , that is, if and only if is orthogonal. Hence, 1) and 2) are equivalent. To prove that 1) implies 3), let ~ :Á: . Then if # ~ b ! for ! : , it follows that

: and

)#) ~ ) ) b )!) ) ) ~ )#) Now suppose that 3) holds. Then im²³ l ker²³ ~ = ~ ker²³ p ker²³ and we wish to show that the first sum is orthogonal. If $ im²³, then $ ~ % b &, where % ker²³ and & ker²³ . Hence, $ ~ $ ~ % b & ~ & and so the orthogonality of % and & implies that )%) b )&) ~ )$) ~ )&) )&) Hence, % ~ and so im²³ ker²³ , which implies that im²³ ~ ker²³ .

Orthogonal Resolutions of the Identity We have seen (Theorem 2.25) that resolutions of the identity b Ä b ~ on = correspond to direct sum decompositions of = . If, in addition, the projections are orthogonal, then the direct sum is an orthogonal sum. Definition An orthogonal resolution of the identity is a resolution of the identity b Ä b ~ in which each projection is orthogonal. The following theorem displays a correspondence between orthogonal direct sum decompositions of = and orthogonal resolutions of the identity. Theorem 10.6 Let = be an inner product space. Orthogonal resolutions of the identity on = correspond to orthogonal direct sum decompositions of = as follows: 1) If b Ä b ~ is an orthogonal resolution of the identity, then = ~ im² ³ p Ä p im² ³ and is orthogonal projection onto im² ³.

Structure Theory for Normal Operators

233

2) Conversely, if = ~ : p Ä p : and if is orthogonal projection onto : , then b Ä b ~ is an orthogonal resolution of the identity. Proof. To prove 1), if b Ä b ~ is an orthogonal resolution of the identity, Theorem 2.25 implies that = ~ im² ³ l Ä l im² ³ However, since the 's are pairwise orthogonal and self-adjoint, it follows that º #Á $» ~ º#Á $» ~ º#Á » ~ and so = ~ im² ³ p Ä p im² ³ For the converse, Theorem 2.25 implies that b Ä b ~ is a resolution of the identity where is projection onto im² ³ along ker² ³ ~ im² ³ ~ im² ³ £

Hence, is orthogonal.

Unitary Diagonalizability We have seen (Theorem 8.10) that a linear operator B²= ³ on a finitedimensional vector space = is diagonalizable if and only if = ~ ; l Ä l ; Of course, each eigenspace ; has an orthonormal basis E , but the union of these bases need not be an orthonormal basis for = . Definition A linear operator B²= ³ is unitarily diagonalizable (when = is complex) and orthogonally diagonalizable (when = is real) if there is an ordered orthonormal basis E ~ ²" Á Ã Á " ³ of = for which the matrix ´ µE is diagonal, or equivalently, if " ~ " for all ~ Á Ã Á . Here is the counterpart of Theorem 8.10 for inner product spaces. Theorem 10.7 Let = be a finite-dimensional inner product space and let B²= ³. The following are equivalent: 1) is unitarily (orthogonally) diagonalizable. 2) = has an orthonormal basis that consists entirely of eigenvectors of .

234

Advanced Linear Algebra

3) = has the form = ~ ; p Ä p ; where Á Ã Á are the distinct eigenvalues of . For simplicity in exposition, we will tend to use the term unitarily diagonalizable for both cases. Since unitarily diagonalizable operators are so well behaved, it is natural to seek a characterization of such operators. Remarkably, there is a simple one, as we will see next.

Normal Operators Operators that commute with their own adjonts are very special. Definition 1) A linear operator on an inner product space = is normal if it commutes with its adjoint: i ~ i 2) A matrix ( C ²- ³ is normal if ( commutes with its adjoint (i . If is normal and E is an ordered orthonormal basis of = , then ´ µE ´ µiE ~ ´ µE ´ i µE ~ ´ i µE and ´ µiE ´ µE ~ ´ i µE ´ µE ~ ´ i µE and so is normal if and only if ´ µE is normal for some, and hence all, orthonormal bases for = . Note that this does not hold for bases that are not orthonormal. Normal operators have some very special properties. Theorem 10.8 Let B²= ³ be normal. 1) The following are also normal: a) O: , if reduces ²:Á : ³ b) i c) c , if is invertible d) ² ³, for any polynomial ²%³ - ´%µ 2) For any #Á $ = , º #Á $» ~ º i #Á i $» and, in particular, ) #) ~ ) i #)

Structure Theory for Normal Operators

235

and so ker² i ³ ~ ker² ³ 3) For any integer , ker² ³ ~ ker² ³ 4) The minimal polynomial ²%³ is a product of distinct prime monic polynomials. 5) # ~ #

¯

i # ~ #

6) If : and ; are submodules of = with relatively prime orders, then : ; . 7) If and are distinct eigenvalues of , then ; ; . Proof. We leave part 1) for the reader. For part 2), normality implies that º #Á $» ~ º i #Á #» ~ º i #Á #» ~ º i #Á i #» We prove part 3) first for the operator ~ i , which is self-adjoint, that is, i ~ ² i ³ i ~ i ~ If # ~ for , then ~ º #Á c #» ~ ºc #Á c #» and so c # ~ . Continuing in this way gives # ~ . Now, if # ~ for , then # ~ ² i ³ # ~ ² i ³ # ~ and so # ~ . Hence, ~ º#Á #» ~ º i #Á #» ~ º #Á #» and so # ~ . For part 4), suppose that ²%³ ~ ²%³²%³ where ²%³ is monic and prime. Then for any # = , ² ³´² ³#µ ~ and since ² ³ is also normal, part 3) implies that ² ³´² ³#µ ~ for all # = . Hence, ² ³² ³ ~ , which implies that ~ . Thus, the prime factors of ²%³ appear only to the first power.

236

Advanced Linear Algebra

Part 5) follows from part 2): ker² c ³ ~ ker´² c ³i µ ~ ker² i c ³ For part 6), if ²:³ ~ ²%³ and ²; ³ ~ ²%³, then there are polynomials ²%³ and ²%³ for which ²%³²%³ b ²%³²%³ ~ and so ² ³² ³ b ² ³² ³ ~ Now, ~ ² ³² ³ annihilates : and ~ ² ³² ³ annihilates ; . Therefore i also annihilates ; and so º:Á ; » ~ º² b ³:Á ; » ~ º :Á ; » ~ º:Á i ; » ~ ¸¹ Part 7) follows from part 6), since ²; ³ ~ % c and ²; ³ ~ % c are relatively prime when £ . Alternatively, for # ; and $ ; , we have º#Á $» ~ º #Á $» ~ º#Á i $» ~ º#Á $» ~ º#Á $» and so £ implies that º#Á $» ~ .

The Spectral Theorem for Normal Operators Theorem 10.8 implies that when - ~ d, the minimal polynomial ²%³ splits into distinct linear factors and so Theorem 8.11 implies that is diagonalizable, that is, = ~ ; l Ä l ; Moreover, since distinct eigenspaces of a normal operator are orthogonal, we have = ~ ; p Ä p ; and so is unitarily diagonalizable. The converse of this is also true. If = has an orthonormal basis E ~ i are ¸# Á Ã Á # ¹ of eigenvectors for , then since ´ µE and ´ i µE ~ ´ µE diagonal, these matrices commute and therefore so do i and . Theorem 10.9 (The spectral theorem for normal operators: complex case) Let = be a finite-dimensional complex inner product space and let B²= ³. The following are equivalent: 1) is normal. 2) is unitarily diagonalizable, that is, = ~ ; p Ä p ;

Structure Theory for Normal Operators

237

3) has an orthogonal spectral resolution (10.1)

~ b Ä b

where b Ä b ~ and is orthogonal for all , in which case, ¸ Á Ã Á ¹ is the spectrum of and and

im² ³ ~ ;

ker² ³ ~ ; £

Proof. We have seen that 1) and 2) are equivalent. To see that 2) and 3) are equivalent, Theorem 8.12 says that = ~ ; l Ä l ; if and only if ~ b Ä b and in this case, im² ³ ~ ;

and

ker² ³ ~ ; £

But ; ; for £ if and only if im² ³ ker² ³ that is, if and only if each is orthogonal. Hence, the direct sum = ~ ; l Ä l ; is an orthogonal sum if and only if each projection is orthogonal. The Real Case If - ~ s, then ²%³ has the form ²%³ ~ ²% c ³Ä²% c ³ ²%³Ä ²%³ where each ²%³ is an irreducible monic quadratic. Hence, the primary cyclic decomposition of = gives = ~ ; p Ä p ; p > p Ä p > where > is cyclic with prime quadratic order ²%³. Therefore, Theorem 8.8 implies that there is an ordered basis 8 for which ´ O> µ8 ~ >

c ?

Theorem 10.10 (The spectral theorem for normal operators: real case) A linear operator on a finite-dimensional real inner product space is normal if and only if

238

Advanced Linear Algebra

= ~ ; p Ä p ; p > p Ä p > where ¸ Á Ã Á ¹ is the spectrum of and each > is an indecomposable twodimensional -invariant subspace with an ordered basis 8 for which ´ µ8 ~ >

c ?

Proof. We need only show that if = has such a decomposition, then is normal. But ´ µ8 ´ µ!8 ~ ² b ³0 ~ ´ µ!8 ´ µ8 and so ´ µ8 is normal. It follows easily that is normal.

Special Types of Normal Operators We now want to introduce some special types of normal operators. Definition Let = be an inner product space. 1) B²= ³ is self-adjoint (also called Hermitian in the complex case and symmetric in the real case) if i ~ 2) B²= ³ is skew self-adjoint (also called skew-Hermitian in the complex case and skew-symmetric in the real case) if i ~ c 3) B²= ³ is unitary in the complex case and orthogonal in the real case if is invertible and i ~ c

There are also matrix versions of these definitions, obtained simply by replacing the operator by a matrix (. Moreover, the operator is self-adjoint if and only if any matrix that represents with respect to an ordered orthonormal basis E is self-adjoint. Similar statements hold for the other types of operators in the previous definition. In some sense, square complex matrices are a generalization of complex numbers and the adjoint (conjugate transpose) is a generalization of the complex conjugate. In looking for a better analogy, we could consider just the diagonal matrices, but this is a bit too restrictive. The next logical choice is the set D of normal matrices. Indeed, among the complex numbers, there are some special subsets: the real numbers, the positive numbers and the numbers on the unit circle. We will soon see that a complex matrix ( is self-adjoint if and only if its complex eigenvalues

Structure Theory for Normal Operators

239

are real. This would suggest that the analog of the set of real numbers is the set of self-adjoint matrices. Also, we will see that a complex matrix is unitary if and only if its eigenvalues have norm , so numbers on the unit circle seem to correspond to the set of unitary matrices. This leaves open the question of which normal matrices correspond to the positive real numbers. These are the positive definite matrices, which we will discuss later in the chapter.

Self-Adjoint Operators Let us consider the basic properties of self-adjoint operators. The quadratic form associated with the linear operator is the function 8 ¢ = ¦ - defined by 8 ²#³ ~ º #Á #» We have seen (Theorem 9.2) that in a complex inner product space, ~ if and only if 8 ~ but this does not hold, in general, for real inner product spaces. However, it does hold for symmetric operators on a real inner product space. Theorem 10.11 Let = be a finite-dimensional inner product space and let Á B²= ³. 1) If and are self-adjoint, then so are the following: a) b b) c , if is invertible c) ² ³, for any real polynomial ²%³ s´%µ 2) A complex operator is Hermitian if and only if 8 ²#³ is real for all #=. 3) If is a complex operator or a real symmetric operator, then ~

¯

8 ~

4) The characteristic polynomial ²%³ of a self-adjoint operator splits over s, that is, all complex roots of ²%³ are real. Hence, the minimal polynomial ²%³ of is the product of distinct monic linear factors over s. Proof. For part 2), if is Hermitian, then º #Á #» ~ º#Á #» ~ º #Á #» and so 8 ²#³ ~ º #Á #» is real. Conversely, if º #Á #» s, then º#Á #» ~ º #Á #» ~ º#Á i #» and so ~ i . For part 3), we need only prove that 8 ~ implies ~ when - ~ s. But if 8 ~ , then

240

Advanced Linear Algebra

~ º ²% b &³Á % b &» ~ º %Á %» b º &Á &» b º %Á &» b º &Á %» ~ º %Á &» b º &Á %» ~ º %Á &» b º%Á &» ~ º %Á &» b º %Á &» ~ º %Á &» and so ~ . For part 4), if is Hermitian (- ~ d) and # ~ #, then # ~ # ~ i # ~ # and so ~ is real. If is symmetric (- ~ s) , we must be a bit careful, since a nonreal root of ²%³ is not an eigenvalue of . However, matrix techniques can come to the rescue here. If ( ~ ´ µE for any ordered orthonormal basis E for = , then ²%³ ~ ( ²%³. Now, ( is a real symmetric matrix, but can be thought of as a complex Hermitian matrix with real entries. As such, it represents a Hermitian linear operator on the complex space d and so, by what we have just shown, all (complex) roots of its characteristic polynomial are real. But the characteristic polynomial of ( is the same, whether we think of ( as a real or a complex matrix and so the result follows.

Unitary Operators and Isometries We now turn to the basic properties of unitary operators. These are the workhorse operators, in that a unitary operator is precisely a normal operator that maps orthonormal bases to orthonormal bases. Note that is unitary if and only if º #Á $» ~ º#Á c $» for all #Á $ = . Theorem 10.12 Let = be a finite-dimensional inner product space and let Á B²= ³. 1) If and are unitary/orthogonal, then so are the following: a) , for dÁ (( ~ b) c) c , if is invertible. 2) is unitary/orthogonal if and only it is an isometric isomorphism. 3) is unitary/orthogonal if and only if it takes some orthonormal basis to an orthonormal basis, in which case it takes all orthonormal bases to orthonormal bases. 4) If is unitary/orthogonal, then the eigenvalues of have absolute value .

Structure Theory for Normal Operators

241

Proof. We leave the proof of part 1) to the reader. For part 2), a unitary/orthogonal map is injective and since = is finite-dimensional, it is bijective. Moreover, for a bijective linear map , we have is an isometry ¯ º #Á $» ~ º#Á $» for all #Á $ = ¯ º#Á i $» ~ º#Á $» for all #Á $ = ¯ i ~ ¯ i ~ c ¯ is unitary/orthogonal For part 3), suppose that is unitary/orthogonal and that E ~ ¸" Á Ã Á " ¹ is an orthonormal basis for = . Then º " Á " » ~ º" Á " » ~ Á and so E is an orthonormal basis for = . Conversely, suppose that E and E are orthonormal bases for = . Then º " Á " » ~ Á ~ º" Á " » which implies that º #Á $» ~ º#Á $» for all #Á $ = unitary/orthogonal.

and so is

For part 4), if is unitary and # ~ #, then º#Á #» ~ º#Á #» ~ º #Á #» ~ º#Á #» and so (( ~ ~ , which implies that (( ~ . We also have the following theorem concerning unitary (and orthogonal) matrices. Theorem 10.13 Let ( be an d matrix over - ~ d or - ~ s. 1) The following are equivalent: a) ( is unitary/orthogonal. b) The columns of ( form an orthonormal set in - . c) The rows of ( form an orthonormal set in - . 2) If ( is unitary, then (det²(³( ~ . If ( is orthogonal, then det²(³ ~ f. Proof. The matrix ( is unitary if and only if ((i ~ 0 , which is equivalent to the rows of ( being orthonormal. Similarly, ( is unitary if and only if (i ( ~ 0 , which is equivalent to the columns of ( being orthonormal. As for part 2), ((i ~ 0

¬

det²(³det²(i ³ ~

from which the result follows.

¬

det²(³det²(³ ~

242

Advanced Linear Algebra

Unitary/orthogonal matrices play the role of change of basis matrices when we restrict attention to orthonormal bases. Let us first note that if 8 ~ ²" Á Ã Á " ³ is an ordered orthonormal basis and # ~ " b Ä b " $ ~ " b Ä b " then º#Á $» ~ b Ä b ~ ´#µ8 h ´$µ8 where the right hand side is the standard inner product in - and so # $ if and only if ´#µ8 ´$µ8 . We can now state the analog of Theorem 2.9. Theorem 10.14 If we are given any two of the following: 1) A unitary/orthogonal d matrix (, 2) An ordered orthonormal basis 8 for - , 3) An ordered orthonormal basis 9 for - , then the third is uniquely determined by the equation ( ~ 48 Á 9 Proof. Let 8 ~ ¸ ¹ be a basis for = . If 9 is an orthonormal basis for = , then º Á » ~ ´ µ9 h ´ µ9 where ´ µ9 is the th column of ( ~ 48Á9 . Hence, ( is unitary if and only if 8 is orthonormal. We leave the rest of the proof to the reader.

Unitary Similarity We have seen that the change of basis formula for operators is given by ´ µ8Z ~ 7 ´ µ8 7 c where 7 is an invertible matrix. What happens when the bases are orthonormal? Definition 1) Two complex matrices ( and ) are unitarily similar (also called unitarily equivalent) if there exists a unitary matrix < for which ) ~ < (< c ~ < (< i The equivalence classes associated with unitary similarity are called unitary similarity classes. 2) Similarly, two real matrices ( and ) are orthogonally similar (also called orthogonally equivalent) if there exists an orthogonal matrix 6 for which ) ~ 6(6c ~ 6(6! The equivalence classes associated with orthogonal similarity are called orthogonal similarity classes.

Structure Theory for Normal Operators

243

The analog of Theorem 2.19 is the following. Theorem 10.15 Let = be an inner product space of dimension . Then two d matrices ( and ) are unitarily/orthogonally similar if and only if they represent the same linear operator B²= ³ with respect to (possibly different) ordered orthonormal bases. In this case, ( and ) represent exactly the same set of linear operators in B²= ³ with respect to ordered orthonormal bases. Proof. If ( and ) represent B²= ³, that is, if ( ~ ´ µ8

and

) ~ ´ µ 9

for ordered orthonormal bases 8 and 9, then ) ~ 48Á9 (49Á8 and according to Theorem 10.14, 48Á9 is unitary/orthogonal. Hence, ( and ) are unitarily/orthogonally similar. Now suppose that ( and ) are unitarily/orthogonally similar, say ) ~ < (< c where < is unitary/orthogonal. Suppose also that ( represents a linear operator B²= ³ for some ordered orthonormal basis 8 , that is, ( ~ ´ µ8 Theorem 10.14 implies that there is a unique ordered orthonormal basis 9 for = for which < ~ 48Á9 . Hence ) ~ 48Á9 ´ µ8 48c Á 9 ~ ´ µ9 and so ) also represents . By symmetry, we see that ( and ) represent the same set of linear operators, under all possible ordered orthonormal bases. We have shown (see the discussion of Schur's theorem) that any complex matrix ( is unitarily similar to an upper triangular matrix, that is, that ( is unitarily upper triangularizable. However, upper triangular matrices do not form a set of canonical forms under unitary similarity. Indeed, the subject of canonical forms for unitary similarity is rather complicated and we will not discuss it in this book, but instead refer the reader to the survey article [28].

Reflections The following defines a very special type of unitary operator.

244

Advanced Linear Algebra

Definition For a nonzero # = , the unique operator /# for which /# # ~ c#Á /# $ ~ $ for all $ º#» is called a reflection or a Householder transformation. It is easy to verify that /# % ~ % c

º%Á #» # º#Á #»

Moreover, /# % ~ c% for % £ if and only if % ~ # for some - and so we can uniquely identify # by the behavior of the reflection on = . If /# is a reflection and if we extend # to an ordered orthonormal basis 8 for = , then ´/# µ8 is the matrix obtained from the identity matrix by replacing the upper left entry by c, v c x ´/# µ8 ~ x w

y { {

Æ

z

Thus, a reflection is both unitary and Hermitian, that is, /#i ~ /#c ~ /# Given two nonzero vectors of equal length, there is precisely one reflection that interchanges these vectors. Theorem 10.16 Let #Á $ = be distinct nonzero vectors of equal length. Then /#c$ is the unique reflection sending # to $ and $ to #. Proof. If )#) ~ )$), then ²# c $³ ²# b $³ and so /#c$ ²# c $³ ~ $ c # /#c$ ²# b $³ ~ # b $ from which it follows that /#c$ ²#³ ~ $ and /#c$ ²$³ ~ #. As to uniqueness, suppose /% is a reflection for which /% ²#³ ~ $. Since /%c ~ /% , we have /% ²$³ ~ # and so /% ²# c $³ ~ c²# c $³ which implies that /% ~ /#c$ . Reflections can be used to characterize unitary operators. Theorem 10.17 Let = be a finite-dimensional inner product space. The following are equivalent for an operator B²= ³:

Structure Theory for Normal Operators

245

1) is unitary/orthogonal 2) is a product of reflections. Proof. Since reflections are unitary/orthogonal and the product of unitary/ orthogonal operators is unitary, it follows that 2) implies 1). For the converse, let be unitary. Let 8 ~ ²" Á Ã Á " ³ be an orthonormal basis for = . Then / " c" ² " ³ ~ " and so if % ~ " c " then ²/% ³" ~ " that is, /% is the identity on º" ». Suppose that we have found reflections /%c Á Ã Á /% for which c /%c Ä/% is the identity on º" Á Ã Á "c ». Then /c " c" ²c " ³ ~ " Moreover, we claim that ²c " c " ³ " for , since ºc " c " Á " » ~ º²/%c Ä/% ³" Á " » ~ º " Á /% Ä/%c " » ~ º " Á " » ~ º" Á " » ~ Hence, if % ~ c " c " , then ²/% Ä/% ³" ~ /% " ~ " and so /% Ä/% is the identity on º" Á Ã Á " ». Thus, for ~ we have /% Ä/% ~ and so ~ /% Ä/% , as desired.

The Structure of Normal Operators The following theorem includes the spectral theorems stated above for real and complex normal operators, along with some further refinements related to selfadjoint and unitary/orthogonal operators. Theorem 10.18 (The structure theorem for normal operators) 1) (Complex case) Let = be a finite-dimensional complex inner product space. a) The following are equivalent for B²= ³: i) is normal ii) is unitarily diagonalizable iii) has an orthogonal spectral resolution ~ b Ä b

246

Advanced Linear Algebra

b) Among the normal operators, the Hermitian operators are precisely those for which all complex eigenvalues are real. c) Among the normal operators, the unitary operators are precisely those for which all complex eigenvalues have norm . 2) (Real case) Let = be a finite-dimensional real inner product space. a) B²= ³ is normal if and only if = ~ ; p Ä p ; p > p Ä p > where ¸ Á Ã Á ¹ is the spectrum of and each > is a twodimensional indecomposable -invariant subspace with an ordered basis 8 for which ´ µ8 ~ >

c ?

b) Among the real normal operators, the symmetric operators are those for which there are no subspaces > in the decomposition of part 2a). Hence, the following are equivalent for B²= ³: i) is symmetric. ii) is orthogonally diagonalizable. iii) has the orthogonal spectral resolution ~ b Ä b c)

Among the real normal operators, the orthogonal operators are precisely those for which the eigenvalues are equal to f and the matrices ´ µ8 described in part 2a) have rows (and columns) of norm , that is, ´ µ8 ~ >

sin cos

ccos sin ?

for some s. Proof. We have proved part 1a). As to part 1b), it is only necessary to look at a diagonal matrix ( representing . This matrix has the eigenvalues of on its main diagonal and so it is Hermitian if and only if the eigenvalues of are real. Similarly, ( is unitary if and only if the eigenvalues of have absolute value equal to . We have proved part 2a). Parts 2b) and 2c) follow by looking at the matrix ( ~ ´ µ8 where 8 ~ 8 . This matrix is symmetric if and only if ( is diagonal, and ( is orthogonal if and only if ~ f and the matrices ´ µ8 have orthonormal rows.

Matrix Versions We can formulate matrix versions of the structure theorem for normal operators.

Structure Theory for Normal Operators

247

Theorem 10.19 (The structure theorem for normal matrices) 1) (Complex case) a) A complex matrix ( is normal if and only if it is unitarily diagonalizable, that is, if and only if there is a unitary matrix < for which < (< i ~ diag² Á Ã Á ³ b) A complex matrix ( is Hermitian if and only if 1a) holds, where all eigenvalues are real. c) A complex matrix ( is unitary if and only if 1a) holds, where all eigenvalues have norm . 2) (Real case) a) A real matrix ( is normal if and only if there is an orthogonal matrix 6 for which 6(6! ~ diag6 Á Ã Á Á >

c ÁÃÁ> ?

c ?7

b) A real matrix ( is symmetric if and only if it is orthogonally diagonalizable, that is, if and only if there is an orthogonal matrix 6 for which 6(6! ~ diag² Á Ã Á ³ c)

A real matrix ( is orthogonal if and only if there is an orthogonal matrix 6 for which 6(6! ~ diag6 Á Ã Á Á >

sin cos

sin ccos ÁÃÁ> sin ? cos

ccos sin ?7

for some Á Ã Á s.

Functional Calculus Let be a normal operator on a finite-dimensional inner product space = and let have spectral resolution ~ b Ä b Since each is idempotent, we have ~ for all . The pairwise orthogonality of the projections implies that ~ ² b Ä b ³ ~ b Ä b More generally, for any polynomial ²%³ over - , ² ³ ~ ² ³ b Ä b ² ³ Note that a polynomial of degree c is uniquely determined by specifying an

248

Advanced Linear Algebra

arbitrary set of of its values at the distinct points Á Ã Á . This follows from the Lagrange interpolation formula c v % c y ²%³ ~ ² ³ w £ c z ~

Therefore, we can define a unique polynomial ²%³ by specifying the values ² ³, for ~ Á Ã Á . For example, for a given , if ²%³ is a polynomial for which ² ³ ~ Á for ~ Á Ã Á , then ² ³ ~ and so each projection is a polynomial function of . As another example, if is invertible and ² ³ ~ c , then ² ³ ~ c b Ä b c ~ c as can easily be verified by direct calculation. Finally, if ² ³ ~ , then since each is self-adjoint, we have ² ³ ~ b Ä b ~ i and so i is a polynomial in . We can extend this idea further by defining, for any function ¢ ¸ Á Ã Á ¹ ¦ the linear operator ² ³ by ² ³ ~ ² ³ b Ä b ² ³ For example, we may define j Á c Á and so on. Notice, however, that since the spectral resolution of is a finite sum, we gain nothing (but convenience) by using functions other than polynomials, for we can always find a polynomial ²%³ for which ² ³ ~ ² ³ for ~ Á Ã Á and so ² ³ ~ ² ³. The study of the properties of functions of an operator is referred to as the functional calculus of . According to the spectral theorem, if = is complex and is normal, then ² ³ is a normal operator whose eigenvalues are ² ³. Similarly, if = is real and is symmetric, then ² ³ is symmetric, with eigenvalues ² ³.

Structure Theory for Normal Operators

249

Commutativity The functional calculus can be applied to the study of the commutativity properties of operators. Here are two simple examples. Theorem 10.20 Let = be a finite-dimensional complex inner product space. For Á B²= ³, we write © to denote the fact that and commute. Let and have spectral resolutions ~ b Ä b ~ b Ä b Then 1) For any B²= ³, ©

¯

© for all

©

¯

© , for all Á

2)

3) If ¢ ¸ Á Ã Á ¹ ¦ - and ¢ ¸ Á Ã Á ¹ ¦ - are injective functions, then ² ³ © ²³

¯

©

Proof. For 1), if © for all , then © and the converse follows from the fact that is a polynomial in . Part 2) is similar. For part 3), © clearly implies ² ³ © ²³. For the converse, let $ ~ ¸ Á Ã Á ¹. Since is injective, the inverse function c ¢ ²$³ ¦ $ is well-defined and c ² ² ³³ ~ . Thus, is a function of ² ³. Similarly, is a function of ²³. It follows that ² ³ © ²³ implies © . Theorem 10.21 Let and be normal operators on a finite-dimensional complex inner product space = . Then and commute if and only if they have the form ~ ²² Á ³³ ~ ²² Á ³³ where ²%³Á ²%³ and ²%Á &³ are polynomials. Proof. If and are polynomials in ~ ² Á ³, then they clearly commute. For the converse, suppose that ~ and let ~ b Ä b and ~ b Ä b be the orthogonal spectral resolutions of and .

250

Advanced Linear Algebra

Then Theorem 10.20 implies that ~ . Hence, ~ ² b Ä b ³ ² b Ä b ³ ~ ² b Ä b ³² b Ä b ³ ~ Á

It follows that for any polynomial ²%Á &³ in two variables, ² Á ³ ~ ² Á ³ Á

So if we choose ²%Á &³ with the property that Á ~ ² Á ³ are distinct, then ² Á ³ ~ Á Á

and we can also choose ²%³ and ²%³ so that ²Á ³ ~ for all and ²Á ³ ~ for all . Then ²² Á ³³ ~ ²Á ³ ~ Á

Á

~ 8 98 9 ~ ~

and similarly, ²² Á ³³ ~ .

Positive Operators One of the most important cases of the functional calculus is ²%³ ~ j%. Recall that the quadratic form associated with a linear operator is 8 ²#³ ~ º #Á #» Definition A self-adjoint linear operator B²= ³ is 1) positive if 8 ²#³ for all # = 2) positive definite if 8 ²#³ for all # £ . Theorem 10.22 A self-adjoint operator on a finite-dimensional inner product space is 1) positive if and only if all of its eigenvalues are nonnegative 2) positive definite if and only if all of its eigenvalues are positive. Proof. If 8 ²#³ and # ~ #, then º #Á #» ~ º#Á #»

Structure Theory for Normal Operators

251

and so . Conversely, if all eigenvalues of are nonnegative, then ~ b Ä b Á and since ~ b Ä b , º #Á #» ~ º #Á #» ~ ) #) Á

and so is positive. Part 2) is proved similarly. If is a positive operator, with spectral resolution ~ b Ä b Á then we may take the positive square root of , j ~ j b Ä b j where j is the nonnegative square root of . It is clear that ²j ³ ~ and it is not hard to see that j is the only positive operator whose square is . In other words, every positive operator has a unique positive square root. Conversely, if has a positive square root, that is, if ~ , for some positive operator , then is positive. Hence, an operator is positive if and only if it has a positive square root. If is positive, then j is self-adjoint and so ²j ³ i j ~ Conversely, if ~ i for some operator , then is positive, since it is clearly self-adjoint and º #Á #» ~ ºi #Á #» ~ º#Á #» Thus, is positive if and only if it has the form ~ i for some operator . (A complex number ' is nonnegative if and only if has the form ' ~ $$ for some complex number $.) Theorem 10.23 Let B²= ³. 1) is positive if and only if it has a positive square root. 2) is positive if and only if it has the form ~ i for some operator . Here is an application of square roots. Theorem 10.24 If and are positive operators and ~ , then is positive.

252

Advanced Linear Algebra

Proof. Since is a positive operator, it has a positive square root j , which is a polynomial in . A similar statement holds for . Therefore, since and commute, so do j and j. Hence, ²j j³ ~ ²j ³ ²j³ ~ Since j and j are self-adjoint and commute, their product is self-adjoint and so is positive.

The Polar Decomposition of an Operator It is well known that any nonzero complex number ' can be written in the polar form ' ~ , where is a positive number and is real. We can do the same for any nonzero linear operator on a finite-dimensional complex inner product space. Theorem 10.25 Let be a nonzero linear operator on a finite-dimensional complex inner product space = . 1) There exist a positive operator and a unitary operator for which ~ . Moreover, is unique and if is invertible, then is also unique. 2) Similarly, there exist a positive operator and a unitary operator for which ~ . Moreover, is unique and if is invertible, then is also unique. Proof. Let us suppose for a moment that ~ . Then i ~ ²³i ~ i i ~ c and so i ~ c ~ Also, if # = , then # ~ ²#³ These equations give us a clue as to how to define and . Let us define to be the unique positive square root of the positive operator i . Then )#) ~ º#Á #» ~ º #Á #» ~ º i #Á #» ~ ) #)

(10.2)

Define on im²³ by ²#³ ~ # for all # = . Equation (10.2) shows that % ~ & implies that % ~ & and so this definition of on im²³ is well-defined. Moreover, is an isometry on im²³, since (10.2) gives

Structure Theory for Normal Operators

253

) ²#³) ~ ) #) ~ )#) Thus, if 8 ~ ¸ Á Ã Á ¹ is an orthonormal basis for im²³, then 8 ~ ¸ Á Ã Á ¹ is an orthonormal basis for ²im²³³ ~ im² ³. Finally, we may extend both orthonormal bases to orthonormal bases for = and then extend the definition of to an isometry on = for which ~ . As for the uniqueness, we have seen that must satisfy ~ i and since has a unique positive square root, we deduce that is uniquely defined. Finally, if is invertible, then so is since ker²³ ker² ³. Hence, ~ c is uniquely determined by . Part 2) can be proved by applying the previous theorem to the map i , to get ~ ² i ³i ~ ²³i ~ c ~ where is unitary. We leave it as an exercise to show that any unitary operator has the form ~ , where is a self-adjoint operator. This gives the following corollary. Corollary 10.26 (Polar decomposition) Let be a nonzero linear operator on a finite-dimensional complex inner product space. Then there is a positive operator and a self-adjoint operator for which has the polar decomposition ~ Moreover, is unique and if is invertible, then is also unique. Normal operators can be characterized using the polar decomposition. Theorem 10.27 Let ~ be a polar decomposition of a nonzero linear operator . Then is normal if and only if ~ . Proof. Since i ~ c ~ and i ~ c ~ c we see that is normal if and only if c ~

254

Advanced Linear Algebra

or equivalently, ~ Now, is a polynomial in and is a polynomial in and so this holds if and only if ~ .

Exercises 1. 2.

Let B²< Á = ³. If is surjective, find a formula for the right inverse of in terms of i . If is injective, find a formula for a left inverse of in terms of i . Hint: Consider i and i . Let B²= ³ where = is a complex vector space and let ~

² b i ³ and ~ ² c i ³

Show that and are self-adjoint and that ~ b and i ~ c What can you say about the uniqueness of these representations of and i? 3. Prove that all of the roots of the characteristic polynomial of a skewHermitian matrix are pure imaginary. 4. Give an example of a normal operator that is neither self-adjoint nor unitary. 5. Prove that if ) #) ~ ) i ²#³) for all # = , where = is complex, then is normal. 6. Let be a normal operator on a complex finite-dimensional inner product space = or a self-adjoint operator on a real finite-dimensional inner product space. a) Show that i ~ ² ³, for some polynomial ²%³ d´%µ. b) Show that for any B²= ³, ~ implies i ~ i . In other words, i commutes with all operators that commute with . 7. Show that a linear operator on a finite-dimensional complex inner product space = is normal if and only if whenever : is an invariant subspace under , so is : . 8. Let = be a finite-dimensional inner product space and let be a normal operator on = . a) Prove that if is idempotent, then it is also self-adjoint. b) Prove that if is nilpotent, then ~ . c) Prove that if ~ , then is idempotent. 9. Show that if is a normal operator on a finite-dimensional complex inner product space, then the algebraic multiplicity is equal to the geometric multiplicity for all eigenvalues of . 10. Show that two orthogonal projections and are orthogonal to each other if and only if im²³ im²³.

Structure Theory for Normal Operators

255

11. Let be a normal operator and let be any operator on = . If the eigenspaces of are -invariant, show that and commute. 12. Prove that if and are normal operators on a finite-dimensional complex inner product space and if ~ for some operator then i ~ i . 13. Prove that if two normal d complex matrices are similar, then they are unitarily similar, that is, similar via a unitary matrix. 14. If is a unitary operator on a complex inner product space, show that there exists a self-adjoint operator for which ~ . 15. Show that a positive operator has a unique positive square root. 16. Prove that if has a square root, that is, if ~ , for some positive operator , then is positive. 17. Prove that if (that is, c is positive) and if is a positive operator that commutes with both and then . 18. Using the 89 factorization, prove the following result, known as the Cholsky decomposition. An invertible linear operator B²= ³ is positive if and only if it has the form ~ i where is upper triangularizable. Moreover, can be chosen with positive eigenvalues, in which case the factorization is unique. 19. Does every self-adjoint operator on a finite-dimensional real inner product space have a square root? 20. Let be a linear operator on d and let Á Ã Á be the eigenvalues of , each one written a number of times equal to its algebraic multiplicity. Show that ( ( tr² i ³

where tr is the trace. Show also that equality holds if and only if is normal. 21. If B²= ³ where = is a real inner product space, show that the Hilbert space adjoint satisfies ² i ³d ~ ² d ³i .

Part II—Topics

Chapter 11

Metric Vector Spaces: The Theory of Bilinear Forms

In this chapter, we study vector spaces over arbitrary fields that have a bilinear form defined on them. Unless otherwise mentioned, all vector spaces are assumed to be finitedimensional. The symbol - denotes an arbitrary field and - denotes a finite field of size .

Symmetric, Skew-Symmetric and Alternate Forms We begin with the basic definition. Definition Let = be a vector space over - . A mapping ºÁ »¢ = d = ¦ - is called a bilinear form if it is linear in each coordinate, that is, if º% b &Á '» ~ º%Á '» b º&Á '» and º'Á % b &» ~ º'Á %» b º'Á &» A bilinear form is 1) symmetric if º%Á &» ~ º&Á %» for all %Á & = . 2) skew-symmetric (or antisymmetric) if º%Á &» ~ cº&Á %» for all %Á & = .

260

Advanced Linear Algebra

3) alternate (or alternating) if º%Á %» ~ for all % = . A bilinear form that is either symmetric, skew-symmetric, or alternate is referred to as an inner product and a pair ²= Á ºÁ »³, where = is a vector space and ºÁ » is an inner product on = , is called a metric vector space or inner product space. As usual, we will refer to = as a metric vector space when the form is understood. 4) A metric vector space = with a symmetric form is called an orthogonal geometry over - . 5) A metric vector space = with an alternate form is called a symplectic geometry over - . The term symplectic, from the Greek for “intertwined,” was introduced in 1939 by the famous mathematician Hermann Weyl in his book The Classical Groups, as a substitute for the term complex. According to the dictionary, symplectic means “relating to or being an intergrowth of two different minerals.” An example is ophicalcite, which is marble spotted with green serpentine. Example 11.1 Minkowski space 44 is the four-dimensional real orthogonal geometry s with inner product defined by º Á » ~ º Á » ~ º3 Á 3 » ~ º4 Á 4 » ~ c º Á » ~ for £ where Á Ã Á 4 is the standard basis for s . As is traditional, when the inner product is understood, we will use the phrase “let = be a metric vector space.” The real inner products discussed in Chapter 9 are inner products in the present sense and have the additional property of being positive definite—a notion that does not even make sense if the base field is not ordered. Thus, a real inner product space is an orthogonal geometry. On the other hand, the complex inner products of Chapter 9, being sesquilinear, are not inner products in the present sense. For this reason, we use the term metric vector space in this chapter, rather than inner product space. If : is a vector subspace of a metric vector space = , then : inherits the metric structure from = . With this structure, we refer to : as a subspace of = . The concepts of being symmetric, skew-symmetric and alternate are not independent. However, their relationship depends on the characteristic of the base field - , as do many other properties of metric vector spaces. In fact, the

Metric Vector Spaces: The Theory of Bilinear Forms

261

next theorem tells us that we do not need to consider skew-symmetric forms per se, since skew-symmetry is always equivalent to either symmetry or alternateness. Theorem 11.1 Let = be a vector space over a field - . 1) If char²- ³ ~ , then alternate ¬ symmetric ¯ skew-symmetric 2) If char²- ³ £ , then alternate ¯ skew-symmetric Also, the only form that is both alternate and symmetric is the zero form: º%Á &» ~ for all %Á & = . Proof. First note that for an alternating form over any base field, ~ º% b &Á % b &» ~ º%Á &» b º&Á %» and so º%Á &» ~ cº&Á %» which shows that the form is skew-symmetric. Thus, alternate always implies skew-symmetric. If char²- ³ ~ , then c ~ and so the definitions of symmetric and skewsymmetric are equivalent, which proves 1). If char²- ³ £ and the form is skew-symmetric, then for any % = , we have º%Á %» ~ cº%Á %» or º%Á %» ~ , which implies that º%Á %» ~ . Hence, the form is alternate. Finally, if the form is alternate and symmetric, then it is also skew-symmetric and so º"Á #» ~ cº"Á #» for all "Á # = , that is, º"Á #» ~ for all "Á # = . Example 11.2 The standard inner product on = ²Á ³, defined by ²% Á Ã Á % ³ h ²& Á Ã Á & ³ ~ % & b Ä b % & is symmetric, but not alternate, since ²Á Á Ã Á ³ h ²Á Á Ã Á ³ ~ £

The Matrix of a Bilinear Form If 8 ~ ² Á Ã Á ³ is an ordered basis for a metric vector space = , then a bilinear form is completely determined by the d matrix of values 48 ~ ²Á ³ ~ ²º Á »³ This is referred to as the matrix of the form (or the matrix of = ) with respect to the ordered basis 8 . Moreover, any d matrix over - is the matrix of some bilinear form on = .

262

Advanced Linear Algebra

Note that if % ~ ' then 48 ´%µ8 ~

v º Á %» y Å w º Á %» z

and ´%µ!8 48 ~ º%Á » Ä º%Á » It follows that if & ~ , then ´%µ!8 48 ´&µ8 ~ º%Á » Ä º%Á »

v w

Å

y z

~ º%Á &»

and this uniquely defines the matrix 48 , that is, if ´%µ!8 (´&µ8 ~ º%Á &» for all %Á & = , then ( ~ 48 . A matrix is alternate if it is skew-symmetric and has 's on the main diagonal. Thus, we can say that a form is symmetric (skew-symmetric, alternate) if and only if the matrix 48 is symmetric (skew-symmetric, alternate). Now let us see how the matrix of a form behaves with respect to a change of basis. Let 9 ~ ² Á Ã Á ³ be an ordered basis for = . Recall from Chapter 2 that the change of basis matrix 49Á8 , whose th column is ´ µ8 , satisfies ´#µ8 ~ 49Á8 ´#µ9 Hence, º%Á &» ~ ´%µ!8 48 ´&µ8 ~ ²´%µ!9 49!Á8 ³48 ²49Á8 ´&µ9 ³ ~ ´%µ9! ²49!Á8 48 49Á8 ³´&µ9 and so 49 ~ 49!Á8 48 49Á8 This prompts the following definition. Definition Two matrices (Á ) C ²- ³ are congruent if there exists an invertible matrix 7 for which ( ~ 7 ! )7 The equivalence classes under congruence are called congruence classes.

Metric Vector Spaces: The Theory of Bilinear Forms

263

Thus, if two matrices represent the same bilinear form on = , they must be congruent. Conversely, if ) ~ 48 represents a bilinear form on = and ( ~ 7 ! )7 where 7 is invertible, then there is an ordered basis 9 for = for which 7 ~ 49 Á 8 and so ( ~ 49!Á8 48 49Á8 Thus, ( ~ 49 represents the same form with respect to 9. Theorem 11.2 Let 8 ~ ² Á Ã Á ³ be an ordered basis for an inner product space = , with matrix 48 ~ ²º Á »³ 1) The form can be recovered from the matrix by the formula º%Á &» ~ ´%µ!8 48 ´&µ8 2) If 9 ~ ² Á Ã Á ³ is also an ordered basis for = , then 49 ~ 49!Á8 48 49Á8 where 49Á8 is the change of basis matrix from 9 to 8 . 3) Two matrices ( and ) represent the same bilinear form on a vector space = if and only if they are congruent, in which case they represent the same set of bilinear forms on = . In view of the fact that congruent matrices have the same rank, we may define the rank of a bilinear form (or of = ) to be the rank of any matrix that represents that form.

The Discriminant of a Form If ( and ) are congruent matrices, then det²(³ ~ det²7 ! )7 ³ ~ det²7 ³ det²)³ and so det²(³ and det²)³ differ by a square factor. The discriminant " of a bilinear form is the set of determinants of all of the matrices that represent the form. Thus, if 8 is an ordered basis for = , then " ~ - det²48 ³ ~ ¸ det²48 ³ £ - ¹

264

Advanced Linear Algebra

Quadratic Forms There is a close link between symmetric bilinear forms on = and quadratic forms on = . Definition A quadratic form on a vector space = is a map 8¢ = ¦ - with the following properties: 1) For all - Á # = , 8²#³ ~ 8²#³ 2) The map º"Á #»8 ~ 8²" b #³ c 8²"³ c 8²#³ is a (symmetric) bilinear form. Thus, every quadratic form 8 on = defines a symmetric bilinear form º"Á #»8 on = . Conversely, if char²- ³ £ and if ºÁ » is a symmetric bilinear form on = , then the function 8²%³ ~

º%Á %»

is a quadratic form 8. Moreover, the bilinear form associated with 8 is the original bilinear form: º"Á #»8 ~ 8²" b #³ c 8²"³ c 8²#³ ~ º" b #Á " b #» c º"Á "» c º#Á #» ~ º"Á #» b º#Á "» ~ º"Á #» Thus, the maps ºÁ » ¦ 8 and 8 ¦ ºÁ »8 are inverses and so there is a one-to-one correspondence between symmetric bilinear forms on = and quadratic forms on = . Put another way, knowing the quadratic form is equivalent to knowing the corresponding bilinear form. Again assuming that char²- ³ £ , if 8 ~ ²# Á Ã Á # ³ is an ordered basis for an orthogonal geometry = and if the matrix of the symmetric form on = is 48 ~ ²Á ³, then for % ~ '% # , 8²%³ ~

º%Á %» ~ ´%µ!8 48 ´%µ8 ~ Á % % Á

and so 8²%³ is a homogeneous polynomial of degree 2 in the coordinates % . (The term “form” means homogeneous polynomial—hence the term quadratic form.)

Metric Vector Spaces: The Theory of Bilinear Forms

265

Orthogonality As we will see, not all metric vector spaces behave as nicely as real inner product spaces and this necessitates the introduction of a new set of terminology to cover various types of behavior. (The base field - is the culprit, of course.) The most striking differences stem from the possibility that º%Á %» ~ for a nonzero vector % = . The following terminology should be familiar. Definition Let = be a metric vector space. A vector % is orthogonal to a vector &, written % &, if º%Á &» ~ . A vector % = is orthogonal to a subset : of = , written % : , if º%Á » ~ for all : . A subset : of = is orthogonal to a subset ; of = , written : ; , if º Á !» ~ for all : and ! ; . The orthogonal complement ? of a subset ? of = is the subspace ? ~ ¸# = # ?¹

Note that regardless of whether the form is symmetric or alternate (and hence skew-symmetric), orthogonality is a symmetric relation, that is, % & implies & %. Indeed, this is precisely why we restrict attention to these two types of bilinear forms. There are two types of degenerate behaviors that a vector may possess: It may be orthogonal to itself or, worse yet, it may be orthogonal to every vector in = . With respect to the former, we have the following terminology. Definition Let = be a metric vector space. 1) A nonzero % = is isotropic (or null) if º%Á %» ~ ; otherwise it is nonisotropic. 2) = is isotropic if it contains at least one isotropic vector. Otherwise, = is nonisotropic (or anisotropic). 3) = is totally isotropic (that is, symplectic) if all vectors in = are isotropic. Note that if # is an isotropic vector, then so is # for all - . This can be expressed by saying that the set 0 of isotropic vectors in = is a cone in = . (A cone in = is a nonempty subset that is closed under scalar multiplication.) With respect to the more severe forms of degeneracy, we have the following terminology. Definition Let = be a metric vector space.

266

Advanced Linear Algebra

1) A vector # = is degenerate if # = . The set = of all degenerate vectors is called the radical of = and denoted by rad²= ³. Thus, rad²= ³ ~ = 2) = is nonsingular, or nondegenerate, if rad²= ³ ~ ¸¹. 3) = is singular, or degenerate, if rad²= ³ £ ¸¹. 4) = is totally singular, or totally degenerate, if rad²= ³ ~ = . Some of the above terminology is not entirely standard, so care should be exercised in reading the literature. Theorem 11.3 A metric vector space = is nonsingular if and only if all representing matrices 48 are nonsingular. A note of caution is in order. If : is a subspace of a metric vector space = , then rad²:³ denotes the set of vectors in : that are degenerate in : , that is, rad²:³ is the radical of : , as a metric vector space in its own right. However, : denotes the set of all vectors in = that are orthogonal to : . Thus, rad²:³ ~ : q : Note also that rad²:³ ~ : q : : q : ~ rad²: ³ and so if : is singular, then so is : . Example 11.3 Recall that = ²Á ³ is the set of all ordered -tuples whose components come from the finite field - . (See Example 11.2.) It is easy to see that the subspace : ~ ¸Á Á Á ¹ of = ²Á ³ has the property that : ~ : . Note also that = ²Á ³ is nonsingular and yet the subspace : is totally singular. The following result explains why we restrict attention to symmetric or alternate forms (which includes skew-symmetric forms). Theorem 11.4 Let = be a vector space with a bilinear form. The following are equivalent: 1) Orthogonality is a symmetric relation, that is, %& ¬ &% 2) The form on = is symmetric or alternate, that is, = is a metric vector space.

Metric Vector Spaces: The Theory of Bilinear Forms

267

Proof. It is clear that orthogonality is symmetric if the form is symmetric or alternate, since in the latter case, the form is also skew-symmetric. For the converse, assume that orthogonality is symmetric. For convenience, let % & mean that º%Á &» ~ º&Á %» and let % = mean that º%Á #» ~ º#Á %» for all # = . If % = for all % = , then = is orthogonal and we are done. So let us \ =. examine vectors % with the property that % We wish to show that \= %

¬

% is isotropic and ²% & ¬ % &³

(11.1)

Note that if the second conclusion holds, then since % %, it follows that % is \ = , there is a ' = for which isotropic. So suppose that % &. Since % º%Á '» £ º'Á %» and so % & if and only if º%Á &»²º%Á '» c º'Á %»³ ~ Now, º%Á &»²º%Á '» c º'Á %»³ ~ º%Á &»º%Á '» c º%Á &»º'Á %» ~ º&Á %»º%Á '» c º%Á &»º'Á %» ~ º%Á º&Á %»' c &º'Á %»» But reversing the coordinates in the last expression gives ºº&Á %»' c &º'Á %»Á %» ~ º&Á %»º'Á %» c º&Á %»º'Á %» ~ and so the symmetry of orthogonality implies that the last expression is and so we have proven (11.1). Let us assume that = is not orthogonal and show that all vectors in = are isotropic, whence = is symplectic. Since = is not orthogonal, there exist \ # and so " \ = and # \ = . Hence, the vectors " and # "Á # = for which " are isotropic and for all & = , &" &#

¬ ¬

&" &#

\ = are isotropic, let $ = . Then $ " and Since all vectors $ for which $ $ # and so $ " and $ #. Now write $ ~ ²$ c "³ b " where $ c " ", since " is isotropic. Since the sum of two orthogonal isotropic vectors is isotropic, it follows that $ is isotropic if $ c " is isotropic. But º$ b "Á #» ~ º"Á #» £ º#Á "» ~ º#Á $ b "»

268

Advanced Linear Algebra

\ = , which implies that $ b " is isotropic. Thus, $ is also and so ²$ b "³ isotropic and so all vectors in = are isotropic.

Orthogonal and Symplectic Geometries If a metric vector space is both orthogonal and symplectic, then the form is both symmetric and skew-symmetric and so º"Á #» ~ º#Á "» ~ cº"Á #» Therefore, when char²- ³ £ , = is orthogonal and symplectic if and only if = is totally degenerate. However, if char²- ³ ~ , then there are orthogonal symplectic geometries that are not totally degenerate. For example, let = ~ span²"Á #³ be a twodimensional vector space and define a form on = whose matrix is 4 ~>

?

Since 4 is both symmetric and alternate, so is the form.

Linear Functionals The Riesz representation theorem says that every linear functional on a finitedimensional real or complex inner product space = is represented by a Riesz vector 9 = , in the sense that ²#³ ~ º#Á 9 » for all # = . A similar result holds for nonsingular metric vector spaces. Let = be a metric vector space over - . Let % = and define the inner product map º h Á %»¢ = ¦ - by º h Á %»# ~ º#Á %» This is easily seen to be a linear functional and so we can define a linear map ¢ = ¦ = i by % ~ º h Á %» The bilinearity of the form ensures that is linear and the kernel of is ker² ³ ~ ¸% = º= Á %» ~ ¸¹¹ ~ = ~ rad²= ³ Hence, is injective (and therefore an isomorphism) if and only if = is nonsingular. Theorem 11.5 (The Riesz representation theorem) Let = be a finitedimensional nonsingular metric vector space. The map ¢ = ¦ = i defined by

Metric Vector Spaces: The Theory of Bilinear Forms

269

% ~ º h Á %» is an isomorphism from = to = i . It follows that for each = i there exists a unique vector % = for which # ~ º#Á %» for all # = . The requirement that = be nonsingular is necessary. As a simple example, if = is totally singular, then no nonzero linear functional could possibly be represented by an inner product. The Riesz representation theorem applies to nonsingular metric vector spaces. However, we can also achieve something useful for singular subspaces : of a nonsingular metric vector space. The reason is that any linear functional : i can be extended to a linear functional on = , where it has a Riesz vector, that is, # ~ º#Á 9 » ~ º h Á 9 »# Hence, also has this form, where its “Riesz vector” is an element of = , but is not necessarily in : . Theorem 11.6 (The Riesz representation theorem for subspaces) Let : be a subspace of a metric vector space = . If either = or : is nonsingular, the linear map ¢ = ¦ : i defined by % ~ º h Á %»O: is surjective and has kernel : . Hence, for any linear functional : i , there is a (not necessarily unique) vector % = for which ~ º Á %» for all : . Moreover, if : is nonsingular, then % can be taken from : , in which case it is unique.

Orthogonal Complements and Orthogonal Direct Sums Definition A metric vector space = is the orthogonal direct sum of the subspaces : and ; , written = ~: p; if = ~ : l ; and : ; . If : is a subspace of a real inner product space, the projection theorem says that the orthogonal complement : of : is a true vector space complement of : , that is, = ~ : p :

270

Advanced Linear Algebra

However, in general metric vector spaces, an orthogonal complement may not be a vector space complement. In fact, Example 11.3 shows that in some cases : ~ : . In other cases, for example, if # is degenerate, then º#» ~ = . However, as we will see, the orthogonal complement of : is a vector space complement if and only if either the sum is correct, = ~ : b : , or the intersection is correct, : q : ~ ¸¹. Note that the latter is equivalent to the nonsingularity of : . Many nice properties of orthogonality in real inner product spaces do carry over to nonsingular metric vector spaces. Moreover, the next result shows that the restriction to nonsingular spaces is not that severe. Theorem 11.7 Let = be a metric vector space. Then = ~ rad²= ³ p : where : is nonsingular and rad²= ³ is totally singular. Proof. If : is any vector space complement of rad²= ³, then rad²= ³ : and so = ~ rad²= ³ p : Also, : is nonsingular since rad²:³ rad²= ³. Here are some properties of orthogonality in nonsingular metric vector spaces. In particular, if either = or : is nonsingular, then the orthogonal complement of : always has the expected dimension, dim²: ³ ~ dim²= ³ c dim²:³ even if : is not well behaved with respect to its intersection with : . Theorem 11.8 Let : be a subspace of a finite-dimensional metric vector space =. 1) If either = or : is nonsingular, then dim²:³ b dim²: ³ ~ dim²= ³ Hence, the following are equivalent: a) = ~ : b : b) : is nonsingular, that is, : q : ~ ¸¹ c) = ~ : p : . 2) If = is nonsingular, then a) : ~ : b) rad²:³ ~ rad²: ³ c) : is nonsingular if and only if : is nonsingular. Proof. For part 1), the map ¢ = ¦ : i of Theorem 11.6 is surjective and has kernel : . Thus, the rank-plus-nullity theorem implies that

Metric Vector Spaces: The Theory of Bilinear Forms

271

dim²: i ³ b dim²: ³ ~ dim²= ³ However, dim²: i ³ ~ dim²:³ and so part 1) follows. For part 2), since rad²:³ ~ : q : : q : ~ rad²: ³ the nonsingularity of : implies the nonsingularity of : . Then part 1) implies that dim²:³ b dim²: ³ ~ dim²= ³ and dim²: ³ b dim²: ³ ~ dim²= ³ Hence, : ~ : and rad²:³ ~ rad²: ³. The previous theorem cannot in general be strengthened. Consider the twodimensional metric vector space = ~ span²"Á #³ where º"Á "» ~ Á º"Á #» ~ Á º#Á #» ~ If : ~ span²"³, then : ~ span²#³. Now, : is nonsingular but : is singular and so 2c) does not hold. Also, rad²:³ ~ ¸¹ and rad²: ³ ~ : and so 2b) fails. Finally, : ~ = £ : and so 2a) fails.

Isometries We now turn to a discussion of structure-preserving maps on metric vector spaces. Definition Let = and > be metric vector spaces. We use the same notation ºÁ » for the bilinear form on each space. A bijective linear map ¢ = ¦ > is called an isometry if º "Á #» ~ º"Á #» for all vectors " and # in = . If an isometry exists from = to > , we say that = and > are isometric and write = > . It is evident that the set of all isometries from = to = forms a group under composition. If = is a nonsingular orthogonal geometry, an isometry of = is called an orthogonal transformation. The set E²= ³ of all orthogonal transformations on = is a group under composition, known as the orthogonal group of = . If = is a nonsingular symplectic geometry, an isometry of = is called a symplectic transformation. The set Sp²= ³ of all symplectic transformations on = is a group under composition, known as the symplectic group of = .

272

Advanced Linear Algebra

Note that, in contrast to the case of real inner product spaces, we must include the requirement that be bijective since this does not follow automatically if = is singular. Here are a few of the basic properties of isometries. Theorem 11.9 Let B²= Á > ³ be a linear transformation between finitedimensional metric vector spaces = and > . 1) Let 8 ~ ¸# Á Ã Á # ¹ be a basis for = . Then is an isometry if and only if is bijective and º # Á # » ~ º# Á # » for all Á . 2) If = is orthogonal and char²- ³ £ , then is an isometry if and only if it is bijective and º #Á #» ~ º#Á #» for all # = . 3) Suppose that ¢ = > is an isometry and = ~ : p :

and > ~ ; p ;

If : ~ ; , then ²: ³ ~ ; . Proof. We prove part 3) only. To see that ²: ³ ~ ; , if ' : and ! ; , then since ; ~ : , we can write ! ~ for some : and so º 'Á !» ~ º 'Á » ~ º'Á » ~ whence ²: ³ ; . But since the dimensions are equal, it follows that ²: ³ ~ ; .

Hyperbolic Spaces A special type of two-dimensional metric vector space plays an important role in the structure theory of metric vector spaces. Definition Let = be a metric vector space. A hyperbolic pair is a pair of vectors "Á # = for which º"Á "» ~ º#Á #» ~ Á º"Á #» ~ Note that º#Á "» ~ if = is orthogonal and º#Á "» ~ c if = is symplectic. In either case, the subspace / ~ span²"Á #³ is called a hyperbolic plane and any space of the form > ~ / p Ä p / where each / is a hyperbolic plane, is called a hyperbolic space. If ²" Á # ³ is a hyperbolic pair for / , then we refer to the basis ²" Á # Á Ã Á " Á # ³

Metric Vector Spaces: The Theory of Bilinear Forms

273

for > as a hyperbolic basis. (In the symplectic case, the usual term is symplectic basis.) Note that any hyperbolic space > is nonsingular. In the orthogonal case, hyperbolic planes can be characterized by their degree of isotropy, so to speak. (In the symplectic case, all spaces are totally isotropic by definition.) Indeed, we leave it as an exercise to prove that a two-dimensional nonsingular orthogonal geometry = is a hyperbolic plane if and only if = contains exactly two one-dimensional totally isotropic (equivalently, totally degenerate) subspaces. Put another way, the cone of isotropic vectors is the union of two one-dimensional subspaces of = .

Nonsingular Completions of a Subspace Let < be a subspace of a nonsingular metric vector space = . If < is singular, it is of interest to find a minimal nonsingular subspace of = containing < . Definition Let = be a nonsingular metric vector space and let < be a subspace of = . A subspace : of = for which < : is called an extension of < . A nonsingular completion of < is an extension of < that is minimal in the family of all nonsingular extensions of < . Theorem 11.10 Let = be a nonsingular finite-dimensional metric vector space over - . We assume that char²- ³ £ when = is orthogonal. 1) Let : be a subspace of = . If # is isotropic and the orthogonal direct sum span²#³ p : exists, then there is a hyperbolic plane / ~ span²#Á '³ for which / p: exists. In particular, if # is isotropic, then there is a hyperbolic plane containing #. 2) Let < be a subspace of = and let < ~ span²# Á Ã Á # ³ p > where > is nonsingular and ¸# Á Ã Á # ¹ are linearly independent in rad²< ³. Then there is a hyperbolic space > ~ / p Ä p / with hyperbolic basis ²# Á ' Á Ã Á # Á ' ³ for which < ~ > p > is a nonsingular proper extension of < . If ¸# Á Ã Á # ¹ is a basis for rad²< ³, then dim²< ³ ~ dim²< ³ b dim²rad²< ³³

274

Advanced Linear Algebra

and we refer to < as a hyperbolic extension of < . If < is nonsingular, we say that < is a hyperbolic extension of itself. Proof. For part 1), the nonsingularity of = implies that : ~ : . Hence, # ¤ : ~ : and so there is an % : for which º#Á %» £ . If = is symplectic, then all vectors are isotropic and so we can take ' ~ ²°º#Á %»³%. If = is orthogonal, let ' ~ # b %. The conditions defining ²#Á '³ as a hyperbolic pair are (since # is isotropic) ~ º#Á '» ~ º#Á # b %» ~ º#Á %» and ~ º'Á '» ~ º# b %Á # b %» ~ º#Á %» b

º%Á %» ~ b

º%Á %»

Since º#Á %» £ , the first of these equations can be solved for and since char²- ³ £ , the second equation can then be solved for . Thus, in either case, there is a vector ' : for which / ~ span²#Á '³ : is hyperbolic. Hence, : : / and since / is nonsingular, that is, / q / ~ ¸¹, we have / q : ~ ¸¹ and so / p : exists. Part 2) is proved by induction on . Note first that all of the vectors # are isotropic. If ~ , then span²# ³ p > exists and so part 1) implies that there is a hyperbolic plane / ~ span²# Á '³ for which / p > exists. Assume that the result is true for independent sets of size less than . Since span²# ³ p span²# Á Ã Á # ³ p > exists, part 1) implies that there exists a hyperbolic plane / ~ span²# Á ' ³ for which / p span²# Á Ã Á # ³ p > exists. Since # Á Ã Á # are in the radical of span²# Á Ã Á # ³ p > , the inductive hypothesis implies that there is a hyperbolic space / p Ä p / with hyperbolic basis ²# Á ' Á Ã Á # Á ' ³ for which the orthogonal direct sum / p Ä p / p > exists. Hence, / p Ä p / p > also exists. We can now prove that the hyperbolic extensions of < are precisely the minimal nonsingular extensions of < . Theorem 11.11 (Nonsingular extension theorem) Let < be a subspace of a nonsingular finite-dimensional metric vector space = . The following are equivalent: 1) ; ~ > p > is a hyperbolic extension of < 2) ; is a minimal nonsingular extension of <

Metric Vector Spaces: The Theory of Bilinear Forms

275

3) ; is a nonsingular extension of < and dim²; ³ ~ dim²< ³ b dim²rad²< ³³ Thus, any two nonsingular completions of < are isometric. Proof. If < ? = where ? is nonsingular, then we may apply Theorem 11.10 to < as a subspace of ? , to obtain a hyperbolic extension A p > of < for which < Ap> ? Thus, every nonsingular extension of < contains a hyperbolic extension of < . Moreover, all hyperbolic extensions of < have the same dimension: dim²> p > ³ ~ dim²< ³ b dim²rad²< ³³ and so no hyperbolic extension of < is properly contained in another hyperbolic extension of < . This proves that 1)–3) are equivalent. The final statement follows from the fact that hyperbolic spaces of the same dimension are isometric.

Extending Isometries to Nonsingular Completions Let = and = Z be isometric nonsingular metric vector spaces and let < ~ rad²< ³ p > be a subspace of = , with nonsingular completion < ~ > p >. If ¢ < ¦ < is an isometry, then it is a simple matter to extend to an isometry from < onto a nonsingular completion of < . To see this, let ²" Á ' Á Ã Á " Á ' ³ be a hyperbolic basis for >. Since ²" Á Ã Á " ³ is a basis for rad²< ³, it follows that ² " Á Ã Á " ³ is a basis for rad² < ³. Hence, we can hyperbolically extend < ~ rad² > ³ p > to get < ~ >Z p > where >Z has hyperbolic basis ² " Á % Á Ã Á " Á % ³. To extend , simply set ' ~ % for all ~ Á Ã Á . Theorem 11.12 Let = and = Z be isometric nonsingular metric vector spaces and let < be a subspace of = , with nonsingular completion < . Any isometry ¢ < ¦ < can be extended to an isometry from < onto a nonsingular completion of < .

The Witt Theorems: A Preview There are two important theorems that are quite easy to prove in the case of real inner product spaces, but require more work in the case of metric vector spaces in general. Let = and = Z be isometric nonsingular metric vector spaces over a field - . We assume that char²- ³ £ if = is orthogonal.

276

Advanced Linear Algebra

The Witt extension theorem says that if : is a subspace of = , then any isometry ¢ : ¦ : = Z can be extended to an isometry from = to = Z . The Witt cancellation theorem says that if = ~ : p :

and = Z ~ ; p ;

then : ; ¬ : ; We will prove these theorems in both the orthogonal and symplectic cases a bit later in the chapter. For now, we simply want to show that it is easy to prove one Witt theorem using the other. Suppose that the Witt extension theorem holds and assume that = ~ : p :

and = Z ~ ; p ;

and : ; . Then any isometry ¢ : ¦ ; can be extended to an isometry from = to = Z . According to Theorem 11.9, we have ²: ³ ~ ; and so : ; . Hence, the Witt cancellation theorem holds. Conversely, suppose that the Witt cancellation theorem holds and let ¢ : ¦ : = Z be an isometry. Since can be extended to a nonsingular completion of : , we may assume that : is nonsingular. Then = ~ : p : Since is an isometry, : is also nonsingular and we can write = Z ~ : p ² :³ Since : : , Witt's cancellation theorem implies that : ² :³ . If ¢ : ¦ ² :³ is an isometry, then the map ¢ = ¦ = Z defined by ²" b #³ ~ " b # for " : and # : is an isometry that extends . Hence Witt's extension theorem holds.

The Classification Problem for Metric Vector Spaces The classification problem for a class of metric vector spaces (such as the orthogonal or symplectic spaces) is the problem of determining when two metric vector spaces in the class are isometric. The classification problem is considered “solved,” at least in a theoretical sense, by finding a set of canonical forms or a complete set of invariants for matrices under congruence.

Metric Vector Spaces: The Theory of Bilinear Forms

277

To see why, suppose that ¢ = ¦ > is an isometry and 8 ~ ²# Á Ã Á # ³ is an ordered basis for = . Then 9 ~ ² # Á Ã Á # ³ is an ordered basis for > and 48 ²= ³ ~ ²º# Á # »³ ~ ²º # Á # »³ ~ 49 ²> ³ Thus, the congruence class of matrices representing = is identical to the congruence class of matrices representing > . Conversely, suppose that = and > are metric vector spaces with the same congruence class of representing matrices. Then if 8 ~ ²# Á Ã Á # ³ is an ordered basis for = , there is an ordered basis 9 ~ ²$ Á Ã Á $ ³ for > for which ²º# Á # »³ ~ 48 ²= ³ ~ 49 ²> ³ ~ ²º$ Á $ »³ Hence, the map ¢ = ¦ > defined by # ~ $ is an isometry from = to > . We have shown that two metric vector spaces are isometric if and only if they have the same congruence class of representing matrices. Thus, we can determine whether any two metric vector spaces are isometric by representing each space with a matrix and determining whether these matrices are congruent, using a set of canonical forms or a set of complete invariants.

Symplectic Geometry We now turn to a study of the structure of orthogonal and symplectic geometries and their isometries. Since the study of the structure (and the structure itself) of symplectic geometries is simpler than that of orthogonal geometries, we begin with the symplectic case. The reader who is interested only in the orthogonal case may omit this section. Throughout this section, let = be a nonsingular symplectic geometry.

The Classification of Symplectic Geometries Among the simplest types of metric vector spaces are those that possess an orthogonal basis. However, it is easy to see that a symplectic geometry = has an orthogonal basis if and only if it is totally degenerate and so no “interesting” symplectic geometries have orthogonal bases. Thus, in searching for an orthogonal decomposition of = , we turn to twodimensional subspaces and this puts us in mind of hyperbolic spaces. Let < be the family of all hyperbolic subspaces of = , which is nonempty since the zero subspace ¸¹ is singular and so has a nonzero hyperbolic extension. Since = is finite-dimensional, < has a maximal member >. Since > is nonsingular, if > £ = , then = ~ > p > where > £ ¸¹. But then if # > is nonzero, there is a hyperbolic extension

278

Advanced Linear Algebra

/ p > of > containing #, which contradicts the maximality of >. Hence, = ~ >. This proves the following structure theorem for symplectic geometries. Theorem 11.13 1) A symplectic geometry has an orthogonal basis if and only if it is totally degenerate. 2) Any nonsingular symplectic geometry = is a hyperbolic space, that is, = ~ / p / p Ä p / where each / is a hyperbolic plane. Thus, there is a hyperbolic basis for = , that is, a basis 8 for which the matrix of the form is

@

v x c x x x ~x x x x

c

Æ c

w

y { { { { { { { {

z

In particular, the dimension of = is even. 3) Any symplectic geometry = has the form = ~ rad²= ³ p > where > is a hyperbolic space and rad²= ³ is a totally degenerate space. The rank of the form is dim²>³ and = is uniquely determined up to isometry by its rank and its dimension. Put another way, up to isometry, there is precisely one symplectic geometry of each rank and dimension. Symplectic forms are represented by alternate matrices, that is, skew-symmetric matrices with zero diagonal. Moreover, according to Theorem 11.13, each d alternate matrix is congruent to a matrix of the form @ ?Ác ~ >

c ?block

Since the rank of ?Ác is , no two such matrices are congruent. Theorem 11.14 The set of d matrices of the form ?Ác is a set of canonical forms for alternate matrices under congruence. The previous theorems solve the classification problem for symplectic geometries by stating that the rank and dimension of = form a complete set of

Metric Vector Spaces: The Theory of Bilinear Forms

279

invariants under congruence and that the set of all matrices of the form ?Ác is a set of canonical forms.

Witt's Extension and Cancellation Theorems We now prove the Witt theorems for symplectic geometries. Theorem 11.15 (Witt's extension theorem) Let = and = Z be isometric nonsingular symplectic geometries over a field - . Then any isometry ¢ : ¦ : = Z on a subspace : of = can be extended to an isometry from = to = Z . Proof. According to Theorem 11.12, we can extend to a nonsingular completion of : , so we may simply assume that : and : are nonsingular. Hence, = ~ : p : and = Z ~ : p ² :³ To complete the extension of to = , we need only choose a hyperbolic basis ² Á Á Ã Á Á ³ for : and a hyperbolic basis ²Z Á Z Á Ã Á Z Á Z ³ for ² :³ and define the extension by setting ~ Z and ~ Z . As a corollary to Witt's extension theorem, we have Witt's cancellation theorem. Theorem 11.16 (Witt's cancellation theorem) Let = and = Z be isometric nonsingular symplectic geometries over a field - . If = ~ : p :

and = Z ~ ; p ;

then : ; ¬ : ;

The Structure of the Symplectic Group: Symplectic Transvections Let us examine the nature of symplectic transformations (isometries) on a nonsingular symplectic geometry = . Recall that for a real vector space, an isometric isomorphism, which corresponds to an isometry in the present context, is the same as an orthogonal map and orthogonal maps are products of reflections (Theorem 10.17). Recall also that a reflection /# is defined as an operator for which

280

Advanced Linear Algebra

/# # ~ c#Á /# $ ~ $ for all $ º#» and that /# % ~ % c

º%Á #» # º#Á #»

In the present context, we do not dare divide by º#Á #», since all vectors are isotropic. So here is the next-best thing. Definition Let = be a nonsingular symplectic geometry over - . Let # = be nonzero and let - . The map #Á ¢ = ¦ = defined by #Á ²%³ ~ % b º%Á #»# is called the symplectic transvection determined by # and . Note that if ~ , then #Á ~ and if £ , then #Á is the identity precisely on the subspace span²#³ of codimension . In the case of a reflection, /# is the identity precisely on span²#³ and = ~ span²#³ p span²#³ However, for a symplectic transvection, #Á is the identity precisely on span²#³ (for £ ) but span²#³ span²#³ . Here are the basic properties of symplectic transvections. Theorem 11.17 Let #Á be a symplectic transvection on = . Then 1) #Á is a symplectic transformation (isometry). 2) #Á ~ if and only if ~ . 3) If % #, then #Á ²%³ ~ %. For £ , % # if and only if #Á ²%³ ~ %. 4) #Á #Á ~ #Áb . c 5) #Á ~ #Ác . 6) For any symplectic transformation , #Á c ~ #Á 7) For - i , #Á ~ #Á

Note that if < is a subspace of = and if "Á is a symplectic transvection on < , then, by definition, " < . However, the formula "Á ²%³ ~ % b º%Á "»" also defines a symplectic transvection on = , where % ranges over = . Moreover, for any ' < , we have "Á ' ~ ' and so "Á is the identity on < .

Metric Vector Spaces: The Theory of Bilinear Forms

281

We now wish to prove that any symplectic transformation on a nonsingular symplectic geometry = is the product of symplectic transvections. The proof is not difficult, but it is a bit lengthy, so we break it up into parts. Our first goal is to show that we can get from any hyperbolic pair to any other hyperbolic pair using a product of symplectic transvections. Let us say that two hyperbolic pairs ²%Á &³ and ²$Á '³ are connected if there is a product of symplectic transvections that carries % to $ and & to ' and write ¢ ²%Á &³ ª ²$Á '³ or ²%Á &³ © ²$Á '³. It is clear that connectedness is an equivalence relation on the set of hyperbolic pairs. Theorem 11.18 In a nonsingular symplectic geometry = , every pair of hyperbolic pairs are connected. Proof. Note first that if º Á !» £ , then £ ! and so

~ b º Á c !»² c !³ ~ c º Á !»² c !³

c!Á

Taking ~ °º Á !» gives c!Á ~ !. Therefore, if ² Á "³ is hyperbolic, then we can always find a vector % for which ² Á "³ © ²!Á %³ namely, % ~

c!Á ".

Also, if both ² Á "³ and ²!Á "³ are hyperbolic, then ² Á "³ © ²!Á "³

since º c !Á "» ~ and so

c!Á "

~ ".

Actually, these statements are still true if º Á !» ~ . For in this case, there is a nonzero vector & for which º Á &» £ and º!Á &» £ . This follows from the fact that there is an = i for which £ and ! £ and so the Riesz vector 9 is such a vector. Therefore, if ² Á "³ is hyperbolic, then ² Á "³ © ²&Á

c&Á "³

© ²!Á &c!ÁZ

c&Á "³

and if both ² Á "³ and ²!Á "³ are hyperbolic, then ² Á "³ © ²&Á "³ © ²!Á "³ Hence, transitivity gives the same result as in the case º Á !» £ . Finally, if ²" Á " ³ and ²# Á # ³ are hyperbolic, then there is a & for which ²" Á " ³ © ²# Á &³ © ²# Á # ³ and so transitivity shows that ²" Á " ³ © ²# Á # ³.

282

Advanced Linear Algebra

We can now show that the symplectic transvections generate the symplectic group. Theorem 11.19 Every symplectic transformation on a nonsingular symplectic geometry = is the product of symplectic transvections. Proof. Let be a symplectic transformation on = . We proceed by induction on ~ dim²= ³. If ~ , then = ~ / ~ span²"Á '³ is a hyperbolic plane and Theorem 11.18 implies that there is a product of symplectic transvections on = for which ¢ ²"Á '³ ª ²"Á '³ This proves the result if ~ . Assume that the result holds for all dimensions less than and let dim²= ³ ~ . Now, = ~/ pA where / ~ span²"Á '³ and A is a symplectic geometry of dimension less than that of = . As before, there is a product of symplectic transvections on = for which ¢ ²"Á '³ ª ²"Á '³ and so O/ ~ O/ Note that c / ~ / and so Theorem 11.9 implies that c ²/ ³ ~ / . Since dim²/ ³ dim²/³, the inductive hypothesis applied to the symplectic transformation c on / implies that there is a product of symplectic transvections on / for which ~ c . As remarked earlier, is also a product of symplectic transvections on = that is the identity on / and so O/ ~ /

and ~ on /

Thus, ~ on both / and on / and so ~ is a product of symplectic transvections on = .

The Structure of Orthogonal Geometries: Orthogonal Bases We have seen that no interesting (that is, not totally degenerate) symplectic geometries have orthogonal bases. By contrast, almost all interesting orthogonal geometries = have orthogonal bases. To understand why, it is convenient to group the orthogonal geometries into two classes: those that are also symplectic and those that are not. The reason is that all orthogonal nonsymplectic geometries have orthogonal bases, as we will see. However, an orthogonal symplectic geometry has an orthogonal basis if and

Metric Vector Spaces: The Theory of Bilinear Forms

283

only if it is totally degenerate. Furthermore, we have seen that if char²- ³ £ , then all orthogonal symplectic geometries are totally degenerate and so all such geometries have orthogonal bases. But if char²- ³ ~ , then there are orthogonal symplectic geometries that are not totally degenerate and therefore do not have orthogonal bases. Thus, if we exclude orthogonal symplectic geometries when char²- ³ ~ , we can say that every orthogonal geometry has an orthogonal basis. If a metric vector space = has an orthogonal basis, the natural next step is to look for an orthonormal basis. However, if = is singular, then there is a nonzero vector # = and such a vector can never be a linear combination of vectors from an orthonormal basis ¸" Á Ã Á " ¹, since the coefficients in such a linear combination are º#Á " » ~ . However, even if = is nonsingular, orthonormal bases do not always exist and the question of how close we can come to such an orthonormal basis depends on the nature of the base field. We will examine this issue in three cases: algebraically closed fields, the field of real numbers and finite fields. We should also mention that even when = has an orthogonal basis, the Gram– Schmidt orthogonalization process may not apply to produce such a basis, because even nonsingular orthogonal geometries may have isotropic vectors, and so division by º"Á "» is problematic. For example, consider an orthogonal hyperbolic plane / ~ span²"Á #³ and assume that char²- ³ £ . Thus, " and # are isotropic and º"Á #» ~ . The vector " cannot be extended to an orthogonal basis using the Gram–Schmidt process, since ¸"Á " b #¹ is orthogonal if and only if ~ . However, / does have an orthogonal basis, namely, ¸" b #Á " c #¹.

Orthogonal Bases Let = be an orthogonal geometry. As we have discussed, if = is also symplectic, then = has an orthogonal basis if and only if it is totally degenerate. Moreover, when char²- ³ £ , all orthogonal symplectic geometries are totally degenerate and so all orthogonal symplectic geometries have an orthogonal basis. If = is orthogonal but not symplectic, then = contains a nonisotropic vector " , the subspace span²" ³ is nonsingular and = ~ span²" ³ p = where = ~ span²" ³ . If = is not symplectic, then we may decompose it to get = ~ span²" ³ p span²" ³ p =

284

Advanced Linear Algebra

This process may be continued until we reach a decomposition = ~ span²" ³ p Ä p span²" ³ p < where < is symplectic as well as orthogonal. (This includes the case < ~ ¸¹.) Let 8 ~ ²" Á Ã Á " ³. If char²- ³ £ , then < is totally degenerate. Thus, if 9 is a basis for < , then the union 8 r 9 is an orthogonal basis for = . If char²- ³ ~ , then < ~ > p rad²< ³, where > is hyperbolic and so = ~ span²" ³ p Ä p span²" ³ p > p rad²< ³ where rad²< ³ is totally degenerate and the " are nonisotropic. If 9 ~ ²% Á & Á Ã Á % Á & ³ is a hyperbolic basis for > and : ~ ²' Á Ã Á ' ³ is an ordered basis for rad²< ³, then the union ; ~ 8 r 9 r : ~ ²" Á Ã Á " Á % Á & Á Ã Á % Á & Á ' Á Ã Á ' ³ is an ordered orthogonal basis for = . However, we can do better (in some sense). The following lemma says that when char²- ³ ~ , a pair of isotropic basis vectors, such as % Á & , can be replaced by a pair of nonisotropic basis vectors, when coupled with a nonisotropic basis vector, such as " . Lemma 11.20 Suppose that char²- ³ ~ . Let > be a three-dimensional orthogonal geometry of the form > ~ span²#³ p span²%Á &³ where # is nonisotropic and / ~ span²%Á &³ is a hyperbolic plane. Then > ~ span²# ³ p span²# ³ p span²# ³ where each # is nonisotropic. Proof. It is straightforward to check that if º#Á #» ~ , then the vectors # ~ " b % b & # ~ " b % # ~ " b ² c ³% b & are linearly independent and mutually orthogonal. Details are left to the reader. Using the previous lemma, we can replace the vectors ¸" Á % Á & ¹ with the nonisotropic vectors ¸# Á #b Á #b ¹, while retaining orthogonality. Moreover, the replacement process can continue until the isotropic vectors are absorbed, leaving an orthogonal basis of nonisotropic vectors.

Metric Vector Spaces: The Theory of Bilinear Forms

285

Let us summarize. Theorem 11.21 Let = be an orthogonal geometry. 1) If = is also symplectic, then = has an orthogonal basis if and only if it is totally degenerate. When char²- ³ £ , all orthogonal symplectic geometries have an orthogonal basis, but this is not the case when char²- ³ ~ . 2) If = is not symplectic, then = has an ordered orthogonal basis 8 ~ ²" Á Ã Á " Á ' Á Ã Á ' ³ for which º" Á " » ~ £ and º' Á '» ~ . Hence, 48 has the diagonal form v x x x 48 ~ x x x

y { { { { { {

Æ Æ

w

z

with ~ rk²48 ³ nonzero entries on the diagonal. As a corollary, we get a nice theorem about symmetric matrices. Corollary 11.22 Let 4 be a symmetric matrix and assume that 4 is not alternate if char²- ³ ~ . Then 4 is congruent to a diagonal matrix.

The Classification of Orthogonal Geometries: Canonical Forms We now want to consider the question of improving upon Theorem 11.21. The diagonal matrices of this theorem do not form a set of canonical forms for congruence. In fact, if Á Ã Á are nonzero scalars, then the matrix of = with respect to the basis 9 ~ ² " Á Ã Á " Á ' Á Ã Á ' ³ is

v x x x 49 ~ x x x w

Æ Æ

y { { { { { {

(11.2)

z

Hence, 48 and 49 are congruent diagonal matrices. Thus, by a simple change of basis, we can multiply any diagonal entry by a nonzero square in - . The determination of a set of canonical forms for symmetric (nonalternate when char²- ³ ~ ) matrices under congruence depends on the properties of the base field. Our plan is to consider three types of base fields: algebraically closed

286

Advanced Linear Algebra

fields, the real field s and finite fields. Here is a preview of the forthcoming results. 1) When the base field - is algebraically closed, there is an ordered basis 8 for which

48 ~ AÁ

v x x x ~x x x

y { { { { { {

Æ Æ

w

z

If = is nonsingular, then 48 is an identity matrix and = has an orthonormal basis. 2) Over the real base field, there is an ordered basis 8 for which

48 ~ ZÁÁ

v x x x x x x ~x x x x x x

Æ c Æ c Æ

w

y { { { { { { { { { { { { z

3) If - is a finite field, there is an ordered basis 8 for which v x x x x 48 ~ ZÁ ²³ ~ x x x x w

Æ Æ

y { { { { { { { { z

where is unique up to multiplication by a square and if char²- ³ ~ , then we can take ~ . Now let us turn to the details.

Algebraically Closed Fields If - is algebraically closed, then for every - , the polynomial % c has a root in - , that is, every element of - has a square root in - . Therefore, we may choose ~ °j in (11.2), which leads to the following result.

Metric Vector Spaces: The Theory of Bilinear Forms

287

Theorem 11.23 Let = be an orthogonal geometry over an algebraically closed field - . Provided that = is not symplectic as well when char²- ³ ~ , then = has an ordered orthogonal basis 8 ~ ²" Á Ã Á " Á ' Á Ã Á ' ³ for which º" Á " » ~ and º' Á ' » ~ . Hence, 48 has the diagonal form

48 ~ AÁ

v x x x ~x x x

Æ

w

Æ

y { { { { { { z

with ones and zeros on the diagonal. In particular, if = is nonsingular, then = has an orthonormal basis. The matrix version of Theorem 11.23 follows. Theorem 11.24 Let I be the set of all d symmetric matrices over an algebraically closed field - . If char²- ³ ~ , we restrict I to the set of all symmetric matrices with at least one nonzero entry on the main diagonal. 1) Any matrix 4 in I is congruent to a unique matrix of the form ZÁ , in fact, ~ rk²4 ³ and ~ c rk²4 ³. 2) The set of all matrices of the form ZÁ for b ~ is a set of canonical forms for congruence on I . 3) The rank of a matrix is a complete invariant for congruence on I .

The Real Field s If - ~ s, we can choose ~ °j( (, so that all nonzero diagonal elements in (11.2) will be either , or c. Theorem 11.25 (Sylvester's law of inertia) Any real orthogonal geometry = has an ordered orthogonal basis 8 ~ ²" Á Ã Á " Á # Á Ã Á # Á ' Á Ã Á ' ³ for which º" Á " » ~ , º# Á # » ~ c and º' Á ' » ~ . Hence, the matrix 48 has the diagonal form

288

Advanced Linear Algebra

48 ~ ZÁÁ

v x x x x x x ~x x x x x x

Æ c Æ c Æ

w

y { { { { { { { { { { { { z

with ones, negative ones and zeros on the diagonal. Here is the matrix version of Theorem 11.25. Theorem 11.26 Let I be the set of all d symmetric matrices over the real field s. 1) Any matrix in I is congruent to a unique matrix of the form ZÁÁ Á for some Á and ~ c c . 2) The set of all matrices of the form ZÁÁ for b b ~ is a set of canonical forms for congruence on I . 3) Let 4 I and let 4 be congruent to AÁÁ . Then b is the rank of 4 and c is the signature of 4 and the triple ²Á Á ³ is the inertia of 4 . The pair ²Á ³, or equivalently the pair ² b Á c ³, is a complete invariant under congruence on I . Proof. We need only prove the uniqueness statement in part 1). Let 8 ~ ²" Á Ã Á " Á # Á Ã Á # Á ' Á Ã Á ' ³ and 9 ~ ²"Z Á Ã Á "ZZ Á #Z Á Ã Á #ZZ 'Z Á Ã Á 'Z Z ³ be ordered bases for which the matrices 48 and 49 have the form shown in Theorem 11.25. Since the rank of these matrices must be equal, we have b ~ Z b Z and so ~ Z . If % span²" Á Ã Á " ³ and % £ , then º%Á %» ~ L " Á " M ~ º" Á " » ~ Á ~ Á

Á

On the other hand, if & span²#Z Á Ã Á #ZZ ³ and & £ , then º&Á &» ~ L #Z Á #Z M ~ Á

Z Z º# Á # »

~ c

Á

~ c

Á

Hence, if & span²#Z Á Ã Á #ZZ Á 'Z Á Ã Á 'Z Z ³ then º&Á &» . It follows that

Metric Vector Spaces: The Theory of Bilinear Forms

289

span²" Á Ã Á " ³ q span²#Z Á Ã Á #ZZ Á 'Z Á Ã Á 'Z Z ³ ~ ¸¹ and so b ² c Z ³ that is, Z . By symmetry, Z and so ~ Z . Finally, since ~ Z , it follows that ~ Z .

Finite Fields To deal with the case of finite fields, we must know something about the distribution of squares in finite fields, as well as the possible values of the scalars º#Á #». Theorem 11.27 Let - be a finite field with elements. 1) If char²- ³ ~ , then every element of - is a square. 2) If char²- ³ £ , then exactly half of the nonzero elements of - are squares, that is, there are ² c ³° nonzero squares in - . Moreover, if % is any nonsquare in - , then all nonsquares have the form %, for some - . Proof. Write - ~ - , let - i be the subgroup of all nonzero elements in - and let ²- i ³ ~ ¸ - i ¹ be the subgroup of all nonzero squares in - . The Frobenius map ¢ - i ¦ ²- i ³ defined by ²³ ~ is a surjective group homomorphism, with kernel ker²³ ~ ¸ - ~ ¹ ~ ¸cÁ ¹ If char²- ³ ~ , then ker²³ ~ ¸¹ and so is bijective and (- i ( ~ (²- i ³ (, which proves part 1). If char²- ³ £ , then (ker²³( ~ and so (- i ( ~ (²- i ³ (, which proves the first part of part 2). We leave proof of the last statement to the reader. Definition A bilinear form on = is universal if for any nonzero - there exists a vector # = for which º#Á #» ~ . Theorem 11.28 Let = be an orthogonal geometry over a finite field - with char²- ³ £ and assume that = has a nonsingular subspace of dimension at least . Then the bilinear form of = is universal. Proof. Theorem 11.21 implies that = contains two linearly independent vectors " and # for which º"Á "» ~ £ Á º#Á #» ~ £ Á º"Á #» ~

290

Advanced Linear Algebra

Given any - , we want to find and for which ~ º" b #Á " b #» ~ b or ~ c If ( ~ ¸ - ¹, then ((( ~ ² b ³°, since there are ² c ³° nonzero squares , along with ~ . If ) ~ ¸ c - ¹, then for the same reasons () ( ~ ² b ³°. It follows that ( q ) cannot be the empty set and so there exist and for which ~ c . Now we can proceed with the business at hand. Theorem 11.29 Let = be an orthogonal geometry over a finite field - and assume that = is not symplectic if char²- ³ ~ . If char²- ³ £ , then let be a fixed nonsquare in - . For any nonzero - , write v x x x x ? ²³ ~ x x x x

y { { { { { { { {

Æ Æ

w

z

where rk²? ²³³ ~ . 1) If char²- ³ ~ , then there is an ordered basis 8 for which 48 ~ ? ²³. 2) If char²- ³ £ , then there is an ordered basis 8 for which 48 equals ? ²³ or ? ²³. Proof. We can dispose of the case char²- ³ ~ quite easily: Referring to (11.2), since every element of - has a square root, we may take ~ ²j ³c . If char²- ³ £ , then Theorem 11.21 implies that there is an ordered orthogonal basis 8 ~ ²" Á Ã Á " Á ' Á Ã Á ' ³ for which º" Á " » ~ £ and º' Á ' » ~ . Hence, 48 has the diagonal form v x x x 48 ~ x x x w

Æ Æ

y { { { { { { z

Metric Vector Spaces: The Theory of Bilinear Forms

291

Now consider the nonsingular orthogonal geometry = ~ span²" Á " ³. According to Theorem 11.28, the form is universal when restricted to = . Hence, there exists a # = for which º# Á # » ~ . Now, # ~ " b " for Á - not both , and we may swap " and " if necessary to ensure that £ . Hence, 8 ~ ²# Á " Á Ã Á " Á ' Á Ã Á ' ³ is an ordered basis for = for which the matrix 48 is diagonal and has a in the upper left entry. We can repeat the process with the subspace = ~ span²# Á # ³. Continuing in this way, we can find an ordered basis 9 ~ ²# Á # Á Ã Á # Á ' Á Ã Á ' ³ for which 49 ~ ? ²³ for some nonzero - . Now, if is a square in - , then we can replace # by ²°j³# to get a basis : for which 4: ~ ? ²³. If is not a square in - , then ~ for some - and so replacing # by ²°³# gives a basis : for which 4: ~ ? ²³. Theorem 11.30 Let I be the set of all d symmetric matrices over a finite field - . If char²- ³ ~ , we restrict I to the set of all symmetric matrices with at least one nonzero entry on the main diagonal. 1) If char²- ³ ~ , then any matrix in I is congruent to a unique matrix of the form ? ²³ and the matrices ¸? ²³ ~ Á Ã Á ¹ form a set of canonical forms for I under congruence. Also, the rank is a complete invariant. 2) If char²- ³ £ , let be a fixed nonsquare in - . Then any matrix I is congruent to a unique matrix of the form ? ²³ or ? ²³. The set ¸? ²³Á ? ²³ ~ Á Ã Á ¹ is a set of canonical forms for congruence on I . (Thus, there are exactly two congruence classes for each rank .)

The Orthogonal Group Having “settled” the classification question for orthogonal geometries over certain types of fields, let us turn to a discussion of the structure-preserving maps, that is, the isometries.

Rotations and Reflections We begin by examining the matrix of an orthogonal transformation. If 8 is an ordered basis for = , then for any %Á & = , º%Á &» ~ ´%µ!8 48 ´&µ8 and so if B²= ³, then º %Á &» ~ ´ %µ!8 48 ´ &µ8 ~ ´%µ!8 ²´ µ!8 48 ´ µ8 ³´&µ8 Hence, is an orthogonal transformation if and only if

292

Advanced Linear Algebra

´ µ!8 48 ´ µ8 ~ 48 Taking determinants gives det²48 ³ ~ det²´ µ8 ³ det²48 ³ Therefore, if = is nonsingular, then det²´ µ8 ³ ~ f Since the determinant is an invariant under similarity, we have the following theorem. Theorem 11.31 Let be an orthogonal transformation on a nonsingular orthogonal geometry = . 1) det²´ µ8 ³ is the same for all ordered bases 8 for = and det²´ µ8 ³ ~ f This determinant is called the determinant of and denoted by det² ³. 2) If det² ³ ~ , then is called a rotation and if det² ³ ~ c, then is called a reflection. 3) The set Eb ²= ³ of rotations is a subgroup of the orthogonal group E²= ³ and the determinant map det¢ E²= ³ ¦ ¸cÁ ¹ is an epimorphism with kernel Eb ²= ³. Hence, if char²- ³ £ , then Eb ²= ³ is a normal subgroup of E²= ³ of index .

Symmetries Recall again that for a real inner product space, a reflection /" is defined as an operator for which /" " ~ c"Á /" $ ~ $ for all $ º"» and that /" % ~ % c

º%Á "» " º"Á "»

In particular, if char²- ³ £ and " = is nonisotropic, then span²"³ is nonsingular and so = ~ span²"³ p span²"³ Then the reflection /" is well-defined and, in the context of general orthogonal geometries, is called the symmetry determined by " and we will denote it by " . We can also write " ~ c p , that is, " ²% b &³ ~ c% b & for all % span²"³ and & span²"³ .

Metric Vector Spaces: The Theory of Bilinear Forms

293

For real inner product spaces, Theorem 10.16 says that if )#) ~ )$) £ , then /#c$ is the unique reflection sending # to $, that is, /#c$ ²#³ ~ $. In the present context, we must be careful, since symmetries are defined for nonisotropic vectors only. Here is what we can say. Theorem 11.32 Let = be a nonsingular orthogonal geometry over a field - , with char²- ³ £ . If "Á # = are nonisotropic vectors with the same (nonzero) “length,” that is, if º"Á "» ~ º#Á #» £ then there exists a symmetry for which " ~ # or " ~ c# Proof. Since " and # are nonisotropic, one of " c # or " b # must also be nonisotropic, for otherwise, since " c # and " b # are orthogonal, their sum " would also be isotropic. If " b # is nonisotropic, then "b# ²" b #³ ~ c²" b #³ and "b# ²" c #³ ~ " c # and so "b# " ~ c#. On the other hand, if " c # is nonisotropic, then "c# ²" c #³ ~ c²" c #³ and "c# ²" b #³ ~ " b # and so "c# " ~ #. Recall that an operator on a real inner product space is unitary if and only if it is a product of reflections. Here is the generalization to nonsingular orthogonal geometries. Theorem 11.33 Let = be a nonsingular orthogonal geometry over a field with char²- ³ £ . A linear transformation on = is an orthogonal transformation if and only if is the product of symmetries on = . Proof. The proof is by induction on ~ dim²= ³. If ~ , then = ~ span²#³ where º#Á #» £ . Let # ~ # for - . Since is unitary º#Á #» ~ º#Á #» ~ º #Á #» ~ º#Á #» and so ~ f. If ~ , then is the identity, which is equal to # . On the other hand, if ~ c then ~ # . In either case, is a product of symmetries.

294

Advanced Linear Algebra

Assume now that the theorem is true for dimensions less than and let dim²= ³ ~ . Let # = be nonisotropic. Since º #Á #» ~ º#Á #» £ , Theorem 11.32 implies the existence of a symmetry on = for which ² #³ ~ # where ~ f . Thus, ~ f on span²#³. Since Theorem 11.9 implies that span²#³ is -invariant, we may apply the induction hypothesis to on span²#³ to get Ospan²#³ ~ $ Ä$ ~ where $ span²#³ and each $ is a symmetry on span²#³ . But each $ can be extended to a symmetry on = by setting $ # ~ #. Assume that is the extension of to = , where ~ on span²#³. Hence, ~ on span²#³ and ~ on span²#³. If ~ , then ~ on = and so ~ , which completes the proof. If ~ c, then ~ # on span²#³ since # is the identity on span²#³ and ~ # on span²#³. Hence, ~ # on = and so ~ # on = .

The Witt Theorems for Orthogonal Geometries We are now ready to consider the Witt theorems for orthogonal geometries. Theorem 11.34 (Witt's cancellation theorem) Let = and > be isometric nonsingular orthogonal geometries over a field - with char²- ³ £ . Suppose that = ~ : p :

and > ~ ; p ;

Then : ; ¬ : ; Proof. First, we prove that it is sufficient to consider the case = ~ > . Suppose that the result holds when = ~ > and that ¢ = ¦ > is an isometry. Then ²:³ p ²: ³ ~ ²: p : ³ ~ = ~ > ~ ; p ; Furthermore, : : ; . We can therefore apply the theorem to > to get : ²: ³ ; as desired. To prove the theorem when = ~ > , assume that = ~ : p : ~ ; p ; where : and ; are nonsingular and : ; . Let ¢ : ¦ ; be an isometry. We proceed by induction on dim²:³.

Metric Vector Spaces: The Theory of Bilinear Forms

295

Suppose first that dim²:³ ~ and that : ~ span² ³. Since º Á » ~ º Á » £ Theorem 11.32 implies that there is a symmetry for which ~ where ~ f . Hence, is an isometry of = for which ; ~ : and Theorem 11.9 implies that ; ~ ²: ³. Thus, O: is the desired isometry. Now suppose the theorem is true for dim²:³ and let dim²:³ ~ . Let ¢ : ¦ ; be an isometry. Since : is nonsingular, we can choose a nonisotropic vector : and write : ~ span² ³ p < , where < is nonsingular. It follows that = ~ : p : ~ span² ³ p < p : and = ~ ; p ; ~ ²span² ³³ p < p ; Now we may apply the one-dimensional case to deduce that < p : < p ; If ¢ < p : ¦ < p ; is an isometry, then < p ²: ³ ~ ²< p : ³ ~ < p ; But < < and since dim²< ³ ~ dim²< ³ , the induction hypothesis implies that : ²: ³ ; . As we have seen, Witt's extension theorem is a corollary of Witt's cancellation theorem. Theorem 11.35 (Witt's extension theorem) Let = and = Z be isometric nonsingular orthogonal geometries over a field - , with char²- ³ £ . Suppose that < is a subspace of = and ¢ < ¦ < = Z is an isometry. Then can be extended to an isometry from = to = Z .

Maximal Hyperbolic Subspaces of an Orthogonal Geometry We have seen that any orthogonal geometry = can be written in the form = ~ < p rad²= ³ where < is nonsingular. Nonsingular spaces are better behaved than singular ones, but they can still possess isotropic vectors.

296

Advanced Linear Algebra

We can improve upon the preceding decomposition by noticing that if " < is isotropic, then Theorem 11.10 implies that span²"³ can be “captured” in a hyperbolic plane / ~ span²"Á %³. Then we can write = ~ / p / < p rad²= ³ where / < is the orthogonal complement of / in < and has “one fewer” isotropic vector. In order to generalize this process, we first discuss maximal totally degenerate subspaces.

Maximal Totally Degenerate Subspaces Let = be a nonsingular orthogonal geometry over a field - , with char²- ³ £ . Suppose that < and < Z are maximal totally degenerate subspaces of = . We claim that dim²< ³ ~ dim²< Z ³. For if dim²< ³ dim²< Z ³, then there is a vector space isomorphism ¢ < ¦ < < Z , which is also an isometry, since < and < Z are totally degenerate. Thus, Witt's extension theorem implies the existence of an isometry ¢ = ¦ = that extends . In particular, c ²< Z ³ is a totally degenerate space that contains < and so c ²< Z ³ ~ < , which shows that dim²< ³ ~ dim²< Z ³. Theorem 11.36 Let = be a nonsingular orthogonal geometry over a field - , with char²- ³ £ . 1) All maximal totally degenerate subspaces of = have the same dimension, which is called the Witt index of = and is denoted by $²= ³. 2) Any totally degenerate subspace of = of dimension $²= ³ is maximal.

Maximal Hyperbolic Subspaces We can prove by a similar argument that all maximal hyperbolic subspaces of = have the same dimension. Let > ~ / p Ä p / and A ~ 2 p Ä p 2 be maximal hyperbolic subspaces of = and suppose that / ~ span²" Á # ³ and 2 ~ span²% Á & ³. We may assume that dim²>³ dim²A³. The linear map ¢ > ¦ A defined by " ~ % Á # ~ & is clearly an isometry from > to >. Thus, Witt's extension theorem implies the existence of an isometry ¢ = ¦ = that extends . In particular, c ²A³ is a hyperbolic space that contains > and so c ²A³ ~ >. It follows that dim²A³ ~ dim²>³.

Metric Vector Spaces: The Theory of Bilinear Forms

297

It is not hard to see that the maximum dimension ²= ³ of a hyperbolic subspace of = is $²= ³, where $²= ³ is the Witt index of = . First, the nonsingular extension of a maximal totally degenerate subspace <$ of = is a hyperbolic space of dimension $²= ³ and so ²= ³ $²= ³. On the other hand, there is a totally degenerate subspace < contained in any hyperbolic space > and so $²= ³, that is, dim²> ³ $²= ³. Hence ²= ³ $²= ³ and so ²= ³ ~ $²= ³. Theorem 11.37 Let = be a nonsingular orthogonal geometry over a field - , with char²- ³ £ . 1) All maximal hyperbolic subspaces of = have dimension $²= ³. 2) Any hyperbolic subspace of dimension $²= ³ must be maximal. 3) The Witt index of a hyperbolic space > is .

The Anisotropic Decomposition of an Orthogonal Geometry If > is a maximal hyperbolic subspace of = , then = ~ > p > Since > is maximal, > is anisotropic, for if " > were isotropic, then the nonsingular extension of > p span²"³ would be a hyperbolic space strictly larger than >. Thus, we arrive at the following decomposition theorem for orthogonal geometries. Theorem 11.38 (The anisotropic decomposition of an orthogonal geometry) Let = ~ be a maximal hyperbolic subspace of < , where > ~ ¸¹ if < has no isotropic vectors. Then = ~ : p > p rad²= ³ where : is anisotropic, > is hyperbolic of dimension $²= ³ and rad²= ³ is totally degenerate.

Exercises 1.

2. 3.

Let < Á > be subspaces of a metric vector space = . Show that a) < > ¬ > < b) < < c) < ~ < Let < Á > be subspaces of a metric vector space = . Show that a) ² ³ ~ < q > b) ²< q > ³ ~ Prove that the following are equivalent: a) = is nonsingular b) º"Á %» ~ º#Á %» for all % = implies " ~ #

298 4. 5.

6.

Advanced Linear Algebra Show that a metric vector space = is nonsingular if and only if the matrix 48 of the form is nonsingular, for every ordered basis 8 . Let = be a finite-dimensional vector space with a bilinear form ºÁ ». We do not assume that the form is symmetric or alternate. Show that the following are equivalent: a) ¸# = º#Á $» ~ for all $ = ¹ ~ b) ¸# = º$Á #» ~ for all $ = ¹ ~ Hint: Consider the singularity of the matrix of the form. Find a diagonal matrix congruent to v w

7.

y c z

Prove that the matrices 0 ~ >

8.

and 4 ~ > ?

?

are congruent over the base field - ~ r of rational numbers. Find an invertible matrix 7 such that 7 ! 0 7 ~ 4 . Let = be an orthogonal geometry over a field - with char²- ³ £ . We wish to construct an orthogonal basis E ~ ²" Á Ã Á " ³ for = , starting with any generating set = ~ ²# Á Ã Á # ³. Justify the following steps, essentially due to Lagrange. We may assume that = is not totally degenerate. a) If º# Á # » £ for some , then let " ~ # . Otherwise, there are indices £ for which º# Á # » £ . Let " ~ # b # . b) Assume we have found an ordered set of vectors E ~ ²" Á Ã Á " ³ that form an orthogonal basis for a subspace = of = and that none of the " 's are isotropic. Then = ~ = p = . c) For each # = , let

$ ~ # c ~

º# Á " » " º" Á " »

Then the vectors $ span = . If = is totally degenerate, take any basis for = and append it to E . Otherwise, repeat step a) on = to get another vector "b and let Eb ~ ²" Á Ã Á "b ³. Eventually, we arrive at an orthogonal basis E for = . 9. Prove that orthogonal hyperbolic planes may be characterized as twodimensional nonsingular orthogonal geometries that have exactly two onedimensional totally isotropic (equivalently: totally degenerate) subspaces. 10. Prove that a two-dimensional nonsingular orthogonal geometry is a hyperbolic plane if and only if its discriminant is - ² c ³. 11. Does Minkowski space contain any isotropic vectors? If so, find them. 12. Is Minkowski space isometric to Euclidean space s ?

Metric Vector Spaces: The Theory of Bilinear Forms

299

13. If ºÁ » is a symmetric bilinear form on = and char²- ³ £ , show that 8²%³ ~ º%Á %»° is a quadratic form. 14. Let = be a vector space over a field - , with ordered basis 8 ~ ²# Á Ã Á # ³. Let ²% Á Ã Á % ³ be a homogeneous polynomial of degree over - , that is, a polynomial each of whose terms has degree . The -form defined by is the function from = to - defined as follows. If # ~ ' # , then ²#³ ~ ² Á Ã Á ³ (We use the same notation for the form and the polynomial.) Prove that forms are the same as quadratic forms. 15. Show that is an isometry on = if and only if 8² #³ ~ 8²#³ where 8 is the quadratic form associated with the bilinear form on = . (Assume that char²- ³ £ .³ 16. Show that a quadratic form 8 on = satisfies the parallelogram law: 8²% b &³ b 8²% c &³ ~ ´8²%³ b 8²&³µ 17. Show that if = is a nonsingular orthogonal geometry over a field - , with char²- ³ £ , then any totally isotropic subspace of = is also a totally degenerate space. 18. Is it true that = ~ rad²= ³ p rad²= ³ ? 19. Let = be a nonsingular symplectic geometry and let #Á be a symplectic transvection. Prove that a) #Á #Á ~ #Áb b) For any symplectic transformation , #Á c ~ #Á c)

For - i , #Á ~ #Á

20.

21. 22. 23. 24.

25.

d) For a fixed # £ , the map ª #Á is an isomorphism from the additive group of - onto the group ¸#Á - ¹ Sp²= ³. Prove that if % is any nonsquare in a finite field - , then all nonsquares have the form %, for some - . Hence, the product of any two nonsquares in - is a square. Formulate Sylvester's law of inertia in terms of quadratic forms on = . Show that a two-dimensional space is a hyperbolic plane if and only if it is nonsingular and contains an isotropic vector. Assume that char²- ³ £ . Prove directly that a hyperbolic plane in an orthogonal geometry cannot have an orthogonal basis when char²- ³ ~ . a) Let < be a subspace of = . Show that the inner product º% b < Á & b < » ~ º%Á &» on the quotient space = °< is well-defined if and only if < rad²= ³. b) If < rad²= ³, when is = °< nonsingular? Let = ~ 5 p : , where 5 is a totally degenerate space.

300

26. 27.

28.

29. 30. 31.

Advanced Linear Algebra a) Prove that 5 ~ rad²= ³ if and only if : is nonsingular. b) If : is nonsingular, prove that : = °rad²= ³. Let dim²= ³ ~ dim²> ³. Prove that = °rad²= ³ > °rad²> ³ implies = >. Let = ~ : p ; . Prove that a) rad²= ³ ~ rad²:³ p rad²; ³ b) = °rad²= ³ :°rad²:³ p ; °rad²; ³ c) dim²rad²= ³³ ~ dim²rad²:³³ b dim²rad²; ³³ d) = is nonsingular if and only if : and ; are both nonsingular. Let = be a nonsingular metric vector space. Because the Riesz representation theorem is valid in = , we can define the adjoint i of a linear map B²= ³ exactly as in the case of real inner product spaces. Prove that is an isometry if and only if it is bijective and unitary (that is, i ~ ). If char²- ³ £ , prove that B²= Á > ³ is an isometry if and only if it is bijective and º #Á #» ~ º#Á #» for all # = . Let 8 ~ ¸# Á Ã Á # ¹ be a basis for = . Prove that B²= Á > ³ is an isometry if and only if it is bijective and º # Á # » ~ º# Á # » for all Á . Let be a linear operator on a metric vector space = . Let 8 ~ ²# Á Ã Á # ³ be an ordered basis for = and let 48 be the matrix of the form relative to 8 . Prove that is an isometry if and only if ´ µ!8 48 ´ µ8 ~ 48

32. Let = be a nonsingular orthogonal geometry and let B²= ³ be an isometry. a) Show that dim²ker² c ³³ ~ dim²im² c ³ ³. b) Show that ker² c ³ ~ im² c ³ . How would you describe ker² c ³ in words? c) If is a symmetry, what is dim²ker² c ³³? d) Can you characterize symmetries by means of dim²ker² c ³³? 33. A linear transformation B²= ³ is called unipotent if c is nilpotent. Suppose that = is a nonisotropic metric vector space and that is unipotent and isometric. Show that ~ . 34. Let = be a hyperbolic space of dimension and let < be a hyperbolic subspace of = of dimension . Show that for each , there is a hyperbolic subspace > of = for which < > = . 35. Let char²- ³ £ . Prove that if ? is a totally degenerate subspace of an orthogonal geometry = , then dim²?³ dim²= ³°. 36. Prove that an orthogonal geometry = of dimension is a hyperbolic space if and only if = is nonsingular, is even and = contains a totally degenerate subspace of dimension °. 37. Prove that a symplectic transformation has determinant equal to .

Chapter 12

Metric Spaces

The Definition In Chapter 9, we studied the basic properties of real and complex inner product spaces. Much of what we did does not depend on whether the space in question is finite-dimensional or infinite-dimensional. However, as we discussed in Chapter 9, the presence of an inner product and hence a metric, on a vector space, raises a host of new issues related to convergence. In this chapter, we discuss briefly the concept of a metric space. This will enable us to study the convergence properties of real and complex inner product spaces. A metric space is not an algebraic structure. Rather it is designed to model the abstract properties of distance. Definition A metric space is a pair ²4 Á ³, where 4 is a nonempty set and ¢ 4 d 4 ¦ s is a real-valued function, called a metric on 4 , with the following properties. The expression ²%Á &³ is read “the distance from % to & .” 1) (Positive definiteness) For all %Á & 4 , ²%Á &³ and ²%Á &³ ~ if and only if % ~ &. 2) (Symmetry) For all %Á & 4 , ²%Á &³ ~ ²&Á %³ 3) (Triangle inequality) For all %Á &Á ' 4 , ²%Á &³ ²%Á '³ b ²'Á &³

As is customary, when there is no cause for confusion, we simply say “let 4 be a metric space.”

302

Advanced Linear Algebra

Example 12.1 Any nonempty set 4 is a metric space under the discrete metric, defined by ²%Á &³ ~ F

if % ~ & if % £ &

Example 12.2 1) The set s is a metric space, under the metric defined for % ~ ²% Á Ã Á % ³ and & ~ ²& Á Ã Á & ³ by ²%Á &³ ~ j²% c & ³ b Ä b ²% c & ³ This is called the Euclidean metric on s . We note that s is also a metric space under the metric ²%Á &³ ~ (% c & ( b Ä b (% c & ( Of course, ²s Á ³ and ²s Á ³ are different metric spaces. 2) The set d is a metric space under the unitary metric ²%Á &³ ~ k(% c & ( b Ä b (% c & ( where % ~ ²% Á Ã Á % ³ and & ~ ²& Á Ã Á & ³ are in d .

Example 12.3 1) The set *´Á µ of all real-valued (or complex-valued) continuous functions on ´Á µ is a metric space, under the metric ² Á ³ ~ sup ( ²%³ c ²%³( %´Áµ

We refer to this metric as the sup metric. 2) The set *´Á µ of all real-valued ²or complex-valued) continuous functions on ´Á µ is a metric space, under the metric

² ²%³Á ²%³³ ~ ( ²%³ c ²%³( %

Example 12.4 Many important sequence spaces are metric spaces. We will often use boldface italic letters to denote sequences, as in % ~ ²% ³ and & ~ ²& ³. 1) The set MB s of all bounded sequences of real numbers is a metric space under the metric defined by ²%Á &³ ~ sup(% c & (

MB d

The set of all bounded complex sequences, with the same metric, is also a metric space. As is customary, we will usually denote both of these spaces by MB .

Metric Spaces

303

2) For , let M be the set of all sequences % ~ ²% ³ of real ²or complex) numbers for which B

(% ( B ~

We define the -norm of % by °

B

)%) ~ 8 (% ( 9 ~

Then M is a metric space, under the metric °

B

²%Á &³ ~ )% c &) ~ 8 (% c & ( 9 ~

The fact that M is a metric follows from some rather famous results about sequences of real or complex numbers, whose proofs we leave as (wellhinted) exercises. ¨ Holder's inequality Let Á and b ~ . If % M and & M , then the product sequence %& ~ ²% & ³ is in M and )%&) )%) )&) that is, B

°

B

(% & ( 8 (% ( 9 ~

~

B

°

8 (& ( 9 ~

A special case of this (with ~ ~ 2) is the Cauchy–Schwarz inequality B

B

B

~

~

~

(% & ( m (% ( m (& ( Minkowski's inequality For , if %Á & M then the sum % b & ~ ²% b & ³ is in M and ) % b & ) ) %) b ) & ) that is, B

8 (% b & ( 9 ~

°

B

8(% ( 9 ~

°

B

b 8 (& ( 9 ~

°

304

Advanced Linear Algebra

If 4 is a metric space under a metric , then any nonempty subset : of 4 is also a metric under the restriction of to : d : . The metric space : thus obtained is called a subspace of 4 .

Open and Closed Sets Definition Let 4 be a metric space. Let % 4 and let be a positive real number. 1) The open ball centered at % , with radius , is )²% Á ³ ~ ¸% 4 ²%Á % ³ ¹ 2) The closed ball centered at % , with radius , is )²% Á ³ ~ ¸% 4 ²%Á % ³ ¹ 3) The sphere centered at % , with radius , is :²% Á ³ ~ ¸% 4 ²%Á % ³ ~ ¹

Definition A subset : of a metric space 4 is said to be open if each point of : is the center of an open ball that is contained completely in : . More specifically, : is open if for all % : , there exists an such that )²%Á ³ : . Note that the empty set is open. A set ; 4 is closed if its complement ; in 4 is open. It is easy to show that an open ball is an open set and a closed ball is a closed set. If % 4 , we refer to any open set : containing % as an open neighborhood of %. It is also easy to see that a set is open if and only if it contains an open neighborhood of each of its points. The next example shows that it is possible for a set to be both open and closed, or neither open nor closed. Example 12.5 In the metric space s with the usual Euclidean metric, the open balls are just the open intervals )²% Á ³ ~ ²% c Á % b ³ and the closed balls are the closed intervals )²% Á ³ ~ ´% c Á % b µ Consider the half-open interval : ~ ²Á µ, for . This set is not open, since it contains no open ball centered at : and it is not closed, since its complement : ~ ²cBÁ µ r ²Á B³ is not open, since it contains no open ball about .

Metric Spaces

305

Observe also that the empty set is both open and closed, as is the entire space s. (Although we will not do so, it is possible to show that these are the only two sets that are both open and closed in s.³ It is not our intention to enter into a detailed discussion of open and closed sets, the subject of which belongs to the branch of mathematics known as topology. In order to put these concepts in perspective, however, we have the following result, whose proof is left to the reader. Theorem 12.1 The collection E of all open subsets of a metric space 4 has the following properties: 1) J E , 4 E 2) If : , ; E then : q ; E 3) If ¸: 2¹ is any collection of open sets, then 2 : E . These three properties form the basis for an axiom system that is designed to generalize notions such as convergence and continuity and leads to the following definition. Definition Let ? be a nonempty set. A collection E of subsets of ? is called a topology for ? if it has the following properties: 1) J EÁ ? E 2) If :Á ; E then : q ; E 3) If ¸: 2¹ is any collection of sets in E , then : E . 2

We refer to subsets in E as open sets and the pair ²?Á E³ as a topological space. According to Theorem 12.1, the open sets (as we defined them earlier) in a metric space 4 form a topology for 4 , called the topology induced by the metric. Topological spaces are the most general setting in which we can define concepts such as convergence and continuity, which is why these concepts are called topological concepts. However, since the topologies with which we will be dealing are induced by a metric, we will generally phrase the definitions of the topological properties that we will need directly in terms of the metric.

Convergence in a Metric Space Convergence of sequences in a metric space is defined as follows. Definition A sequence ²% ³ in a metric space 4 converges to % 4 , written ²% ³ ¦ %, if lim ²% Á %³ ~

¦B

Equivalently, ²% ³ ¦ % if for any , there exists an 5 such that

306

Advanced Linear Algebra

5 ¬ ²% Á %³ or equivalently, 5 ¬ % )²%Á ³ In this case, % is called the limit of the sequence ²% ³.

If 4 is a metric space and : is a subset of 4 , by a sequence in : , we mean a sequence whose terms all lie in : . We next characterize closed sets and therefore also open sets, using convergence. Theorem 12.2 Let 4 be a metric space. A subset : 4 is closed if and only if whenever ²% ³ is a sequence in : and ²% ³ ¦ %, then % : . In loose terms, a subset : is closed if it is closed under the taking of sequential limits. Proof. Suppose that : is closed and let ²% ³ ¦ %, where % : for all . Suppose that % ¤ : . Then since % : and : is open, there exists an for which % )²%Á ³ : . But this implies that )²%Á ³ q ¸% ¹ ~ J which contradicts the fact that ²% ³ ¦ %. Hence, % : . Conversely, suppose that : is closed under the taking of limits. We show that : is open. Let % : and suppose to the contrary that no open ball about % is contained in : . Consider the open balls )²%Á °³, for all . Since none of these balls is contained in : , for each , there is an % : q )²%Á °³. It is clear that ²% ³ ¦ % and so % : . But % cannot be in both : and : . This contradiction implies that : is open. Thus, : is closed.

The Closure of a Set Definition Let : be any subset of a metric space 4 . The closure of : , denoted by cl²:³, is the smallest closed set containing : . We should hasten to add that, since the entire space 4 is closed and since the intersection of any collection of closed sets is closed (exercise), the closure of any set : does exist and is the intersection of all closed sets containing : . The following definition will allow us to characterize the closure in another way. Definition Let : be a nonempty subset of a metric space 4 . An element % 4 is said to be a limit point, or accumulation point, of : if every open ball centered at % meets : at a point other than % itself. Let us denote the set of all limit points of : by M²:³. Here are some key facts concerning limit points and closures.

Metric Spaces

307

Theorem 12.3 Let : be a nonempty subset of a metric space 4 . 1) % M²:³ if and only if there is a sequence ²% ³ in : for which % £ % for all and ²% ³ ¦ %. 2) : is closed if and only if M²:³ : . In words, : is closed if and only if it contains all of its limit points. 3) cl²:³ ~ : r M²:³. 4) % cl²:³ if and only if there is a sequence ²% ³ in : for which ²% ³ ¦ %. Proof. For part 1), assume first that % M²:³. For each , there exists a point % £ % such that % )²%Á °³ q : . Thus, we have ²% Á %³ ° and so ²% ³ ¦ %. For the converse, suppose that ²% ³ ¦ %, where % £ % : . If )²%Á ³ is any ball centered at %, then there is some 5 such that 5 implies % )²%Á ³. Hence, for any ball )²%Á ³ centered at %, there is a point % £ % such that % : q )²%Á ³. Thus, % is a limit point of : . As for part 2), if : is closed, then by part 1), any % M²:³ is the limit of a sequence ²% ³ in : and so must be in : . Hence, M²:³ : . Conversely, if M²:³ : , then : is closed. For if ²% ³ is any sequence in : and ²% ³ ¦ %, then there are two possibilities. First, we might have % ~ % for some , in which case % ~ % : . Second, we might have % £ % for all , in which case ²% ³ ¦ % implies that % M²:³ : . In either case, % : and so : is closed under the taking of limits, which implies that : is closed. For part 3), let ; ~ : r M²:³. Clearly, : ; . To show that ; is closed, we show that it contains all of its limit points. So let % M²; ³. Hence, there is a sequence ²% ³ ; for which % £ % and ²% ³ ¦ %. Of course, each % is either in : , or is a limit point of : . We must show that % ; , that is, that % is either in : or is a limit point of : . Suppose for the purposes of contradiction that % ¤ : and % ¤ M²:³. Then there is a ball )²%Á ³ for which )²%Á ³ q : £ J. However, since ²% ³ ¦ %, there must be an % )²%Á ³. Since % cannot be in : , it must be a limit point of : . Referring to Figure 12.1, if ²% Á %³ ~ , then consider the ball )²% Á ²c³°³. This ball is completely contained in )²%Á ³ and must contain an element & of : , since its center % is a limit point of : . But then & : q )²%Á ³, a contradiction. Hence, % : or % M²:³. In either case, % ; ~ : r M²:³ and so ; is closed. Thus, ; is closed and contains : and so cl²:³ ; . On the other hand, ; ~ : r M²:³ cl²:³ and so cl²:³ ~ ; .

308

Advanced Linear Algebra

Figure 12.1 For part 4), if % cl²:³, then there are two possibilities. If % : , then the constant sequence ²% ³, with % ~ % for all %, is a sequence in : that converges to %. If % ¤ : , then % M²:³ and so there is a sequence ²% ³ in : for which % £ % and ²% ³ ¦ %. In either case, there is a sequence in : converging to %. Conversely, if there is a sequence ²% ³ in : for which ²% ³ ¦ %, then either % ~ % for some , in which case % : cl²:³, or else % £ % for all , in which case % M²:³ cl²:³.

Dense Subsets The following concept is meant to convey the idea of a subset : 4 being “arbitrarily close” to every point in 4 . Definition A subset : of a metric space 4 is dense in 4 if cl²:³ ~ 4 . A metric space is said to be separable if it contains a countable dense subset. Thus, a subset : of 4 is dense if every open ball about any point % 4 contains at least one point of : . Certainly, any metric space contains a dense subset, namely, the space itself. However, as the next examples show, not every metric space contains a countable dense subset. Example 12.6 1) The real line s is separable, since the rational numbers r form a countable dense subset. Similarly, s is separable, since the set r is countable and dense. 2) The complex plane d is separable, as is d for all . 3) A discrete metric space is separable if and only if it is countable. We leave proof of this as an exercise.

Metric Spaces

309

Example 12.7 The space MB is not separable. Recall that MB is the set of all bounded sequences of real numbers (or complex numbers) with metric ²%Á &³ ~ sup(% c & (

To see that this space is not separable, consider the set : of all binary sequences : ~ ¸²% ³ % ~ or for all ¹ This set is in one-to-one correspondence with the set of all subsets of o and so is uncountable. (It has cardinality 2L L .³ Now, each sequence in : is certainly bounded and so lies in MB . Moreover, if % £ & MB , then the two sequences must differ in at least one position and so ²%Á &³ ~ . In other words, we have a subset : of MB that is uncountable and for which the distance between any two distinct elements is . This implies that the balls in the uncountable collection ¸)² Á °³ :¹ are mutually disjoint. Hence, no countable set can meet every ball, which implies that no countable set can be dense in MB . Example 12.8 The metric spaces M are separable, for . The set : of all sequences of the form ~ ² Á Ã Á Á Á Ã ³ for all , where the 's are rational, is a countable set. Let us show that it is dense in M . Any % M satisfies B

(% ( B ~

Hence, for any , there exists an 5 such that B

(% ( ~5 b

Since the rational numbers are dense in s, we can find rational numbers for which (% c ( 5 for all ~ Á Ã Á 5 . Hence, if ~ ² Á Ã Á 5 Á Á Ã ³, then 5

B

~

~5 b

²%Á ³ ~ (% c ( b (% (

b ~

which shows that there is an element of : arbitrarily close to any element of M . Thus, : is dense in M and so M is separable.

310

Advanced Linear Algebra

Continuity Continuity plays a central role in the study of linear operators on infinitedimensional inner product spaces. Definition Let ¢ 4 ¦ 4 Z be a function from the metric space ²4 Á ³ to the metric space ²4 Z Á Z ³. We say that is continuous at % 4 if for any , there exists a such that ²%Á % ³ ¬ Z ² ²%³Á ²% ³³ or, equivalently, 4)²% Á ³5 )² ²% ³Á ³ (See Figure 12.2.) A function is continuous if it is continuous at every % 4 .

Figure 12.2 We can use the notion of convergence to characterize continuity for functions between metric spaces. Theorem 12.4 A function ¢ 4 ¦ 4 Z is continuous if and only if whenever ²% ³ is a sequence in 4 that converges to % 4 , then the sequence ² ²% ³³ converges to ²% ³, in short, ²% ³ ¦ % ¬ ² ²% ³³ ¦ ²% ³ Proof. Suppose first that is continuous at % and let ²% ³ ¦ % . Then, given , the continuity of implies the existence of a such that ²)²% Á ³³ )² ²% ³Á ³ Since ²% ³ ¦ %, there exists an 5 such that % )²% Á ³ for 5 and so 5 ¬ ²% ³ )² ²% ³Á ³ Thus, ²% ³ ¦ ²% ³. Conversely, suppose that ²% ³ ¦ % implies ² ²% ³³ ¦ ²% ³. Suppose, for the purposes of contradiction, that is not continuous at % . Then there exists an

Metric Spaces

311

such that for all , 4)²% Á ³5 \ )² ²% ³Á ³ Thus, for all , 4) 6% Á

\ )² ²% ³Á ³ 75

and so we may construct a sequence ²% ³ by choosing each term % with the property that % ) 6% Á

7, but ²% ³ ¤ )² ²% ³Á ³

Hence, ²% ³ ¦ % , but ²% ³ does not converge to ²% ³. This contradiction implies that must be continuous at % . The next theorem says that the distance function is a continuous function in both variables. Theorem 12.5 Let ²4 Á ³ be a metric space. If ²% ³ ¦ % and ²& ³ ¦ &, then ²% Á & ³ ¦ ²%Á &³. Proof. We leave it as an exercise to show that (²% Á & ³ c ²%Á &³( ²% Á %³ b ²& Á &³ But the right side tends to as ¦ B and so ²% Á & ³ ¦ ²%Á &³.

Completeness The reader who has studied analysis will recognize the following definitions. Definition A sequence ²% ³ in a metric space 4 is a Cauchy sequence if for any , there exists an 5 for which Á 5 ¬ ²% Á % ³

We leave it to the reader to show that any convergent sequence is a Cauchy sequence. When the converse holds, the space is said to be complete. Definition Let 4 be a metric space. 1) 4 is said to be complete if every Cauchy sequence in 4 converges in 4 . 2) A subspace : of 4 is complete if it is complete as a metric space. Thus, : is complete if every Cauchy sequence ² ³ in : converges to an element in :. Before considering examples, we prove a very useful result about completeness of subspaces.

312

Advanced Linear Algebra

Theorem 12.6 Let 4 be a metric space. 1) Any complete subspace of 4 is closed. 2) If 4 is complete, then a subspace : of 4 is complete if and only if it is closed. Proof. To prove 1), assume that : is a complete subspace of 4 . Let ²% ³ be a sequence in : for which ²% ³ ¦ % 4 . Then ²% ³ is a Cauchy sequence in : and since : is complete, ²% ³ must converge to an element of : . Since limits of sequences are unique, we have % : . Hence, : is closed. To prove part 2), first assume that : is complete. Then part 1) shows that : is closed. Conversely, suppose that : is closed and let ²% ³ be a Cauchy sequence in : . Since ²% ³ is also a Cauchy sequence in the complete space 4 , it must converge to some % 4 . But since : is closed, we have ²% ³ ¦ % : . Hence, : is complete. Now let us consider some examples of complete (and incomplete) metric spaces. Example 12.9 It is well known that the metric space s is complete. (However, a proof of this fact would lead us outside the scope of this book.³ Similarly, the complex numbers d are complete. Example 12.10 The Euclidean space s and the unitary space d are complete. Let us prove this for s . Suppose that ²% ³ is a Cauchy sequence in s , where % ~ ²%Á Á Ã Á %Á ³ Thus,

²% Á % ³ ~ ²%Á c %Á ³ ¦ as Á ¦ B ~

and so, for each coordinate position , ²%Á c %Á ³ ²% Á % ³ ¦ which shows that the sequence ²%Á ³~Á2ÁÃ of th coordinates is a Cauchy sequence in s. Since s is complete, we must have ²%Á ³ ¦ & as ¦ B If & ~ ²& Á Ã Á & ³, then

²% Á &³ ~ ²%Á c & ³ ¦ as ¦ B ~

and so ²% ³ ¦ & s . This proves that s is complete.

Metric Spaces

313

Example 12.11 The metric space ²*´Á µÁ ³ of all real-valued (or complexvalued) continuous functions on ´Á µ, with metric ² Á ³ ~ sup ( ²%³ c ²%³( %´Áµ

is complete. To see this, we first observe that the limit with respect to is the uniform limit on ´Á µ, that is ² Á ³ ¦ if and only if for any , there is an 5 for which 5 ¬ ( ²%³ c ²%³( for all % ´Á µ Now let ² ³ be a Cauchy sequence in ²*´Á µÁ ³. Thus, for any , there is an 5 for which Á 5 ¬ ( ²%³ c ²%³( for all % ´Á µ

(12.1)

This implies that, for each % ´Á µ, the sequence ² ²%³³ is a Cauchy sequence of real (or complex) numbers and so it converges. We can therefore define a function on ´Á µ by ²%³ ~ lim ²%³ ¦B

Letting ¦ B in (12.1), we get 5 ¬ ( ²%³ c ²%³( for all % ´Á µ Thus, ²%³ converges to ²%³ uniformly. It is well known that the uniform limit of continuous functions is continuous and so ²%³ *´Á µ. Thus, ² ²%³³ ¦ ²%³ *´Á µ and so ²*´Á µÁ ³ is complete. Example 12.12 The metric space ²*´Á µÁ ³ of all real-valued (or complexvalued) continuous functions on ´Á µ, with metric

² ²%³Á ²%³³ ~ ( ²%³ c ²%³(%

is not complete. For convenience, we take ´Á µ ~ ´Á µ and leave the general case for the reader. Consider the sequence of functions ²%³ whose graphs are shown in Figure 12.3. (The definition of ²%³ should be clear from the graph.³

314

Advanced Linear Algebra

Figure 12.3 We leave it to the reader to show that the sequence ² ²%³³ is Cauchy, but does not converge in ²*´Á µÁ ³. (The sequence converges to a function that is not continuous.) Example 12.13 The metric space MB is complete. To see this, suppose that ²% ³ is a Cauchy sequence in MB , where % ~ ²%Á Á %Á2 Á Ã ³ Then, for each coordinate position , we have (%Á c %Á ( sup (%Á c %Á ( ¦ as Á ¦ B

(12.2)

Hence, for each , the sequence ²%Á ³ of th coordinates is a Cauchy sequence in s (or d). Since s (or d) is complete, we have ²%Á ³ ¦ & as ¦ B for each coordinate position . We want to show that & ~ ²& ³ MB and that ²% ³ ¦ &. Letting ¦ B in ²12.2) gives sup (%Á c & ( ¦ as ¦ B

(12.3)

and so, for some , (%Á c & ( for all and so (& ( b (%Á ( for all But since % MB , it is a bounded sequence and therefore so is ²& ³. That is, & ~ ²& ³ MB . Since ²12.3) implies that ²% ³ ¦ &, we see that MB is complete.

Metric Spaces

315

Example 12.14 The metric space M is complete. To prove this, let ²% ³ be a Cauchy sequence in M , where % ~ ²%Á Á %Á2 Á Ã ³ Then, for each coordinate position , B

(%Á c %Á ( (%Á c %Á ( ~ ²% Á % ³ ¦ ~

which shows that the sequence ²%Á ³ of th coordinates is a Cauchy sequence in s (or d). Since s (or d) is complete, we have ²%Á ³ ¦ & as ¦ B We want to show that & ~ ²& ³ M and that ²% ³ ¦ & . To this end, observe that for any , there is an 5 for which

Á 5 ¬ (%Á c %Á ( ~

for all . Now we let ¦ B, to get

5 ¬ (%Á c & ( ~

for all . Letting ¦ B, we get, for any 5 , B

(%Á c & ( ~

which implies that ²% ³ c & M and so & ~ & c ²% ³ b ²% ³ M addition, ²% ³ ¦ &.

and in

As we will see in the next chapter, the property of completeness plays a major role in the theory of inner product spaces. Inner product spaces for which the induced metric space is complete are called Hilbert spaces.

Isometries A function between two metric spaces that preserves distance is called an isometry. Here is the formal definition. Definition Let ²4 Á ³ and ²4 Z Á Z ³ be metric spaces. A function ¢ 4 ¦ 4 Z is called an isometry if Z ² ²%³Á ²&³³ ~ ²%Á &³

316

Advanced Linear Algebra

for all %Á & 4 . If ¢ 4 ¦ 4 Z is a bijective isometry from 4 to 4 Z , we say that 4 and 4 Z are isometric and write 4 4 Z . Theorem 12.7 Let ¢ ²4 Á ³ ¦ ²4 Z Á Z ³ be an isometry. Then 1) is injective 2) is continuous 3) c ¢ ²4 ³ ¦ 4 is also an isometry and hence also continuous. Proof. To prove 1), we observe that ²%³ ~ ²&³ ¯ Z ² ²%³Á ²&³³ ~ ¯ ²%Á &³ ~ ¯ % ~ & To prove 2), let ²% ³ ¦ % in 4 . Then Z ² ²% ³Á ²%³³ ~ ²% Á %³ ¦ as ¦ B and so ² ²% ³³ ¦ ²%³, which proves that is continuous. Finally, we have ² c ² ²%³³Á c ² ²&³³ ~ ²%Á &³ ~ Z ² ²%³Á ²&³³ and so c ¢ ²4 ³ ¦ 4 is an isometry.

The Completion of a Metric Space While not all metric spaces are complete, any metric space can be embedded in a complete metric space. To be more specific, we have the following important theorem. Theorem 12.8 Let ²4 Á ³ be any metric space. Then there is a complete metric space ²4 Z Á Z ³ and an isometry ¢ 4 ¦ 4 4 Z for which 4 is dense in 4 Z . The metric space ²4 Z Á Z ³ is called a completion of ²4 Á ³. Moreover, ²4 Z Á Z ³ is unique, up to bijective isometry. Proof. The proof is a bit lengthy, so we divide it into various parts. We can simplify the notation considerably by thinking of sequences ²% ³ in 4 as functions ¢ o ¦ 4 , where ²³ ~ % .

Cauchy Sequences in 4 The basic idea is to let the elements of 4 Z be equivalence classes of Cauchy sequences in 4 . So let CS²4 ³ denote the set of all Cauchy sequences in 4 . If Á CS²4 ³, then, intuitively speaking, the terms ²³ get closer together as ¦ B and so do the terms ²³. Therefore, it seems reasonable that ² ²³Á ²³³ should approach a finite limit as ¦ B. Indeed, since (² ²³Á ²³³ c ² ²³Á ²³³( ² ²³Á ²³³ b ²²³Á ²³³ ¦ as Á ¦ B it follows that ² ²³Á ²³³ is a Cauchy sequence of real numbers, which implies that

Metric Spaces

317

(12.4)

lim ² ²³Á ²³³ B

¦B

(That is, the limit exists and is finite.³

Equivalence Classes of Cauchy Sequences in 4 We would like to define a metric Z on the set CS²4 ³ by Z ² Á ³ ~ lim ² ²³Á ²³³ ¦B

However, it is possible that lim ² ²³Á ²³³ ~

¦B

for distinct sequences and , so this does not define a metric. Thus, we are led to define an equivalence relation on CS²4 ³ by ¯ lim ² ²³Á ²³³ ~ ¦B

Let CS²4 ³ be the set of all equivalence classes of Cauchy sequences and define, for Á CS²4 ³, Z ² Á ³ ~ lim ² ²³Á ²³³ ¦B

(12.5)

where and . To see that Z is well-defined, suppose that Z and Z . Then since Z and Z , we have (² Z ²³Á Z ²³³ c ² ²³Á ²³³( ² Z ²³Á ²³³ b ²Z ²³Á ²³³ ¦ as ¦ B. Thus, Z and Z ¬ lim ² Z ²³Á Z ²³³ ~ lim ² ²³Á ²³³ ¦B

¦B

¬ Z ² Z Á Z ³ ~ Z ² Á ³ which shows that Z is well-defined. To see that Z is a metric, we verify the triangle inequality, leaving the rest to the reader. If Á and are Cauchy sequences, then ² ²³Á ²³³ ² ²³Á ²³³ b ²²³Á ²³³ Taking limits gives lim ² ²³Á ²³³ lim ² ²³Á ²³³ b lim ²²³Á ²³³

¦B

¦B

¦B

318

Advanced Linear Algebra

and so Z ² Á ³ Z ² Á ³ b Z ²Á ³

Embedding ²4 Á ³ in ²4 Z Á Z ³ For each % 4 , consider the constant Cauchy sequence ´%µ, where ´%µ²³ ~ % for all . The map ¢ 4 ¦ 4 Z defined by % ~ ´%µ is an isometry, since Z ² %Á &³ ~ Z ²´%µÁ ´&µ³ ~ lim ²´%µ²³Á ´&µ²³³ ~ ²%Á &³ ¦B

Moreover, 4 is dense in 4 Z . This follows from the fact that we can approximate any Cauchy sequence in 4 by a constant sequence. In particular, let 4 Z . Since is a Cauchy sequence, for any , there exists an 5 such that Á 5 ¬ ² ²³Á ²³³ Now, for the constant sequence ´ ²5 ³µ we have Z 4´ ²5 ³µÁ 5 ~ lim ² ²5 ³Á ²³³ ¦B

and so 4 is dense in 4 Z .

²4 Z Á Z ³ Is Complete Suppose that Á Á 3 Á Ã is a Cauchy sequence in 4 Z . We wish to find a Cauchy sequence in 4 for which Z ² Á ³ ~ lim ² ²³Á ²³³ ¦ as ¦ B ¦B

Since 4 Z and since 4 is dense in 4 Z , there is a constant sequence ´ µ ~ ² Á Á Ã ³ for which Z ² Á ´ µ³

Metric Spaces

319

We can think of as a constant approximation to , with error at most ° . Let be the sequence of these constant approximations: ²³ ~ This is a Cauchy sequence in 4 . Intuitively speaking, since the 's get closer to each other as ¦ B, so do the constant approximations. In particular, we have ² Á ³ ~ Z ²´ µÁ ´ µ³ Z ²´ µÁ ³ b Z ² Á ³ b Z ² Á ´ µ³ b Z ² Á ³ b ¦ as Á ¦ B. To see that converges to , observe that Z ² Á ³ Z ² Á ´ µ³ b Z ²´ µÁ ³ ~

b lim ² Á ³ ¦B

b lim ² Á ²³³ ¦B

Now, since is a Cauchy sequence, for any , there is an 5 such that Á 5 ¬ ² Á ³ In particular, 5 ¬ lim ² Á ³ ¦B

and so 5 ¬ Z ² Á ³

b

which implies that ¦ , as desired.

Uniqueness Finally, we must show that if ²4 Z Á Z ³ and ²4 ZZ Á ZZ ³ are both completions of ²4 Á ³, then 4 Z 4 ZZ . Note that we have bijective isometries ¢ 4 ¦ 4 4 Z and ¢ 4 ¦ 4 4 ZZ Hence, the map ~ c ¢ 4 ¦ 4 is a bijective isometry from 4 onto 4 , where 4 is dense in 4 Z . ²See Figure 12.4.³

320

Advanced Linear Algebra

Figure 12.4 Our goal is to show that can be extended to a bijective isometry from 4 Z to 4 ZZ . Let % 4 Z . Then there is a sequence ² ³ in 4 for which ² ³ ¦ %. Since ² ³ is a Cauchy sequence in 4 , ²² ³³ is a Cauchy sequence in 4 4 ZZ and since 4 ZZ is complete, we have ²² ³³ ¦ & for some & 4 ZZ . Let us define ²%³ ~ &. To see that is well-defined, suppose that ² ³ ¦ % and ² ³ ¦ %, where both sequences lie in 4 . Then ZZ ²² ³Á ² ³³ ~ Z ² Á ³ ¦ as ¦ B and so ²² ³³ and ²² ³³ converge to the same element of 4 ZZ , which implies that ²%³ does not depend on the choice of sequence in 4 converging to %. Thus, is well-defined. Moreover, if 4 , then the constant sequence ´µ converges to and so ²³ ~ lim ²³ ~ ²³, which shows that is an extension of . To see that is an isometry, suppose that ² ³ ¦ % and ² ³ ¦ &. Then ²² ³³ ¦ ²%³ and ²² ³³ ¦ ²&³ and since ZZ is continuous, we have ZZ ²²%³Á ²&³³ ~ lim ZZ ²² ³Á ² ³³ ~ lim Z ² Á ³ ~ Z ²%Á &³ ¦B

¦B

Thus, we need only show that is surjective. Note first that 4 ~ im²³ im²³. Thus, if im²³ is closed, we can deduce from the fact that 4 is dense in 4 ZZ that im²³ ~ 4 ZZ . So, suppose that ²²% ³³ is a sequence in im²³ and ²²% ³³ ¦ ' . Then ²²% ³³ is a Cauchy sequence and therefore so is ²% ³. Thus, ²% ³ ¦ % 4 Z . But is continuous and so ²²% ³³ ¦ ²%³, which implies that ²%³ ~ ' and so ' im²³. Hence, is surjective and 4 Z 4 ZZ .

Metric Spaces

321

Exercises 1.

Prove the generalized triangle inequality ²% Á % ³ ²% Á % ³ b ²% Á %3 ³ b Ä b ²%c Á % ³

2.

a)

Use the triangle inequality to prove that (²%Á &³ c ²Á ³( ²%Á ³ b ²&Á ³

b) Prove that (²%Á '³ c ²&Á '³( ²%Á &³ 3. 4.

5.

6. 7. 8. 9.

Let : MB be the subspace of all binary sequences (sequences of 's and 's). Describe the metric on : . Let 4 ~ ¸Á ¹ be the set of all binary -tuples. Define a function ¢ : d : ¦ s by letting ²%Á &³ be the number of positions in which % and & differ. For example, ´²³Á ²³µ ~ . Prove that is a metric. (It is called the Hamming distance function and plays an important role in the theory of error-correcting codes.³ Let B. a) If % ~ ²% ³ M show that % ¦ b) Find a sequence that converges to but is not an element of any M for B. a) Show that if % ~ ²% ³ M , then % M for all . b) Find a sequence % ~ ²% ³ that is in M for , but is not in M . Show that a subset : of a metric space 4 is open if and only if : contains an open neighborhood of each of its points. Show that the intersection of any collection of closed sets in a metric space is closed. Let ²4 Á ³ be a metric space. The diameter of a nonempty subset : 4 is ²:³ ~ sup ²%Á &³ %Á&:

A set : is bounded if ²:³ B. a) Prove that : is bounded if and only if there is some % 4 and s for which : )²%Á ³. b) Prove that ²:³ ~ if and only if : consists of a single point. c) Prove that : ; implies ²:³ ²; ³. d) If : and ; are bounded, show that : r ; is also bounded. 10. Let ²4 Á ³ be a metric space. Let Z be the function defined by Z ²%Á &³ ~

²%Á &³ b ²%Á &³

322

Advanced Linear Algebra

Show that ²4 Á Z ³ is a metric space and that 4 is bounded under this metric, even if it is not bounded under the metric . b) Show that the metric spaces ²4 Á ³ and ²4 Á Z ³ have the same open sets. 11. If : and ; are subsets of a metric space ²4 Á ³, we define the distance between : and ; by a)

²:Á ; ³ ~

12. 13. 14. 15. 16.

17. 18. 19. 20. 21.

inf ²%Á &³

%:Á!;

a) Is it true that ²:Á ; ³ ~ if and only if : ~ ; ? Is a metric? b) Show that % cl²:³ if and only if ²¸%¹Á :³ ~ . Prove that % 4 is a limit point of : 4 if and only if every neighborhood of % meets : in a point other than % itself. Prove that % 4 is a limit point of : 4 if and only if every open ball )²%Á ³ contains infinitely many points of : . Prove that limits are unique, that is, ²% ³ ¦ %, ²% ³ ¦ & implies that % ~ &. Let : be a subset of a metric space 4 . Prove that % cl²:³ if and only if there exists a sequence ²% ³ in : that converges to %. Prove that the closure has the following properties: a) : cl²:³ b) cl²cl²:³³ ~ : c) cl²: r ; ³ ~ cl²:³ r cl²; ³ d) cl²: q ; ³ cl²:³ q cl²; ³ Can the last part be strengthened to equality? a) Prove that the closed ball )²%Á ³ is always a closed subset. b) Find an example of a metric space in which the closure of an open ball )²%Á ³ is not equal to the closed ball )²%Á ³. Provide the details to show that s is separable. Prove that d is separable. Prove that a discrete metric space is separable if and only if it is countable. Prove that the metric space 8´Á µ of all bounded functions on ´Á µ, with metric ² Á ³ ~ sup ( ²%³ c ²%³( %´Áµ

is not separable. 22. Show that a function ¢ ²4 Á ³ ¦ ²4 Z Á Z ³ is continuous if and only if the inverse image of any open set is open, that is, if and only if c ²< ³ ~ ¸% 4 ²%³ < ¹ is open in 4 whenever < is an open set in 4 Z . 23. Repeat the previous exercise, replacing the word open by the word closed. 24. Give an example to show that if ¢ ²4 Á ³ ¦ ²4 Z Á Z ³ is a continuous function and < is an open set in 4 , it need not be the case that ²< ³ is open in 4 Z .

Metric Spaces

323

25. Show that any convergent sequence is a Cauchy sequence. 26. If ²% ³ ¦ % in a metric space 4 , show that any subsequence ²% ³ of ²% ³ also converges to %. 27. Suppose that ²% ³ is a Cauchy sequence in a metric space 4 and that some subsequence ²% ³ of ²% ³ converges. Prove that ²% ³ converges to the same limit as the subsequence. 28. Prove that if ²% ³ is a Cauchy sequence, then the set ¸% ¹ is bounded. What about the converse? Is a bounded sequence necessarily a Cauchy sequence? 29. Let ²% ³ and ²& ³ be Cauchy sequences in a metric space 4 . Prove that the sequence ~ ²% Á & ³ converges. 30. Show that the space of all convergent sequences of real numbers ²or complex numbers) is complete as a subspace of MB . 31. Let F denote the metric space of all polynomials over d, with metric ²Á ³ ~ sup (²%³ c ²%³( %´Áµ

Is F complete? 32. Let : MB be the subspace of all sequences with finite support (that is, with a finite number of nonzero terms). Is : complete? 33. Prove that the metric space { of all integers, with metric ²Á ³ ~ ( c (, is complete. 34. Show that the subspace : of the metric space *´Á µ (under the sup metric) consisting of all functions *´Á µ for which ²³ ~ ²³ is complete. 35. If 4 4 Z and 4 is complete, show that 4 Z is also complete. 36. Show that the metric spaces *´Á µ and *´Á µ, under the sup metric, are isometric. 37. Prove Ho¨lder's inequality B

°

B

(% & ( 8 (% ( 9 ~

~

°

B

8 (& ( 9 ~

as follows: a) Show that ~ !c ¬ ! ~ c b) Let " and # be positive real numbers and consider the rectangle 9 in s with corners ²Á ³, ²"Á ³, ²Á #³ and ²"Á #³, with area "#. Argue geometrically (that is, draw a picture) to show that "

"# !c ! b

#

c

and so "# c)

" # b

Now let ? ~ '(% ( B and @ ~ '(& ( B. Apply the results of part b) to

324

Advanced Linear Algebra

"~

(% ( Á ? °

#~

(& ( @ °

and then sum on to deduce Ho¨lder's inequality. 38. Prove Minkowski's inequality °

B

8(% b & ( 9 ~

°

B

8 (% ( 9 ~

B

°

b 8 (& ( 9 ~

as follows: a) Prove it for ~ first. b) Assume . Show that (% b & ( (% ((% b & (c b (& ((% b & (c c)

Sum this from ~ to and apply Ho¨lder's inequality to each sum on the right, to get

(% b & ( ~

H8(% ( 9 ~

°

b 8(& ( 9 ~

°

°

I8(% b & ( 9 ~

Divide both sides of this by the last factor on the right and let ¦ B to deduce Minkowski's inequality. 39. Prove that M is a metric space.

Chapter 13

Hilbert Spaces

Now that we have the necessary background on the topological properties of metric spaces, we can resume our study of inner product spaces without qualification as to dimension. As in Chapter 9, we restrict attention to real and complex inner product spaces. Hence - will denote either s or d.

A Brief Review Let us begin by reviewing some of the results from Chapter 9. Recall that an inner product space = over - is a vector space = , together with an inner product ºÁ »¢ = d = ¦ - . If - ~ s, then the inner product is bilinear and if - ~ d, the inner product is sesquilinear. An inner product induces a norm on = , defined by )#) ~ jº#Á #» We recall in particular the following properties of the norm. Theorem 13.1 1) (The Cauchy-Schwarz inequality) For all "Á # = , (º"Á #»( )") )#) with equality if and only if " ~ # for some - . 2) (The triangle inequality) For all "Á # = , )" b #) )") b )#) with equality if and only if " ~ # for some - . 3) (The parallelogram law) )" b #) b )" c #) ~ )") b )#)

We have seen that the inner product can be recovered from the norm, as follows.

326

Advanced Linear Algebra

Theorem 13.2 1) If = is a real inner product space, then º"Á #» ~

²)" b #) c )" c #) ³

2) If = is a complex inner product space, then º"Á #» ~

²)" b #) c )" c #) ³ b ²)" b #) c )" c #) ³

The inner product also induces a metric on = defined by ²"Á #³ ~ )" c #) Thus, any inner product space is a metric space. Definition Let = and > be inner product spaces and let B²= Á > ³. 1) is an isometry if it preserves the inner product, that is, if º "Á #» ~ º"Á #» for all "Á # = . 2) A bijective isometry is called an isometric isomorphism. When ¢ = ¦ > is an isometric isomorphism, we say that = and > are isometrically isomorphic. It is easy to see that an isometry is always injective but need not be surjective, even if = ~ > . Theorem 13.3 A linear transformation B²= Á > ³ is an isometry if and only if it preserves the norm, that is, if and only if )#) ~ )#) for all # = . The following result points out one of the main differences between real and complex inner product spaces. Theorem 13.4 Let = be an inner product space and let B²= ³. 1) If º #Á $» ~ for all #Á $ = , then ~ . 2) If = is a complex inner product space and 8 ²#³ ~ º #Á #» ~ for all # = , then ~ . 3) Part 2) does not hold in general for real inner product spaces.

Hilbert Spaces Since an inner product space is a metric space, all that we learned about metric spaces applies to inner product spaces. In particular, if ²% ³ is a sequence of

Hilbert Spaces

327

vectors in an inner product space = , then ²% ³ ¦ % if and only if )% c %) ¦ as ¦ B The fact that the inner product is continuous as a function of either of its coordinates is extremely useful. Theorem 13.5 Let = be an inner product space. Then 1) ²% ³ ¦ %Á ²& ³ ¦ & ¬ º% Á & » ¦ º%Á &» 2) ²% ³ ¦ % ¬ )% ) ¦ )%)

Complete inner product spaces play an especially important role in both theory and practice. Definition An inner product space that is complete under the metric induced by the inner product is said to be a Hilbert space. Example 13.1 One of the most important examples of a Hilbert space is the space M . Recall that the inner product is defined by B

º%Á &» ~ % & ~

(In the real case, the conjugate is unnecessary.³ The metric induced by this inner product is °2

B

²%Á &³ ~ )% c &) ~ 8(% c & ( 9 ~

which agrees with the definition of the metric space M given in Chapter 12. In other words, the metric in Chapter 12 is induced by this inner product. As we saw in Chapter 12, this inner product space is complete and so it is a Hilbert space. (In fact, it is the prototype of all Hilbert spaces, introduced by David Hilbert in 1912, even before the axiomatic definition of Hilbert space was given by John von Neumann in 1927.³ The previous example raises the question whether the other metric spaces M ² £ ), with distance given by B

°

²%Á &³ ~ )% c &) ~ 8 (% c & ( 9

(13.1)

~

are complete inner product spaces. The fact is that they are not even inner product spaces! More specifically, there is no inner product whose induced metric is given by (13.1). To see this, observe that, according to Theorem 13.1,

328

Advanced Linear Algebra

any norm that comes from an inner product must satisfy the parallelogram law ) % b & ) b ) % c & ) ~ )%) b )& ) But the norm in (13.1) does not satisfy this law. To see this, take % ~ ²Á Á Á Ã ³ and & ~ ²Á cÁ Á Ã ³. Then )% b &) ~ Á )% c &) ~ and )%) ~ ° Á )&) ~ ° Thus, the left side of the parallelogram law is and the right side is h 2° , which equals if and only if ~ . Just as any metric space has a completion, so does any inner product space. Theorem 13.6 Let = be an inner product space. Then there exists a Hilbert space / and an isometry ¢ = ¦ / for which = is dense in / . Moreover, / is unique up to isometric isomorphism. Proof. We know that the metric space ²= Á ³, where is induced by the inner product, has a unique completion ²= Z Á Z ³, which consists of equivalence classes of Cauchy sequences in = . If ²% ³ ²% ³ = Z and ²& ³ ²& ³ = Z , then we set ²% ³ b ²& ³ ~ ²% b & ³Á ²% ³ ~ ²% ³ and º²% ³Á ²& ³» ~ lim º% Á & » ¦B

It is easy to see that since ²% ³ and ²& ³ are Cauchy sequences, so are ²% b & ³ and ²% ³. In addition, these definitions are well-defined, that is, they are independent of the choice of representative from each equivalence class. For instance, if ²% V ³ ²% ³, then lim )% c % V ) ~

¦B

and so (º% Á & » c º% V Á & »( ~ (º% c % V Á & »( )% c % V ))& ) ¦ (The Cauchy sequence ²& ³ is bounded.³ Hence, º²% ³Á ²& ³» ~ lim º% Á & » ~ lim º% V Á & » ~ º²% V ³Á ²&³» ¦B

¦B

We leave it to the reader to show that = Z is an inner product space under these operations.

Hilbert Spaces

329

Moreover, the inner product on = Z induces the metric Z , since º²% c & ³Á ²% c & ³» ~ lim º% c & Á % c & » ¦B

~ lim ²% Á & ³ ¦B

~ Z ²²% ³Á ²& ³³ Hence, the metric space isometry ¢ = ¦ = Z is an isometry of inner product spaces, since º %Á &» ~ Z ² %Á &³ ~ ²%Á &³ ~ º%Á &» Thus, = Z is a complete inner product space and = is a dense subspace of = Z that is isometrically isomorphic to = . We leave the issue of uniqueness to the reader. The next result concerns subspaces of inner product spaces. Theorem 13.7 1) Any complete subspace of an inner product space is closed. 2) A subspace of a Hilbert space is a Hilbert space if and only if it is closed. 3) Any finite-dimensional subspace of an inner product space is closed and complete. Proof. Parts 1) and 2) follow from Theorem 12.6. Let us prove that a finitedimensional subspace : of an inner product space = is closed. Suppose that ²% ³ is a sequence in : , ²% ³ ¦ % and % ¤ : . Let 8 ~ ¸ Á Ã Á ¹ be an orthonormal Hamel basis for : . The Fourier expansion

~ º%Á » ~

in : has the property that % c £ but º% c Á » ~ º%Á » c º Á » ~ Thus, if we write & ~ % c and & ~ % c : , the sequence ²& ³, which is in : , converges to a vector & that is orthogonal to : . But this is impossible, because & & implies that )& c &) ~ )& ) b )&) )&) ¦ ° This proves that : is closed. To see that any finite-dimensional subspace : of an inner product space is complete, let us embed : (as an inner product space in its own right) in its completion : Z . Then : (or rather an isometric copy of : ) is a finite-dimensional

330

Advanced Linear Algebra

subspace of a complete inner product space : Z and as such it is closed. However, : is dense in : Z and so : ~ : Z , which shows that : is complete.

Infinite Series Since an inner product space allows both addition of vectors and convergence of sequences, we can define the concept of infinite sums, or infinite series. Definition Let = be an inner product space. The th partial sum of the sequence ²% ³ in = is

~ % b Ä b %

If the sequence ² ³ of partial sums converges to a vector = , that is, if )

c ) ¦ as ¦ B

then we say that the series % converges to and write B

% ~

~

We can also define absolute convergence. Definition A series % is said to be absolutely convergent if the series B

)% ) ~

converges. The key relationship between convergence and absolute convergence is given in the next theorem. Note that completeness is required to guarantee that absolute convergence implies convergence. Theorem 13.8 Let = be an inner product space. Then = is complete if and only if absolute convergence of a series implies convergence. Proof. Suppose that = is complete and that )% ) B. Then the sequence of partial sums is a Cauchy sequence, for if , we have

)

c

)

~ i % i )% ) ¦ ~b

~b

Hence, the sequence ² ³ converges, that is, the series % converges. Conversely, suppose that absolute convergence implies convergence and let ²% ³ be a Cauchy sequence in = . We wish to show that this sequence converges. Since ²% ³ is a Cauchy sequence, for each , there exists an 5

Hilbert Spaces

331

with the property that Á 5 ¬ )% c % )

Clearly, we can choose 5 5 Ä, in which case )%5b c %5 )

and so B

B

B ~

)%5b c %5 ) ~

Thus, according to hypothesis, the series B

²%5b c %5 ³ ~

converges. But this is a telescoping series, whose th partial sum is %5b c %5 and so the subsequence ²%5 ³ converges. Since any Cauchy sequence that has a convergent subsequence must itself converge, the sequence ²% ³ converges and so = is complete.

An Approximation Problem Suppose that = is an inner product space and that : is a subset of = . It is of considerable interest to be able to find, for any % = , a vector in : that is closest to % in the metric induced by the inner product, should such a vector exist. This is the approximation problem for = . Suppose that % = and let ~ inf )% c ) :

Then there is a sequence

for which ~ )% c

as shown in Figure 13.1.

)

¦

332

Advanced Linear Algebra

Figure 13.1 Let us see what we can learn about this sequence. First, if we let & ~ % c then according to the parallelogram law,

,

)& b & ) b )& c & ) ~ ²)& ) b )& ) ³ or )& c & ) ~ ²)& ) b )& ) ³ c h

& b & h

(13.2)

Now, if the set : is convex, that is, if %Á & : ¬ % b ² c ³& : for all (in words, : contains the line segment between any two of its points), then ² b ³° : and so h

& b & h ~ h% c

b

h

Thus, (13.2) gives )& c & ) ²)& ) b )& ) ³ c ¦ as Á ¦ B. Hence, if : is convex, then the sequence ²& ³ ~ ²% c Cauchy sequence and therefore so is ² ³.

³

is a

If we also require that : be complete, then the Cauchy sequence ² ³ converges to a vector % V : and by the continuity of the norm, we must have )% c % V) ~ . Let us summarize and add a remark about uniqueness. Theorem 13.9 Let = be an inner product space and let : be a complete convex subset of = . Then for any % = , there exists a unique % V : for which )% c % V) ~ inf )% c ) :

The vector % V is called the best approximation to % in : .

Hilbert Spaces

333

Proof. Only the uniqueness remains to be established. Suppose that )% c % V) ~ ~ )% c %Z ) Then, by the parallelogram law,

)% V c %Z ) ~ )²% c %Z ³ c ²% c %³ V)

~ )% c % V) b )% c %Z ) c )% c % V c %Z )

~ )% c % V) b )% c %Z ) c h% c

% V b %Z h 2

b c ~ and so % V ~ % Z . Since any subspace : of an inner product space = is convex, Theorem 13.9 applies to complete subspaces. However, in this case, we can say more. Theorem 13.10 Let = be an inner product space and let : be a complete subspace of = . Then for any % = , the best approximation to % in : is the unique vector %Z : for which % c %Z : . Proof. Suppose that % c %Z : , where %Z : . Then for any : , we have % c %Z c %Z and so

)% c ) ~ )% c %Z ) b )%Z c ) )% c %Z )

Hence %Z ~ % V is the best approximation to % in : . Now we need only show that %c% V : , where % V is the best approximation to % in : . For any : , a little computation reminiscent of completing the square gives )% c )2 ~ º% c Á % c » ~ )%) c º%Á » c º Á %» b ) ) ~ )%) b ) ) 8 c ~ ) % ) b ) ) 8 c

~ ) % ) b ) ) e c

º%Á » ) )

º%Á » ) )

º%Á » ) )

º%Á » ) )

98 c

e c

Now, this is smallest when ~

c

º%Á » ) )

9

º%Á » ) )

(º%Á »( ) )

9c

(º%Á »( ) )

334

Advanced Linear Algebra

in which case )% c ) ~ )%) c

(º%Á »( ) )

Replacing % by % c % V gives )% c % V c ) ~ )% c % V) c

(º% c %Á V »(

) )

But % V is the best approximation to % in : and since % V c : we must have )% c % V c ) )% c % V)

Hence, (º% c %Á V »(

) )

~

or equivalently, º% c %Á V »~ Hence, % c % V : . According to Theorem 13.9, if : is a complete subspace of an inner product space = , then for any % = , we may write %~% V b ²% c %³ V where % V : and % c % V : . Hence, = ~ : b : and since : q : ~ ¸¹, we also have = ~ : p : . This is the projection theorem for arbitrary inner product spaces. Theorem 13.11 (The projection theorem) If : is a complete subspace of an inner product space = , then = ~ : p : In particular, if : is a closed subspace of a Hilbert space / , then / ~ : p :

Theorem 13.12 Let : , ; and ; Z be subspaces of an inner product space = . 1) If = ~ : p ; then ; ~ : . 2) If : p ; ~ : p ; Z then ; ~ ; Z . Proof. If = ~ : p ; , then ; : by definition of orthogonal direct sum. On the other hand, if ' : , then ' ~ b !, for some : and ! ; . Hence, ~ º'Á » ~ º Á » b º!Á » ~ º Á »

Hilbert Spaces and so 1 ) .

335

~ , implying that ' ~ ! ; . Thus, : ; . Part 2) follows from part

Let us denote the closure of the span of a set : of vectors by cspan²:³. Theorem 13.13 Let / be a Hilbert space. 1) If ( is a subset of / , then cspan²(³ ~ ( 2) If : is a subspace of / , then cl²:³ ~ : 3) If 2 is a closed subspace of / , then 2 ~ 2 Proof. We leave it as an exercise to show that ´cspan²(³µ ~ ( . Hence / ~ cspan²(³ p ´cspan²(³µ ~ cspan²(³ p ( But since ( is closed, we also have / ~ ( p ( and so by Theorem 13.12, cspan²(³ ~ ( . The rest follows easily from part 1 ) . In the exercises, we provide an example of a closed subspace 2 of an inner product space = for which 2 £ 2 . Hence, we cannot drop the requirement that / be a Hilbert space in Theorem 13.13. Corollary 13.14 If ( is a subset of a Hilbert space / , then span²(³ is dense in / if and only if ( ~ ¸¹. Proof. As in the previous proof, / ~ cspan²(³ p ( and so ( ~ ¸¹ if and only if / ~ cspan²(³.

Hilbert Bases We recall the following definition from Chapter 9. Definition A maximal orthonormal set in a Hilbert space / is called a Hilbert basis for / . Zorn's lemma can be used to show that any nontrivial Hilbert space has a Hilbert basis. Again, we should mention that the concepts of Hilbert basis and Hamel basis (a maximal linearly independent set) are quite different. We will show

336

Advanced Linear Algebra

later in this chapter that any two Hilbert bases for a Hilbert space have the same cardinality. Since an orthonormal set E is maximal if and only if E ~ ¸¹, Corollary 13.14 gives the following characterization of Hilbert bases. Theorem 13.15 Let E be an orthonormal subset of a Hilbert space / . The following are equivalent: 1) E is a Hilbert basis 2) E ~ ¸¹ 3) E is a total subset of / , that is, cspan²E³ ~ / . Part 3) of this theorem says that a subset of a Hilbert space is a Hilbert basis if and only if it is a total orthonormal set.

Fourier Expansions We now want to take a closer look at best approximations. Our goal is to find an explicit expression for the best approximation to any vector % from within a closed subspace : of a Hilbert space / . We will find it convenient to consider three cases, depending on whether : has finite, countably infinite, or uncountable dimension.

The Finite-Dimensional Case Suppose that E ~ ¸" Á Ã Á " ¹ is an orthonormal set in a Hilbert space / . Recall that the Fourier expansion of any % / , with respect to E , is given by

% V ~ º%Á " »" ~

where º%Á " » is the Fourier coefficient of % with respect to " . Observe that º% c %Á V " » ~ º%Á " » c º%Á V " » ~ and so % c % V span²E³. Thus, according to Theorem 13.9, the Fourier expansion % V is the best approximation to % in span²E³. Moreover, since %c% V% V, we have )% V ) ~ ) % ) c )% c % V) )%)

and so )% V ) )% ) with equality if and only if % ~ % V, which happens if and only if % span²E³. Let us summarize.

Hilbert Spaces

337

Theorem 13.16 Let E ~ ¸" Á Ã Á " ¹ be a finite orthonormal set in a Hilbert space / . For any % / , the Fourier expansion % V of % is the best approximation to % in span²E³. We also have Bessel's inequality )% V ) )% ) or equivalently,

(º%Á " »( )%)

(13.3)

~

with equality if and only if % span²E³.

The Countably Infinite-Dimensional Case In the countably infinite case, we will be dealing with infinite sums and so questions of convergence will arise. Thus, we begin with the following. Theorem 13.17 Let E ~ ¸" Á " Á Ã ¹ be a countably infinite orthonormal set in a Hilbert space / . The series B

(13.4)

" ~

converges in / if and only if the series B

( (

(13.5)

~

converges in s. If these series converge, then they converge unconditionally (that is, any series formed by rearranging the order of the terms also converges). Finally, if the series (13.4) converges, then B

B

i " i ~ ( ( ~

~

Proof. Denote the partial sums of the first series by the second series by . Then for )

c

)

and the partial sums of

~ i " i ~ ( ( ~ ( c ( ~b

~b

Hence ² ³ is a Cauchy sequence in / if and only if ² ³ is a Cauchy sequence in s. Since both / and s are complete, ² ³ converges if and only if ² ³ converges. If the series (13.5) converges, then it converges absolutely and hence unconditionally. (A real series converges unconditionally if and only if it

338

Advanced Linear Algebra

converges absolutely.³ But if (13.5) converges unconditionally, then so does (13.4). The last part of the theorem follows from the continuity of the norm. Now let E ~ ¸" Á " Á Ã ¹ be a countably infinite orthonormal set in / . The Fourier expansion of a vector % / is defined to be the sum B

% V ~ º%Á " »"

(13.6)

~

To see that this sum converges, observe that for any , (13.3) gives

(º%Á " »( )%) ~

and so B

(º%Á " »( )%) ~

which shows that the series on the left converges. Hence, according to Theorem 13.17, the Fourier expansion (13.6) converges unconditionally. Moreover, since the inner product is continuous, º% c %Á V " » ~ º%Á " » c º%Á V " » ~ and so % c % V ´span²E³µ ~ ´cspan²E³µ . Hence, % V is the best approximation to % in cspan²E³. Finally, since % c % V% V, we again have )% V ) ~ ) % ) c )% c % V) )%)

and so )% V ) )% ) with equality if and only if % ~ % V, which happens if and only if % cspan²E³. Thus, the following analog of Theorem 13.16 holds. Theorem 13.18 Let E ~ ¸" Á " Á Ã ¹ be a countably infinite orthonormal set in a Hilbert space / . For any % / , the Fourier expansion B

% V ~ º%Á " »" ~

of % converges unconditionally and is the best approximation to % in cspan²E³. We also have Bessel's inequality )% V ) )% )

Hilbert Spaces

339

or equivalently, B

(º%Á " »( )%) ~

with equality if and only if % cspan²E³.

The Arbitrary Case To discuss the case of an arbitrary orthonormal set E ~ ¸" 2¹, let us first define and discuss the concept of the sum of an arbitrary number of terms. (This is a bit of a digression, since we could proceed without all of the coming details c but they are interesting.³ Definition Let A ~ ¸% 2¹ be an arbitrary family of vectors in an inner product space = . The sum % 2

is said to converge to a vector % = and we write % ~ %

(13.7)

2

if for any , there exists a finite set : 2 for which ; :Á ; finite ¬ i % c %i

;

For those readers familiar with the language of convergence of nets, the set F ²2³ of all finite subsets of 2 is a directed set under inclusion (for every (Á ) F ²2³ there is a * F ²2³ containing ( and ) ) and the function : ¦ % :

is a net in / . Convergence of ²13.7) is convergence of this net. In any case, we will refer to the preceding definition as the net definition of convergence. It is not hard to verify the following basic properties of net convergence for arbitrary sums. Theorem 13.19 Let A ~ ¸% 2¹ be an arbitrary family of vectors in an inner product space = . If % ~ % and & ~ & 2

then

2

340

Advanced Linear Algebra

1) (Linearity) ²% b & ³ ~ % b & 2

for any Á 2) (Continuity) º% Á &» ~ º%Á &» and º&Á % » ~ º&Á %» 2

2

The next result gives a useful “Cauchy-type” description of convergence. Theorem 13.20 Let A ~ ¸% 2¹ be an arbitrary family of vectors in an inner product space = . 1) If the sum % 2

converges, then for any , there exists a finite set 0 2 such that 1 q 0 ~ JÁ 1 finite ¬ i % i 1

2) If = is a Hilbert space, then the converse of 1) also holds. Proof. For part 1), given , let : 2 , : finite, be such that ; :Á ; finite ¬ i% c %i ;

2

If 1 q : ~ J, 1 finite, then i % i ~ i²% b % c %³ c ²% c %³i 1

1

:

:

i% c %i b i% c %i 1 r:

:

b ~ 2 2

As for part 2), for each , let 0 2 be a finite set for which 1 q 0 ~ JÁ 1 finite ¬ i% i 1

and let & ~ % 0

Hilbert Spaces

341

Then ²& ³ is a Cauchy sequence, since )& c & ) ~ i % c % i ~ i % c % i 0

0

0 c0

0 c0

i % i b i % i b ¦ 0 c0 0 c0

Since = is assumed complete, we have ²& ³ ¦ & . Now, given , there exists an 5 such that 5 ¬ )& c &) ~ i% c &i 0

2

Setting ~ max¸5 Á °¹ gives for ; 0 Á ; finite, i % c & i ~ i % c & b % i ;

0

; c0

i % c & i b i % i 0

; c0

b

and so 2 % converges to &. The following theorem tells us that convergence of an arbitrary sum implies that only countably many terms can be nonzero so, in some sense, there is no such thing as a nontrivial uncountable sum. Theorem 13.21 Let A ~ ¸% 2¹ be an arbitrary family of vectors in an inner product space = . If the sum % 2

converges, then at most a countable number of terms % can be nonzero. Proof. According to Theorem 13.20, for each , we can let 0 2 , 0 finite, be such that 1 q 0 ~ JÁ 1 finite ¬ i % i 1

Let 0 ~ 0 . Then 0 is countable and ¤ 0 ¬ ¸¹ q 0 ~ J for all ¬ )% )

for all ¬ % ~

342

Advanced Linear Algebra

Here is the analog of Theorem 13.17. Theorem 13.22 Let E ~ ¸" 2¹ be an arbitrary orthonormal family of vectors in a Hilbert space / . The two series " and ( ( 2

2

converge or diverge together. If these series converge, then i " i ~ ( ( 2

2

Proof. The first series converges if and only if for every , there exists a finite set 0 2 such that

1 q 0 ~ JÁ 1 finite ¬ i " i 1

or equivalently, 1 q 0 ~ JÁ 1 finite ¬ ( ( 1

and this is precisely what it means for the second series to converge. We leave proof of the remaining statement to the reader. The following is a useful characterization of arbitrary sums of nonnegative real terms. Theorem 13.23 Let ¸ 2¹ be a collection of nonnegative real numbers. Then ~ sup 1 finite 1 2

2

1

provided that either of the preceding expressions is finite. Proof. Suppose that sup ~ 9 B

1 finite 1 2 1

Then, for any , there exists a finite set : 2 such that 9 9 c :

(13.8)

Hilbert Spaces

343

Hence, if ; 2 is a finite set for which ; : , then since , 9 9 c ;

:

and so i9 c i ;

which shows that converges to 9. Finally, if the sum on the left of (13.8) converges, then the supremum on the right is finite and so (13.8) holds. The reader may have noticed that we have two definitions of convergence for countably infinite series: the net version and the traditional version involving the limit of partial sums. Let us write B

% and % ob

~

for the net version and the partial sum version, respectively. Here is the relationship between these two definitions. Theorem 13.24 Let / be a Hilbert space. If % / , then the following are equivalent: 1) 2)

% converges (net version) to %

ob B

% converges unconditionally to %

~

Proof. Assume that 1) holds. Suppose that is any permutation of ob . Given any , there is a finite set : ob for which ; :Á ; finite ¬ i % c %i ;

Let us denote the set of integers ¸Á Ã Á ¹ by 0 and choose a positive integer such that ²0 ³ : . Then for we have ²0 ³ ²0 ³ : ¬ i %²³ c %i ~ % c % ~ ²0 ³ and so 2) holds.

344

Advanced Linear Algebra

Next, assume that 2) holds, but that the series in 1) does not converge. Then there exists an such that for any finite subset 0 ob , there exists a finite subset 1 with 1 q 0 ~ J for which i % i 1

From this, we deduce the existence of a countably infinite sequence 1 of mutually disjoint finite subsets of ob with the property that max²1 ³ ~ 4 b ~ min²1b ³ and i % i 1

Now we choose any permutation ¢ ob ¦ ob with the following properties 1) ²´ Á 4 µ³ ´ Á 4 µ 2) if 1 ~ ¸Á Á Ã Á Á" ¹, then ² ³ ~ Á Á ² b ³ ~ Á2 Á Ã Á ² b " c ³ ~ Á" The intention in property 2) is that for each , takes a set of consecutive integers to the integers in 1 . For any such permutation , we have b" c

i %²³ i ~ i % i ~

1

which shows that the sequence of partial sums of the series B

%²³ ~

is not Cauchy and so this series does not converge. This contradicts 2) and shows that 2) implies at least that 1) converges. But if 1) converges to & / , then since 1) implies 2) and since unconditional limits are unique, we have & ~ %. Hence, 2) implies 1). Now we can return to the discussion of Fourier expansions. Let E ~ ¸" 2¹ be an arbitrary orthonormal set in a Hilbert space / . Given any % / , we may apply Theorem 13.16 to all finite subsets of E , to deduce

Hilbert Spaces

345

that sup (º%Á " »( )%)

1 finite 1 2

1

and so Theorem 13.23 tells us that the sum (º%Á " »( 2

converges. Hence, according to Theorem 13.22, the Fourier expansion % V ~ º%Á " »" 2

of % also converges and )% V) ~ (º%Á " »( 2

Note that, according to Theorem 13.21, % V is a countably infinite sum of terms of the form º%Á " »" and so is in cspan²E³. The continuity of infinite sums with respect to the inner product (Theorem 13.19) implies that º% c %Á V " » ~ º%Á " » c º%Á V " » ~ and so % c % V ´span²E³µ ~ ´cspan²E³µ . Hence, Theorem 3.9 tells us that % V is the best approximation to % in cspan²E³. Finally, since % c % V% V, we again have )% V ) ~ ) % ) c )% c % V) )%)

and so )% V ) )% ) with equality if and only if % ~ % V, which happens if and only if % cspan²E³. Thus, we arrive at the most general form of a key theorem about Hilbert spaces. Theorem 13.25 Let E ~ ¸" 2¹ be an orthonormal family of vectors in a Hilbert space / . For any % / , the Fourier expansion % V ~ º%Á " »" 2

of % converges in / and is the unique best approximation to % in cspan²E³. Moreover, we have Bessel's inequality )% V ) )% )

346

Advanced Linear Algebra

or equivalently, (º%Á " »( )%) 2

with equality if and only if % cspan²E³.

A Characterization of Hilbert Bases Recall from Theorem 13.15 that an orthonormal set E ~ ¸" 2¹ in a Hilbert space / is a Hilbert basis if and only if cspan²E³ ~ / Theorem 13.25, then leads to the following characterization of Hilbert bases. Theorem 13.26 Let E ~ ¸" 2¹ be an orthonormal family in a Hilbert space / . The following are equivalent: 1) E is a Hilbert basis (a maximal orthonormal set) 2) E ~ ¸¹ 3) E is total (that is, cspan²E³ ~ / ) 4) % ~ % V for all % / 5) Equality holds in Bessel's inequality for all % / , that is, ) % ) ~ )% V) for all % / 6) Parseval's identity º%Á &» ~ º%Á V V&» holds for all %Á & / , that is, º%Á &» ~ º%Á " »º&Á " » 2

Proof. Parts 1), 2) and 3) are equivalent by Theorem 13.15. Part 4) implies part 3), since % V cspan²E³ and 3) implies 4) since the unique best approximation of any % cspan²E³ is itself and so % ~ % V. Parts 3) and 5) are equivalent by Theorem 13.25. Parseval's identity follows from part 4) using Theorem 13.19. Finally, Parseval's identity for & ~ % implies that equality holds in Bessel's inequality.

Hilbert Dimension We now wish to show that all Hilbert bases for a Hilbert space / have the same cardinality and so we can define the Hilbert dimension of / to be that cardinality.

Hilbert Spaces

347

Theorem 13.27 All Hilbert bases for a Hilbert space / have the same cardinality. This cardinality is called the Hilbert dimension of / , which we denote by hdim²/³. Proof. If / has a finite Hilbert basis, then that set is also a Hamel basis and so all finite Hilbert bases have size dim²/³. Suppose next that 8 ~ ¸ 2¹ and 9 ~ ¸ 1 ¹ are infinite Hilbert bases for / . Then for each , we have ~ º Á » 1

where 1 is the countable set ¸ º Á » £ ¹. Moreover, since no can be orthogonal to every , we have 2 1 ~ 1 . Thus, since each 1 is countable, we have (1 ( ~ e 1 e L (2 ( ~ (2 ( 2

By symmetry, we also have (2 ( (1 ( and so the Schro¨der–Bernstein theorem implies that (1 ( ~ (2 (. Theorem 13.28 Two Hilbert spaces are isometrically isomorphic if and only if they have the same Hilbert dimension. Proof. Suppose that hdim²/ ³ ~ hdim²/ ³. Let E ~ ¸" 2¹ be a Hilbert basis for / and E ~ ¸# 2¹ a Hilbert basis for / . We may define a map ¢ / ¦ / as follows: 4 " 5 ~ # 2

2

We leave it as an exercise to verify that is a bijective isometry. The converse is also left as an exercise.

A Characterization of Hilbert Spaces We have seen that any vector space = is isomorphic to a vector space ²- ) ³ of all functions from ) to - that have finite support. There is a corresponding result for Hilbert spaces. Let 2 be any nonempty set and let M ²2³ ~ D ¢ 2 ¦ dc ( ²³( BE 2

The functions in M ²2³ are referred to as square summable functions. ²We can also define a real version of this set by replacing d by s.³ We define an inner product on M ²2³ by º Á » ~ ²³²³ 2

The proof that M ²2³ is a Hilbert space is quite similar to the proof that

348

Advanced Linear Algebra

M ~ M ²o³ is a Hilbert space and the details are left to the reader. If we define M ²2³ by ²³ ~ Á ~ F

if ~ if £

then the collection E ~ ¸ 2¹ is a Hilbert basis for M ²2³, of cardinality (2 (. To see this, observe that º Á » ~ ²³ ²³ ~ Á 2

and so E is orthonormal. Moreover, if M ²2³, then ²³ £ for only a countable number of 2 , say ¸ Á Á Ã ¹. If we define Z by B

Z ~ ² ³ ~ Z

Z

then cspan²E³ and ²³ ~ ²³ for all 2 , which implies that ~ Z . This shows that M ²2³ ~ cspan²E³ and so E is a total orthonormal set, that is, a Hilbert basis for M ²2³. Now let / be a Hilbert space, with Hilbert basis 8 ~ ¸" 2¹. We define a map ¢ / ¦ M ²2³ as follows. Since 8 is a Hilbert basis, any % / has the form % ~ º%Á " »" 2

Since the series on the right converges, Theorem 13.22 implies that the series (º%Á " »( 2

converges. Hence, another application of Theorem 13.22 implies that the following series converges: ²%³ ~ º%Á " » 2

It follows from Theorem 13.19 that is linear and it is not hard to see that it is also bijective. Notice that ²" ³ ~ and so takes the Hilbert basis 8 for / to the Hilbert basis E for M ²2³.

Hilbert Spaces

349

Notice also that

)²%³) ~ º²%³Á ²%³» ~ (º%Á " »( ~ i º%Á " »" i ~ )%) 2

2

and so is an isometric isomorphism. We have proved the following theorem. Theorem 13.29 If / is a Hilbert space of Hilbert dimension and if 2 is any set of cardinality , then / is isometrically isomorphic to M ²2³.

The Riesz Representation Theorem We conclude our discussion of Hilbert spaces by discussing the Riesz representation theorem. As it happens, not all linear functionals on a Hilbert space have the form “take the inner product withÃ ,” as in the finitedimensional case. To see this, observe that if & / , then the function & ²%³ ~ º%Á &» is certainly a linear functional on / . However, it has a special property. In particular, the Cauchy–Schwarz inequality gives, for all % / , (& ²%³( ~ (º%Á &»( )%))&) or, for all % £ , (& ²%³( )& ) )%) Noticing that equality holds if % ~ & , we have (& ²%³( ~ )& ) %£ )%)

sup

This prompts us to make the following definition, which we do for linear transformations between Hilbert spaces (this covers the case of linear functionals). Definition Let ¢ / ¦ / be a linear transformation from / to / . Then is said to be bounded if )%) B %£ )%)

sup

If the supremum on the left is finite, we denote it by ) ) and call it the norm of .

350

Advanced Linear Algebra

Of course, if ¢ / ¦ - is a bounded linear functional on / , then ( ²%³( %£ )%)

) ) ~ sup

The set of all bounded linear functionals on a Hilbert space / is called the continuous dual space, or conjugate space, of / and denoted by / i . Note that this differs from the algebraic dual of / , which is the set of all linear functionals on / . In the finite-dimensional case, however, since all linear functionals are bounded (exercise), the two concepts agree. (Unfortunately, there is no universal agreement on the notation for the algebraic dual versus the continuous dual. Since we will discuss only the continuous dual in this section, no confusion should arise.³ The following theorem gives some simple reformulations of the definition of norm. Theorem 13.30 Let ¢ / ¦ / be a bounded linear transformation. 1) ) ) ~ sup ) %) )%)~

2) ) ) ~ sup ) %) )%)

3) ) ) ~ inf¸ s ) %) )%) for all % /¹ The following theorem explains transformations.

the

importance

of bounded linear

Theorem 13.31 Let ¢ / ¦ / be a linear transformation. The following are equivalent: 1) is bounded 2) is continuous at any point % / 3) is continuous. Proof. Suppose that is bounded. Then ) % c % ) ~ ) ²% c % ³) ) ))% c % ) ¦ as % ¦ % . Hence, is continuous at % . Thus, 1) implies 2). If 2) holds, then for any & / , we have ) % c &) ~ ) ²% c & b % ³ c ²% ³) ¦ as % ¦ &, since is continuous at % and % c & b % ¦ % as & ¦ %. Hence, is continuous at any & / and 3) holds. Finally, suppose that 3) holds. Thus, is continuous at and so there exists a such that )%) ¬ ) % )

Hilbert Spaces

351

In particular, )%) ~ ¬

) %) )%)

and so )%) ~ ¬ ) % ) ~ ¬

) ² %³) ) % ) ¬ ) %) )%)

Thus, is bounded. Now we can state and prove the Riesz representation theorem. Theorem 13.32 (The Riesz representation theorem) Let / be a Hilbert space. For any bounded linear functional on / , there is a unique ' / such that ²%³ ~ º%Á ' » for all % / . Moreover, )' ) ~ ) ). Proof. If ~ , we may take ' ~ , so let us assume that £ . Hence, 2 ~ ker² ³ £ / and since is continuous, 2 is closed. Thus / ~ 2 p 2 Now, the first isomorphism theorem, applied to the linear functional ¢ / ¦ - , implies that /°2 - (as vector spaces). In addition, Theorem 3.5 implies that /°2 2 and so 2 - . In particular, dim²2 ³ ~ . For any ' 2 , we have % 2 ¬ ²%³ ~ ~ º%Á '» Since dim²2 ³ ~ , all we need do is find £ ' 2 for which ²'³ ~ º'Á '» for then ²'³ ~ ²'³ ~ º'Á '» ~ º'Á '» for all - , showing that ²%³ ~ º%Á '» for % 2 as well. But if £ ' 2 , then ' ~

²'³ ' º'Á '»

has this property, as can be easily checked. The fact that )' ) ~ ) ) has already been established.

352

Advanced Linear Algebra

Exercises 1. 2. 3.

4. 5. 6.

7.

Prove that the sup metric on the metric space *´Á µ of continuous functions on ´Á µ does not come from an inner product. Hint: let ²!³ ~ and ²!³ ~ ²! c a³°² c a³ and consider the parallelogram law. Prove that any Cauchy sequence that has a convergent subsequence must itself converge. Let = be an inner product space and let ( and ) be subsets of = . Show that a) ( ) ¬ ) ( b) ( is a closed subspace of = c) ´cspan²(³µ ~ ( Let = be an inner product space and : = . Under what conditions is : ~ : ? Prove that a subspace : of a Hilbert space / is closed if and only if : ~ : . Let = be the subspace of M consisting of all sequences of real numbers with the property that each sequence has only a finite number of nonzero terms. Thus, = is an inner product space. Let 2 be the subspace of = consisting of all sequences % ~ ²% ³ in = with the property that '% ° ~ . Show that 2 is closed, but that 2 £ 2 . Hint: For the latter, show that 2 ~ ¸¹ by considering the sequences " ~ ²Á Ã Á cÁ Ã ³, where the term c is in the th coordinate position. Let E ~ ¸" Á " Á Ã ¹ be an orthonormal set in / . If % ~ ' " converges, show that B

)%) ~ ( ( ~

8.

Prove that if an infinite series B

% ~

9.

converges absolutely in a Hilbert space / , then it also converges in the sense of the “net” definition given in this section. Let ¸ 2¹ be a collection of nonnegative real numbers. If the sum on the left below converges, show that ~ sup 2

1 finite 1 2 1

10. Find a countably infinite sum of real numbers that converges in the sense of partial sums, but not in the sense of nets. 11. Prove that if a Hilbert space / has infinite Hilbert dimension, then no Hilbert basis for / is a Hamel basis. 12. Prove that M ²2³ is a Hilbert space for any nonempty set 2 .

Hilbert Spaces

353

13. Prove that any linear transformation between finite-dimensional Hilbert spaces is bounded. 14. Prove that if / i , then ker² ³ is a closed subspace of / . 15. Prove that a Hilbert space is separable if and only if hdim²/³ L . 16. Can a Hilbert space have countably infinite Hamel dimension? 17. What is the Hamel dimension of M ²o³? 18. Let and be bounded linear operators on / . Verify the following: a) ) ) ~ (() ) b) ) b ) ) ) b ) ) c) )) ) ))) 19. Use the Riesz representation theorem to show that / i / for any Hilbert space / .

Chapter 14

Tensor Products

In the preceding chapters, we have seen several ways to construct new vector spaces from old ones. Two of the most important such constructions are the direct sum < l = and the vector space B²< Á = ³ of all linear transformations from < to = . In this chapter, we consider another very important construction, known as the tensor product.

Universality We begin by describing a general type of universality that will help motivate the definition of tensor product. Our description is strongly related to the formal notion of a universal pair in category theory, but we will be somewhat less formal to avoid the need to formally define categorical concepts. Accordingly, the terminology that we shall introduce is not standard, but does not contradict any standard terminology. Referring to Figure 14.1, consider a set ( and two functions and , with domain (. f A S

W

g X Figure 14.1

Suppose that there exists a function ¢ : ¦ ? for which this diagram commutes, that is, ~ k This is sometimes expressed by saying that can be factored through . What does this say about the relationship between the functions and ?

356

Advanced Linear Algebra

Let us think of the “information” about ( contained in a function ¢ ( ¦ ) as the way in which distinguishes elements of ( using labels from ) . The relationship above implies that ²³ £ ²³ ¬ ²³ £ ²³ and this can be phrased by saying that whatever ability has to distinguish elements of ( is also possessed by . Put another way, except for labeling differences, any information about ( that is contained in is also contained in . If happens to be injective, then the only difference between and is the values of the labels. That is, the two functions have the same information about (. However, in general, is not required to be injective and so may contain more information than . Now consider a family I of sets and a family < ~ ¸¢ ( ¦ ? ? I ¹ Assume that : I and ¢ ( ¦ : < . If the diagram in Figure 14.1 commutes for all < , then the information contained in every function in < is also contained in . Moreover, since < , the function cannot contain more information than is contained in the entire family and so we conclude that contains exactly the same information as is contained in the entire family < . In this sense, ¢ ( ¦ : is universal among all functions ¢ ( ¦ ? in < . In this way, a single function ¢ ( ¦ : , or more precisely, a single pair ²:Á ³, can capture a mathematical concept as described by a family of functions. Some examples from linear algebra are basis for a vector space, quotient space, direct sum and bilinearity (as we will see). Let us make a formal definition. Definition Referring to Figure 14.2, let ( be a set and let I be a family of sets. Let < ~ ¸¢ ( ¦ ? ? I ¹ be a family of functions, all of which have domain ( and range a member of I . Let > ~ ¸ ¢ ? ¦ @ ?Á @ I ¹ be a family of functions with domain and range in I . We assume that > has the following structure: 1) > contains the identity function : for each member of I .

Tensor Products

357

2) > is closed under composition of functions, which is an associative operation. 3) For any > and < , the composition k is defined and belongs to <.

S1 A

f1 f2

S2

f3

W1

W2

W3 S3

Figure 14.2 We refer to > as the measuring family and its members as measuring functions. A pair ²:Á ¢ ( ¦ :³, where : I and < has the universal property for the family < as measured by >, or is a universal pair for ²< Á >³, if for every ¢ ( ¦ ? in < , there is a unique ¢ : ¦ ? in > for which the diagram in Figure 14.1 commutes, that is, for which ~ k or equivalently, any < can be factored through . The unique measuring function is called the mediating morphism for . Note the requirement that the mediating morphism be unique. Universal pairs are essentially unique, as the following describes. Theorem 14.1 Let ²:Á ¢ ( ¦ :³ and ²; Á ¢ ( ¦ ; ³ be universal pairs for ²< Á >³. Then there is a bijective measuring function > for which : ~ ; . In fact, the mediating morphism of with respect to and the mediating morphism of with respect to are isomorphisms. Proof. With reference to Figure 14.3, there are mediating morphisms ¢ : ¦ ; and ¢ ; ¦ : for which ~ k ~k Hence, ~ ² k ³ k ~ ² k ³ k However, referring to the third diagram in Figure 14.3, both k ¢ : ¦ : and the identity map ¢ : ¦ : are mediating morphisms for and so the uniqueness

358

Advanced Linear Algebra

of mediating morphisms implies that k ~ . Similarly k ~ and so and are inverses of one another, making the desired bijection. f

A

S W

g

g

A

T V

f

T

f

A

S

S VW L

f S

Figure 14.3

Examples of Universality Now let us look at some examples of the universal property. Let Vect²- ³ denote the family of all vector spaces over the base field - . (We use the term family informally to represent what in set theory is formally referred to as a class. A class is a “collection” that is too large to be considered a set. For example, Vect²- ³ is a class.) Example 14.1 (Bases) Let 8 be a nonempty set and let 1) I ~ Vect²- ³ 2) < ~ set functions from 8 to members of < 3) > ~ linear transformations If =8 is a vector space with basis 8 , then the pair ²=8 Á ¢ 8 ¦ =8 ³, where is the inclusion map # ~ #, is universal for ²< Á >³. To see this, note that the condition that < can be factored through , ~ k is equivalent to the statement that # ~ # for each basis vector # 8 . But this uniquely defines a linear transformation . In fact, the universality of the pair ²=8 Á ³ is precisely the statement that a linear transformation is uniquely determined by assigning its values arbitrarily on a basis 8 , the function doing the arbitrary assignment in this context. Note also that Theorem 14.1 implies that if ²> Á ¢ 8 ¦ > ³ is also universal for ²< Á >³, then there is a bijective mediating morphism from =8 to > , that is, > and =8 are isomorphic. Example 14.2 (Quotient spaces and canonical projections) Let = be a vector space and let 2 be a subspace of = . Let 1) I ~ Vect²- ³

Tensor Products

359

2) < ~ linear maps with domain = , whose kernels contain 2 3) > ~ linear transformations Theorem 3.4 says precisely that the pair ²= °2Á ¢ = ¦ = °2³, where is the canonical projection map, has the universal property for < as measured by >. Example 14.3 (Direct sums) Let < and = be vector spaces over - . Let 1) I ~ Vect²- ³ 2) < ~ ordered pairs ² ¢ < ¦ > Á ¢ = ¦ > ³ of linear transformations 3) > ~ linear transformations Here we have a slight variation on the definition of universal pair: In this case, < is a family of pairs of functions. For > and ² Á ³ < , we set k ² Á ³ ~ ² k Á k ³ Then the pair ²< ^ = Á ² Á ³¢ ²< Á = ³ ¦ < ^ = ³, where " ~ ²"Á ³

and # ~ ²Á #³

are called the canonical injections, has the universal property for ²< Á >³. To see this, observe that for any pair ² Á ³¢ ²< Á = ³ ¦ > in < , the condition ² Á ³ ~ k ² Á ³ is equivalent to ² Á ³ ~ ² k Á k ³ or ²"Á ³ ~ ²"³

and

²Á #³ ~ ²#³

But these conditions define a unique linear transformation ¢ < ^ = ¦ > . Thus, bases, quotient spaces and direct sums are all examples of universal pairs and it should be clear from these examples that the notion of universal property is, well, universal. In fact, it happens that the most useful definition of tensor product is through a universal property, which we now explore.

Bilinear Maps The universality that defines tensor products rests on the notion of a bilinear map. Definition Let < , = and > be vector spaces over - . Let < d = be the cartesian product of < and = as sets. A set function ¢ < d = ¦ >

360

Advanced Linear Algebra

is bilinear if it is linear in both variables separately, that is, if ²" b "Z Á #³ ~ ²"Á #³ b ²"Z Á #³ and ²"Á # b #Z ³ ~ ²"Á #³ b ²"Á #Z ³ The set of all bilinear functions from < d = to > is denoted by hom- ²< Á = Â > ³. A bilinear function ¢ < d = ¦ - with values in the base field - is called a bilinear form on < d = . Note that bilinearity can also be expressed in matrix language as follows: If ~ ² Á Ã Á ³ - Á

~ ² Á Ã Á ³ -

" ~ ²" Á Ã Á " ³ < Á

# ~ ²# Á Ã Á # ³ =

and

then ¢ < d = ¦ > is bilinear if ²"! Á #! ³ ~ - ! where - ~ ´ ²" Á # ³µÁ . It is important to emphasize that, in the definition of bilinear function, < d = is the cartesian product of sets, not the direct product of vector spaces. In other words, we do not consider any algebraic structure on < d = when defining bilinear functions, so expressions like ²%Á &³ b ²'Á $³

and ²%Á &³

are meaningless. In fact, if = is a vector space, there are two classes of functions from = d = to > : the linear maps B²= d = Á > ³, where = d = ~ = ^ = is the direct product of vector spaces, and the bilinear maps hom²= Á = Â > ³, where = d = is just the cartesian product of sets. We leave it as an exercise to show that these two classes have only the zero map in common. In other words, the only map that is both linear and bilinear is the zero map. We made a thorough study of bilinear forms on a finite-dimensional vector space = in Chapter 11 (although this material is not assumed here). However, bilinearity is far more important and far-reaching than its application to metric vector spaces, as the following examples show. Indeed, both multiplication and evaluation are bilinear. Example 14.4 (Multiplication is bilinear) If ( is an algebra, the product map ¢ ( d ( ¦ ( defined by

Tensor Products

361

²Á ³ ~ is bilinear, that is, multiplication is linear in each position. Example 14.5 (Evaluation is bilinear) If = and > are vector spaces, then the evaluation map ¢ B²= Á > ³ d = ¦ > defined by ² Á #³ ~ # is bilinear. In particular, the evaluation map ¢ = i d = ¦ - defined by ² Á #³ ~ # is a bilinear form on = i d = . Example 14.6 If = and > are vector spaces, and = i and > i , then the product map ¢ = d > ¦ - defined by ²#Á $³ ~ ²#³²$³ is bilinear. Dually, if # = and $ > , then the map ¢ = i d > i ¦ defined by ² Á ³ ~ ²#³²$³ is bilinear. It is precisely the tensor product that will allow us to generalize the previous example. In particular, if B²< Á > ³ and B²= Á > ³, then we would like to consider a “product” map ¢ < d = ¦ > defined by ²"Á #³ ~ ²"³ ? ²#³ The tensor product n is just the thing to replace the question mark, because it has the desired bilinearity property, as we will see. In fact, the tensor product is bilinear and nothing else, so it is exactly what we need!

Tensor Products Let < and = be vector spaces. Our guide for the definition of the tensor product < n = will be the desire to have a universal property for bilinear functions, as measured by linearity. Referring to Figure 14.4, we want to define a vector space ; and a bilinear map !¢ < d = ¦ ; so that any bilinear map with domain < d = can be factored through !. Intuitively speaking, ! is the most “general” or “universal” bilinear map with domain < d = : It is bilinear and nothing more.

362

Advanced Linear Algebra

U

t

V f

bilinear

T W

linear

bilinear

W Figure 14.4 Definition Let < d = be the cartesian product of two vector spaces over - . Let I ~ Vect²- ³. Let < ~ ¸hom- ²< Á = Â > ³ > I ¹ >

be the family of all bilinear maps from < d = to any vector space > . The measuring family > is the family of all linear transformations. A pair ²; Á !¢ < d = ¦ ; ³ is universal for bilinearity if it is universal for ²< Á >³, that is, if for every bilinear map ¢ < d = ¦ > , there is a unique linear transformation ¢ ; ¦ > for which ~ k! The map is called the mediating morphism for . We can now define the tensor product via this universal property. Definition Let < and = be vector spaces over a field - . Any universal pair ²; Á !¢ < d = ¦ ; ³ for bilinearity is called a tensor product of < and = . The vector space ; is denoted by < n = and sometimes referred to by itself as the tensor product. The map ! is called the tensor map and the elements of < n = are called tensors. It is customary to use the symbol n to denote the image of any ordered pair ²"Á #³ under the tensor map, that is, " n # ~ !²"Á #³ for any " < and # = . A tensor of the form " n # is said to be decomposable, that is, the decomposable tensors are the images under the tensor map. Since universal pairs are unique up to isomorphism, we may refer to “the” tensor product of vector spaces. Note also that the tensor product n is not a product in the sense of a binary operation on a set. In fact, even when = ~ < , the tensor product " n " is not in < , but rather in < n < .

Tensor Products

363

As we will see, there are other, more constructive ways to define the tensor product. Since we have adopted the universal pair definition, the other ways to define tensor product are, for us, constructions rather than definitions. Let us examine some of these constructions.

Construction I: Intuitive but Not Coordinate Free The universal property for bilinearity captures the essence of bilinearity and the tensor map is the most “general” bilinear function on < d = . To see how this universality can be achieved in a constructive manner, let ¸ 0¹ be a basis for < and let ¸ 1 ¹ be a basis for = . Then a bilinear map ! on < d = is uniquely determined by assigning arbitrary values to the “basis” pairs ² Á ³ and extending by bilinearity, that is, if " ~ and # ~ , then !²"Á #³ ~ !4 Á 5 ~ !² Á ³ Now, the tensor map !, being the most general bilinear map, must do this and nothing more. To achieve this goal, we define the tensor map ! on the pairs ² Á ³ in such a way that the images !² Á ³ do not interact, and then extend by bilinearity. In particular, for each ordered pair ² Á ³, we invent a new formal symbol, written n , and define ; to be the vector space with basis : ~ ¸ n 0Á 1 ¹ The tensor map is defined by setting !² Á ³ ~ n and extending by bilinearity. Thus, !²"Á #³ ~ !4 Á 5 ~ ² n ³ To see that the pair ²; Á !³ is the tensor product of < and = , if ¢ < d = ¦ > is bilinear, the universality condition ~ k ! is equivalent to ² n ³ ~ ² Á ³ which does indeed uniquely define a linear map ¢ ; ¦ > . Hence, ²; Á !³ has the universal property for bilinearity and so we can write ; ~ < n = and refer to ! as the tensor map. Note that while the set : ~ ¸ n ¹ is a basis for ; (by definition), the set ¸" n # " < Á # = ¹ of decomposable tensors spans ; , but is not linearly independent. This does cause some initial confusion during the learning process. For example, one cannot define a linear map on < n = by assigning values arbitrarily to the decomposable tensors, nor is it always easy to tell when a tensor " n # is

364

Advanced Linear Algebra

equal to . We will consider the latter issue in some detail a bit later in the chapter. The fact that : is a basis for < n = gives the following. Theorem 14.2 For finite-dimensional vector spaces < and = , dim²< n = ³ ~ dim²< ³ h dim²= ³

Construction II: Coordinate Free The previous construction of the tensor product is reasonably intuitive, but has the disadvantage of not being coordinate free. The following approach does not require the choice of a basis. Let -< d= be the vector space over - with basis < d = . Let : be the subspace of -< d= generated by all vectors of the form ²"Á $³ b ²#Á $³ c ²" b #Á $³

(14.1)

²"Á #³ b ²"Á $³ c ²"Á # b $³

(14.2)

and

where Á - and "Á # and $ are in the appropriate spaces. Note that these vectors are precisely what we must “identify” as the zero vector in order to enforce bilinearity. Put another way, these vectors are if the ordered pairs are replaced by tensors according to our previous construction. Accordingly, the quotient space < n=

-< d= :

is also sometimes taken as the definition of the tensor product of < and = . (Strictly speaking, we should not be using the symbol < n = until we have shown that this is the tensor product.) The elements of < n = have the form 4 ²" Á # ³5 b : ~ ²" Á # ³ b : ! However, since ²"Á #³ c ²"Á #³ : and ²"Á #³ c ²"Á #³ : , we can absorb the scalar in either coordinate, that is, ´²"Á #³ b :µ ~ ²"Á #³ b : ~ ²"Á #³ b : and so the elements of < n = can be written simply as ´²" Á # ³ b :µ It is customary to denote the coset ²"Á #³ b : by " n #, and so any element of

Tensor Products

365

< n = has the form " n # as in the previous construction. The tensor map !¢ < d = ¦ < n = is defined by !²"Á #³ ~ " n # ~ ²"Á #³ b : This map is bilinear, since !²" b #Á $³ ~ ²" b #Á $³ b : ~ ´²"Á $³ b ²#Á $³µ b : ~ ´²"Á $³ b :µ b ´ ²#Á $³ b :µ ~ !²"Á $³ b !²#Á $³ and similarly for the second coordinate. We next prove that the pair ²< n = Á !¢ < d = ¦ < n = ³ is universal for bilinearity when < n = is defined as a quotient space -< d= °: . Theorem 14.3 Let < and = be vector spaces. The pair ²< n = Á !¢ < d = ¦ < n = ³ is the tensor product of < and = . Proof. Consider the diagram in Figure 14.5. Here -< d= is the vector space with basis < d = .

t U

j

V f

FU

S

U

V

V

V

W

W Figure 14.5 Since k ²"Á #³ ~ ²"Á #³ ~ ²"Á #³ b : ~ " n # ~ !²"Á #³ we have !~k The universal property of vector spaces described in Example 14.1 implies that

366

Advanced Linear Algebra

there is a unique linear transformation ¢ -< d= ¦ > for which k~ Note that sends the vectors (14.1) and (14.2) that generate : to the zero vector and so : ker²³. For example, ´²"Á $³ b ²#Á $³ c ²" b #Á $³µ ~ ´²"Á $³ b ²#Á $³ c ²" b #Á $³µ ~ ²"Á $³ b ²#Á $³ c ²" b #Á $³ ~ ²"Á $³ b ²#Á $³ c ²" b #Á $³ ~ and similarly for the second coordinate. Hence, Theorem 3.4 (the universal property described in Example 14.2) implies that there exists a unique linear transformation ¢ < n = ¦ > for which k ~ Hence, k!~ kk~k~ As to uniqueness, if Z k ! ~ , then Z ´²"Á #³ b :µ ~ ²"Á #³ ~ ´²"Á #³ b :µ and since the cosets ²"Á #³ b : generate -< d= °: , we conclude that Z ~ . Thus, is the mediating morphism and ²< n = Á !³ is universal for bilinearity. Let us take a moment to compare the two previous constructions. Let ¸ 0¹ and ¸ 1 ¹ be bases for < and = , respectively. Let ²; Z Á !Z ³ be the tensor product as constructed using these two bases and let ²; Á !³ ~ ²-< d= °:Á !³ be the tensor product construction using quotient spaces. Since both of these pairs are universal for bilinearity, Theorem 14.1 implies that the mediating morphism for ! with respect to !Z , that is, the map ¢ ; Z ¦ ; defined by ² n ³ ~ ² Á ³ b : is a vector space isomorphism. Therefore, the basis ¸² n ³¹ of ; Z is sent to the set ¸² Á ³ b :¹, which is therefore a basis for ; . In other words, given any two bases ¸ 0¹ and ¸ 1 ¹ for < and = , respectively, the tensors n form a basis for < n = , regardless of which construction of the tensor product we use. Therefore, we are free to think of n either as a formal symbol belonging to a basis for < n = or as the coset ² Á ³ b : belonging to a basis for < n = .

Tensor Products

367

Bilinearity on < d = Equals Linearity on < n = The universal property for bilinearity says that to each bilinear function ¢ < d = ¦ > , there corresponds a unique linear function ¢ < n = ¦ > , called the mediating morphism for . Thus, we can define the mediating morphism map ¢ hom²< Á = Â > ³ ¦ B²< n = Á > ³ by setting ~ . In other words, is the unique linear map for which ² ³²" n #³ ~ ²"Á #³ Observe that is itself linear, since if Á hom²< Á = Â > ³, then ´² ³ b ²³µ²" n #³ ~ ²"Á #³ b ²"Á #³ ~ ² b ³²"Á #³ and so ² ³ b ²³ is the mediating morphism for b , that is, ² ³ b ²³ ~ ² b ³ Also, is surjective, since if ¢ < n = ¦ > is any linear map, then ~ k !¢ < d = ¦ > is bilinear and has mediating morphism , that is, ~ . Finally, is injective, for if ~ , then ~ k ! ~ . We have established the following result. Theorem 14.4 Let < , = and > be vector spaces over - . Then the mediating morphism map ¢ hom²< Á = Â > ³ ¦ B²< n = Á > ³, where is the unique linear map satisfying ~ k !, is an isomorphism and so ¢ hom²< Á = Â > ³ B²< n = Á > ³

When Is a Tensor Product Zero? Armed with the universal property of bilinearity, we can now discuss some of the basic properties of tensor products. Let us first consider the question of when a tensor " n # is zero. The bilinearity of the tensor product gives n # ~ ² b ³ n # ~ n # b n # and so n # ~ . Similarly, " n ~ . Now suppose that " n # ~

where we may assume that none of the vectors " and # are . Let ¢ < d = ¦ > be a bilinear map and let ¢ < n = ¦ > be its mediating morphism, that is, k ! ~ . Then

368

Advanced Linear Algebra

~ 4 " n # 5 ~ ² k !³²" Á # ³ ~ ²" Á # ³

The key point is that this holds for any bilinear function ¢ < d = ¦ > . In particular, let < i and = i and define by ²"Á #³ ~ ²"³ ²#³ which is easily seen to be bilinear. Then the previous display becomes ²" ³ ²# ³ ~

If, for example, the vectors " are linearly independent, we can take to be a dual vector "i to get ~ "i ²" ³ ²# ³ ~ ²# ³

and since this holds for all linear functionals = i , it follows that # ~ . We have proved the following useful result. Theorem 14.5 If " Á Ã Á " are linearly independent vectors in < and # Á Ã Á # are arbitrary vectors in = , then " n # ~

¬

# ~ for all

In particular, " n # ~ if and only if " ~ or # ~ .

Coordinate Matrices and Rank If 8 ~ ¸" 0¹ is a basis for < and 9 ~ ¸# 1 ¹ is a basis for = , then any vector ' < n = has a unique expression as a sum ' ~ Á ²" n # ³ 0 1

where only a finite number of the coefficients Á are nonzero. In fact, for a fixed ' < n = , we may reindex the bases so that

' ~ Á ²" n # ³ ~ ~

where none of the rows or columns of the matrix 9 ~ ²Á ³ consists only of 's. The matrix 9 ~ ²Á ³ is called a coordinate matrix of ' with respect to the bases 8 and 9. Note that a coordinate matrix 9 is determined only up to the order of its rows and columns. We could remove this ambiguity by considering ordered bases,

Tensor Products

369

but this is not necessary for our discussion and adds a complication, since the bases may be infinite. Suppose that M ~ ¸$ 0¹ and N ~ ¸% 1 ¹ are also bases for < and = , respectively, and that

' ~

Á ²$

n % ³

~ ~

where : ~ ² Á ³ is a coordinate matrix of ' with respect to these bases. We claim that the coordinate matrices 9 and : have the same rank, which can then be defined as the rank of the tensor ' < n = . Each $ Á Ã Á $ is a finite linear combination of basis vectors in 8 , perhaps involving some of " Á Ã Á " and perhaps involving other vectors in 8 . We can further reindex 8 so that each $ is a linear combination of the vectors 8 Z ~ ²" Á Ã Á " ³, where and set < ~ span²" Á Ã Á " ³ Next, extend ²$ Á Ã Á $ ³ to a basis M Z ~ ²$ Á Ã Á $ Á $b Á Ã Á $ ³ for < . (Since we no longer need the rest of the basis M , we have commandeered the symbols $b Á Ã Á $ , for simplicity.) Hence

$ ~ Á " for ~ Á Ã Á ~

where ( ~ ²Á ³ is invertible of size d . Now repeat this process on the second coordinate. Reindex the basis 9 so that the subspace = ~ span²# Á Ã Á # ³ contains % Á Ã Á % and extend to a basis N Z ~ ²% Á Ã Á % Á %b Á Ã Á % ³ for = . Then

% ~ Á # for ~ Á Ã Á ~

where ) ~ ²Á ³ is invertible of size d . Next, write

' ~ Á ²" n # ³ ~ ~

by setting Á ~ for or . Thus, the d matrix 9 ~ ²Á ³ comes from 9 by adding c rows of 's to the bottom and then c columns of 's. In particular, 9 and 9 have the same rank.

370

Advanced Linear Algebra

The expression for ' in terms of the basis vectors $ Á Ã Á $ and % Á Ã Á % can also be extended using coefficients to

' ~

Á ²$

n % ³

~ ~

where the d matrix : ~ ²

Á ³

has the same rank as : .

Now at last, we can compute. First, bilinearity gives

$ n % ~ Á Á " n # ~ ~

and so

' ~

Á ²$

n % ³ ~

~ ~

~ ~

Á 8 Á Á " ~ ~

~ 8 ²Á ~ ~

n # 9

~ ~

Á ³Á 9"

n #

~ 8 ²(! : ³Á Á 9" n # ~ ~

~

~ 2(! : ) 3Á " n # ~ ~

Thus

Á ²" n # ³ ~ ' ~ ²(! : )³Á ²" n # ³ ~ ~

~ ~

and so 9 ~ (! : ) . Since ( and ) are invertible, we deduce that rk²9³ ~ rk²9 ³ ~ rk²: ³ ~ rk²:³ as desired. Moreover, in block matrix terms, we can write 9 9 ~ >

?block

and

: : ~ >

! (Á i

i ? i block

and

) ) ~ > Á i

?block

and if we write (! ~ >

then 9 ~ (! : ) implies that

i i ?block

Tensor Products

371

! 9 ~ (Á :)Á

We shall soon have use for the following special case. If

' ~ " n # ~ $ n % ~

(14.3)

~

then 9 ~ : ~ 0 and so

$ ~ Á " for ~ Á Ã Á ~

and

% ~ Á # for ~ Á Ã Á ~

where if (Á ~ ²Á ³ and )Á ~ ²Á ³, then ! 0 ~ (Á )Á

The Rank of a Decomposable Tensor Recall that a tensor of the form " n # is said to be decomposable. If ¸" 0¹ is a basis for < and ¸# 1 ¹ is a basis for = , then any decomposable vector has the form " n # ~ ²" n # ³ Á

Hence, the rank of a decomposable vector is , since the rank of a matrix whose ²Á ³th entry is is .

Characterizing Vectors in a Tensor Product There are several useful representations of the tensors in < n = . Theorem 14.6 Let ¸" 0¹ be a basis for < and let ¸# 1 ¹ be a basis for = . By an “essentially unique” sum, we mean unique up to order and presence of zero terms. 1) Every ' < n = has an essentially unique expression as a finite sum of the form Á " n # Á

where Á - and the tensors " n # are distinct.

372

Advanced Linear Algebra

2) Every ' < n = has an essentially unique expression as a finite sum of the form " n &

where & = and the " 's are distinct. 3) Every ' < n = has an essentially unique expression as a finite sum of the form % n #

where % < and the # 's are distinct. 4) Every nonzero ' < n = has an expression of the form

' ~ % n & ~

where the % 's are distinct, the & 's are distinct and the sets ¸% ¹ < and ¸& ¹ = are linearly independent. As to uniqueness, is the rank of ' and so it is unique. Also, the equation

% n & ~ $ n ' ~

~

where the $ 's are distinct, the ' 's are distinct and ¸$ ¹ < and ¸' ¹ = are linearly independent, holds if and only if there exist invertible d matrices ( ~ ²Á ³ and ) ~ ²Á ³ for which (! ) ~ 0 and

$ ~ Á %

and ' ~ Á &

~

~

for ~ Á Ã Á . Proof. Part 1) merely expresses the fact that ¸" n # ¹ is a basis for < n = . From part 2), we write Á " n # ~ @" n Á # A ~ " n & Á

Uniqueness follows from Theorem 14.5. Part 3) is proved similarly. As to part 4), we start with the expression from part 2):

" n & ~

where we may assume that none of the & 's are . If the set ¸& ¹ is linearly independent, we are done. If not, then we may suppose (after reindexing if

Tensor Products

373

necessary) that c

& ~ & ~

Then

c

c

" n & ~ " n & b 8" n & 9 ~

~ c

~

c

~ " n & b " n & ~ c

~

~ ²" b " ³ n & ~

But the vectors ¸" b " c ¹ are linearly independent. This reduction can be repeated until the second coordinates are linearly independent. Moreover, the identity matrix 0 is a coordinate matrix for ' and so ~ rk²0 ³ ~ rk²'³. As to uniqueness, one direction was proved earlier; see (14.3) and the other direction is left to the reader. The proof of Theorem 14.6 shows that if ' £ and '~

n !

0

where < and ! = , then if the multiset ¸ independent, we can rewrite ' in the form '~

0¹ is not linearly

n !Z

0

where ¸ 0 ¹ is linearly independent. Then we can do the same for the second coordinate to arrive so at the representation rk²%³

' ~ % n & ~

where the multisets ¸% ¹ and ¸& ¹ are linearly independent sets. Therefore, rk²%³ (0 ( and so the rank of ' is the smallest integer for which ' can be written as a sum of decomposable tensors. This is often taken as the definition of the rank of a tensor. However, we caution the reader that there is another meaning to the word rank when applied to a tensor, namely, it is the number of indices required to write the tensor. Thus, a scalar has rank , a vector has rank , the tensor ' above has rank and a tensor of the form

374

Advanced Linear Algebra

'~

n ! n "

0

has rank .

Defining Linear Transformations on a Tensor Product One of the simplest and most useful ways to define a linear transformation on the tensor product < n = is through the universal property, for this property says precisely that a bilinear function on < d = gives rise to a unique (and well-defined) linear transformation on < n = . The proof of the following theorem illustrates this well. Theorem 14.7 Let < and = be vector spaces. There is a unique linear transformation ¢ < i n = i ¦ ²< n = ³i defined by ² n ³ ~ p where ² p ³²" n #³ ~ ²"³²#³ Moreover, is an embedding and is an isomorphism if < and = are finitedimensional. Thus, the tensor product n of linear functionals is (via this embedding) a linear functional on tensor products. Proof. Informally, for fixed and , the function ²"Á #³ ¦ ²"³²#³ is bilinear in " and # and so there is a unique linear map p taking " n # to ²"³²#³. The function ² Á ³ ¦ p is bilinear in and since ² b ³ p ~ ² p ³ b ² p ³ and so there is a unique linear map taking n to p . More formally, for fixed and , the map - Á ¢ < d = ¦ - defined by - Á ²"Á #³ ~ ²"³²#³ is bilinear and so the universal property of tensor products implies that there exists a unique p ²< n = ³i for which ² p ³²" n #³ ~ ²"³²#³ Next, the map .¢ < i d = i ¦ ²< n = ³i defined by .² Á ³ ~ p

Tensor Products

375

is bilinear since, for example, ´² b ³ p µ²" n #³ ~ ² b ³²"³ h ²#³ ~ ²"³²#³ b ²"³²#³ ~ ´² p ³ b ² p ³µ²" n #³ which shows that . is linear in its first coordinate. Hence, the universal property implies that there exists a unique linear map ¢ < i n = i ¦ ²< n = ³i for which ² n ³ ~ p To see that is an injection, if < i n = i is nonzero, then we may write in the form

~ n ~

where the < i are nonzero and ¸ ¹ = i is linearly independent. If ²³ ~ , then for any " < and # = , we have

~ ²³²" n #³ ~ ² n ³²" n #³ ~ ²"³²#³ ~

~

Hence, for each nonzero " < , the linear functional

²"³ ~

is the zero map and so the linear independence of ¸ ¹ implies that ²"³ ~ for all . Since " is arbitrary, it follows that ~ for all and so ~ . Finally, in the finite-dimensional case, the map is a bijection since dim²< i n = i ³ ~ dim²²< n = ³i ³ B

Combining the isomorphisms of Theorem 14.4 and Theorem 14.7, we have, for finite-dimensional vector spaces < and = , < i n = i ²< n = ³i hom²< Á = Â - ³

The Tensor Product of Linear Transformations We wish to generalize Theorem 14.7 to arbitrary linear transformations. Let B²< Á < Z ³ and B²= Á = Z ³. While the product ²"³²#³ does not make sense, the tensor product " n # does and is bilinear in " and #, that is, the following function is bilinear:

376

Advanced Linear Algebra

²"Á #³ ~ " n # The same argument that we used in the proof of Theorem 14.7 will work here. Namely, the map ²"Á #³ ª " n # from < d = to < Z n = Z is bilinear in " and # and so there is a unique linear map ² p ³¢ < n = ¦ < Z n = Z for which ² p ³²" n #³ ~ " n # The function ¢ B²< Á < Z ³ d B²= Á = Z ³ ¦ B²< n = Á < Z n = Z ³ defined by ² Á ³ ~ p is bilinear, since ²² b ³ p ³²" n #³ ~ ² b ³²"³ n # ~ ² " b "³ n # ~ ´ " n #µ b ´" n #µ ~ ² p ³²" n #³ b ² p ³²" n #³ ~ ²² p ³ b ² p ³³²" n #³ and similarly for the second coordinate. Hence, there is a unique linear transformation ¢ B²< Á < Z ³ n B²= Á = Z ³ ¦ B²< n = Á < Z n = Z ³ satisfying ² n ³ ~ p that is, ´ ² n ³µ²" n #³ ~ " n # To see that is injective, if B²< Á < Z ³ n B²= Á = Z ³ is nonzero, then we may write

~ n ~ Z

where the B²< Á < ³ are nonzero and the set ¸ ¹ B²= Á = Z ³ is linearly independent. If ²³ ~ , then for all " < and # = we have

~ ²³²" n #³ ~ ² n ³²" n #³ ~ ²"³ n ²#³ ~

~

Since £ , it follows that £ for some and so we may choose a " < such that ²"³ £ for some . Moreover, we may assume, by reindexing if

Tensor Products

377

necessary, that the set ¸ ²"³Á Ã Á ²"³¹ is a maximal linearly independent subset of ¸ ²"³Á Ã Á ²"³¹. Hence, for each , we have

²"³ ~ Á ²"³ ~

and so

~ ²"³ n ²#³ ~

~ ²"³ n ²#³ b @Á ²"³A n ²#³ ~

~b ~

~ ²"³ n ²#³ b Á ²"³ n ²#³! ~

~b ~

~ ²"³ n ²#³ b ²"³ n @ Á ²#³A ~

~

~b

~ ²"³ n @ ²#³ b Á ²#³A ~

~b

Thus, the linear independence of ¸ ²"³Á Ã Á ²"³¹ implies that for each ,

²#³ b Á ²#³ ~ ~b

for all # = and so

b Á ~ ~b

But this contradicts the fact that the set ¸ ¹ is linearly independent. Hence, it cannot happen that ²³ ~ for £ and so is injective. The embedding of B²< Á < Z ³ n B²= Á = Z ³ into B²< n = Á < Z n = Z ³ means that each n can be thought of as the linear transformation p from < n = to < Z n = Z , defined by ² p ³²" n #³ ~ " n # In fact, the notation n is often used to denote both the tensor product of vectors (linear transformations) and the linear map p , and we will do this as well. In summary, we can say that the tensor product n of linear transformations is (up to isomorphism) a linear transformation on tensor products.

378

Advanced Linear Algebra

Theorem 14.8 There is a unique linear transformation ¢ B²< Á < Z ³ n B²= Á = Z ³ ¦ B²< n = Á < Z n = Z ³ defined by ² n ³ ~ p where ² p ³²" n #³ ~ " n # Moreover, is an embedding and is an isomorphism if all vector spaces are finite-dimensional. Thus, the tensor product n of linear transformations is (via this embedding) a linear transformation on tensor products. Let us note a few special cases of the previous theorem.

Corollary 14.9 Let us use the symbol ? Æ @ to denote the fact that there is an embedding of ? into @ that is an isomorphism if ? and @ are finitedimensional. 1) Taking < Z ~ - gives

< i n B²= Á = Z ³ Æ B²< n = Á = Z ³ where ² n ³²" n #³ ~ ²"³²#³ for ~ = Z )

B²< Á < Z ³ n > Æ B²< Á < Z n > ³ where ² n $³²"³ ~ " n $ 4) Taking < Z ~ - and = ~ - gives (letting > ~ = Z )

 Æ B²< Á > ³ where ² n $³²"³ ~ ²"³$

Tensor Products

379

Change of Base Field The tensor product provides a convenient way to extend the base field of a vector space that is more general than the complexification of a real vector space, discussed earlier in the book. We refer to a vector space over a field - as an - -space and write =- . Actually, there are several approaches to “upgrading” the base field of a vector space. For instance, suppose that 2 is an extension field of - , that is, - 2 . If ¸ ¹ is a basis for =- , then every % =- has the form % ~ where - . We can define a 2 -space =2 simply by taking all formal linear combinations of the form % ~ where 2 . Note that the dimension of =2 as a 2 -space is the same as the dimension of =- as an - -space. Also, =2 is an - -space (just restrict the scalars to - ) and as such, the inclusion map ¢ =- ¦ =2 sending % =- to ²%³ ~ % =2 is an - -monomorphism. The approach described in the previous paragraph uses an arbitrarily chosen basis for =- and is therefore not coordinate free. However, we can give a coordinate-free approach using tensor products as follows. Since 2 is a vector space over - , we can form the tensor product >- ~ 2 n - =It is customary to include the subscript - on n - to denote the fact that the tensor product is taken with respect to the base field - . (All relevant maps are - -bilinear and - -linear.) However, since =- is not a 2 -space, the only tensor product of 2 and =- that makes sense is the - -tensor product and so we will drop the subscript - . The tensor product >- is an - -space by definition of tensor product, but we can make it into a 2 -space as follows. For 2 , the temptation is to “absorb” the scalar into the first coordinate, ² n #³ ~ ² ³ n # but we must be certain that this is well-defined, that is, n#~n$

¬

² ³ n # ~ ² ³ n $

But for a fixed , the map ² Á #³ ª ² ³ n # is bilinear and so the universal property of tensor products implies that there is a unique linear map n # ª ² ³ n #, which we define to be scalar multiplication by .

380

Advanced Linear Algebra

To be absolutely clear, we have two distinct vector spaces: the - -space >- ~ 2 n =- defined by the tensor product and the 2 -space >2 ~ 2 n =with scalar multiplication by elements of 2 defined as absorption into the first coordinate. The spaces >- and >2 are identical as sets and as abelian groups. It is only the “permission to multiply by” that is different. Accordingly, we can recover >- from >2 simply by restricting scalar multiplication to scalars from -. Thus, we can speak of “- -linear” maps from =- into >2 , with the expected meaning, that is, ²" b #³ ~ " b # for all scalars Á - . If the dimension of 2 as a vector space over - is , then dim- ²>- ³ ~ dim- ²2 n =- ³ ~ h dim- ²=- ³ As to the dimension of >2 , it is not hard to see that if ¸ ¹ is a basis for =- , then ¸ n ¹ is a basis for >2 . Hence dim2 ²>2 ³ ~ dim- ²=- ³ The map ¢ =- ¦ >- defined by # ~ n # is easily seen to be injective and - -linear and so >- contains an isomorphic copy of =- . We can also think of as mapping =- into >2 , in which case is called the 2 -extension map of =- . This map has a universal property of its own, as described in the next theorem. Theorem 14.10 The - -linear 2 -extension map ¢ =- ¦ 2 n =- has the universal property for the family of all - -linear maps from =- into a 2 -space, as measured by 2 -linear maps. Specifically, for any - -linear map ¢ =- ¦ @ , where @ is a 2 -space, there exists a unique 2 -linear map ¢ 2 n =- ¦ @ for which the diagram in Figure 14.6 commutes, that is, for which k~ Proof. If such a 2 -linear map ¢ 2 n =- ¦ @ is to exist, then it must satisfy, for any 2 , ² n #³ ~ ² n #³ ~ ²#³ ~ ²#³ This shows that if exists, it is uniquely determined by . As usual, when searching for a linear map on a tensor product such as 2 n =- , we look for a bilinear map. The map ¢ ²2 d =- ³ ¦ @ defined by ² Á #³ ~ ²#³

Tensor Products

381

is bilinear and so there exists a unique - -linear map for which ² n #³ ~ ²#³ It is easy to see that is also 2 -linear, since if 2 , then ´² n #³µ ~ ² n #³ ~ ²#³ ~ ² n #³

P

VF

K

VF W

f Y Figure 14.6 Theorem 14.10 is the key to describing how to extend an - -linear map to a 2 linear map. Figure 14.7 shows an - -linear map ¢ = ¦ > between - -spaces = and > . It also shows the 2 -extensions for both spaces, where 2 n = and 2 n > are 2 -spaces.

W

V

W PW

PV KV

W

KW

Figure 14.7 If there is a unique 2 -linear map that makes the diagram in Figure 14.7 commute, then this would be the obvious choice for the extension of the - linear map to a 2 -linear map. Consider the - -linear map ~ ²> k ³¢ = ¦ 2 n > into the 2 -space 2 n > . Theorem 14.10 implies that there is a unique 2 -linear map ¢ 2 n = ¦ 2 n > for which k = ~ that is, k = ~ > k Now, satisfies

382

Advanced Linear Algebra

² n #³ ~ ² n #³ ~ ² k = ³²#³ ~ ²> k ³²#³ ~ ² n #³ ~ n # ~ ²2 n ³² n #³ and so ~ 2 n . Theorem 14.11 Let = and > be - -spaces, with 2 -extension maps = and > , respectively. (See Figure 14.7.) Then for any - -linear map ¢ = ¦ > , the map 2 n ¢ 2 n = ¦ 2 n > is the unique 2 -linear map that makes the diagram in Figure 14.7 commute, that is, for which k ~ ²2 n ³ k

Multilinear Maps and Iterated Tensor Products The tensor product operation can easily be extended to more than two vector spaces. We begin with the extension of the concept of bilinearity. Definition If = Á Ã Á = and > are vector spaces over - , a function ¢ = d Ä d = ¦ > is said to be multilinear if it is linear in each coordinate separately, that is, if ²" Á Ã Á "c Á # b #Z Á "b Á Ã Á " ³ ~ ²" Á Ã Á "c Á #Á "b Á Ã Á " ³ b ²" Á Ã Á "c Á #Z Á "b Á Ã Á "³ for all ~ Á Ã Á . A multilinear function of variables is also referred to as an -linear function. The set of all -linear functions as defined above will be denoted by hom²= Á Ã Á = Â > ³. A multilinear function from = d Ä d = to the base field - is called a multilinear form or -form. Example 14.7 1) If ( is an algebra, then the product map ¢ ( d Ä d ( ¦ ( defined by ² Á Ã Á ³ ~ Ä is -linear. 2) The determinant function det¢ C ¦ - is an -linear form on the columns of the matrices in C . The tensor product is defined via its universal property. Definition As pictured in Figure 14.8, let = d Ä d = be the cartesian product of vector spaces over - . A pair ²; Á !¢ = d Ä d = ¦ ; ³ is universal for multilinearity if for every multilinear map ¢ = d Ä d = ¦ > , there is a unique linear transformation ¢ ; ¦ > for which

Tensor Products

383

~ k! The map is called the mediating morphism for . If ²; Á !³ is universal for multilinearity, then ; is called the tensor product of = Á Ã Á = and denoted by = n Ä n = . The map ! is called the tensor map.

V1uuVn

t

f

V1Vn W W

Figure 14.8 As we have seen, the tensor product is unique up to isomorphism. The basis construction and coordinate-free construction given earlier for the tensor product of two vector spaces carry over to the multilinear case. In particular, let 8 ~ ¸Á 1 ¹ be a basis for = for ~ Á Ã Á . For each ordered -tuple ²Á Á Ã Á Á ³, construct a new formal symbol Á n Ä n Á and define ; to be the vector space with basis : ~ ¸Á n Ä n Á 1 ¹ The tensor map !¢ = d Ä d = ¦ ; is defined by setting !²Á Á Ã Á Á ³ ~ Á n Ä n Á and extending by multilinearity. This uniquely defines a multilinear map ! that is universal for multilinear functions from = d Ä d = . Indeed, if ¢ = d Ä d = ¦ > is multilinear, the condition ~ k ! is equivalent to ²Á n Ä n Á ³ ~ ²Á Á Ã Á Á ³ which uniquely defines a linear map ¢ ; ¦ > . Hence, ²; Á !³ has the universal property for multilinearity. Alternatively, we may take the coordinate-free quotient space approach as follows. Definition Let = Á Ã Á = be vector spaces over - and let < be the vector space with basis = d Ä d = . Let : be the subspace of < generated by all vectors of the form

384

Advanced Linear Algebra

²# Á Ã Á #c Á "Á #b Á Ã Á # ³ b ²# Á Ã Á #c Á "Z Á #b Á Ã Á #³ c ²# Á Ã Á #c Á " b "Z Á #b Á Ã Á # ³ for Á - , "Á "Z = and # = for £ . The quotient space < °: is the tensor product of = Á Ã Á = and the tensor map is the map !²# Á Ã Á # ³ ~ ²# Á Ã Á # ³ b :

As before, we denote the coset ²# Á Ã Á # ³ b : by # n Ä n # and so any element of = n Ä n = is a sum of decomposable tensors, that is, # n Ä n # where the vector space operations are linear in each variable. Here are some of the basic properties of multiple tensor products. Proof is left to the reader. Theorem 14.12 The tensor product has the following properties. Note that all vector spaces are over the same field - . 1) (Associativity) There exists an isomorphism ¢ ²= n Ä n = ³ n ²> n Ä n > ³ ¦ = n Ä n = n > n Ä n > for which ´²# n Ä n # ³ n ²$ n Ä n $ ³µ ~ # n Ä n # n $ n Ä n $ In particular, ²< n = ³ n > < n ²= n > ³ < n = n > 2) (Commutativity) Let be any permutation of the indices ¸Á Ã Á ¹. Then there is an isomorphism ¢ = n Ä n = ¦ =²³ n Ä n =²³ for which ²# n Ä n # ³ ~ #²³ n Ä n #²³ 3) There is an isomorphism ¢ - n = ¦ = for which ² n #³ ~ # and similarly, there is an isomorphism ¢ = n - ¦ = for which ²# n ³ ~ # Hence, - n = = = n - . The analog of Theorem 14.4 is the following.

Tensor Products

385

Theorem 14.13 Let = Á Ã Á = and > be vector spaces over - . Then the mediating morphism map ¢ hom²= Á Ã Á = Â > ³ ¦ B²= n Ä n = Á > ³ defined by the fact that is the unique mediating morphism for is an isomorphism. Thus, hom²= Á Ã Á = Â > ³ B²= n Ä n = Á > ³ Moreover, if all vector spaces are finite-dimensional, then

dim´hom²= Á Ã Á = Â > ³µ ~ dim²> ³ h dim²= ³

~

Theorem 14.8 and its corollary can also be extended. Theorem 14.14 The linear transformation ¢ B²< Á <Z ³ n Ä n B²< Á <Z ³ ¦ B²< n Ä n < Á <Z n Ä n <Z ³ defined by ² n Ä n ³²" n Ä n " ³ ~ " n Ä n " is an embedding and is an isomorphism if all vector spaces are finitedimensional. Thus, the tensor product n Ä n of linear transformations is (via this embedding) a linear transformation on tensor products. Two important special cases of this are

<i n Ä n <i Æ ²< n Ä n < ³i where ² n Ä n ³²" n Ä n " ³ ~ ²" ³Ä ²" ³ and

<i n Ä n <i n = Æ B²< n Ä n < Á = ³ where ² n Ä n n #³²" n Ä n " ³ ~ ²" ³Ä ²" ³#

Tensor Spaces Let = be a finite-dimensional vector space. For nonnegative integers and , the tensor product

386

Advanced Linear Algebra

; ²= ³ ~ = n Ä n = n = i n Ä n = i ~ = n n ²= i ³n factors

factors

is called the space of tensors of type ²Á ³, where is the contravariant type and is the covariant type. If ~ ~ , then ; ²= ³ ~ - , the base field. Here we use the notation = n for the -fold tensor product of = with itself. We will also write = d for the -fold cartesian product of = with itself. Since = = ii , we have ; ²= ³ ~ = n n ²= i ³n ²²= i ³n n = n ³i hom- ²²= i ³d d = d Á - ³ which is the space of all multilinear functionals on = i d Ä d = i d = dÄd= factors

factors

In fact, tensors of type ²Á ³ are often defined as multilinear functionals in this way. Note that dim²; ²= ³³ ~ ´dim²= ³µb Also, the associativity and commutativity of tensor products allows us to write b ; ²= ³ n ; ²= ³ ~ ;b ²= ³

at least up to isomorphism. Tensors of type ²Á ³ are called contravariant tensors ; ²= ³ ~ ; ²= ³ ~ = nÄn= factors

and tensors of type ²Á ³ are called covariant tensors ; ²= ³ ~ ; ²= ³ ~ =inÄn=i factors

Tensors with both contravariant and covariant indices are called mixed tensors. In general, a tensor can be interpreted in a variety of ways as a multilinear map on a cartesian product, or a linear map on a tensor product. Indeed, the interpretation we mentioned above that is sometimes used as the definition is only one possibility. We simply need to decide how many of the contravariant factors and how many of the covariant factors should be “active participants” and how many should be “passive participants.”

Tensor Products

387

More specifically, consider a tensor of type ²Á ³, written # n Ä n # n Ä n # n n Ä n n Ä n ; ²= ³ where and . Here we are choosing the first vectors and the first linear functionals as active participants. This determines the number of arguments of the map. In fact, we define a map from the cartesian product = i d Ä d = i d = dÄd= factors

factors

to the tensor product = n Ä n = n =inÄn=i c factors

c factors

of the remaining factors by ²# n Ä n # n n Ä n ³² Á Ã Á Á % Á Ã Á %³ ~ ²# ³Ä ²# ³ ²% ³Ä ²% ³#b n Ä n # n b n Ä n In words, the first group # n Ä n # of (active) vectors interacts with the first group Á Ã Á of arguments to produce the scalar ²# ³Ä ²# ³. The first group n Ä n of (active) functionals interacts with the second group % Á Ã Á % of arguments to produce the scalar ²% ³Ä ²% ³. The remaining (passive) vectors #b n Ä n # and functionals b n Ä n are just “copied” to the image tensor. It is easy to see that this map is multilinear and so there is a unique linear map from the tensor product = i n Ä n = i n = nÄn= factors

factors

to the tensor product = n Ä n = n =inÄn=i c factors

c factors

defined by ²# p Ä p # p p Ä p ³² n Ä n n % n Ä n % ³ ~ ²# ³Ä ²# ³ ²% ³Ä ²% ³#b n Ä n # n b n Ä n Moreover, the map ¢ = n n ²= i ³n ¦ B²²= i ³n n = n Á = n²c³ n ²= i ³n²c³ ³ defined by ²# n Ä n # n n Ä n ³ ~ # p Ä p # p p Ä p

388

Advanced Linear Algebra

is an isomorphism, since if # p Ä p # p p Ä p is the zero map then ²# ³Ä ²# ³ ²% ³Ä ²% ³#b n Ä n # n b n Ä n ~ for all = i and % = , which implies that # n Ä n # n n Ä n ~ As usual, we denote the map # p Ä p # p p Ä p by # n Ä n # n n Ä n Theorem 11.15 For and , ; ²= ³ B²²= i ³n n = n Á = n²c³ n ²= i ³n²c³ ³

When ~ and ~ , we get ; ²= ³ B²²= i ³n n = n Á - ³ ~ ²²= i ³n n = n ³i as before. Let us look at some special cases. For ~ we have ; ²= ³ B²²= i ³n Á = n²c³ ³ where ²# n Ä n # ³² n Ä n ³ ~ ²# ³Ä ²# ³#b n Ä n # When ~ ~ , we get for ~ and ~ , ; ²= ³ B²- n = Á = n - ³ B²= ³ where ²# n ³²$³ ~ ²$³# and for ~ and ~ , ; ²= ³ B²= i n - Á - n = i ³ B²= i Á = i ³ where ²# n ³²³ ~ ²#³ Finally, when ~ ~ , we get a multilinear form ²# n ³²Á $³ ~ ²#³ ²$³ Consider also a tensor n of type ²Á ³. When ~ ~ we get a multilinear functional n ¢ ²= d = ³ ¦ - defined by

Tensor Products

389

² n ³²#Á $³ ~ ²#³²$³ This is just a bilinear form on = .

Contraction Covariant and contravariant factors can be “combined” in the following way. Consider the map c ¢ = d d ²= i ³d ¦ ;c ²= ³

defined by ²# Á Ã Á # Á Á Ã Á ³ ~ ²# ³²# n Ä n # n n Ä n ³ This is easily seen to be multilinear and so there is a unique linear map c ¢ ; ²= ³ ¦ ;c ²= ³

defined by ²# n Ä n # n n Ä n ³ ~ ²# ³²# n Ä n # n n Ä n ³ This is called the contraction in the contravariant index and covariant index . Of course, contraction in other indices (one contravariant and one covariant) can be defined similarly. Example 14.8 Let dim²= ³ and consider the tensor space ; ²= ³, which is isomorphic to B²= ³ via the map ²# n ³²$³ ~ ²$³# For a “decomposable” linear operator of the form # n as defined above with # £ and £ , we have ker²# n ³ ~ ker² ³, which has codimension . Hence, if ²$³²#³ ~ ²# n ³²$³ £ , then = ~ º$» l ker² ³ ~ º$» l ; where ; is the eigenspace of # n associated with the eigenvalue . In particular, if ²#³ £ , then ²# n ³²#³ ~ ²#³# and so # is an eigenvector for the nonzero eigenvalue ²#³. Hence, = ~ º#» l ; ~ ; ²#³ l ; and so the trace of # n is

390

Advanced Linear Algebra

tr²# n ³ ~ ²#³ ~ ²# n ³ where is the contraction map.

The Tensor Algebra of = Consider the contravariant tensor spaces ; ²= ³ ~ ; ²= ³ ~ = n For ~ we take ; ²= ³ ~ - . The external direct sum B

; ²= ³ ~ ; ²= ³ ~

of these tensor spaces is a vector space with the property that ; ²= ³ n ; ²= ³ ~ ; b ²= ³ This is an example of a graded algebra, where ; ²= ³ are the elements of grade . The graded algebra ; ²= ³ is called the tensor algebra over = . (We will formally define graded structures a bit later in the chapter.) Since ; ²= ³ ~ = i n Ä n = i ~ ; ²= i ³ factors

there is no need to look separately at ; ²= ³.

Special Multilinear Maps The following definitions describe some special types of multilinear maps. Definition 1) A multilinear map ¢ = d ¦ > is symmetric if interchanging any two coordinate positions changes nothing, that is, if ²# Á Ã Á # Á Ã Á # Á Ã Á # ³ ~ ²# Á Ã Á #Á Ã Á #Á Ã Á #³ for any £ . 2) A multilinear map ¢ = d ¦ > is antisymmetric or skew-symmetric if interchanging any two coordinate positions introduces a factor of c, that is, if ²# Á Ã Á # Á Ã Á # Á Ã Á # ³ ~ c ²# Á Ã Á #Á Ã Á #Á Ã Á #³ for £ .

Tensor Products

391

3) A multilinear map ¢ = ¦ > is alternate or alternating if # ~ # for some £

¬

²# Á Ã Á # ³ ~

As in the case of bilinear forms, we have some relationships between these concepts. In particular, if char²- ³ ~ , then alternate ¬ symmetric ¯ skew-symmetric and if char²- ³ £ , then alternate ¯ skew-symmetric A few remarks about permutations are in order. A permutation of the set 5 ~ ¸Á Ã Á ¹ is a bijective function ¢ 5 ¦ 5 . We denote the group (under composition) of all such permutations by : . This is the symmetric group on symbols. A cycle of length is a permutation of the form ² Á Á Ã Á ³, which sends " to "b for " ~ Á Ã Á c and also sends to . All other elements of 5 are left fixed. Every permutation is the product (composition) of disjoint cycles. A transposition is a cycle ²Á ³ of length . Every cycle (and therefore every permutation) is the product of transpositions. In general, a permutation can be expressed as a product of transpositions in many ways. However, no matter how one represents a given permutation as such a product, the number of transpositions is either always even or always odd. Therefore, we can define the parity of a permutation : to be the parity of the number of transpositions in any decomposition of as a product of transpositions. The sign of a permutation is defined by sg²³ ~ F

c

has even parity has odd parity

If sg²³ ~ , then is an even permutation and if sg²³ ~ c, then is an odd permutation. The sign of is often written ²c³ . With these facts in mind, it is apparent that is symmetric if and only if ²# Á Ã Á # ³ ~ ²#²1) Á Ã Á #²³ ³ for all permutations : and that is skew-symmetric if and only if ²# Á Ã Á # ³ ~ ²c³ ²#²1) Á Ã Á #²³ ³ for all permutations : . A word of caution is in order with respect to the notation above, which is very convenient albeit somewhat prone to confusion. It is intended that a permutation permutes the coordinate positions in , not the indices (despite appearances). Suppose, for example, that ¢ s d s ¦ ? and that ¸ Á ¹ is a basis for s .

392

Advanced Linear Algebra

If ~ ² ³, then applied to ² Á ³ gives ² Á ³ and not ² Á ³, since permutes the two coordinate positions in ²# Á # ³.

Graded Algebras We need to pause for a few definitions that are useful in discussing tensor algebras. An algebra ( over - is said to be a graded algebra if as a vector space over - , ( can be written in the form B

( ~ ( ~

for subspaces ( of (, and where multiplication behaves nicely, that is, ( ( (b The elements of ( are said to be homogeneous of degree . If ( is written ~ b Ä b for ( , £ , then is called the homogeneous component of of degree . The ring of polynomials - ´%µ provides a prime example of a graded algebra, since B

- ´%µ ~ - ´%µ ~

where - ´%µ is the subspace of - ´%µ consisting of all scalar multiples of % . More generally, the ring - ´% Á Ã Á % µ of polynomials in several variables is a graded algebra, since it is the direct sum of the subspaces of homogeneous polynomials of degree . (A polynomial is homogeneous of degree if each term has degree . For example, ~ % % b % % % is homogeneous of degree .)

The Symmetric and Antisymmetric Tensor Algebras Our discussion of symmetric and antisymmetric tensors will benefit by a discussion of a few definitions and setting a bit of notation at the outset. Let - ´ Á Ã Á µ denote the vector space of all homogeneous polynomials of degree (together with the zero polynomial) in the independent variables Á Ã Á . As is sometimes done in this context, we denote the product in - ´ Á Ã Á µ by v , for example, writing as v v . The algebra of all polynomials in Á Ã Á is denoted by - ´ Á Ã Á µ.

Tensor Products

393

We will also need the counterpart of - ´ Á Ã Á µ in which multiplication acts anticommutatively, that is, ~ c . Definition Let , ~ ² Á Ã Á ³ be a sequence of independent variables. For , let -c ´ Á Ã Á µ be the vector space over - with basis 7 ²,³ ~ ¸ Ä Ä ¹ consisting of all words of length over , that are in ascending order. Let -c ´ Á Ã Á µ ~ - , which we identify with - by identifying with - . Define a product on the direct sum

- c ´ Á Ã Á µ ~ -c ~

as follows. First, the product w of monomials ~ % Ä% -c and ~ & Ä& - c is defined as follows: 1) If % Ä% & Ä& has a repeated factor then w ~ . 2) Otherwise, reorder % Ä% & Ä& in ascending order, say ' Ä'b , via the permutation and set w ~ ²c³ ' Ä'b Extend the product by distributivity to - c ´ Á Ã Á µ. The resulting product makes - c ´ Á Ã Á µ into a (noncommutative) algebra over - . This product is called the wedge product or exterior product on - c ´ Á Ã Á µ. For example, by definition of wedge product, w w ~ c w w Let 8 ~ ¸ Á Ã Á ¹ be a basis for = . It will be convenient to group the decomposable basis tensors n Ä n according to their index multiset. Specifically, for each multiset 4 ~ ¸ Á Ã Á ¹ with , let .4 be the set of all tensors n Ä n where ² Á Ã Á ³ is a permutation of ¸ Á Ã Á ¹. For example, if 4 ~ ¸Á Á ¹, then .4 ~ ¸ n n Á n n Á n n ¹ If # ; ²= ³ has the form # ~ ÁÃÁ n Ä n ÁÃÁ

where ÁÃÁ £ , then let .4 ²#³ be the subset of .4 whose elements appear

394

Advanced Linear Algebra

in the sum for #. For example, if # ~ n n b n n b n n then .¸ÁÁ¹ ²#³ ~ ¸ n n Á n n ¹ Let :4 ²#³ denote the sum of the terms of # associated with .4 ²#³. For example, :¸ÁÁ¹ ²#³ ~ n n b n n Thus, # can be written in the form p s ! ! # ~ :4 ²#³ ~ t 4 4 q!.4 ²#³ where the sum is over a collection of multisets 4 with :4 ²#³ £ . Note also that ! £ since ! .4 ²#³. Finally, let "4 ~ n Ä n be the unique member of .4 for which Ä . Now we can get to the business at hand.

Symmetric and Antisymmetric Tensors Let : be the symmetric group on ¸Á Ã Á ¹. For each : , the multilinear map ¢ = d ¦ ; ²= ³ defined by ²% Á Ã Á % ³ ~ % n Ä n % determines a unique linear operator on ; ²= ³ for which ²% n Ä n % ³ ~ % n Ä n % For example, if ~ and ~ ² ³, then ²³ ²# n # n # ³ ~ # n # n # Let ¸ Á Ã Á ¹ be a basis for = . Since is a bijection of the basis 8 ~ ¸ n Ä n 8 ¹ it follows that is an isomorphism of ; ²= ³. Note also that is a permutation of each .4 , that is, the sets .4 are invariant under . Definition Let = be a finite-dimensional vector space.

Tensor Products

395

1) A tensor ! ; ²= ³ is symmetric if ! ~ ! for all permutations : . The set of all symmetric tensors :; ²= ³ ~ ¸! ; ²= ³ ! ~ ! for all : ¹ is a subspace of ; ²= ³, called the symmetric tensor space of degree over = . 2) A tensor ! ; ²= ³ is antisymmetric if ! ~ ²c³ ! The set of all antisymmetric tensors (; ²= ³ ~ ¸! ; ²= ³ ! ~ ²c³ ! for all : ¹ is a subspace of ; ²= ³, called the antisymmetric tensor space or exterior product space of degree over = . We can develop the theory of symmetric and antisymmetric tensors in tandem. Accordingly, let us write (anti)symmetric to denote a tensor that is either symmetric or antisymmetrtic. Since for any Á ! .4 , there is a permutation taking (anti)symmetric tensor # must have .4 ²#³ ~ .4 and so

to !, an

# ~ :4 ²#³ ~ 8 ! !9 4

4

!.4

Since is a permutation of .4 , it follows that # is symmetric if and only if ²:4 ²#³³ ~ :4 ²#³ for all : and this holds if and only if the coefficients ! of :4 ²#³ are equal, say ! ~ 4 for all ! .4 . Hence, the symmetric tensors are precisely the tensors of the form # ~ 84 !9 4

!.4

The tensor # is antisymmetric if and only if ²:4 ²#³³ ~ ²c³ :4 ²#³

(14.4)

In this case, the coefficients ! of :4 ²#³ differ only by sign. Before examining this more closely, we observe that 4 must be a set. For if 4 has an element of multiplicity greater than , we can split .4 into two disjoint parts:

396

Advanced Linear Algebra

Z ZZ .4 ~ .4 r .4 Z where .4 are the tensors that have in positions and : Z .4 ~ ¸ n Ä n

position

nÄn

position

n Ä n ¹

Z ZZ Then ² ³ fixes each element of .4 and sends the elements of .4 to other ZZ elements of .4 . Hence, applying ² ³ to the corresponding decomposition of :4 ²#³: Z ZZ :4 ²#³ ~ :4 b :4

gives Z ZZ Z ZZ c²:4 b :4 ³ ~ c:4 ~ ² ³ :4 ~ :4 b ² ³ :4 Z and so :4 ~ , whence :4 ²#³ ~ . Thus, 4 is a set.

Now, since for any : , .4 ~ ¸ ! ! .4 ¹ equation (14.4) implies that ²c³ ! ! ~ 8 ! !9 ~ ! ! ~ c ! ! !.4

!.4

!.4

!.4

which holds if and only if c ! ~ ²c³ ! , or equivalently, ! ~ ²c³ ! for all ! .4 and : . Choosing " ~ "4 ~ n Ä n , where Ä , as standard-bearer, if "Á! denotes the permutation for which "Á! ²"³ ~ !, then ! ~ ²c³"Á! " Thus, # is antisymmetric if and only if it has the form # ~ 84 ²c³"Á! !9 4

!.4

where 4 ~ " £ and the sum is over a family of sets. In summary, the symmetric tensors are

Tensor Products

397

# ~ 84 !9 4

!.4

where 4 is a multiset and the antisymmetric tensors are # ~ 84 ²c³"Á! !9 4

!.4

where 4 is a set. We can simplify these expressions considerably by representing the inside sums more succinctly. In the symmetric case, define a surjective linear map ¢ ; ²= ³ ¦ - ´ Á Ã Á µ by ² n Ä n ³ ~ v Ä v and extending by linearity. Since takes every member of .4 to the same monomial " ~ v Ä v , where Ä , we have # ~ 884 !99 ~ 4 (.4 ( "4 4

4

!.4

In the antisymmetric case, define a surjective linear map ¢ ; ²= ³ ¦ -c ´ Á Ã Á µ by ² n Ä n ³ ~ w Ä w and extending by linearity. Since ! ~ ²c³"Á! "4 we have # ~ 84 ²c³"Á! !9 4

!.4

~ 84 "4 9 4

!.4

~ 4 (.4 ( "4 4

398

Advanced Linear Algebra

Thus, in both cases, # ~ 4 (.4 ( "4 4

where "4 ~ n Ä n with Ä and "4 ~ v Ä v

or

"4 ~ w Ä w

depending on whether # is symmetric or antisymmetric. However, in either case, the monomials "4 are linearly independent for distinct multisets/sets 4 . Therefore, if # ~ then 4 (.4 ( ~ for all multisets/sets 4 . Hence, if char²- ³ ~ , then 4 ~ and so # ~ . This shows that the restricted maps O:; ²= ³ and O(; ²= ³ are isomorphisms. Theorem 14.16 Let = be a finite-dimensional vector space over a field - with char²- ³ ~ . 1) The symmetric tensor space :; ²= ³ is isomorphic to the algebra - ´ Á Ã Á µ of homogeneous polynomials, via the isomorphism 4 ÁÃÁ n Ä n 5 ~ ÁÃÁ ² v Ä v ³ 2) For , the antisymmetric tensor space (; ²= ³ is isomorphic to the algebra -c ´ Á Ã Á µ of anticommutative homogeneous polynomials of degree , via the isomorphism 4 ÁÃÁ n Ä n 5 ~ ÁÃÁ ² w Ä w ³

The direct sum B

:; ²= ³ ~ :; ²= ³ - ´ Á Ã Á µ ~

is called the symmetric tensor algebra of = and the direct sum

(; ²= ³ ~ (; ²= ³ - c ´ Á Ã Á µ ~

is called the antisymmetric tensor algebra or the exterior algebra of = . These vector spaces are graded algebras, where the product is defined using the vector space isomorphisms described in the previous theorem to move the products of - ´ Á Ã Á µ and - c ´ Á Ã Á µ to :; ²= ³ and (; ²= ³, respectively. Thus, restricting the domains of the maps gives a nice description of the symmetric and antisymmetric tensor algebras, when char²- ³ ~ . However, there are many important fields, such as finite fields, that have nonzero characteristic. We can proceed in a different, albeit somewhat less appealing,

Tensor Products

399

manner that holds regardless of the characteristic of the base field. Namely, rather than restricting the domain of in order to get an isomorphism, we can factor out by the kernel of . Consider a tensor p s ! ! # ~ :4 ²#³ ~ t 4 4 q!.4 ²#³ Since sends elements of different groups .4 ²#³ ~ ¸! Á Ã Á ! ¹ to different monomials in - ´ Á Ã Á µ or -c ´ Á Ã Á µ, it follows that # ker² ³ if and only if ²:4 ²#³³ ~ for all 4 , that is, if and only if ! ! b Ä b ! ! ~ In the symmetric case, is constant on .4 ²#³ and so # ker² ³ if and only if ! b Ä b ! ~ In the antisymmetric case, ! ~ ²c³Á ! where Á ²! ³ ~ ! and so # ker² ³ if and only if ²c³Á ! b Ä b ²c³Á ! ~ In both cases, we solve for ! and substitute into :4 ²#³. In the symmetric case, ! ~ c! c Ä c ! and so :4 ²#³ ~ ! ! b Ä b ! ! ~ ! ²! c ! ³ b Ä b ! ²! c ! ³ In the antisymmetric case, ! ~ c²c³Á ! c Ä c ²c³Á ! and so :4 ²#³ ~ ! ! b Ä b ! ! ~ ! ²²c³Á ! c ! ³ b Ä b ! ²²c³Á ! c ! ³ Since ! 8 , it follows that :4 ²#³ and therefore #, is in the span of tensors of the form ²!³ c ! in the symmetric case and ²c³ ²!³ c ! in the antisymmetric case, where : and ! 8 . Hence, in the symmetric case, ker² ³ 0 º ²!³ c ! ! 8 Á : » and since ² ²!³ c !³ ~ , it follows that ker² ³ ~ 0 . In the antisymmetric case,

400

Advanced Linear Algebra

ker² ³ 0 º²c³ ²!³ c ! ! 8Á : » and since ²²c³ ²!³ c !³ ~ , it follows that ker² ³ ~ 0 . We now have quotient-space characterizations of the symmetric and antisymmetric tensor spaces that do not place any restriction on the characteristic of the base field. Theorem 14.17 Let = be a finite-dimensional vector space over a field - . 1) The surjective linear map ¢ ; ²= ³ ¦ - ´ Á Ã Á µ defined by 4 ÁÃÁ n Ä n 5 ~ ÁÃÁ v Ä v has kernel 0 ~ º ²!³ c ! ! 8 Á : » and so ; ²= ³ - ´ Á Ã Á µ 0 The vector space ; ²= ³°0 is also referred to as the symmetric tensor space of degree of = . 2) The surjective linear map ¢ ; ²= ³ ¦ -c ´ Á Ã Á µ defined by 4 ÁÃÁ n Ä n 5 ~ ÁÃÁ w Ä w has kernel 0 ~ º²c³ ²!³ c ! ! 8Á : » and so ; ²= ³ -c ´ Á Ã Á µ 0 The vector space ; ²= ³°0 is also referred to as the antisymmetric tensor space or exterior product space of degree of = . The isomorphic exterior spaces (; ²= ³ and ; ²= ³°0 are usually denoted by = and the isomorphic exterior algebras (; ²= ³ and ; ²= ³°0 are usually denoted by = . Theorem 14.18 Let = be a vector space of dimension .

Tensor Products

401

1) The dimension of the symmetric tensor space :; ²= ³ is equal to the number of monomials of degree in the variables Á Ã Á and this is dim²:; ²= ³³ ~ 6

bc 7

2) The dimension of the exterior tensor space ²= ³ is equal to the number of words of length in ascending order over the alphabet , ~ ¸ Á Ã Á ¹ and this is dim² ²= ³³ ~ 6 7

Proof. For part 1), the dimension is equal to the number of multisets of size taken from an underlying set ¸ Á Ã Á ¹ of size . Such multisets correspond bijectively to the solutions, in nonnegative integers, of the equation % b Ä b % ~ where % is the multiplicity of in the multiset. To count the number of solutions, invent two symbols % and °. Then any solution % ~ to the previous equation can be described by a sequence of %'s and °'s consisting of %'s followed by one °, followed by %'s and another °, and so on. For example, if ~ and ~ , the solution b b b ~ corresponds to the sequence %%%°%°°%% Thus, the solutions correspond bijectively to sequences consisting of %'s and c °'s. To count the number of such sequences, note that such a sequence can be formed by considering b c “blanks” and selecting of these blanks for the %'s. This can be done in 6

bc 7

ways.

The Universal Property We defined tensor products through a universal property, which as we have seen is a powerful technique for determining the properties of tensor products. It is easy to show that the symmetric tensor spaces are universal for symmetric multilinear maps and the antisymmetric tensor spaces are universal for antisymmetric multilinear maps. Theorem 14.19 Let = be a finite-dimensional vector space with basis ¸ Á Ã Á ¹. 1) The pair ²- ´% Á Ã Á % µÁ !³, where !¢ = d ¦ - ´% Á Ã Á % µ is the multilinear map defined by

402

Advanced Linear Algebra

!² Á Ã Á ³ ~ v Ä v is universal for symmetric -linear maps with domain = d ; that is, for any symmetric -linear map ¢ = d ¦ < where < is a vector space, there is a unique linear map ¢ - ´% Á Ã Á % µ ¦ < for which ² v Ä v ³ ~ ² Á Ã Á ³ 2) The pair ²-c ´% Á Ã Á % µÁ !³, where !¢ = d ¦ -c ´% Á Ã Á % µ is the multilinear map defined by !² Á Ã Á ³ ~ w Ä w is universal for antisymmetric -linear maps with domain = d ; that is, for any antisymmetric -linear map ¢ = d ¦ < where < is a vector space, there is a unique linear map ¢ -c ´% Á Ã Á % µ ¦ < for which ² w Ä w ³ ~ ² Á Ã Á ³ Proof. For part 1), the property ² v Ä v ³ ~ ² Á Ã Á ³ does indeed uniquely define a linear transformation , provided that it is welldefined. However, v Ä v ~ v Ä v if and only if the multisets ¸ Á Ã Á ¹ and ¸ Á Ã Á ¹ are the same, which implies that ² Á Ã Á ³ ~ ² Á Ã Á ³, since is symmetric. For part 2), since is antisymmetric, it is completely determined by the fact that it is alternate and by its values on the basis of ascending words w Ä w . Accordingly, the condition ² w Ä w ³ ~ ² Á Ã Á ³ uniquely defines a linear transformation .

The Symmetrization Map When char²- ³ ~ , we can define a linear map :¢ ; ²= ³ ¦ :; ²= ³, called the symmetrization map, by :! ~

! [ :

Since ~ , we have

Tensor Products

²:!³ ~

403

! ~ ! ~ ! ~ :! [ : [ : [ :

and so :! is symmetric. The reason for the factor °[ is that if # is a symmetric tensor, then # ~ # and so :# ~

# ~ # ~ # [ : [ :

that is, the symmetrization map fixes all symmetric tensors. It follows that for any tensor ! ; ²= ³, : ! ~ :! Thus, : is idempotent and is therefore the projection map of ; ²= ³ onto im²:³ ~ :; ²= ³.

The Determinant The universal property for antisymmetric multilinear maps has the following corollary. Corollary 14.20 Let = be a vector space of dimension over a field - . Let , ~ ² Á Ã Á ³ be an ordered basis for = . Then there is at most one antisymmetric -linear form ¢ = d ¦ - for which ² Á Ã Á ³ ~ Proof. According to the universal property for antisymmetric -linear forms, for every antisymmetric -linear form ¢ = d ¦ - satisfying ² Á Ã Á ³ ~ , there is a unique linear map ¢ = ¦ - for which ² w Ä w ³ ~ ² Á Ã Á ³ ~ But = has dimension and so there is only one linear map ¢ = ¦ with ² w Ä w ³ ~ . Therefore, if and are two such forms, then ~ ~ , from which it follows that ~ k ! ~ k ! ~

We now wish to construct an antisymmetric form ¢ = d ¦ - , which is unique by the previous theorem. Let 8 be a basis for = . For any # = , write ´#µ8Á for the th coordinate of the coordinate matrix ´#µ8 . Thus, # ~ ´#µ8Á

For clarity, and since we will not change the basis, let us write ´#µ for ´#µ8Á .

404

Advanced Linear Algebra

Consider the map ¢ = d ¦ - defined by ²# Á Ã Á # ³ ~ ²c³ ´# µ Ä´# µ :

Then is multilinear since ²# b " Á Ã Á # ³ ~ ²c³ ´# b " µ Ä´#µ :

~ ²c³ ²´# µ b ´" µ ³Ä´# µ() :

~ ²c³ ´# µ Ä´# µ :

b ²c³ ´" µ Ä´# µ :

~ ²# Á Ã Á # ³ b ²" Á # Á Ã Á # ³ and similarly for any coordinate position. To see that is alternating, and therefore antisymmetric since char²- ³ £ , suppose for instance that # ~ # . For any permutation : , let Z ~ ² ³ Then Z % ~ % for % £ Á and Z ~

and

Z ~

Hence, Z £ . Also, since ²Z ³Z ~ , if the sets ¸Á Z ¹ and ¸Á Z ¹ intersect, then they are identical. Thus, the distinct sets ¸Á Z ¹ form a partition of : . It follows that ²# Á # Á # Á Ã Á # ³ ~ ²c³ ´# µ ´# µ Ä´#µ :

~

Z

>²c³ ´# µ ´# µ Ä´# µ b ²c³ ´# µZ ´#µZ Ä´#µZ ?

pairs ¸ÁZ ¹

But ´# µ ´# µ ~ ´# µZ ´# µZ Z

and since ²c³ ~ c²c³ , the sum of the two terms involving the pair ¸Á Z ¹ is . Hence, ²# Á # Á Ã Á # ³ ~ . A similar argument holds for any coordinate pair.

Tensor Products

405

Finally, ² Á Ã Á ³ ~ ²c³ ´ µ Ä´ µ :

~ ²c³ Á ÄÁ :

~ Thus, the map is the unique antisymmetric -linear form on = d for which ² Á Ã Á ³ ~ . Under the ordered basis ; ~ ² Á Ã Á ³, we can view = as the space - of coordinate vectors and view = d as the space 4 ²- ³ of d matrices, via the isomorphism ²# Á Ã Á # ³ ª

v ´# µ Å w ´# µ

Ä

´# µ y Å Ä ´# µ z

where all coordinate matrices are with respect to ; . With this viewpoint, becomes an antisymmetric -form on the columns of a matrix ( ~ ²Á ³ given by ²(³ ~ ²c³ Á ÄÁ :

This is called the determinant of the matrix (.

Properties of the Determinant Let us explore some of the properties of the determinant function. Theorem 14.21 If ( 4 ²- ³, then ²(³ ~ ²(! ³ Proof. We have ²(³ ~ ²c³ Á ÄÁ : c

~ ²c³ c Á Äc Á c :

~ ²c³ Á ÄÁ :

~ ²(! ³ as desired.

406

Advanced Linear Algebra

Theorem 14.22 If (Á ) 4 ²- ³, then ²()³ ~ ²(³²)³ Proof. Consider the map ( ¢ 4 ²- ³ ¦ - defined by ( ²?³ ~ ²(?³ We can consider ( as a function on the columns of ? and think of it as a composition

( ¢ ²? ²³ Á Ã Á ? ²³ ³ ª ²(? ²³ Á Ã Á (? ²³ ³ ª ²(?³ Each step in this map is multilinear and so ( is multilinear. It is also clear that ( is antisymmetric and so ( is a scalar multiple of the determinant function, say ( ²?³ ~ ²?³. Then ²(?³ ~ ( ²?³ ~ ²?³ Setting ? ~ 0 gives ²(³ ~ and so ²(?³ ~ ²(³²?³ as desired. Theorem 14.23 A matrix ( 4 ²- ³ is invertible if and only if ²(³ £ . Proof. If 7 4 ²- ³ is invertible, then 7 7 c ~ 0 and so ²7 ³²7 c ³ ~ which shows that ²7 ³ £ and ²7 c ³ ~ °²7 ³. Conversely, any matrix ( 4 ²- ³ is equivalent to a diagonal matrix ( ~ 7 +8 where 7 and 8 are invertible and + is diagonal with 's and 's on the main diagonal. Hence, ²(³ ~ ²7 ³²+³²8³ and so if ²(³ £ , then ²+³ £ , which happens if and only if + ~ 0 , whence ( is invertible.

Exercises 1. 2. 3.

Show that if ¢ > ¦ ? is a linear map and ¢ < d = ¦ > is bilinear, then k ¢ < d = ¦ ? is bilinear. Show that the only map that is both linear and -linear (for ) is the zero map. Find an example of a bilinear map ¢ = d = ¦ > whose image im² ³ ~ ¸ ²"Á #³ "Á # = ¹ is not a subspace of > .

Tensor Products 4.

407

Let 8 ~ ¸" 0¹ be a basis for < and let 9 ~ ¸# 1 ¹ be a basis for = . Show that the set : ~ ¸" n # 0Á 1 ¹

is a basis for < n = by showing that it is linearly independent and spans. Prove that the following property of a pair ²> Á ¢ < d = ¦ > ³ with bilinear characterizes the tensor product ²< n = Á !¢ < d = ¦ < n = ³ up to isomorphism, and thus could have been used as the definition of tensor product: For a pair ²> Á ¢ < d = ¦ > ³ with bilinear if ¸" ¹ is a basis for < and ¸# ¹ is a basis for = , then ¸²" Á # ³¹ is a basis for > . 6. Prove that < n = = n < . 7. Let ? and @ be nonempty sets. Use the universal property of tensor products to prove that can be extended to a linear function ¢ < n = ¦ > . Deduce that the function can be extended in a unique way to a bilinear map V ¢ < d = ¦ > . Show that all bilinear maps are obtained in this way. 10. Let : Á : be subspaces of < . Show that

5.

²: n = ³ q ²: n = ³ ²: q : ³ n = 11. Let : < and ; = be subspaces of vector spaces < and = , respectively. Show that ²: n = ³ q ²< n ; ³ : n ; 12. Let : Á : < and ; Á ; = be subspaces of < and = , respectively. Show that ²: n ; ³ q ²: n ; ³ ²: q : ³ n ²; n ; ³ 13. Find an example of two vector spaces < and = and a nonzero vector % < n = that has at least two distinct (not including order of the terms) representations of the form

% ~ " n # ~

where the " 's are linearly independent and so are the # 's. 14. Let ? denote the identity operator on a vector space ? . Prove that = p > ~ = n> .

408

Advanced Linear Algebra

15. Suppose that ¢ < ¦ = Á ¢ = ¦ > and ¢ < Z ¦ =2 Á ¢ =2 ¦ > Z . Prove that ² k ³ p ² k ³ ~ ² p ³ k ² p ³ 16. Connect the two approaches to extending the base field of an - -space = to 2 (at least in the finite-dimensional case) by showing that - n - 2 ²2³ . 17. Prove that in a tensor product < n < for which dim²< ³ not all vectors have the form " n # for some "Á # < . Hint: Suppose that "Á # < are linearly independent and consider " n # b # n ". 18. Prove that for the block matrix 4 ~>

(

) * ?block

we have ²4 ³ ~ ²(³²*³. 19. Let (Á ) 4 ²- ³. Prove that if either ( or ) is invertible, then the matrices ( b ) are invertible except for a finite number of 's.

The Tensor Product of Matrices 20. Let ( ~ ²Á ³ be the matrix of a linear operator B²= ³ with respect to the ordered basis 7 ~ ²" Á Ã Á " ³. Let ) ~ ²Á ³ be the matrix of a linear operator B²= ³ with respect to the ordered basis 8 ~ ²# Á Ã Á # ³. Consider the ordered basis 9 ~ ²" n # ³ ordered lexicographically; that is " n # "M n # if M or ~ M and . Show that the matrix of n with respect to 9 is p Á ) r ) ( n ) ~ r Á Å q Á )

21. 22. 23. 24. 25. 26.

Á ) Á ) Å Á )

Ä Ä

Á ) s Á ) u u Å Ä Á ) tblock

This matrix is called the tensor product, Kronecker product or direct product of the matrix ( with the matrix ) . Show that the tensor product is not, in general, commutative. Show that the tensor product ( n ) is bilinear in both ( and ) . Show that ( n ) ~ if and only if ( ~ or ) ~ . Show that a) ²( n )³! ~ (! n ) ! b) ²( n )³i ~ (i n ) i (when - ~ d) Show that if "Á # - , then (as row vectors) "! # ~ "! n #. Suppose that (Á Á )Á Á *Á and +Á are matrices of the given sizes. Prove that ²( n )³²* n +³ ~ ²(*³ n ²)+³ Discuss the case ~ ~ .

Tensor Products

409

27. Prove that if ( and ) are nonsingular, then so is ( n ) and ²( n )³c ~ (c n ) c 28. Prove that tr²( n )³ ~ tr²(³ h tr²)³. 29. Suppose that - is algebraically closed. Prove that if ( has eigenvalues Á Ã Á and ) has eigenvalues Á Ã Á , both lists including multiplicity, then ( n ) has eigenvalues ¸ Á ¹, again counting multiplicity. 30. Prove that det²(Á n )Á ³ ~ ²det²(Á ³³ ²det²)Á ³³ .

Chapter 15

Positive Solutions to Linear Systems: Convexity and Separation

It is of interest to determine conditions that guarantee the existence of positive solutions to homogeneous systems of linear equations (% ~ where ( CÁ ²s³. Definition Let # ~ ² Á Ã Á ³ s . 1) # is nonnegative, written # , if for all (The term positive is also used for this property.) The set of all nonnegative vectors in s is the nonnegative orthant in s À 2) # is strictly positive, written # , if # is nonnegative but not , that is, if for all and for some The set sb of all strictly positive vectors in s is the strictly positive orthant in s À 3) # is strongly positive, written # , if for all The set sbb of all strongly positive vectors in s is the strongly positive orthant in s À We are interested in conditions under which the system (% ~ has strictly positive or strongly positive solutions. Since the strictly and strongly positive orthants in s are not subspaces of s, it is difficult to use strictly linear methods in studying this issue: we must also use geometric methods, in particular, methods of convexity.

412

Advanced Linear Algebra

Let us pause briefly to consider an important application of strictly positive solutions to a system (% ~ . If ? ~ ²% Á Ã Á % ³ is a strictly positive solution to this system, then so is the vector &~

?~ ²% Á Ã Á % ³ ~ ² Á Ã Á ³ '% '%

which is a probability distribution, that is, and ~ . Moreover, if ? is a strongly positive solution, then & has the property that each probability is positive. Now, the product (& is the expected value of the columns of ( with respect to the probability distribution &. Hence, (% ~ has a strictly (strongly) positive solution if and only if there is a strictly (strongly) positive probability distribution for which the columns of ( have expected value . If each column of ( represents the possible payoffs from a game of chance, where each row is a different possible outcome of the game, then the game is fair when the expected value of the columns is . Thus, (% ~ has a strictly (strongly) positive solution ? if and only if the game with payoffs ( and probabilities ? is fair. As another (related) example, in discrete option pricing models of mathematical finance, the absence of arbitrage opportunities in the model is equivalent to the fact that a certain vector describing the gains in a portfolio does not intersect the strictly positive orthant in s . As we will see in this chapter, this is equivalent to the existence of a strongly positive solution to a homogeneous system of equations. This solution, when normalized to a probability distribution, is called a martingale measure. Of course, the equation (% ~ has a strictly positive solution if and only if ker²(³ contains a strictly positive vector, that is, if and only if ker²(³ ~ RowSpace²(³ meets the strictly positive orthant in s . Thus, we wish to characterize the subspaces : of s for which : meets the strictly positive orthant in s , in symbols, : q sb £ J for these are precisely the row spaces of the matrices ( for which (% ~ has a strictly positive solution. A similar statement holds for strongly positive solutions. Looking at the real plane s , we can divine the answer with a picture. A onedimensional subspace : of s has the property that its orthogonal complement : meets the strictly positive orthant (quadrant) in s if and only if : is the %axis, the &-axis or a line with negative slope. For the case of the strongly

Positive Solutions to Linear Systems: Convexity and Separation

413

positive orthant, : must have negative slope. Our task is to generalize this to s . This will lead us to the following results, which are quite intuitive in s and s : : q sbb £ J

: q sb ~J

(15.1)

: q sbb ~ J

(15.2)

¯

and : q sb £ J

¯

Let us translate these statements into the language of the matrix equation (% ~ . If : ~ RowSpace²(³, then : ~ ker²(³ and so we have ker²(³ q sbb £ J

¯

RowSpace²(³ q sb ~J

and ker²(³ q sb £ J

¯

~J RowSpace²(³ q sbb

Now, RowSpace²(³ q sb ~ ¸#( #( ¹ and RowSpace²(³ q sbb ~ ¸#( #( ¹ and so these statements become (% ~ has a strongly positive solution

¯

¸#( #( ¹ ~ J

and (% ~ has a strictly positive solution

¯

¸#( #( ¹ ~ J

We can rephrase these results in the form of a theorem of the alternative, that is, a theorem that says that exactly one of two conditions holds. Theorem 15.1 Let ( CÁ ²s³. 1) Exactly one of the following holds: a) (" ~ for some strongly positive " s . b) #( for some # s . 2) Exactly one of the following holds: a) (" ~ for some strictly positive " s . b) #( for some # s . Before proving Theorem 15.1, we require some background.

Convex, Closed and Compact Sets We shall need the following concepts.

414

Advanced Linear Algebra

Definition 1) Let % Á Ã Á % s . Any linear combination of the form ! % b Ä b ! % where ! and ! b Ä b ! ~ is called a convex combination of the vectors % Á Ã Á % . 2) A subset ? s is convex if whenever %Á & ? , then the line segment between % and & also lies in ? , in symbols, ¸!% b ² c !³& ! ¹ ? 3) A subset ? s is closed if whenever ²% ³ is a convergent sequence of elements of ? , then the limit is also in ? . 4) A subset ? s is compact if it is both closed and bounded. 5) A subset ? s is a cone if % ? implies that % ? for all . We will also have need of the following facts from analysis. 1) A continuous function that is defined on a compact set ? in s takes on maximum and minimum values at some points within the set ? . 2) A subset ? of s is compact if and only if every sequence in ? has a subsequence that converges to a point in ? . Theorem 15.2 Let ? and @ be subsets of s . Define ? b @ ~ ¸ b ?Á @ ¹ 1) If ? and @ are convex, then so is ? b @ 2) If ? is compact and @ is closed, then ? b @ is closed. Proof. For 1), let % b & and % b & be in ? b @ . The line segment between these two points is !²% b & ³ b ² c !³²% b & ³ ~ ´!% b ² c !³% µ b ´!& b ² c !³& µ ? b @ for ! and so ? b @ is convex. For part 2), let % b & be a convergent sequence in ? b @ . Suppose that % b & ¦ ' . We must show that ' ? b @ . Since % is a sequence in the compact set ? , it has a convergent subsequence % whose limit % lies in ? . Since % b & ¦ ' and % ¦ % we can conclude that & ¦ ' c %. Since @ is closed, it follows that ' c % @ and so ' ~ % b ²' c %³ ? b @ .

Convex Hulls We will also have use for the notion of the smallest convex set containing a given set.

Positive Solutions to Linear Systems: Convexity and Separation

415

Definition The convex hull of a set : ~ ¸% Á Ã Á % ¹ of vectors in s is the smallest convex set in s that contains : . We will denote the convex hull of : by 9²:³. Here is a characterization of convex hulls. Theorem 15.3 Let : ~ ¸% Á Ã Á % ¹ be a set of vectors in s . Then the convex hull 9²:³ is the set " of all convex combinations of vectors in : , that is, 9²:³ ~ " D! % b Ä b ! % ! Á ! ~ E Proof. Clearly, if + is a convex set that contains : , then + also contains ". Hence " 9²:³. To prove the reverse inclusion, we need only show that " is convex, since then : " implies that 9²:³ ". So let ? ~ ! % b Ä b ! % @ ~ % b Ä b % be in ". If b ~ and Á then ? b @ ~ ²! % b Ä b ! % ³ b ² % b Ä b ~ ²! b ³% b Ä b ²! b ³%

% ³

But this is also a convex combination of the vectors in : , because ! b

² b ³ h max² Á ! ³ ~ max² Á ! ³

and

²! b ³ ~ ! b ~

~

~b ~

~

Thus, ? b @ ". Theorem 15.4 The convex hull 9²:³ of a finite set : ~ ¸% Á Ã Á % ¹ of vectors in s is a compact set. Proof. The set + ~ D²! Á Ã Á ! ³ ! Á ! ~ E is closed and bounded in s and therefore compact. Define a function ¢ + ¦ s as follows: If ! ~ ²! Á Ã Á ! ³, then ²!³ ~ ! % b Ä b ! % To see that is continuous, let , if ) c !) °4 then

~ ² Á Ã Á

³

and let 4 ~ max²)% )³. Given

416

Advanced Linear Algebra

( c ! ( ) c !)

4

and so ) ² ³ c ²!³) ~ )² c ! ³% b Ä b ² c ! ³% ) ( c ! ()% ) b Ä b ( c ! ()% ) 4 ) c !) ~ Finally, since ²+³ ~ 9²:³, it follows that 9²:³ is compact.

Linear and Affine Hyperplanes We next discuss hyperplanes in s . A linear hyperplane in s is an ² c ³dimensional subspace of s . As such, it is the solution set of a linear equation of the form % b Ä b % ~ or º5 Á %» ~ where 5 ~ ² Á Ã Á ³ is nonzero and % ~ ²% Á Ã Á % ³. Geometrically speaking, this is the set of all vectors in s that are perpendicular (normal) to the vector 5 . An affine hyperplane, or just hyperplane, in s is a linear hyperplane that has been translated by a vector. Thus, it is the solution set to an equation of the form ²% c ³ b Ä b ²% c ³ ~ or equivalently, º5 Á %» ~ where ~ b Ä b . We denote this hyperplane by >²5 Á ³ ~ ¸% s º5 Á %» ~ ¹ Note that the hyperplane >²5 Á )5 ) ³ ~ ¸% s º5 Á %» ~ )5 ) ¹ contains the point 5 , which is the point of >²5 Á )5 ) ³ closest to the origin, since Cauchy's inequality gives )5 ) ~ º5 Á %» )5 ))%) and so )5 ) )%) for all % >²5 Á )5 ) ³. Moreover, we leave it as an

Positive Solutions to Linear Systems: Convexity and Separation

417

exercise to show that any hyperplane has the form >²5 Á )5 ) ³ for an appropriate vector 5 . A hyperplane defines two closed half-spaces >b ²5 Á ³ ~ ¸% s º5 Á %» ¹ >c ²5 Á ³ ~ ¸% s º5 Á %» ¹ and two disjoint open half-spaces >kb ²5 Á ³ ~ ¸% s º5 Á %» ¹ >kc ²5 Á ³ ~ ¸% s º5 Á %» ¹ It is clear that >b ²5 Á ³ q >c ²5 Á ³ ~ >²5 Á ³ k and that the sets >kb ²5 Á ³, >c ²5 Á ³ and >²5 Á ³ form a partition of s .

If 5 s and ? s , we let º5 Á ?» ~ ¸º5 Á %» % ?¹ and write º5 Á ?» to denote the fact that º5 Á %» for all % ? . Definition Two subsets ? and @ of s are strictly separated by a hyperplane >²5 Á ³ if ? lies in one open half-space determined by >²5 Á ³ and @ lies in the other open half-space; in symbols, one of the following holds: 1) º5 Á ?» º5 Á @ » 2) º5 Á @ » º5 Á ?» Note that 1) holds for 5 and if and only if 2) holds for c5 and c , and so we need only consider one of the conditions to demonstrate that two sets ? and @ are not strictly separated. Specifically, if 1) fails for all 5 and , then the condition ºc5 Á @ » c ºc5 Á ?» also fails for all 5 and and so 2) also fails for all 5 and , whence ? and @ are not strictly separated. Definition Two subsets ? and @ of s are strongly separated by a hyperplane >²5 Á ³ if there is an for which one of the following holds: 1) º5 Á ?» c b º5 Á @ » 2) º5 Á @ » c b º5 Á ?»

418

Advanced Linear Algebra

As before, we need only consider one of the conditions to show that two sets are not strongly separated. Note also that if º5 Á %» º5 Á @ » for s, then % s and @ s are stongly separated by the hyperplane >65 Á

b º5 Á %» 7

Separation Now that we have the preliminaries out of the way, we can get down to some theorems. The first is a well-known separation theorem that is the basis for many other separation theorems. It says that if a closed convex set * s does not contain a vector , then * can be strongly separated from . Theorem 15.5 Let * be a closed convex subset of s . 1) * contains a unique vector 5 of minimum norm, that is, there is a unique vector 5 * for which ) 5 ) )% ) for all % *Á % £ 5 . 2) If ¤ * , then * lies in the closed half-space >b 25 Á )5 ) b º5 Á »3 that is, º5 Á *» )5 ) b º5 Á » º5 Á » where 5 is the unique vector of minimum norm in the closed convex set * c ~ ¸ c *¹ Hence, * and are strongly separated by the hyperplane >8 5 Á

)5 ) b º5 Á » 9

Proof. For part 1), if * then this is the unique vector of minimum norm, so we may assume that ¤ * . It follows that no two distinct elements of * can be negative scalar multiples of each other. For if % and % were in * , where Á then taking ! ~ c°² c ³ gives ~ !% b ² c !³% * which is false.

Positive Solutions to Linear Systems: Convexity and Separation

419

We first show that * contains a vector 5 of minimum norm. Recall that the Euclidean norm (distance) is a continuous function. Although * need not be compact, if we choose a real number such that the closed ball ) ²³ ~ ¸' s )' ) ¹ intersects * , then that intersection * Z ~ * q ) ²³ is both closed and bounded and so is compact. The norm function therefore achieves its minimum on * Z , say at the point 5 * Z * . It is clear that if )#) )5 ) for some # * , then # * Z , in contradiction to the minimality of 5 . Hence, 5 is a vector of minimum norm in * . We establish uniqueness first for closed line segments ´"Á #µ in s . If " ~ # where , then )!" b ² c !³#) ~ (! b ² c !³()#) is smallest when ! ~ for and ! ~ for . Assume that " and # are not scalar multiples of each other and suppose that % £ & in ´"Á #µ have minimum norm . If ' ~ ²% b &³° then since % and & are also not scalar multiples of each other, the Cauchy-Schwarz inequality is strict and so )% b & ) ~ ²P%P b º%Á &» b P&P ³ ² b )%))&)³ ~

P'P ~

which contradicts the minimality of . Thus, ´"Á #µ has a unique point of minimum norm. Finally, if % * also has minimum norm, then 5 and % are points of minimum norm in the line segment ´5 Á %µ * and so % ~ 5 . Hence, * has a unique element of minimum norm. For part 2), suppose the result is true when ¤ * . Then ¤ * implies that ¤ * c and so if 5 * c has smallest norm, then º5 Á * c » )5 ) Therefore, º5 Á *» )5 ) b º5 Á » º5 Á » and so * and are strongly separated by the hyperplane

420

Advanced Linear Algebra

>25 Á ²°³)5 ) b º5 Á »3 Thus, we need only prove part 2) for ~ , that is, we need only prove that º5 Á *» )5 ) If there is a nonzero % * for which º5 Á %» )5 ) then )5 ) )%) and º5 Á %» ~ )5 ) c for some . Then for the open line segment ²!³ ~ !5 b ² c !³% with ! , we have P ²!³P ~ )!5 b ² c !³%) ~ ! P5 P b !² c !³º5 Á %» b ² c !³ P%P ²! c ! ³P5 P c !² c !³ b ² c !³ P%P ~ ²cP5 P b b )%) ³! b ²)5 ) c c P%P ³! b )%) Let ²!³ denote the final expression above, which is a quadratic in !. It is easy to see that ²!³ has its minimum at the interior point of the line segment ´5 Á %µ corresponding to ! ~

c)5 ) b b )%) c)5 ) b b )%)

and so ) ²! ³) ²! ³ ²³ ~ )5 ), which is a contradiction. The next result brings us closer to our goal by replacing a single vector with a subspace : disjoint from * . However, we must also require that * be bounded, and therefore compact. Theorem 15.6 Let * be a compact convex subset of s and let : be a subspace of s such that * q : ~ J. Then there exists a nonzero 5 : such that º5 Á %» )5 ) for all % * . Hence, the hyperplane >²5 Á )5 ) °³ strongly separates : and *. Proof. Theorem 15.2 implies that the set : b * is closed and convex. Furthermore, * q : ~ J implies that ¤ : b * and so Theorem 15.5 implies that : b * can be strongly separated from the origin. Hence, there is a nonzero 5 s such that

Positive Solutions to Linear Systems: Convexity and Separation

421

º5 Á » b º5 Á » ~ º5 Á b » )5 ) for all : and * . But if º5 Á » £ for some : , then we can replace by an appropriate scalar multiple of in order to make the left side of this inequality negative, which is impossible. Hence, º5 Á » ~ for all : , that is, 5 : and º5 Á *» )5 )

We can now prove (15.1) and (15.2). Theorem 15.7 Let : be a subspace of s . 1) : q sb ~ J if and only if : q sbb £ J 2) : q sbb ~ J if and only if : q sb £J Proof. In both cases, one direction is easy. It is clear that there cannot exist vectors " sbb and # sb that are orthogonal. Hence, : q sb and : q sbb cannot both be nonempty and so : q sbb £ J implies : q sb ~ J. Also, : q sbb and : q sb cannot both be nonempty and so : q sb £ J implies that : q sbb ~ J. For the converse in part 1), to prove that : q sb ~ J

¬

: q sbb £ J

a good candidate for an element of : q sbb would be a normal to a hyperplane that separates : from a subset of sb . Note that our separation theorems do not allow us to separate : from sb , because sb is not compact. So consider instead the convex hull " of the standard basis vectors Á Ã Á in sb : " ~ ¸! b Ä b ! ! Á '! ~ ¹ which is compact. Moreover, " sb implies that " q : ~ J and so Theorem 15.6 implies that there is a nonzero vector 5 ~ ² Á Ã Á ³ : such that º5 Á » )5 ) for all "À Taking ~ gives ~ º5 Á » )5 ) and so 5 : q sbb , which is therefore nonempty. To prove part 2), again we note that there cannot exist orthogonal vectors and so : q sbb and : q sb cannot both be nonempty. " sbb and # sb Thus, : q sb £ J implies that : q sbb ~ J.

422

Advanced Linear Algebra

To finish the proof of part 2), we must prove that : q sbb ~ J

: q sb £J

¬

Let 8 ~ ¸) Á Ã Á ) ¹ be a basis for : . Then 5 ~ ² Á Ã Á ³ : if and only if 5 ) for all . In matrix terms, if 4 ~ ²Á ³ ~ ²) ) Ä ) ³ has rows 9 Á Ã Á 9 , then 5 : if and only if 5 4 ~ , that is, 9 b Ä b 9 ~ Now, : contains a strictly positive vector 5 ~ ² Á Ã Á ³ if and only if this equation holds, where for all and for some . Moreover, we may assume without loss of generality that ' ~ , or equivalently, that is in the convex hull 9 of the row space of 4 . Hence, : q sb £ J

¯

9

¬

9

Thus, we wish to prove that : q sbb ~ J or equivalently, ¤9

¬

: q sbb £ J

Now we have something to separate. Since 9 is closed and convex, Theorem 15.5 implies that there is a nonzero vector ) ~ ² Á Ã Á ³ s for which º)Á 9» )) ) Consider the vector # ~ ) b Ä b ) : The th coordinate of # is Á b Ä b Á ~ º)Á 9 » )) ) and so # is strongly positive. Hence, # : q sbb , which is therefore nonempty.

Inhomogeneous Systems We now turn our attention to inhomogeneous systems (% ~ The following lemma is required.

Positive Solutions to Linear Systems: Convexity and Separation

423

Lemma 15.8 Let ( CÁ ²s³. Then the set 9 ~ ¸(& & s Á & ¹ is a closed, convex cone. Proof. We leave it as an exercise to prove that 9 is a convex cone and omit the proof that 9 is closed. Theorem 15.9 (Farkas's lemma) Let ( CÁ ²s³ and let s be nonzero. Then exactly one of the following holds: 1) There is a strictly positive solution " s to the system (% ~ . 2) There is a vector # s for which #( and º#Á » . Proof. Suppose first that 1) holds. If 2) also holds, then ²#(³" ~ #²("³ ~ º#Á » However, #( and " imply that ²#(³" . This contradiction implies that 2) cannot hold. Assume now that 1) fails to hold. By Lemma 15.8, the set * ~ ¸(& & s Á & ¹ s is closed and convex. The fact that 1) fails to hold is equivalent to ¤ * . Hence, there is a hyperplane that strongly separates and * . All we require is that and * be strictly separated, that is, for some s and # s , º#Á %» º#Á » for all % * Since * , it follows that and so º#Á » . Also, the first inequality is equivalent to º#Á (&» , that is, º(! #Á &» for all & s Á & . We claim that this implies that (! # cannot have any positive coordinates and thus #( . For if the th coordinate ²(! #³ is positive, then taking & ~ for we get ²(! #³ which does not hold for large . Thus, 2) holds. In the exercises, we ask the reader to show that the previous result cannot be improved by replacing #( in statement 2) with #( .

Exercises 1.

Show that any hyperplane has the form >²5 Á )5 ) ³ for an appropriate vector 5 .

424 2. 3. 4.

Advanced Linear Algebra If ( is an d matrix prove that the set ¸(% % s Á % ¹ is a convex cone in s . If ( and ) are strictly separated subsets of s and if ( is finite, prove that ( and ) are strongly separated as well. Let = be a vector space over a field - with char²- ³ £ . Show that a subset ? of = is closed under the taking of convex combinations of any two of its points if and only if ? is closed under the taking of arbitrary convex combinations, that is, for all ,

% Á Ã Á % ?Á ~ Á ¬ % ? ~

5. 6À

~

Explain why an ² c ³-dimensional subspace of s is the solution set of a linear equation of the form % b Ä b % ~ . Show that >b ²5 Á ³ q >c ²5 Á ³ ~ >²5 Á ³ k and that >kb ²5 Á ³, >c ²5 Á ³ and >²5 Á ³ are pairwise disjoint and k ²5 Á ³ r >²5 Á ³ ~ s >kb ²5 Á ³ r >c

7. 8. 9. 10. 11. 12. 13. 14.

A function ; ¢ s ¦ s is affine if it has the form ; ²#³ ~ # b for s , where B²s Á s ³. Prove that if * s is convex, then so is ; ²*³ s . Find a cone in s that is not convex. Prove that a subset ? of s is a convex cone if and only if %Á & ? implies that % b & ? for all Á . Prove that the convex hull of a set ¸% Á Ã Á % ¹ in s is bounded, without using the fact that it is compact. Suppose that a vector % s has two distinct representations as convex combinations of the vectors # Á Ã Á # . Prove that the vectors # c # Á Ã Á # c # are linearly dependent. Suppose that * is a nonempty convex subset of s and that >²5 Á ³ is a hyperplane disjoint from * . Prove that * lies in one of the open half-spaces determined by >²5 Á ³. Prove that the conclusion of Theorem 15.6 may fail if we assume only that * is closed and convex. Find two nonempty convex subsets of s that are strictly separated but not strongly separated. Prove that ? and @ are strongly separated by >²5 Á ³ if and only if º5 Á %Z » for all %Z ? and º5 Á &Z » for all &Z @ where ? ~ ? b )²Á ³ and @ ~ @ b )²Á ³ and where )²Á ³ is the closed unit ball.

Positive Solutions to Linear Systems: Convexity and Separation

425

15. Show that Farkas's lemma cannot be improved by replacing #( in statement 2) with #( . Hint: A nice counterexample exists for ~ Á ~ .

Chapter 16

Affine Geometry

In this chapter, we will study the geometry of a finite-dimensional vector space = , along with its structure-preserving maps. Throughout this chapter, all vector spaces are assumed to be finite-dimensional.

Affine Geometry The cosets of a quotient space have a special geometric name. Definition Let : be a subspace of a vector space = . The coset # b : ~ ¸# b

:¹

is called a flat in = with base : and flat representative #. We also refer to # b : as a translate of : . The set 7²= ³ of all flats in = is called the affine geometry of = . The dimension dim²7²= ³³ of 7²= ³ is defined to be dim²= ³. While a flat may have many flat representatives, it only has one base since % b : ~ & b ; implies that % & b ; and so % b : ~ & b ; ~ % b ; , whence : ~ ; . Definition The dimension of a flat % b : is dim²:³. A flat of dimension is called a -flat. A -flat is a point, a -flat is a line and a -flat is a plane. A flat of dimension dim²7²= ³³ c is called a hyperplane. Definition Two flats ? ~ % b : and @ ~ & b ; are said to be parallel if : ; or ; : . This is denoted by ? @ . We will denote subspaces of = by the letters :Á ; Á Ã and flats in = by ?Á @ Á Ã . Here are some of the basic intersection properties of flats.

428

Advanced Linear Algebra

Theorem 16.1 Let : and ; be subspaces of = and let ? ~ % b : and @ ~ & b ; be flats in = . 1) The following are equivalent: a) some translate of ? is in @ : $ b ? @ for some $ = b) some translate of : is in ; : # b : ; for some # = c) : ; 2) The following are equivalent: a) ? and @ are translates: $ b ? ~ @ for some $ = b) : and ; are translates: # b : ~ ; for some # = c) : ~ ; 3) ? q @ £ JÁ : ; ¯ ? @ 4) ? q @ £ JÁ : ~ ; ¯ ? ~ @ 5) If ? @ then ? @ , @ ? or ? q @ ~ J 6) ? @ if and only if some translation of one of these flats is contained in the other. Proof. If 1a) holds, then c& b $ b % b : ; and so 1b) holds. If 1b) holds, then # ; and so : ~ ²# b :³ c # ; and so 1c) holds. If 1c) holds, then & c % b ? ~ & b : & b ; ~ @ and so 1a) holds. Part 2) is proved in a similar manner. For part 3), : ; implies that # b ? @ for some # = and so if ' ? q @ then # b ' @ and so # @ , which implies that ? @ . Conversely, if ? @ then part 1) implies that : ; . Part 4) follows similarly. We leave proof of 5) and 6) to the reader.

Affine Combinations Let ? be a nonempty subset of = . It is well known that 1) ? is a subspace of = if and only if ? is closed under linear combinations, or equivalently, ? is closed under linear combinations of any two vectors in ? . 2) The smallest subspace of = containing ? is the set of all linear combinations of elements of ? . In different language, the linear hull of ? is equal to the linear span of ? . We wish to establish the corresponding properties of affine subspaces of = , beginning with the counterpart of a linear combination. Definition Let = be a vector space and let % = . A linear combination % b Ä b % where - and ~ is called an affine combination of the vectors % . Let us refer to a nonempty subset ? of = as affine closed if ? is closed under any affine combination of vectors in ? and two-affine closed if ? is closed

Affine Geometry

429

under affine combinations of any two vectors in ? . These are not standard terms. The line containing two distinct vectors %Á & = is the set %& ~ ¸% b ² c ³& - ¹ ~ & b º% c &» of all affine combinations of % and &. Thus, a subset ? of = is two-affine closed if and only if ? contains the line through any two of its points. Theorem 16.2 Let = be a vector space over a field - with char²- ³ £ . Then a subset ? of = is affine closed if and only if it is two-affine closed. Proof. The theorem is proved by induction on the number of terms in an affine combination. The case ~ holds by assumption. Assume the result true for affine combinations with fewer than terms and consider the affine combination ' ~ % b Ä b % where . There are two cases to consider. If either of and is not equal to , say £ , write ' ~ % b ² c ³>

% b Ä b % ? c c

and if ~ ~ , then since char²- ³ £ 2, we may write ' ~ > % b % ? b 3 % 3 b Ä b % In either case, the inductive hypothesis applies to the expression inside the square brackets and then to ' . The requirement char²- ³ £ is necessary, for if - ~ { , then the subset ? ~ ¸²Á ³Á ²Á ³Á ²Á ³¹ of - is two-affine closed but not affine closed. We can now characterize flats. Theorem 16.3 A nonempty subset ? of a vector space = is a flat if and only if ? is affine closed. Moreover, if char²- ³ £ , then ? is a flat if and only if ? is two-affine closed. Proof. Let ? ~ % b : be a flat and let % ~ % b ? , where : . If ' ~ , then % ~ ²% b ³ ~ % b

%b:

and so ? is affine closed. Conversely, suppose that ? is affine closed, let % ? and let : ~ ? c %. If - and : then

430

Advanced Linear Algebra

b

~ ²% c %³ b ²% c %³ ~ % b % c ² b ³%

for % ? . Since the sum of the coefficients of % , % and % in the last expression is , it follows that

b

b % ~ % b % c ² b c ³% ?

and so b ? c % ~ : . Thus, : is a subspace of = and ? ~ % b : is a flat. The rest follows from Theorem 16.2.

Affine Hulls The following definition is the analog of the subspace spanned by a collection of vectors. Definition Let ? be a nonempty set of vectors in = . 1) The affine hull of ? , denoted by affhull²?³, is the smallest flat containing ?. 2) The affine span of ? , denoted by affspan²?³, is the set of all affine combinations of vectors in ? . Theorem 16.4 Let ? be a nonempty subset of = . Then affhull²?³ ~ affspan²?³ ~ % b span²? c %³ or equivalently, for a subspace : of = , % b : ~ affspan²?³

¯

: ~ span²? c %³

Also, dim²affspan²?³³ ~ dim²span²? c %³³ Proof. Theorem 16.3 implies that affspan²?³ affhull²?³ and so it is sufficient to show that ( ~ affspan²?³ is a flat, or equivalently, that for any & (, the set : ~ ( c & is a subspace of = . To this end, let

& ~ Á % ~

Then any two elements of : have the form & c & and & c & , where

& ~ Á % ~

are in (. But if Á ! - , then

and

& ~ 2Á % ~

Affine Geometry

431

' ²& c &³ b !²& c &³

~ 4 Á b !2Á c ² b !³Á 5% ~

~ 4 Á b !2Á c ² b ! c ³Á 5% c & ~

which is in ( c & ~ : , since the last sum is an affine sum. Hence, : is a subspace of = . We leave the rest of the proof to the reader.

The Lattice of Flats The intersection of subspaces is a subspace, although it may be trivial. For flats, if the intersection is not empty, then it is also a flat. However, since the intersection of flats may be empty, the set 7²= ³ does not form a lattice under intersection. However, we can easily fix this. Theorem 16.5 Let = be a vector space. The set 7 ²= ³ ~ 7²= ³ r ¸J¹ of all flats in = , together with the empty set, is a complete lattice in which meet is intersection. In particular: 1) 7 ²= ³ is closed under arbitrary intersection. In fact, if < ~ ¸% b : 2¹ has nonempty intersection, then

< ~ ²% b : ³ ~ % b : 2

2

2

for some % < . In other words, the base of the intersection is the intersection of the bases. 2) The join < of the family < ~ ¸% b : 2¹ is the intersection of all flats containing the members of < . Also, < ~ affhull4 < 5 3) If ? ~ % b : and @ ~ & b ; are flats in = , then ? v @ ~ % b ²º% c &» b : b ; ³ If ? q @ £ J, then ? v @ ~ % b ²: b ; ³ Proof. For part 1), if % ²% b : ³ 2

then % b : ~ % b : for all 2 and so

432

Advanced Linear Algebra

²% b : ³ ~ ²% b : ³ ~ % b : 2

2

2

We leave proof of part 2) to the reader. For part 3), since %Á & ? v @ , it follows that ?v@ ~%b< ~&b< for some subspace < of = . Thus, % c & < . Also, % b : % b < implies that : < and similarly ; < , whence : b ; < and so if > ~ º% c &» b : b ; , then > < . Hence, % b > % b < ~ ? v @ . On the other hand, ? ~%b: %b> and @ ~ & b ; ~ % c ²% c &³ b ; % b > and so ? v @ % b > . Thus, ? v @ ~ % b > . If ? q @ £ J, then we may take the flat representatives for ? and @ to be any element ' ? q @ , in which case part 1) gives ? v @ ~ ' b ²º' c '» b : b ; ³ ~ ' b : b ; and since % ? v @ , we also have ? v @ ~ % b : b ; . We can now describe the dimension of the join of two flats. Theorem 16.6 Let ? ~ % b : and @ ~ & b ; be flats in = . 1) If ? q @ £ J, then dim²? v @ ³ ~ dim²: b ; ³ ~ dim²?³ b dim²@ ³ c dim²? q @ ³ 2) If ? q @ ~ J, then dim²? v @ ³ ~ dim²: b ; ³ b Proof. We have seen that if ? q @ £ J, then ?v@ ~%b: b; and so dim²? v @ ³ ~ dim²: b ; ³ On the other hand, if ? q @ ~ J, then ? v @ ~ % b ²º% c &» b : b ; ³ and since dim²º% c &»³ ~ , we get

Affine Geometry

433

dim²? v @ ³ ~ dim²: b ; ³ b Finally, we have dim²: b ; ³ ~ dim²:³ b dim²; ³ c dim²: q ; ³ and Theorem 16.5 implies that dim²? q @ ³ ~ dim²: q ; ³

Affine Independence We now discuss the affine counterpart of linear independence. Theorem 16.7 Let ? be a nonempty set of vectors in = . The following are equivalent: 1) For all % ? , the set ²? c %³ ± ¸¹ is linearly independent. 2) For all % ? , % ¤ affhull²? ± ¸%¹³ 3) For any vectors % ? , % ~ Á ~

~ for all

¬

4) For affine combinations of vectors in ? , % ~ %

¬

~

for all

5) When ? ~ ¸% Á Ã Á % ¹ is finite, dim²affhull²?³³ ~ c A set ? of vectors satisfying any (any hence all) of these conditions is said to be affinely independent. Proof. If 1) holds but there is an affine combination equal to %,

% ~ % ~

where % £ % for all , then

²% c %³ ~ ~

434

Advanced Linear Algebra

Since is nonzero for some , this contradicts 1). Hence, 1) implies 2). Suppose that 2) holds and % ~ where ~ . If some , say , is nonzero then % ~ c ² ° ³% affhull²? ± ¸% ¹³

which contradicts 2) and so ~ for all . Hence, 2) implies 3). If 3) holds and the affine combinations satisfy % ~ % then ² c ³% ~ and since ² c ³ ~ c ~ , it follows that ~ for all . Hence, 4) holds. Thus, it is clear that 3) and 4) are equivalent. If 3) holds and ²% c %³ ~ for % £ % , then % c 4 5% ~ and so 3) implies that ~ for all . Finally, suppose that ? ~ ¸% Á Ã Á % ¹. Since dim²affhull²?³³ ~ dim²º? c % »³ it follows that 5) holds if and only if ²? c % ³ ± ¸¹, which has size c , is linearly independent. Affinely independent sets enjoy some of the basic properties of linearly independent sets. For example, a nonempty subset of an affinely independent set is affinely independent. Also, any nonempty set ? contains an affinely independent set. Since the affine hull / ~ affhull²?³ of an affinely independent set ? is not the affine hull of any proper subset of ? , we deduce that ? is a minimal affine spanning set of its affine hull.

Affine Geometry

435

Affine Bases and Barycentric Coordinates We have seen that a set ? is affinely independent if and only if the set ?% ~ ²? c %³ ± ¸¹ is linearly independent. We have also seen that for a subsapce : of = , % b : ~ affspan²?³

¯

: ~ span²?% ³

Therefore, if by analogy, we define a subset 8 of a flat ( ~ % b : to be an affine basis for ( if 8 is affinely independent and affspan²8 ³ ~ (, then 8 is an affine basis for % b : if and only if 8% is a basis for : . Theorem 16.8 Let ( ~ % b : be a flat of dimension . Let 8 ~ ²% Á Ã Á % ³ be an ordered basis for : and let ²8 b %³ r ¸%¹ ~ ²% b %Á Ã Á % b %Á %³ be an ordered affine basis for (. Then every # ( has a unique expression as an affine combination # ~ % b Ä b % b b % The coefficients are called the barycentric coordinates of # with respect to the ordered affine basis 8 b %. For example, in s , a plane is a flat of the form ( ~ % b º# Á # » where 8 ~ ²# Á # ³ is an ordered basis for a two-dimensional subspace of s . Then ²8 b %³ r ¸%¹ ~ ²# b %Á # b %Á %³ ~ ² Á Á ³ are barycentric coordinates for the plane, that is, any # ( has the form b b where b b ~ .

Affine Transformations Now let us discuss some properties of maps that preserve affine structure. Definition A function ¢ = ¦ = that preserves affine combinations, that is, for which ~ ¬ 4 % 5 ~ ²% ³

is called an affine transformation (or affine map, or affinity). We should mention that some authors require that be bijective in order to be an affine map. The following theorem is the analog of Theorem 16.2.

436

Advanced Linear Algebra

Theorem 16.9 If char²- ³ £ , then a function ¢ = ¦ = is an affine transformation if and only if it preserves affine combinations of every pair of its points, that is, if and only if ²% b ² c ³&³ ~ ²%³ b ² c ³ ²&³

Thus, if char²- ³ £ , then a map is an affine transformation if and only if it sends the line through % and & to the line through ²%³ and ²&³. It is clear that linear transformations are affine transformations. So are the following maps. Definition Let # = . The affine map ;# ¢ = ¦ = defined by ;# ²%³ ~ % b # for all % = , is called translation by #. It is not hard to see that any composition ;# k , where B²= ³, is affine. Conversely, any affine map must have this form. Theorem 16.10 A function ¢ = ¦ = is an affine transformation if and only if it is a linear operator followed by a translation, ~ ;# k where # = and B²= ³. Proof. We leave proof that ;# k is an affine transformation to the reader. Let be an affine map and suppose that ~ c' . Then ;' k ~ . Moreover, letting ~ ;' k , we have ²" b #³ ~ ²" b #³ b ' ~ ²" b # b ² c c ³³ b ' ~ " b # c ² c c ³' b ' ~ " b # and so is linear. Corollary 16.11 1) The composition of two affine transformations is an affine transformation. 2) An affine transformation ~ ;# k is bijective if and only if is bijective. 3) The set aff²= ³ of all bijective affine transformations on = is a group under composition of maps, called the affine group of = . Let us make a few group-theoretic remarks about aff²= ³. The set trans²= ³ of all translations of = is a subgroup of aff²= ³. We can define a function ¢ aff²= ³ ¦ B²= ³ by ²;# k ³ ~

Affine Geometry

437

It is not hard to see that is a well-defined group homomorphism from aff²= ³ onto B²= ³, with kernel trans²= ³. Hence, trans²= ³ is a normal subgroup of aff²= ³ and aff²= ³ B²= ³ trans²= ³

Projective Geometry If dim²= ³ ~ , the join (affine hull) of any two distinct points in = is a line. On the other hand, it is not the case that the intersection of any two lines is a point, since the lines may be parallel. Thus, there is a certain asymmetry between the concepts of points and lines in = . This asymmetry can be removed by constructing the projective plane. Our plan here is to very briefly describe one possible construction of projective geometries of all dimensions. By way of motivation, let us consider Figure 16.1.

Figure 16.1 Note that / is a hyperplane in a 3-dimensional vector space = and that ¤ / . Now, the set 7²/³ of all flats of = that lie in / is an affine geometry of dimension . (According to our definition of affine geometry, / must be a vector space in order to define 7²/³. However, we hereby extend the definition of affine geometry to include the collection of all flats contained in a flat of = .³ Figure 16.1 shows a one-dimensional flat ? and its linear span º?», as well as a zero-dimensional flat @ and its span º@ ». Note that, for any flat ? in / , we have dim²º?»³ ~ dim²?³ b Note also that if 3 and 3 are any two distinct lines in / , the corresponding

438

Advanced Linear Algebra

planes º3 » and º3 » have the property that their intersection is a line through the origin, even if the lines are parallel. We are now ready to define projective geometries. Let = be a vector space of any dimension and let / be a hyperplane in = not containing the origin. To each flat ? in / , we associate the subspace º?» of = generated by ? . Thus, the linear span function 7 ¢ 7²/³ ¦ I ²= ³ maps affine subspaces ? of / to subspaces º?» of = . The span function is not surjective: Its image is the set of all subspaces that are not contained in the base subspace 2 of the flat / . The linear span function is one-to-one and its inverse is intersection with / , 7 c < ~ < q / for any subspace < not contained in 2 . The affine geometry 7²/³ is, as we have remarked, somewhat incomplete. In the case dim²/³ ~ , every pair of points determines a line but not every pair of lines determines a point. Now, since the linear span function 7 is injective, we can identify 7²/³ with its image 7 ²7²/³³, which is the set of all subspaces of = not contained in the base subspace 2 . This view of 7²/³ allows us to “complete” 7²/³ by including the base subspace 2 . In the three-dimensional case of Figure 16.1, the base plane, in effect, adds a projective line at infinity. With this inclusion, every pair of lines intersects, parallel lines intersecting at a point on the line at infinity. This two-dimensional projective geometry is called the projective plane. Definition Let = be a vector space. The set I ²= ³ of all subspaces of = is called the projective geometry of = . The projective dimension pdim²:³ of : I ²= ³ is defined as pdim²:³ ~ dim²:³ c The projective dimension of F²= ³ is defined to be pdim²= ³ ~ dim²= ³ c . A subspace of projective dimension is called a projective point and a subspace of projective dimension is called a projective line. Thus, referring to Figure 16.1, a projective point is a line through the origin and, provided that it is not contained in the base plane 2 , it meets / in an affine point. Similarly, a projective line is a plane through the origin and, provided that it is not 2 , it will meet / in an affine line. In short, span²affine point³ ~ line through the origin ~ projective point span²affine line³ ~ plane through the origin ~ projective line The linear span function has the following properties.

Affine Geometry

439

Theorem 16.12 The linear span function 7 ¢ 7²/³ ¦ I ²= ³ from the affine geometry 7²/³ to the projective geometry I ²= ³ defined by 7 ? ~ º?» satisfies the following properties: 1) The linear span function is injective, with inverse given by 7 c < ~ < q / for all subspaces < not contained in the base subspace 2 of / . 2) The image of the span function is the set of all subspaces of = that are not contained in the base subspace 2 of / . 3) ? @ if and only if º?» º@ » 4) If ? are flats in / with nonempty intersection, then span4 ? 5 ~ span²? ³ 2

2

5) For any collection of flats in / , span8 ? 9 ~ span²? ³ 2

2

6) The linear span function preserves dimension, in the sense that pdim²span²?³³ ~ dim²?³ 7) ? @ if and only if one of º?» q 2 and º@ » q 2 is contained in the other. Proof. To prove part 1), let % b : be a flat in / . Then % / and so / ~ % b 2 , which implies that : 2 . Note also that º% b :» ~ º%» b : and ' º% b :» q / ~ ²º%» b :³ q ²% b 2³ ¬ ' ~ % b ~ % b for some : , 2 and - . This implies that ² c ³% 2 , which implies that either % 2 or ~ . But % / implies % ¤ 2 and so ~ , which implies that ' ~ % b % b : . In other words, º% b :» q / % b : Since the reverse inclusion is clear, we have º% b :» q / ~ % b : This establishes 1). To prove 2), let < be a subspace of = that is not contained in 2 . We wish to show that < is in the image of the linear span function. Note first that since < 2 and dim²2³ ~ dim²= ³ c , we have < b 2 ~ = and so dim²< q 2³ ~ dim²< ³ b dim²2³ c dim²< b 2³ ~ dim²< ³ c

440

Advanced Linear Algebra

Now let £ % < c 2 . Then % ¤ 2 ¬ º%» b 2 ~ = ¬ % b / for some £ - Á 2 ¬ % / Thus, % < q / for some £ - . Hence, the flat % b ²< q 2³ lies in / and dim²% b ²< q 2³³ ~ dim²< q 2³ ~ dim²< ³ c which implies that span²% b ²< q 2³³ ~ º%» b ²< q 2³ lies in < and has the same dimension as < . In other words, span²% b ²< q 2³³ ~ º%» b ²< q 2³ ~ < We leave proof of the remaining parts of the theorem as exercises.

Exercises 1. 2.

Show that if % Á Ã Á % = , then the set : ~ ¸' % ' ~ ¹ is a subspace of = . Prove that if ? = is nonempty then affhull²?³ ~ % b º? c %»

Prove that the set ? ~ ¸²Á ³Á ²Á ³Á ²Á ³¹ in ²{ ³ is closed under the formation of lines, but not affine hulls. 4. Prove that a flat contains the origin if and only if it is a subspace. 5. Prove that a flat ? is a subspace if and only if for some % ? we have % ? for some £ - . 6. Show that the join of a collection 9 ~ ¸% b : 2¹ of flats in = is the intersection of all flats that contain all flats in 9 . 7. Is the collection of all flats in = a lattice under set inclusion? If not, how can you “fix” this? 8. Suppose that ? ~ % b : and @ ~ & b ; . Prove that if dim²?³ ~ dim²@ ³ and ? @ , then : ~ ; . 9. Suppose that ? ~ % b : and @ ~ & b ; are disjoint hyperplanes in = . Show that : ~ ; . 10. (The parallel postulate) Let ? be a flat in = and # ¤ ? . Show that there is exactly one flat containing #, parallel to ? and having the same dimension as ? . 11. a) Find an example to show that the join ? v @ of two flats may not be the set of all lines connecting all points in the union of these flats. b) Show that if ? and @ are flats with ? q @ £ J, then ? v @ is the union of all lines %& where % ? and & @ . 12. Show that if ? @ and ? q @ ~ J, then 3.

dim²? v @ ³ ~ max¸dim²?³Á dim²@ ³¹ b

Affine Geometry

441

13. Let dim²= ³ ~ . Prove the following: a) The join of any two distinct points is a line. b) The intersection of any two nonparallel lines is a point. 14. Let dim²= ³ ~ . Prove the following: a) The join of any two distinct points is a line. b) The intersection of any two nonparallel planes is a line. c) The join of any two lines whose intersection is a point is a plane. d) The intersection of two coplanar nonparallel lines is a point. e) The join of any two distinct parallel lines is a plane. f) The join of a line and a point not on that line is a plane. g) The intersection of a plane and a line not on that plane is a point. 15. Prove that ¢ = ¦ = is a surjective affine transformation if and only if ~ k ;$ for some $ = and B²= ³. 16. Verify the group-theoretic remarks about the group homomorphism ¢ aff²= ³ ¦ B²= ³ and the subgroup trans²= ³ of aff²= ³.

Chapter 17

Singular Values and the Moore–Penrose Inverse

Singular Values Let < and = be finite-dimensional inner product spaces over d or s and let B²< Á = ³. The spectral theorem applied to i can be of considerable help in understanding the relationship between and its adjoint i . This relationship is shown in Figure 17.1. Note that < and = can be decomposed into direct sums < ~(l)

and

= ~* l+

in such a manner that ¢ ( ¦ * and i ¢ * ¦ ( act symmetrically in the sense that ¢ " ª

#

and

i ¢ # ª

"

Also, both and i are zero on ) and +, respectively. We begin by noting that i B²< ³ is a positive Hermitian operator. Hence, if ~ rk² ³ ~ rk² i ³, then < has an ordered orthonormal basis 8 ~ ²" Á Ã Á " Á "b Á Ã Á " ³ of eigenvectors for i , where the corresponding eigenvalues can be arranged so that Ä ~ b ~ Ä ~ The set ²"b Á Ã Á " ³ is an ordered orthonormal basis for ker² i ³ ~ ker² ³ and so ²" Á Ã Á " ³ is an ordered orthonormal basis for ker² ³ ~ im² i ³.

444

Advanced Linear Algebra

im(W*) u1 ONB of eigenvectors for W*W

im(W) v1

W(uk)=skvk W (vk)=skuk

ur ur+1

W

vr+1

un

W

vm

ker(W)

ONB of eigenvectors for WW*

vr

ker(W*) Figure 17.1

For ~ Á Ã Á , the positive numbers of . If we set ~ for , then

~ j are called the singular values

i " ~

"

for ~ Á Ã Á . We can achieve some “symmetry” here between and i by setting # ~ ²° ³ " for each , giving #

"

" ~ F and i # ~ F

The vectors # Á Ã Á # are orthonormal, since if Á , then º# Á # » ~

º " Á " » ~

º i " Á " » ~

º" Á " » ~ Á

Hence, ²# Á Ã Á # ³ is an orthonormal basis for im² ³ ~ ker² i ³ , which can be extended to an orthonormal basis 9 ~ ²# Á Ã Á # ³ for = , the extension ²#b Á Ã Á # ³ being an orthonormal basis for ker² i ³. Moreover, since i # ~

"

~

#

the vectors # Á Ã Á # are eigenvectors for i with the same eigenvalues ~ as for i . This completes the picture in Figure 17.1.

Singular Values and the Moore–Penrose Inverse

445

Theorem 17.1 Let < and = be finite-dimensional inner product spaces over d or s and let B²< Á = ³ have rank . Then there are ordered orthonormal bases 8 and 9 for < and = , respectively, for which 8 ~ ² " Á Ã Á " Á "b Á Ã Á " ³ ONB for im² i ³

ONB for ker² ³

and 9 ~ ² # Á Ã Á # Á #b Á Ã Á # ³ ONB for im² ³

ONB for ker² i ³

Moreover, for , " ~ i # ~ where

# "

are called the singular values of , defined by i " ~

" Á

for . The vectors " Á Ã Á " are called the right singular vectors for and the vectors # Á Ã Á # are called the left singular vectors for . The matrix version of the previous discussion leads to the well-known singularvalue decomposition of a matrix. Let ( CÁ ²- ³ and let 8 ~ ²" Á Ã Á " ³ and 9 ~ ²# Á Ã Á # ³ be the orthonormal bases from < and = , respectively, in Theorem 17.1, for the operator ( . Then ´ µ8Á9 ~ ' ~ diag² Á

Á Ã Á Á Á Ã Á ³

A change of orthonormal bases from the standard bases to 9 and : gives ( ~ ´( µ; Á; ~ 49Á; ´( µ8Á9 4; Á8 ~ 7 '8i where 7 ~ 49Á; and 8 ~ 48Á; are unitary/orthogonal. This is the singularvalue decomposition of (. As to uniqueness, if ( ~ 7 '8i , where 7 and 8 are unitary and ' is diagonal, with diagonal entries , then (i ( ~ ²7 '8i ³i 7 '8i ~ 8'i '8i and since 'i ' ~ diag² Á Ã Á ³, it follows that the 's are eigenvalues of (i (, that is, they are the squares of the singular values along with a sufficient number of 's. Hence, ' is uniquely determined by (, up to the order of the diagonal elements.

446

Advanced Linear Algebra

We state without proof the following uniqueness facts and refer the reader to [48] for details. If and if the eigenvalues are distinct, then 7 is uniquely determined up to multiplication on the right by a diagonal matrix of the form + ~ diag²' Á Ã Á ' ³ with (' ( ~ . If , then 8 is never uniquely determined. If ~ ~ , then for any given 7 there is a unique 8. Thus, we see that, in general, the singular-value decomposition is not unique.

The Moore–Penrose Generalized Inverse Singular values lead to a generalization of the inverse of an operator that applies to all linear transformations. The setup is the same as in Figure 17.1. Referring to that figure, we are prompted to define a linear transformation b ¢ = ¦ < by b # ~ F

"

for for

since then ² b ³Oº" ÁÃÁ" » ~ ² b ³Oº"b ÁÃÁ" » ~ and ² b ³Oº# ÁÃÁ# » ~ ² b ³Oº#b ÁÃÁ# » ~ Hence, if ~ ~ , then b ~ c . The transformation b is called the Moore–Penrose generalized inverse or Moore–Penrose pseudoinverse of . We abbreviate this as MP inverse. Note that the composition b is the identity on the largest possible subspace of < on which any composition of the form could be the identity, namely, the orthogonal complement of the kernel of . A similar statement holds for the composition b . Hence, b is as “close” to an inverse for as is possible. We have said that if is invertible, then b ~ c . More is true: If is injective, then b ~ and so b is a left inverse for . Also, if is surjective, then b is a right inverse for . Hence the MP inverse b generalizes the onesided inverses as well. Here is a characterization of the MP inverse. Theorem 17.2 Let B²< Á = ³. The MP inverse b of is completely characterized by the following four properties: 1) b ~ 2) b b ~ b 3) b is Hermitian 4) b is Hermitian

Singular Values and the Moore–Penrose Inverse

447

Proof. We leave it to the reader to show that b does indeed satisfy conditions 1)–4) and prove only the uniqueness. Suppose that and satisfy 1)–4) when substituted for b . Then ~ ~ ² ³ i ~ i i ~ ² ³ i i ~ i i i i ~ ² ³ i i i ~ i i ~ ~ and ~ ~ ² ³ i ~ i i ~ i ² ³i ~ i i i i ~ i i ² ³i ~ i i ~ ~ which shows that ~ . The MP inverse can also be defined for matrices. In particular, if ( 4Á ²- ³, then the matrix operator ( has an MP inverse (b . Since this is a linear transformation from - to - , it is just multiplication by a matrix (b ~ ) . This matrix ) is the MP inverse for ( and is denoted by (b . Since (b ~ (b and () ~ ( ) , the matrix version of Theorem 17.2 implies that (b is completely characterized by the four conditions 1) 2) 3) 4)

((b ( ~ ( (b ((b ~ (b ((b is Hermitian (b ( is Hermitian

Moreover, if ( ~ < '<i is the singular-value decomposition of (, then

448

Advanced Linear Algebra

(b ~ < 'Z <i where 'Z is obtained from ' by replacing all nonzero entries by their multiplicative inverses. This follows from the characterization above and also from the fact that for , < 'Z <i # ~ < 'Z ~

c <

~

c "

and for , < 'Z <i # ~ < 'Z ~

Least Squares Approximation Let us now discuss the most important use of the MP inverse. Consider the system of linear equations (% ~ # where ( 4Á ²- ³. (As usual, - ~ d or - ~ s.) This system has a solution if and only if # im²( ³. If the system has no solution, then it is of considerable practical importance to be able to solve the system (% ~ V# where V# is the unique vector in im²( ³ that is closest to #, as measured by the unitary (or Euclidean) distance. This problem is called the linear least squares problem. Any solution to the system (% ~ V# is called a least squares solution to the system (% ~ #. Put another way, a least squares solution to (% ~ # is a vector % for which )(% c #) is minimized. Suppose that $ and ' are least squares solutions to (% ~ #. Then ($ ~ V# ~ (' and so $ c ' ker²(³. (We will write ( for ( .) Thus, if $ is a particular least squares solution, then the set of all least squares solutions is $ b ker²(³. Among all solutions, the most interesting is the solution of minimum norm. Note that if there is a least squares solution $ that lies in ker²(³ , then for any ' ker²(³, we have )$ b ' ) ~ )$) b )' ) )$) and so $ will be the unique least squares solution of minimum norm. Before proceeding, we recall (Theorem 9.14) that if : is a subspace of a finitedimensional inner product space = , then the best approximation to a vector # = from within : is the unique vector V# : for which # c V# : . Now we can see how the MP inverse comes into play.

Singular Values and the Moore–Penrose Inverse

449

Theorem 17.3 Let ( 4Á ²- ³. Among the least squares solutions to the system (% ~ V# there is a unique solution of minimum norm, given by (b #, where (b is the MP inverse of (. Proof. A vector $ is a least squares solution if and only if ($ ~ V#. Using the characterization of the best approximation V#, we see that $ is a solution to ($ ~ V# if and only if ($ c # im²(³ Since im²(³ ~ ker²(i ³ this is equivalent to (i ²($ c #³ ~ or (i ($ ~ (i # This system of equations is called the normal equations for (% ~ #. Its solutions are precisely the least squares solutions to the system (% ~ #. To see that $ ~ (b # is a least squares solution, recall that, in the notation of Figure 17.1, ((b # ~ F

#

and so (i (²(b # ³ ~ F

(i #

~ ( i #

and since 9 ~ ²# Á Ã Á # ³ is a basis for = , we conclude that (b # satisfies the normal equations. Finally, since (b # ker²(³ , we deduce by the preceding remarks that (b # is the unique least squares solution of minimum norm.

Exercises 1. 2.

Let B²< ³. Show that the singular values of i are the same as those of . Find the singular values and the singular value decomposition of the matrix (~>

3.

?

Find (b . Find the singular values and the singular value decomposition of the matrix

450

Advanced Linear Algebra

(~>

4. 5.

?

Find (b . Hint: Is it better to work with (i ( or ((i ? Let ? ~ ²% % Ä % ³! be a column matrix over d. Find a singular-value decomposition of ? . Let ( 4Á ²- ³ and let ) 4bÁb ²- ³ be the square matrix )~>

(i

( ?block

Show that, counting multiplicity, the nonzero eigenvalues of ) are precisely the singular values of ( together with their negatives. Hint: Let ( ~ < '<i be a singular-value decomposition of ( and try factoring ) into a product < : <

6.

7.

8.

< ?> '

'i ?> <i

<i ?

What are the eigenvalues of the middle factor on the right? (Try b b and c b .) Use the results of the previous exercise to show that a matrix ( 4Á ²- ³, its adjoint (i , its transpose (! and its conjugate ( all have the same singular values. Show also that if < and < Z are unitary, then ( and < (< Z have the same singular values. Let ( 4 ²- ³ be nonsingular. Show that the following procedure produces a singular-value decomposition ( ~ < '<i of (. a) Write ( ~ < +< i where + ~ diag² Á Ã Á ³ and the 's are positive and the columns of < form an orthonormal basis of eigenvectors for (. (We never said that this was a practical procedure.) ° ° b) Let ' ~ diag² Á Ã Á ³ where the square roots are nonnegative. Also let < ~ < and U ~ (i < 'c . If ( ~ ²Á ³ is an d matrix, then the Frobenius norm of ( is °

)( )- ~ Show that )()- ~ (.

8 Á 9 Á

is the sum of the squares of the singular values of

Chapter 18

An Introduction to Algebras

Motivation We have spent considerable time studying the structure of a linear operator B- ²= ³ on a finite-dimensional vector space = over a field - . In our studies, we defined the - ´%µ-module = and used the decomposition theorems for modules over a principal ideal domain to dissect this module. We concentrated on an individual operator , rather than the entire vector space B- ²= ³. In fact, we have made relatively little use of the fact that B- ²= ³ is an algebra under composition. In this chapter, we give a brief introduction to the theory of algebras, of which B- ²= ³ is the most general, in the sense of Theorem 18.2 below.

Associative Algebras An algebra is a combination of a ring and a vector space, with an axiom that links the ring product with scalar multiplication. Definition An (associative) algebra ( over a field - , or an - -algebra, is a nonempty set (, together with three operations, called addition (denoted by b ), multiplication (denoted by juxtaposition) and scalar multiplication (also denoted by juxtaposition), for which the following properties hold: 1) ( is a vector space over - under addition and scalar multiplication. 2) ( is a ring with identity under addition and multiplication. 3) If - and Á (, then ²³ ~ ²³ ~ ²³ An algebra is finite-dimensional if it is finite-dimensional as a vector space. An algebra is commutative if ( is a commutative ring. An element ( is invertible if there is ( for which ~ ~ . Our definition requires that ( have a multiplicative identity. Such algebras are called unital algebras. Algebras without unit are also of great importance, but

452

Advanced Linear Algebra

we will not study them here. Also, in this chapter, we will assume that all algebras are associative. Nonassociative algebras, such as Lie algebras and Jordan algebras, are important as well.

The Center of an Algebra Definition The center of an - -algebra ( is the set A²(³ ~ ¸ ( % ~ % for all % (¹ of all elements of ( that commute with every element of (. The center of an algebra is never trivial since it contains a copy of - : ¸ - ¹ A²(³ Definition An - -algebra ( is central if its center is as small as possible, that is, if A²(³ ~ ¸ - ¹

From a Vector Space to an Algebra If = is a vector space over a field - and if 8 ~ ¸ 0¹ is a basis for = , then it is natural to wonder whether we can form an - -algebra simply by defining a product for the basis elements and then using the distributive laws to extend the product to = . In particular, we choose a set of constants Á with the property that for each pair ²Á ³, only finitely many of the Á are nonzero. Then we set ~ Á

and make multiplication bilinear, that is,

8 9 ~ ~

~

8 9 ~ ~

~

and 8 9 ~ ~

~

for - . It is easy to see that this does define a nonunital associative algebra ( provided that

An Introduction to Algebras

453

² ³ ~ ² ³ for all Á Á 0 and that ( is commutative if and only if ~ for all Á 0 . The constants Á are called the structure constants for the algebra (. To get a unital algebra, we can take for a given 0 , the structure constants to be Á ~ Á ~ Á in which case is the multiplicative identity. (An alternative is to adjoin a new element to the basis and define its structure constants in this way.)

Examples The following examples will make it clear why algebras are important. Example 18.1 If - , are fields, then , is a vector space over - . This vector space structure, along with the ring structure of , , is an algebra over - . Example 18.2 The ring - ´%µ of polynomials is an algebra over - . Example 18.3 The ring C ²- ³ of all d matrices over a field - is an algebra over - , where scalar multiplication is defined by 4 ~ ²Á ³Á

-

¬

4 ~ ²Á ³

Example 18.4 The set B- ²= ³ of all linear operators on a vector space = over a field - is an - -algebra, where addition is addition of functions, multiplication is composition of functions and scalar multiplication is given by ²³²#³ ~ ´#µ The identity map B- ²= ³ is the multiplicative identity and the zero map B- ²= ³ is the additive identity. This algebra is also denoted by End- ²= ³, since the linear operators on = are also called endomorphisms of = . Example 18.5 If . is a group and - is a field, then we can form a vector space - ´.µ over - by taking all formal - -linear combinations of elements of . and treating . as a basis for - ´.µ. This vector space can be made into an - -algebra where the structure constants are determined by the group product, that is, if ~ " , then Á ~ Á" . The group identity ~ is the algebra identity since ~ and so Á ~ Á and similarly, Á ~ Á . The resulting associative algebra - ´.µ is called the group algebra over - . Specifically, the elements of - ´.µ have the form

454

Advanced Linear Algebra

% ~ b Ä b where - and .. If &~

bÄb

then we can include additional terms with coefficients and reindex if necessary so that we may assume that ~ and ~ for all . Then the sum in - ´.µ is given by

8 9 b 8 9 ~ ² b ³ ~

~

~

Also, the product is given by

8 98 9 ~ ~

~

Á

and the scalar product is

8 9 ~ ~

~

The Usual Suspects Algebras have substructures and structure-preserving maps, as do groups, rings and other algebraic structures. Subalgebras Definition Let ( be an - -algebra. A subalgebra of ( is a subset ) of ( that is a subring of ( (with the same identity as () and a subspace of (. The intersection of subalgebras is a subalgebra and so the family of all subalgebras of ( is a complete lattice, where meet is intersection and the join of a family < of subalgebras is the intersection of all subalgebras of ( that contain the members of < . The subalgebra generated by a nonempty subset ? of an algebra ( is the smallest subalgebra of ( that contains ? and is easily seen to be the set of all linear combinations of finite products of elements of ? , that is, the subspace spanned by the products of finite subsets of elements of ? : º?»alg ~ º% Ä% % ?» Alternatively, º?»alg is the set of all polynomials in the variables in ? . In particular, the algebra generated by a single element % ( is the set of all polynomials in % over - .

An Introduction to Algebras

455

Ideals and Quotients In defining the notion of an ideal of an algebra (, we must consider the fact that ( may be noncommutative. Definition A (two-sided) ideal of an associative algebra ( is a nonempty subset 0 of ( that is closed under addition and subtraction, that is, Á 0

¬

b Á c 0

and also left and right multiplication by elements of (, that is, 0Á

Á (

¬

0

The ideal generated by a nonempty subset ? of ( is the smallest ideal containing ? and is equal to

º?»ideal ~ H % % ?Á Á (I ~

Definition An algebra ( is simple if 1) The product in ( is not trivial, that is, £ for at least one pair of elements Á ( 2) ( has no proper nonzero ideals. Definition If 0 is an ideal in (, then the quotient algebra is the quotient ring/quotient space (°0 ~ ¸ b 0 (¹ with operations ² b 0³ b ² b 0³ ~ ² b ³ b 0 ² b 0³² b 0³ ~ b 0 ² b 0³ ~ b 0 where Á ( and - . These operations make (°0 an - -algebra. Homomorphisms Definition If ( and ) are - -algebras a map ¢ ( ¦ ) is an algebra homomorphism if it is a ring homomorphism as well as a linear transformation, that is, ² b Z ³ ~ b Z Á

²Z ³ ~ ²³²Z ³Á

and ²³ ~ ²³ for - .

~

456

Advanced Linear Algebra

The usual terms monomorphism, epimorphism, isomorphism, embedding, endomorphism and automorphism apply to algebras with the analogous meaning as for vector spaces and modules. Example 18.6 Let = be an -dimensional vector space over - . Fix an ordered basis 8 for = . Consider the map ¢ B²= ³ ¦ C ²- ³ defined by ²³ ~ ´µ8 where ´µ8 is the matrix representation of with respect to the ordered basis 8 . This map is a vector space isomorphism and since ´µ8 ~ ´ µ8 ´µ8 it is also an algebra isomorphism.

Another View of Algebras If ( is an algebra over - , then ( contains a copy of - . Specifically, we define a function ¢ - ¦ ( by ~ for all - , where ( is the multiplicative identity. The elements are in the center of (, since for any (, ²³ ~ ²³ ~ and ²³ ~ ²³ ~ Thus, ¢ - ¦ A²(³. To see that is a ring homomorphism, we have ²- ³ ~ - h ~ ² b ³ ~ ² b ³ ~ b ~ ²³ b ² ³ ² ³ ~ ² ³ ~ ² ³ ~ ² ³ ~ ² h ² ³³ ~ ² h ³² ³ ~ ²³² ³ Moreover, if ~ and £ , then ~ c ²³ ~ - h ~ and so provided that £ in (, we have ~ . Thus, is an embedding. Theorem 18.1 1) If ( is an associative algebra over - and if £ in (, then the map ¢ - ¦ A²(³ defined by ~ is an embedding of the field - into the center A²(³ of the ring ( . Thus, can be embedded as a subring of A²(³.

An Introduction to Algebras

457

2) Conversely, if 9 is a ring with identity and if - A²9³ is a field, then 9 is an - -algebra with scalar multiplication defined by the product in 9 . One interesting consequence of this theorem is that a ring 9 whose center does not contain a field is not an algebra over any field - . This happens, for example, with the ring { .

The Regular Representation of an Algebra An algebra homomorphism ¢ ( ¦ B- ²= ³ is called a representation of the algebra ( in B- ²= ³. A representation is faithful if it is injective, that is, if is an embedding. In this case, ( is isomorphic to a subalgebra of B- ²= ³. Actually, the endomorphism algebras B- ²= ³ are the most general algebras possible, in the sense that any algebra ( has a faithful representation in some endomorphism algebra. Theorem 18.2 Any associative - -algebra ( is isomorphic to a subalgebra of the endomorphism algebra B- ²(³. In fact, if is the left multiplication map defined by % ~ % then the map ¢ ( ¦ B ²(³ is an algebra embedding, called the left regular representation of (. When dim²(³ ~ B, we can select an ordered basis 8 for ( and represent the elements of B- ²(³ by matrices. This gives an embedding of ( into the matrix algebra C ²- ³, called the left regular matrix representation of ( with respect to the ordered basis 8 . Example 18.7 Let . ~ ¸Á Á Ã Á c ¹ be a finite cyclic group. Let 8 ~ ²Á Á Ã Á c ³ be an ordered basis for the group algebra - ´.µ. The multiplication map that is multiplcation by is a shifting of 8 (with wraparound) and so the matrix representation of is the matrix whose columns are obtained from the identity matrix by shifting columns to the right (with wrap around). For example, v x x ´ µ8 ~ x x Å w

Å

Ä Ä Æ Æ Ä

These matrices are called circulant matrices.

Å

y { { { { Å z

458

Advanced Linear Algebra

Since the endomorphism algebras B- ²= ³ are of obvious importance, let us examine them a bit more closely. Theorem 18.3 Let = be a vector space over a field - . 1) The algebra B- ²= ³ has center A ~ ¸ - ¹ and so B- ²= ³ is central. 2) The set 0 of all elements of B- ²= ³ that have finite rank is an ideal of B- ²= ³ and is contained in all other ideals of B- ²= ³. 3) B- ²= ³ is simple if and only if = is finite-dimensional. Proof. We leave the proof of parts 1) and 3) as exercises. For part 2), we leave it to the reader to show that 0 is an ideal of B- ²= ³. Let 1 be a nonzero ideal of B- ²= ³. Let B- ²= ³ have rank . Then there is a basis 8 ~ 8 r 8 (a disjoint union) and a nonzero $ = for which 8 is a finite set, ²8 ³ ~ ¸¹ and ²³ ~ $ for all 8 . Thus, is a linear combination over - of endomorphisms defined by ²³ ~ $Á

²8 ± ¸¹³ ~ ¸¹

Hence, we need only show that 1 . If 1 is nonzero, then there is an 8 for which ~ " £ . If B- ²= ³ is defined by ~ Á

²8 ± ¸¹³ ~ ¸¹

and B- ²= ³ is defined by " ~ $Á

²8 ± ¸"¹³ ~ ¸¹

then ²³ ~ $Á

²8 ± ¸¹³ ~ ¸¹

and so ~ 1 .

Annihilators and Minimal Polynomials If ( is an - -algebra an (, then it may happen that satisfies a nonzero polynomial ²%³ - ´%µ. This always happens, in particular, if ( is finitedimensional, since in this case the powers Á Á Á Ã must be linearly dependent and so there is a nonzero polynomial in that is equal to . Definition Let ( be an - -algebra. An element ( is algebraic if there is a nonzero polynomial ²%³ - ´%µ for which ²³ ~ . If is algebraic, the

An Introduction to Algebras

459

monic polynomial ²%³ of smallest degree that is satisfied by is called the minimal polynomial of . If ( is algebraic over - , then the subalgebra generated by over - is - ´µ ~ ¸²³ ²%³ - ´%µÁ deg²³ deg² ³¹ and this is isomorphic to the quotient algebra - ´µ

- ´%µ º ²%³»

where º ²%³» is the ideal generated by the minimal polynomial of . We leave the details of this as an exercise. The minimal polynomial can be used to tell when an element is invertible. Theorem 18.4 1) The minimal polynomial ²%³ of ( generates the annihilator of , that is, the ideal ann²³ ~ ¸ ²%³ - ´%µ ²³ ~ ¹ of all polynomials that annihilate . 2) The element ( is invertible if and only if ²%³ has nonzero constant term. Proof. We prove only the second statement. If is invertible but ²%³ ~ %²%³ then ~ ²³ ~ ²³. Multiplying by c gives ²³ ~ , which contradicts the minimality of deg² ²%³³. Conversely, if ²%³ ~ b % b Ä b c %c b % where £ , then ~ b b Ä b c c b and so c ² b b Ä b c c b c ³ ~ and so c ~

c ² b b Ä b c c b c ³

460

Advanced Linear Algebra

Theorem 18.5 If ( is a finite-dimensional - -algebra, then every element of ( is algebraic. There are infinite-dimensional algebras in which all elements are algebraic. Proof. The first statement has been proved. To prove the second, let us consider the complex field d as a r-algebra. The set ( of algebraic elements of d is a field, known as the field of algebraic numbers. These are the complex numbers that are roots of some nonzero polynomial with rational (or integral) coefficients. To see this, if (, then the subalgebra r´µ is finite-dimensional. Also, r´µ is a field. To prove this, first note that since d is a field, the minimal polynomial of any nonzero ( is irreducible, for if ²%³ ~ ²%³²%³, then ~ ²³²³ and so one of ²³ and ²³ is , which implies that ²%³ ~ ²%³ or ²%³ ~ ²%³. Since ²%³ is irreducible, it has nonzero constant term and so the inverse of is a polynomial in , that is, c r´µ. Of course, r´µ is closed under addition and multiplication and so r´µ is a subfield of d. Thus, d is an algebra over r´µ. By similar reasoning, if (, then the minimal polynomial of over r´µ is irreducible and so c r´µ´µ. Since r´µ´µ ~ r´Á µ is the set of all polynomials in the “variables” and , it is closed under addition and multiplication as well. Hence, r´Á µ is a finitedimensional algebra over r´µ, as well as a subfield of d. Now, dimr ²r´Á µ³ ~ dimr´µ ²r´Á µ³ h dimr ²r´µ³ and so r´Á µ is finite-dimensional over r. Hence, the elements of r´Á µ are algebraic over r, that is, r´Á µ (. But r´Á µ contains c Á b Á c and and so ( is a field. We claim that ( is not finite-dimensional over r. This follows from the fact that for every prime , the polynomial % c is irreducible over r (by Eisenstein's criterion). Hence, if is a complex root of % c , then has minimal polynomial % c over r and so the dimension of r´µ over r is . Hence, ( cannot be finite-dimensional.

The Spectrum of an Element Let ( be an algebra. A nonzero element ( is a left zero divisor if ~ for some £ and a right zero divisor if ~ for some £ . In the exercises, we ask the reader to show that an algebraic element is a left zero divisor if and only if it is a right zero divisor. Theorem 18.6 Let ( be a algebra. An algebraic element ( is invertible if and only if it is not a zero divisor. Proof. If is invertible and ~ , then multiplying by c gives ~ . Conversely, suppose that is not invertible but ~ implies ~ . Then

An Introduction to Algebras

461

²%³ ~ %²%³ for some nonzero polynomial ²%³ and so ~ ²³, which implies that ²³ ~ , a contradiction to the minimality of ²%³. We have seen that the eigenvalues of a linear operator on a finite-dimensional vector space are the roots of the minimal polynomial of , or equivalently, the scalars for which c is not invertible. By analogy, we can define the eigenvalues of an element of an algebra (. Theorem 18.7 Let ( be an algebra and let ( be algebraic. An element - is a root of the minimal polynomial ²%³ if and only if c is not invertible in (. Proof. If c is not invertible, then c ²%³ ~ %²%³ and since ²% b ³ is satisfied by c , it follows that %²%³ ~ c ²%³ ²% b ³ Hence, ²% c ³ ²%³. Alternatively, if c is not invertible, then there is a nonzero ( such that ² c ³ ~ , that is, ~ . Hence, for any polynomial ²%³ we have ²³ ~ ²³ . Setting ²%³ ~ ²%³ gives ²³ ~ . Conversely, if ²³ ~ , then ²%³ ~ ²% c ³²%³ and so ~ ² c ³²³, which shows that c is a zero divisor and therefore not invertible. Definition Let A be an - -algebra and let ( be algebraic. The roots of the minimal polynomial of are called the eigenvalues of . The set of all eigenvalues of Spec²³ ~ ¸ - ²³ ~ ¹ is called the spectrum of . Note that ( is invertible if and only if ¤ Spec²³. Theorem 18.8 (The spectral mapping theorem) Let ( be an algebra over an algebraically closed field - . Let ( and let ²%³ - ´%µ. Then Spec²²³³ ~ ²Spec²³³ ~ ¸²³ Spec²³¹ Proof. We leave it as an exercise to show that ²Spec²³³ Spec²²³³. For the reverse inclusion, let Spec²²³³ and suppose that ²%³ c ~ ²% c ³ Ä²% c ³ Then

462

Advanced Linear Algebra

²³ c ~ ² c ³ Ä² c ³ and since the left-hand side is not invertible, neither is one of the factors c , whence Spec²³. But ² ³ c ~ and so ~ ² ³ ²Spec²³³. Hence, Spec²²³³ ²Spec²³³. Theorem 18.9 Let ( be an algebra over an algebraically closed field - . If Á (, then Spec²³ ~ Spec²³ Proof. If £ ¤ Spec²³, then c is invertible and a simple computation gives ² c ³´² c ³c c µ ~ and so c is invertible and ¤ Spec²³. If ¤ Spec²³, then is invertible. We leave it as an exercise to show that this implies that is also invertible and so ¤ Spec²³. Thus, Spec²³ Spec²³ and by symmetry, equality must hold.

Division Algebras Some important associative algebras ( have the property that all nonzero elements are invertible and yet ( is not a field since it is not commutative. Definition An associative algebra + over a field - is a division algebra if every nonzero element has a multiplicative inverse. Our goal in this section is to classify all finite-dimensional division algebras over the real field s, over any algebraically closed field - and over any finite field. The classification of finite-dimensional division algebras over the rational field r is quite complicated and we will not treat it here.

The Quaternions Perhaps the most famous noncommutative division algebra is the following. Define a real vector space i with basis 8 ~ ¸Á Á Á ¹ To make i into an - -algebra, define the product of basis vectors as follows: 1) 2) 3) 4)

% ~ % ~ % for all % 8 ~ ~ ~ c ~ Á ~ Á ~ ~ cÁ ~ cÁ ~ c

An Introduction to Algebras

463

Note that 3) can be stated as follows: The product of two consecutive elements Á Á is the next element (with wraparound). Also, 4) says that &% ~ c%& for %Á & ¸Á Á ¹. This product is extended to all of i by distributivity. We leave it to the reader to verify that i is a division algebra, called Hamilton's quaternions, after their discoverer William Rowan Hamilton ( 1805-1865). (Readers familiar with group theory will recognize the quaternion group 8 ~ ¸fÁ fÁ fÁ f¹.) The quaternions have applications in geometry, computer science and physics.

Finite-Dimensional Division Algebras over an Algebraically Closed Field It happens that there are no interesting finite-dimensional division algebras over an algebraically closed field. Theorem 18.10 If + is a finite-dimensional division algebra over an algebraically closed field - then + ~ - . Proof. Let + have minimal polynomial ²%³. Since a division algebra has no zero divisors, ²%³ must be irreducible over - and so must be linear. Hence, ²%³ ~ % c and so ~ - .

Finite-Dimensional Division Algebras over a Finite Field The finite-dimensional division algebras over a finite field are also easily described: they are all commutative and so are finite fields. The proof, however, is a bit more challenging. To understand the proof, we need two facts: the class equation and some information about the complex roots of unity. So let us briefly describe what we need. The Class Equation Those who have studied group theory have no doubt encountered the famous class equation. Let . be a finite group. Each . can be thought of as a permutation of . defined by % ~ %c for all % .. The set of all conjugates %c of % is denoted by ¸%¹. and so ¸%¹. ~ ¸ % .¹ This set is also called a conjugacy class in . . Now, the following are equivalent:

464

Advanced Linear Algebra

% ~ % %c ~ % c c % c ~ % c *. ²%³ where *. ²%³ ~ ¸ . % ~ %¹ is the centralizer of %. But c *. ²%³ if and only if and are in the same coset of *. ²%³. Thus, there is a one-to-one correspondence between the conjugates of % and the cosets of *. ²%³. Hence, b¸%¹. b ~ ². ¢ *. ²%³³ Since the distinct conjugacy classes form a partition of . (because conjugacy is an equivalence relation), we have (.( ~ b¸%¹. b ~ ². ¢ *. ²%³³ %:

%:

where : is a set consisting of exactly one element from each conjugacy class ¸%¹. . Note that a conjugacy class ¸%¹. has size if and only if %c ~ % for all ., that is, % ~ % for all . and these are precisely the elements in the center A².³ of . . Hence, the previous equation can be written in the form (.( ~ (A².³( b ². ¢ *. ²%³³ %: Z

where : Z is a set consisting of exactly one element from each conjugacy class ¸%¹. of size greater than . This is the class equation for . . The Complex Roots of Unity If is a positive integer, then the complex th roots of unity are the complex solutions to the equation % c ~ The set < of complex th roots of unity is a cyclic group of order . To see this, note first that < is an abelian group since Á < implies that < and c < . Also, since % c has no multiple roots, < has order . Now, in any finite abelian group ., if is the maximum order of all elements of ., then ~ for all .. Thus, if no element of < has order , then and every . satisfies the equation % c ~ , which has fewer than solutions. This contradiction implies that some element of < must have order and so < is cyclic.

An Introduction to Algebras

465

The elements of < that generate < , that is, the elements of order are called the primitive th roots of unity. We denote the set of primitive th roots of unity by + . Hence, if + , then + ~ ¸ ²Á ³ ~ ¹ has size ²³, where is the Euler phi function. (The value ²³ is defined to be the number of positive integers less than or equal to and relatively prime to .) The th cyclotomic polynomial is defined by 8 ²%³ ~ ²% c $³ $+

Thus, deg²8 ²%³³ ~ ²³ Since every th root of unity is a primitive th root of unity for some and since every primitive th root of unity for is also an th root of unity, we deduce that < ~ +

where the union is a disjoint one. It follows that % c ~ 8 ²%³

Finally, we show that 8 ²%³ is monic and has integer coefficients by induction on . It is clear from the definition that 8 ²%³ is monic. Since 8 ²%³ ~ % c , the result is true for ~ . If is a prime, then all nonidentity th roots of unity are primitive and so 8 ²%³ ~

% c ~ %c b %c2 b Ä b % b %c

and the result holds for ~ . Assume the result holds for all proper divisors of . Then % c ~ 8 ²%³8 ²%³ ~ 8 ²%³9²%³

By the induction hypothesis, 9²%³ has integer coefficients and it follows that 8 ²%³ must also have integer coefficients. Wedderburn's Theorem Now we can prove Wedderburn's theorem.

466

Advanced Linear Algebra

Theorem 18.11 (Wedderburn's theorem) If + is a finite division algebra, then + is a field. Proof. We must show that + is commutative. Let . ~ +i be the multiplicative group of all nonzero elements of +. The class equation is (+i ( ~ (A²+i ³( b ²+i ¢ *² ³³ where the sum is taken over one representative from each conjugacy class of size greater than . If we assume for the purposes of contradiction that + is not commutative, that is, that A²+i ³ £ +i , then the sum on the far right is not an empty sum and so (* i ² ³( (+i ( for some +i . The sets A²+³ and *² ³ are subalgebras of + and, in fact, A²+³ is a commutative division algebra; that is, a field. Let (A²+³( ~ ' . Since A²+³ *² ³, we may view *² ³ and + as vector spaces over A²+³ and so (*² ³( ~ ' ² ³

and

(+( ~ '

for integers ² ³ . The class equation now gives ' c ~ ' c b ² ³

' c ' ² ³ c

and since ' ² ³ c ' c , it follows that ² ³ . If 8 ²%³ is the th cyclotomic polynomial, then 8 ²'³ divides ' c . But 8 ²'³ also divides each summand on the far right above, since its roots are not roots of ' ² ³ c . It follows that 8 ²'³ ' c . On the other hand, 8 ²'³ ~ ²' c ³ +

and since + implies that (' c ( ' c , we have a contradiction. Hence A²+i ³ ~ +i and + is commutative, that is, + is a field.

Finite-Dimensional Real Division Algebras We now consider the finite-dimensional division algebras over the real field s. In 1877, Frobenius proved that there are only three such division algebras. Theorem 18.12 (Frobenius, 1877) If + is a finite-dimensional division algbera over s, then + ~ sÁ

+~d

or + ~ i

Proof. Note first that the minimal polynomial ²%³ of any + is either linear, in which case s or irreducible quadratic ²%³ ~ % b % b with c . Completing the square gives

An Introduction to Algebras

467

~ ²³ ~ b b ~ 6 b 7 b ² c ³ Hence, any + has the form ~ 6 b 7 c ~ b ! where ! s and either ~ or . Hence, s but ¤ s. Thus, every element of + is the sum of an element of s and an element of the set +Z ~ ¸ + ¹ that is, as sets: + ~ s b +Z Also, s q +Z ~ ¸¹. To see that +Z is a subspace of +, let "Á # +Z . We wish to show that " b # +Z . If # ~ " for some s, then " b # ~ ² b ³" +Z . So assume that " and # are linearly independent. Then " and # are nonzero and so also nonreal. Now, " b # and " c # cannot both be real, since then " and # would be real. We have seen that "b#~b and "c#~ b where Á s, at least one of or is nonzero and Á . Then ²" b #³ b ²" c #³ ~ ² b ³ b ² b ³ and so " b # ~ b b b

b b

Collecting the real part on one side gives b ~ " b # c ² b b

b ³

Now, if we knew that Á and were linearly independent over s we could conclude that ~ ~ and so ²" b #³ ~

and ²" c #³ ~

which shows that " b # and " c # are in +Z . To see that ¸Á Á ¹ is linearly independent, it is equivalent to show that ¸"Á #Á ¹ is linearly independent. But if

468

Advanced Linear Algebra

# ~ " b for Á s, then # ~ " b " b and since " ¤ s, it follows that ~ and so ~ or ~ . But £ since # ¤ s and £ since ¸"Á #¹ are linearly independent. Thus, +Z is a subspace of + and + ~ s l +Z We now look at +Z , which is a real vector space. If +Z ~ ¸¹, then + ~ s and we are done, so assume otherwise. If +Z is nonzero, then ~ c where s. Hence, ~ c +Z satisfies ~ c. If +Z ~ s ~ ¸ s¹ then + ~ s l s ~ d and we are done. If not, then s is a proper subspace of +Z . In the quaternion field, there is an element for which b ~ . So we seek a +Z ± s with this property. To this end, define a bilinear form on +Z by º"Á #» ~ c²"# b #"³ Then it is easy to see that this form is a real inner product on +Z (positive definite, symmetric and bilinear). Hence, if s is a proper subspace of +Z , then + Z ~ s p : where p denotes the orthogonal direct sum. If " : is nonzero, then " ~ c for s and so if ~ "c , then ~ c

and

b ~

Now, s is a subspace of : and so + Z ~ s p s p ; Setting ~ , we have cºÁ » ~ b ~ b ~ and cºÁ » ~ b ~ b ~ and so ; and we can write +Z ~ s p s p s p <

An Introduction to Algebras

469

Now, if < £ ¸¹, then there is a " < for which " ~ c and " ~ c" " ~ c" " ~ c" The third equation is "²³ ~ c²³" and so "²³ ~ c²³" ~ " ~ c" whence " ~ , which is false. Hence, < ~ ¸¹ and + ~ s l s l s l s ~ i This completes the proof.

Exercises 1.

Prove that the subalgebra generated by a nonempty subset ? of an algebra ( is the subspace spanned by the products of finite subsets of elements of ?: º?»alg ~ º% Ä% % ?»

2. 3. 4. 5. 6. 7.

8.

Verify that the group algebra - ´.µ is indeed an associative algebra over - . Show that the kernel of an algebra homomorphism is an ideal. Let ( be a finite-dimensional algebra over - and let ) be a subalgebra. Show that if ) is invertible, then c ) . If ( is an algebra and : ( is nonempty, define the centralizer *( ²:³ of : to be the set of elements of ( that commute with all elements of : . Prove that *( ²:³ is a subalgebra of (. Show that { is not an algebra over any field. Let ( ~ - ´µ be the algebra generated over - by a single algebraic element . Show that ( is isomorphic to the quotient algebra - ´%µ°º ²%³», where º ²%³» is the ideal generated by ²%³ - ´%µ. What can you say about ²%³? What is the dimension of (? What happens if is not algebraic? Let . ~ ¸ ~ Á Ã Á ¹ be a finite group. For % - ´.µ of the form % ~ b Ä b

let ; ²%³ ~ b Ä b . Prove that ; ¢ - ´.µ ¦ - is an algebra homomorphism, where - is an algebra over itself. 9. Prove the first isomorphism theorem of algebras: A homomorphism ¢ ( ¦ ) of - -algebras induces an isomorphism ¢ (°ker²³ im²³ defined by ²ker²³³ ~ . 10. Prove that the quaternion field is an - -algebra and a field. Hint: For % ~ b b b £

470

Advanced Linear Algebra ( ~ ) consider % ~ c c c

11. Describe the left regular representation of the quaternions using the ordered basis 8 ~ ²Á Á Á ³. 12. Let : be the group of permutations (bijective functions) of the ordered set ? ~ ²% Á Ã Á % ³, under composition. Verify the following statements. Each : defines a linear isomorphism on the vector space = with basis ? over a field - . This defines an algebra homomorphism ¢ - ´: µ ¦ B- ²= ³ with the property that ²³ ~ . What does the matrix representation of a : look like? Is the representation faithful? 13. Show that the center of the algebra B- ²= ³ is A ~ ¸ - ¹ 14. Show that B- ²= ³ is simple if and only if dim²= ³ B. 15. Prove that for , the matrix algebras C ²- ³ are central and simple. 16. An element ( is left-invertible if there is a ( for which ~ , in which case is called a left inverse of . Similarly, ( is rightinvertible if there is a ( for which ~ , in which case is called a right inverse of . Left and right inverses are called one-sided inverses and an ordinary inverse is called a two-sided inverse. Let ( be algebraic over - . a) Prove that ~ for some £ if and only if ~ for some £ . Does necessarily equal ? b) Prove that if has a one-sided inverse , then is a two-sided inverse. Does this hold if is not algebraic? Hint: Consider the algebra ( ~ B- ²- ´%µ³. c) Let Á ( be algebraic. Show that is invertible if and only if and are invertible, in which case is also invertible.

Chapter 19

The Umbral Calculus

In this chapter, we give a brief introduction to an area called the umbral calculus. This is a linear-algebraic theory used to study certain types of polynomial functions that play an important role in applied mathematics. We give only a brief introduction to the subject, emphasizing the algebraic aspects rather than the applications. For more on the umbral calculus, may we suggest The Umbral Calculus, by Roman ´1984µ? One bit of notation: The lower factorial numbers are defined by ²³ ~ ² c ³Ä² c b ³

Formal Power Series We begin with a few remarks concerning formal power series. Let < denote the algebra of formal power series in the variable !, with complex coefficients. Thus, < is the set of all formal sums of the form B

²!³ ~ !

(19.1)

~

where d (the complex numbers). Addition and multiplication are purely formal: B

B

B

! b ! ~ ² b ³! ~

~

~

and B

B

B

4 ! 54 ! 5 ~ 4 c 5 ! ~

~

~

~

The order ² ³ of is the smallest exponent of ! that appears with a nonzero coefficient. The order of the zero series is defined to be b B. Note that a series

472

Advanced Linear Algebra

has a multiplicative inverse, denoted by c , if and only if ² ³ ~ . We leave it to the reader to show that ² ³ ~ ² ³ b ²³ and ² b ³ min¸² ³Á ²³¹ If is a sequence in < with ² ³ ¦ B as ¦ , then for any series B

²!³ ~ ! ~

we may substitute for ! to get the series B

²!³ ~ ²!³ ~

which is well-defined since the coefficient of each power of ! is a finite sum. In particular, if ² ³ , then ² ³ ¦ B and so the composition B

² k ³²!³ ~ ² ²!³³ ~ ²!³ ~

is well-defined. It is easy to see that ² k ³ ~ ²³² ³. If ² ³ ~ , then has a compositional inverse, denoted by and satisfying ² k ³²!³ ~ ² k ³²!³ ~ ! A series with ² ³ ~ is called a delta series. The sequence of powers of a delta series forms a pseudobasis for < , in the sense that for any < , there exists a unique sequence of constants for which B

²!³ ~ ²!³ ~

Finally, we note that the formal derivative of the series (19.1) is given by B

C! ²!³ ~ Z ²!³ ~ !c ~

The operator C! is a derivation, that is, C! ² ³ ~ C! ² ³ b C! ²³

The Umbral Calculus

473

The Umbral Algebra Let F ~ d´%µ denote the algebra of polynomials in a single variable % over the complex field. One of the starting points of the umbral calculus is the fact that any formal power series in < can play three different roles: as a formal power series, as a linear functional on F and as a linear operator on F . Let us first explore the connection between formal power series and linear functionals. Let F i denote the vector space of all linear functionals on F . Note that F i is the algebraic dual space of F , as defined in Chapter 2. It will be convenient to denote the action of 3 F i on ²%³ F by º3 ²%³» (This is the “bra-ket” notation of Paul Dirac.) The vector space operations on F i then take the form º3 b 4 ²%³» ~ º3 ²%³» b º4 ²%³» and º3 ²%³» ~ º3 ²%³»Á d Note also that since any linear functional on F is uniquely determined by its values on a basis for F Á the functional 3 F i is uniquely determined by the values º3 % » for . Now, any formal series in < can be written in the form B

²!³ ~ ~

! !

and we can use this to define a linear functional ²!³ by setting º ²!³ % » ~ for . In other words, the linear functional ²!³ is defined by B

º ²!³ % » ! ! ~

²!³ ~

where the expression ²!³ on the left is just a formal power series. Note in particular that º! % » ~ !Á where Á is the Kronecker delta function. This implies that º! ²%³» ~ ²³ ²³ and so ! is the functional “ th derivative at .” Also, ! is evaluation at .

474

Advanced Linear Algebra

As it happens, any linear functional 3 on F has the form ²!³. To see this, we simply note that if B

º3 % » ! ! ~

3 ²!³ ~ then

º3 ²!³ % » ~ º3 % » for all and so as linear functionals, 3 ~ 3 ²!³. Thus, we can define a map ¢ F i ¦ < by ²3³ ~ 3 ²!³. Theorem 19.1 The map ¢ F i ¦ < defined by ²3³ ~ 3 ²!³ is a vector space isomorphism from F i onto < . Proof. To see that is injective, note that 3 ²!³ ~ 4 ²!³ ¬ º3 % » ~ º4 % » for all ¬ 3 ~ 4 Moreover, the map is surjective, since for any < , the linear functional 3 ~ ²!³ has the property that ²3³ ~ 3 ²!³ ~ ²!³. Finally, B

º3 b 4 % » ! ! ~

²3 b 4 ³ ~

B

B º3 % » º4 % » ! b ! ! ! ~ ~

~

~ ²3³ b ²4 ³

From now on, we shall identify the vector space F i with the vector space < , using the isomorphism ¢ F i ¦ < . Thus, we think of linear functionals on F simply as formal power series. The advantage of this approach is that < is more than just a vector space—it is an algebra. Hence, we have automatically defined a multiplication of linear functionals, namely, the product of formal power series. The algebra < , when thought of as both the algebra of formal power series and the algebra of linear functionals on F , is called the umbral algebra. Let us consider an example. Example 19.1 For d, the evaluation functional F i is defined by º ²%³» ~ ²³

The Umbral Calculus

475

In particular, º % » ~ and so the formal power series representation for this functional is B

B º % » ! ~ ! ~ ! ! ! ~ ~

²!³ ~

which is the exponential series. If ! is evaluation at , then ! ! ~ ²b³! and so the product of evaluation at and evaluation at is evaluation at b . When we are thinking of a delta series < as a linear functional, we refer to it as a delta functional. Similarly, an invertible series < is referred to as an invertible functional. Here are some simple consequences of the development so far. Theorem 19.2 1) For any < , B

º ²!³ % » ! [ ~

²!³ ~ 2) For any F ,

º! ²%³» % [

²%³ ~ 3) For any Á < ,

º ²!³²!³ % » ~ 4 5 º ²!³ % »º²!³ %c » ~

4) ² ²!³³ deg ²%³ ¬ º ²!³ ²%³» ~ 5) If ² ³ ~ for all , then B

L ²!³c²%³M ~ º ²!³ ²%³» ~

where the sum on the right is a finite one. 6) If ² ³ ~ for all , then º ²!³ ²%³» ~ º ²!³ ²%³» for all ¬ ²%³ ~ ²%³ 7) If deg ²%³ ~ for all , then º ²!³ ²%³» ~ º²!³ ²%³» for all ¬ ²!³ ~ ²!³

476

Advanced Linear Algebra

Proof. We prove only part 3). Let B

²!³ ~ ~

B ! and ²!³ ~ ! ! ! ~

Then B

²!³²!³ ~ 4 ~

4 5 c 5! ! ~

and applying both sides of this ²as linear functionals) to % gives º ²!³²!³ % » ~ 4 5 c ~

The result now follows from the fact that part 1) implies ~ º ²!³ % » and c ~ º²!³ %c ». We can now present our first “umbral” result. Theorem 19.3 For any ²!³ < and ²%³ F , º ²!³ %²%³» ~ ºC! ²!³ ²%³» Proof. By linearity, we need only establish this for ²%³ ~ % . But if B

²!³ ~ ~

! !

then B

!c c% M ! ² c ³ ~

ºC! ²!³ % » ~ L B

cÁ ² c ³! ~

~

~ b ~ º ²!³ %b »

Let us consider a few examples of important linear functionals and their power series representations. Example 19.2 1) We have already encountered the evaluation functional ! , satisfying º! ²%³» ~ ²³

The Umbral Calculus

477

2) The forward difference functional is the delta functional ! c , satisfying º! c ²%³» ~ ²³ c ²³ 3) The Abel functional is the delta functional !e! , satisfying º!e! ²%³» ~ Z ²³ 4) The invertible functional ² c !³c satisfies º² c !³c ²%³» ~

B

²"³ c" "

as can be seen by setting ²%³ ~ % and expanding the expression ² c !³c . 5) To determine the linear functional satisfying

º ²!³ ²%³» ~ ²"³ "

we observe that B

B º ²!³ % » b !! c ! ~ ! ~ ! ² b ³[ ! ~ ~

²!³ ~

The inverse !°²! c ³ of this functional is associated with the Bernoulli polynomials, which play a very important role in mathematics and its applications. In fact, the numbers ) ~ L

!

! c% M c

are known as the Bernoulli numbers.

Formal Power Series as Linear Operators We now turn to the connection between formal power series and linear operators on F . Let us denote the th derivative operator on F by ! . Thus, ! ²%³ ~ ²³ ²%³ We can then extend this to formal series in !, B

²!³ ~ ~

! !

(19.2)

478

Advanced Linear Algebra

by defining the linear operator ²!³¢ F ¦ F by B

²!³²%³ ~ ~

´! ²%³µ ~ ²³ ²%³ ! !

the latter sum being a finite one. Note in particular that ²!³% ~ 4 5 %c ~

(19.3)

With this definition, we see that each formal power series < plays three roles in the umbral calculus, namely, as a formal power series, as a linear functional and as a linear operator. The two notations º ²!³ ²%³» and ²!³²%³ will make it clear whether we are thinking of as a functional or as an operator. It is important to note that ~ in < if and only if ~ as linear functionals, which holds if and only if ~ as linear operators. It is also worth noting that ´ ²!³²!³µ²%³ ~ ²!³´²!³²%³µ and so we may write ²!³²!³²%³ without ambiguity. In addition, ²!³²!³²%³ ~ ²!³ ²!³²%³ for all Á < and F . When we are thinking of a delta series as an operator, we call it a delta operator. The following theorem describes the key relationship between linear functionals and linear operators of the form ²!³. Theorem 19.4 If Á < , then º ²!³²!³ ²%³» ~ º ²!³ ²!³²%³» for all polynomials ²%³ F . Proof. If has the form (19.2), then by (19.3), º! (!³% » ~ L! c4 5 %c M ~ ~ º ²!³ % » ~

²19.4)

By linearity, this holds for % replaced by any polynomial ²%³. Hence, applying this to the product gives º ²!³²!³ ²%³» ~ º! ²!³²!³²%³» ~ º! ²!³´²!³²%³µ» ~ º ²!³ ²!³²%³»

Equation (19.4) shows that applying the linear functional (!³ is equivalent to applying the operator ²!³ and then following by evaluation at % ~ .

The Umbral Calculus

479

Here are the operator versions of the functionals in Example 19.2. Example 19.3 1) The operator ! satisfies B

! % ~ 4 5 %c ~ ²% b ³ ! ~ ~

! % ~ and so

! ²%³ ~ ²% b ³ for all F . Thus ! is a translation operator. 2) The forward difference operator is the delta operator ! c , where ²! c )²%³ ~ ²% b ³ c ²³ 3) The Abel operator is the delta operator !e! , where !e! ²%³ ~ Z ²% b ³ 4) The invertible operator ² c !³c satisfies ² c !³c ²%³ ~

B

²% b "³c" "

!

5) The operator ² c )°! is easily seen to satisfy %b ! c ²%³ ~ ²"³ " ! %

We have seen that all linear functionals on F have the form ²!³, for < . However, not all linear operators on F have this form. To see this, observe that deg ´ ²!³²%³µ deg ²%³ but the linear operator ¢ F ¦ F defined by ²²%³³ ~ %²%³ does not have this property. Let us characterize the linear operators of the form ²!³. First, we need a lemma. Lemma 19.5 If ; is a linear operator on 7 and ; ²!³ ~ ²!³; for some delta series ²!³, then deg²; ²%³³ deg²²%³³. Proof. For any , deg²; % ³ c ~ deg² ²!³; % ³ ~ deg²; ²!³% ³ and so

480

Advanced Linear Algebra

deg²; % ³ ~ deg²; ²!³% ³ b Since deg² ²!³% ³ ~ c we have the basis for an induction. When ~ we get deg²; ³ ~ . Assume that the result is true for c . Then deg²; % ³ ~ deg²; ²!³% ³ b c b ~

Theorem 19.6 The following are equivalent for a linear operator ; ¢ F ¦ F . 1) ; has the form ²!³, that is, there exists an < for which ; ~ ²!³, as linear operators. 2) ; commutes with the derivative operator, that is, ; ! ~ !; . 3) ; commutes with any delta operator ²!³, that is, ; ²!³ ~ ²!³; . 4) ; commutes with any translation operator, that is, ; ! ~ ! ; . Proof. It is clear that 1) implies 2). For the converse, let B

º! ; % » ! ! ~

²!³ ~ Then

º²!³ % » ~ º! ; % » Now, since ; commutes with !, we have º! ; % » ~ º! ! ; % » ~ º! ; ! % » ~ ²³ º! ; %c » ~ ²³ º! ²!³%c » ~ º! ²!³% » and since this holds for all and we get ; ~ ²!³. We leave the rest of the proof as an exercise.

Sheffer Sequences We can now define the principal object of study in the umbral calculus. When referring to a sequence ²%³ in F , we shall always assume that deg ²%³ ~ for all . Theorem 19.7 Let be a delta series, let be an invertible series and consider the geometric sequence Á Á Á Á Ã in < . Then there is a unique sequence conditions

²%³

in F satisfying the orthogonality

The Umbral Calculus

º²!³ ²!³

²%³»

~ [Á

481

(19.5)

for all Á . Proof. The uniqueness follows from Theorem 19.2. For the existence, if we set ²%³

~ Á % ~

and B

²!³ ²!³ ~ Á ! ~

where Á £ , then (19.5) is B

[Á ~ L Á ! c Á % M ~ B

~

~ Á Á º! c% » ~ ~

~ Á Á [ ~

Taking ~ we get Á ~

Á

For ~ c we have ~ cÁc Ác ² c ³[ b cÁ Á [ and using the fact that Á ~ °Á we can solve this for Ác . By successively taking ~ Á c Á c Á Ã we can solve the resulting equations for the coefficients Á of the sequence ²%³. Definition The sequence ²%³ in (19.5) is called the Sheffer sequence for the ordered pair ²²!³Á ²!³³. We shorten this by saying that ²%³ is Sheffer for ²²!³Á ²!³³. Two special types of Sheffer sequences deserve explicit mention. Definition The Sheffer sequence for a pair of the form ²Á ²!³³ is called the associated sequence for ²!³. The Sheffer sequence for a pair of the form ²²!³Á !³ is called the Appell sequence for ²!³.

482

Advanced Linear Algebra

Note that the sequence

²%³

is Sheffer for ²²!³Á ²!³³ if and only if

º²!³ ²!³

²%³»

~ [Á

which is equivalent to º ²!³ ²!³ ²%³» ~ [Á which, in turn, is equivalent to saying that the sequence ²%³ ~ ²!³ ²%³ is the associated sequence for ²!³. Theorem 19.8 The sequence ²%³ is Sheffer for ²²!³Á ²!³³ if and only if the sequence ²%³ ~ ²!³ ²%³ is the associated sequence for ²!³. Before considering examples, we wish to describe several characterizations of Sheffer sequences. First, we require a key result. Theorem 19.9 (The expansion theorems) Let 1) For any < ,

²%³

be Sheffer for ²²!³Á ²!³³.

B

º²!³ ²%³» ²!³ ²!³ ! ~

²!³ ~ 2) For any F ,

º²!³ ²!³ ²%³ !

²%³ ~

²%³

Proof. Part 1) follows from Theorem 19.2, since B

B º²!³ ²%³» º²!³ ²%³» ²!³ ²!³c ²%³M ~ !Á ! ! ~ ~

L

~ º²!³

²%³»

Part 2) follows in a similar way from Theorem 19.2. We can now begin our characterization of Sheffer sequences, starting with the generating function. The idea of a generating function is quite simple. If ²%³ is a sequence of polynomials, we may define a formal power series of the form B

²%³ ! ! ~

²!Á %³ ~

This is referred to as the (exponential) generating function for the sequence ²%³. (The term exponential refers to the presence of ! in this series. When this is not present, we have an ordinary generating function.³ Since the series is a formal one, knowing ²!Á %³ is equivalent (in theory, if not always in practice)

The Umbral Calculus

483

to knowing the polynomials ²%³. Moreover, a knowledge of the generating function of a sequence of polynomials can often lead to a deeper understanding of the sequence itself, that might not be otherwise easily accessible. For this reason, generating functions are studied quite extensively. For the proofs of the following characterizations, we refer the reader to Roman ´1984µ. Theorem 19.10 (Generating function) 1) The sequence ²%³ is the associated sequence for a delta series ²!³ if and only if B

²&³ ! ! ~

& ²!³ ~

where ²!³ is the compositional inverse of ²!³. 2) The sequence ²%³ is Sheffer for ²²!³Á ²!³³ if and only if B ²&³ & ²!³ ~ ! ! ² ²!³³ ~

The sum on the right is called the generating function of ²%³. Proof. Part 1) is a special case of part 2). For part 2), the expression above is equivalent to B &! ²&³ ~ ²!³ ²!³ ! ~

which is equivalent to B

&! ~ ~

²&³

!

²!³ ²!³

But if ²%³ is Sheffer for ² ²!³Á ²!³³, then this is just the expansion theorem for &! . Conversely, this expression implies that B &! ²&³ ~ º

²%³» ~ ~

and so º²!³ ²!³ ² Á ³.

²%³»

²&³

!

º²!³ ²!³

~ [Á , which says that

We can now give a representation for Sheffer sequences. Theorem 19.11 (Conjugate representation)

²%³»

²%³

is Sheffer for

484

Advanced Linear Algebra

1) A sequence ²%³ is the associated sequence for ²!³ if and only if

º ²!³ % »% [ ~

²%³ ~ 2) A sequence

²%³

is Sheffer for ²²!³Á ²!³³ if and only if

²%³

º² ²!³³c ²!³ % »% [ ~

~

Proof. We need only prove part 2). We know that ²²!³Á ²!³³ if and only if

²%³

is Sheffer for

B ²&³ & ²!³ ~ ! ! ² ²!³³ ~

But this is equivalent to N

B ²&³ & ²!³ % O ~ L ! c% M ~ ! ² ²!³³ ~

²&³

Expanding the exponential on the left gives B

B º² ²!³³c ²!³ % » ²&³ & ~ L ! c% M ~ [ ! ~ ~

²&³

Replacing & by % gives the result. Sheffer sequences can also be characterized by means of linear operators. Theorem 19.12 ²Operator characterization) 1) A sequence ²%³ is the associated sequence for ²!³ if and only if a) ²³ ~ Á b) ²!³ ²%³ ~ c ²%³ for 2) A sequence ²%³ is Sheffer for ²²!³Á ²!³³ for some invertible series ²!³ if and only if ²!³ ²%³ ~

c ²%³

for all . Proof. For part 1), if ²%³ is associated with ²!³, then ²³ ~ º! ²%³» ~ º ²!³ ²%³» ~ [Á and

The Umbral Calculus

485

º ²!³ ²!³ ²%³» ~ º ²!³b ²%³» ~ [Áb ~ ² c ³[cÁ ~ º ²!³ c ²%³» and since this holds for all we get 1b). Conversely, if 1a) and 1b) hold, then º ²!³ ²%³» ~ º! ²!³ ²%³» ~ ²³ c ²³ ~ ²³ cÁ ~ [Á and so ²%³ is the associated sequence for ²!³. As for part 2), if

²%³

is Sheffer for ²²!³Á ²!³³, then

º²!³ ²!³ ²!³ ²%³» ~ º²!³ ²!³b ²%³» ~ [Áb ~ ² c ³[cÁ ~ º²!³ ²!³ c ²%³» and so ²!³ ²%³ ~

c ²%³,

as desired. Conversely, suppose that

²!³ ²%³ ~

c ²%³

and let ²%³ be the associated sequence for ²!³. Let ; be the invertible linear operator on = defined by ;

²%³

~ ²%³

Then ; ²!³ ²%³ ~ ;

c ²%³

~ c ²%³ ~ ²!³ ²%³ ~ ²!³;

²%³

and so Lemma 19.5 implies that ; ~ ²!³ for some invertible series ²!³. Then º²!³ ²!³

and so

²%³

²%³»

~ º ²!³ ²!³ ²%³» ~ º! ²!³ ²%³» ~ ²³ c ²³ ~ ²³ cÁ ~ [Á

is Sheffer for ²²!³Á ²!³³.

We next give a formula for the action of a linear operator ²!³ on a Sheffer sequence.

486

Advanced Linear Algebra

Theorem 19.13 Let ²%³ be a Sheffer sequence for ²²!³Á ²!³³ and let ²%³ be associated with ²!³. Then for any ²!³ we have ²!³ ²%³ ~ 4 5º²!³ ~

²%³»c ²%³

Proof. By the expansion theorem B

º²!³ ²%³» ²!³ ²!³ ! ~

²!³ ~ we have B

º²!³ ²%³» ²!³ ²!³ ²%³ ! ~

²!³ ²%³ ~ B

º²!³ ²%³» ²³ c ²%³ ! ~

~

which is the desired formula. Theorem 19.14 1) (The binomial identity) A sequence ²%³ is the associated sequence for a delta series ²!³ if and only if it is of binomial type, that is, if and only if it satisfies the identity ²% b &³ ~ 4 5 ²&³c ²%³ ~

for all & d. 2) (The Sheffer identity) A sequence some invertible ²!³ if and only if ²%

²%³

is Sheffer for ²²!³Á ²!³³ for

b &³ ~ 4 5 ²&³ ~

c ²%³

for all & d, where ²%³ is the associated sequence for ²!³. Proof. To prove part 1), if ²%³ is an associated sequence, then taking ²!³ ~ &! in Theorem 19.13 gives the binomial identity. Conversely, suppose that the sequence ²%³ is of binomial type. We will use the operator characterization to show that ²%³ is an associated sequence. Taking % ~ & ~ we have for ~ , ²³ ~ ²³ ²³ and so ²³ ~ . Also,

The Umbral Calculus

487

²³ ~ ²³ ²³ b ²³ ²³ ~ ²³ and so ²³ ~ . Assuming that ²³ ~ for ~ Á Ã Á c we have ²³ ~ ²³ ²³ b ²³ ²³ ~ ²³ and so ²³ ~ . Thus, ²³ ~ Á . Next, define a linear functional ²!³ by º ²!³ ²%³» ~ Á Since º ²!³ » ~ º ²!³ ²%³» ~ and º ²!³ ²%³» ~ £ we deduce that ²!³ is a delta series. Now, the binomial identity gives º ²!³ &! ²%³» ~ 4 5 ²&³º ²!³ c ²%³» ~ ~ 4 5 ²&³cÁ ~

~ c ²&³ and so º&! ²!³ ²%³» ~ º&! c ²%³» and since this holds for all &, we get ²!³ ²%³ ~ c ²%³. Thus, ²%³ is the associated sequence for ²!³. For part 2), if ²%³ is a Sheffer sequence, then taking ²!³ ~ &! in Theorem 19.13 gives the Sheffer identity. Conversely, suppose that the Sheffer identity holds, where ²%³ is the associated sequence for ²!³. It suffices to show that ²!³ ²%³ ~ ²%³ for some invertible ²!³. Define a linear operator ; by ;

²%³

~ ²%³

Then &! ;

²%³

~ &! ²%³ ~ ²% b &³

and by the Sheffer identity, ; &! ²%³ ~ 4 5 ²&³; ~

²%³ ~ 4 5 ²&³c ²%³ c ~

and the two are equal by part 1). Hence, ; commutes with &! and is therefore of the form ²!³, as desired.

488

Advanced Linear Algebra

Examples of Sheffer Sequences We can now give some examples of Sheffer sequences. While it is often a relatively straightforward matter to verify that a given sequence is Sheffer for a given pair ²²!³Á ²!³³, it is quite another matter to find the Sheffer sequence for a given pair. The umbral calculus provides two formulas for this purpose, one of which is direct, but requires the usually very difficult computation of the series ² ²!³°!³c . The other is a recurrence relation that expresses each ²%³ in terms of previous terms in the Sheffer sequence. Unfortunately, space does not permit us to discuss these formulas in detail. However, we will discuss the recurrence formula for associated sequences later in this chapter. Example 19.4 The sequence ²%³ ~ % is the associated sequence for the delta series ²!³ ~ !. The generating function for this sequence is B

&! ~ ~

& ! [

and the binomial identity is the well-known binomial formula ²% b &³ ~ 4 5 % &c ~

Example 19.5 The lower factorial polynomials ²%³ ~ %²% c ³Ä²% c b ³ form the associated sequence for the forward difference functional ²!³ ~ ! c discussed in Example 19.2. To see this, we simply compute, using Theorem 19.12. Since ²³ is defined to be , we have ²³ ~ Á . Also, ²! c ³²%³ ~ ²% b ³ c ²%³ ~ ´²% b ³%²% c ³Ä²% c b ³µ c ´%²% c ³Ä²% c b ³µ ~ %²% c ³Ä²% c b ³´²% b ³ c ²% c b ³µ ~ %²% c ³Ä²% c b ³ ~ ²%³c The generating function for the lower factorial polynomials is B

²&³ ! [ ~

& log²b!³ ~

The Umbral Calculus

489

which can be rewritten in the more familiar form B & ² b !³& ~ 4 5 ! ~

Of course, this is a formal identity, so there is no need to make any restrictions on !. The binomial identity in this case is ²% b &³ ~ 4 5²%³ ²&³c ~

which can also be written in the form 4

%b& % & 5 ~ 4 54 5 c ~

This is known as the Vandermonde convolution formula. Example 19.6 The Abel polynomials ( ²%Â ³ ~ %²% c ³c form the associated sequence for the Abel functional ²!³ ~ !e! also discussed in Example 19.2. We leave verification of this to the reader. The generating function for the Abel polynomials is B

&²& c ³c ! [ ~

& ²!³ ~

Taking the formal derivative of this with respect to & gives B

²& c ³²& c ³c ! [ ~

²!³& ²!³ ~

which, for & ~ , gives a formula for the compositional inverse of the series ²!³ ~ !! , B

²c³ c ! ²c³[ ~

²!³ ~

Example 19.7 The famous Hermite polynomials / ²%³ form the Appell sequence for the invertible functional

²!³ ~ ! °

490

Advanced Linear Algebra

We ask the reader to show that ²%³ is the Appell sequence for ²!³ if and only if ²%³ ~ ²!³c % . Using this fact, we get ²³ c / ²%³ ~ c! ° % ~ ²c ³ % [ The generating function for the Hermite polynomials is B

/ ²&³ ! [ ~

&!c! ° ~ and the Sheffer identity is

/ ²% b &³ ~ 4 5/ ²%³&c ~

We should remark that the Hermite polynomials, as defined in the literature, often differ from our definition by a multiplicative constant. ²³

Example 19.8 The well-known and important Laguerre polynomials 3 ²%³ of order form the Sheffer sequence for the pair ²!³ ~ ² c !³cc Á ²!³ ~

! !c

It is possible to show ²although we will not do so here³ that

[ b 4 5²c%³ [ c ~

3²³ ²%³ ~

The generating function of the Laguerre polynomials is ²³

B 3 ²%³ &!°²!c³ ~ ! b ² c !³ [ ~

As with the Hermite polynomials, some definitions of the Laguerre polynomials differ by a multiplicative constant. We presume that the few examples we have given here indicate that the umbral calculus applies to a significant range of important polynomial sequences. In Roman ´1984µ, we discuss approximately 30 different sequences of polynomials that are (or are closely related to) Sheffer sequences.

Umbral Operators and Umbral Shifts We have now established the basic framework of the umbral calculus. As we have seen, the umbral algebra plays three roles: as the algebra of formal power series in a single variable, as the algebra of all linear functionals on F and as the

The Umbral Calculus

491

algebra of all linear operators on F that commute with the derivative operator. Moreover, since < is an algebra, we can consider geometric sequences Á Á Á Á Ã in < , where ²³ ~ and ² ³ ~ . We have seen by example that the orthogonality conditions º²!³ ²!³

²%³»

~ [Á

define important families of polynomial sequences. While the machinery that we have developed so far does unify a number of topics from the classical study of polynomial sequences (for example, special cases of the expansion theorem include Taylor's expansion, the Euler– MacLaurin formula and Boole's summation formula), it does not provide much new insight into their study. Our plan now is to take a brief look at some of the deeper results in the umbral calculus, which center on the interplay between operators on F and their adjoints, which are operators on the umbral algebra < ~ F i. We begin by defining two important operators on F associated with each Sheffer sequence. Definition Let ²%³ be Sheffer for ²²!³Á ²!³³. The linear operator Á ¢ F ¦ F defined by Á ²% ³ ~

²%³

is called the Sheffer operator for the pair ²²!³Á ²!³³, or for the sequence ²%³. If ²%³ is the associated sequence for ²!³, the Sheffer operator ²% ³ ~ ²%³ is called the umbral operator for ²!³, or for ²%³. Definition Let ²%³ be Sheffer for ²²!³Á ²!³³. The linear operator Á ¢ F ¦ F defined by Á ´ ²%³µ ~

b ²%³

is called the Sheffer shift for the pair ²²!³Á ²!³³, or for the sequence ²%³ is the associated sequence for ²!³, the Sheffer operator ´ ²%³µ ~ b ²%³ is called the umbral shift for ²!³, or for ²%³.

²%³.

If

492

Advanced Linear Algebra

It is clear that each Sheffer sequence uniquely determines a Sheffer operator and vice versa. Hence, knowing the Sheffer operator of a sequence is equivalent to knowing the sequence.

Continuous Operators on the Umbral Algebra It is clearly desirable that a linear operator ; on the umbral algebra < pass under infinite sums, that is, that B

B

; 4 ²!³5 ~ ; ´ ²!³µ ~

(19.6)

~

whenever the sum on the left is defined, which is precisely when ² ²!³³ ¦ B as ¦ B. Not all operators on < have this property, which leads to the following definition. Definition A linear operator ; on the umbral algebra < is continuous if it satisfies ²19.6). The term continuous can be justified by defining a topology on < . However, since no additional topological concepts will be needed, we will not do so here. Note that in order for (19.6) to make sense, we must have ²; ´ ²!³µ³ ¦ B. It turns out that this condition is also sufficient. Theorem 19.15 A linear operator ; on < is continuous if and only if (19.7)

² ³ ¦ B ¬ ²; ² ³³ ¦ B

Proof. The necessity is clear. Suppose that (19.7) holds and that ² ³ ¦ B. For any , we have B

L; ²!³c% M ~ L; ²!³c% M b L; ²!³c% M (19.8) ~

~

Since 8 ²!³9 ¦ B

(19.7) implies that we may choose large enough that 4; ²!³5

and ²; ´ ²!³µ³ for

The Umbral Calculus

493

Hence, (19.8) gives B

L; ²!³c% M ~ L; ²!³c% M ~

~

~ L ; ´ ²!³µc% M ~ B

~ L ; ´ ²!³µc% M ~

which implies the desired result.

Operator Adjoints If ¢ F ¦ F is a linear operator on F , then its (operator) adjoint d is an operator on F i ~ < defined by d ´²!³µ ~ ²!³ k In the symbolism of the umbral calculus, this is º d ²!³ ²%³» ~ º²!³ ²%³» (We have reduced the number of parentheses used to aid clarity.³ Let us recall the basic properties of the adjoint from Chapter 3. Theorem 19.16 For Á B²F ³, 1) ² b ³d ~ d b d 2) ² ³d ~ d for any d 3) ²³d ~ d d 4) ² c ³d ~ ² d ³c for any invertible B²F ³

Thus, the map ¢ B²F ³ ¦ B²< ³ that sends ¢ F ¦ F to its adjoint d ¢ < ¦ < is a linear transformation from B²F ³ to B²< ³. Moreover, since d ~ implies that º²!³ ²%³» ~ for all ²!³ < and ²%³ F , which in turn implies that ~ , we deduce that is injective. The next theorem describes the range of . Theorem 19.17 A linear operator ; B²< ³ is the adjoint of a linear operator B B²F ³ if and only if ; is continuous. Proof. First, suppose that ; ~ d for some B²F ³ and let ² ²!³³ ¦ B. If , then for all we have º d ²!³ % » ~ º ²!³ % » and so it is only necessary to take large enough that ² ²!³³ deg ²% ³ for all , whence

494

Advanced Linear Algebra

º d ²!³ % » ~ for all and so ² d ²!³³ . Thus, ² d ²!³³ ¦ B and d is continuous. For the converse, assume that ; is continuous. If ; did have the form d , then º; ! % » ~ º d ! % » ~ º! % » and since % ~

º! % » % [

we are prompted to define by % ~

º; ! % » % [

This makes sense since ²; ! ³ ¦ B as ¦ B and so the sum on the right is a finite sum. Then º d ! % » ~ º! % » ~

º; ! % » º! % » ~ º; ! % » [

which implies that ; ! ~ d ! for all . Finally, since ; and d are both continuous, we have ; ~ d .

Umbral Operators and Automorphisms of the Umbral Algebra Figure 19.1 shows the map , which is an isomorphism from the vector space B²F ³ onto the space of all continuous linear operators on < . We are interested in determining the images under this isomorphism of the set of umbral operators and the set of umbral shifts, as pictured in Figure 19.1.

The Umbral Calculus

495

Figure 19.1 Let us begin with umbral operators. Suppose that is the umbral operator for the associated sequence ²%³, with delta series ²!³ < . Then ºd ²!³ % » ~ º ²!³ % » ~ º ²!³ ²%³» ~ [Á ~ º! %» for all and . Hence, d ²!³ ~ ! and the continuity of d implies that d ! ~ ²!³ More generally, for any ²!³ < , d ²!³ ~ ² ²!³³

(19.9)

In words, d is composition by ²!³. From (19.9), we deduce that d is a vector space isomorphism and that d ´²!³²!³µ ~ ² ²!³³² ²!³³ ~ d ²!³d²!³ Hence, d is an automorphism of the umbral algebra < . It is a pleasant fact that this characterizes umbral operators. The first step in the proof of this is the following, whose proof is left as an exercise. Theorem 19.18 If ; is an automorphism of the umbral algebra, then ; preserves order, that is, ²; ²!³³ ~ ² ²!³³. In particular, ; is continuous. Theorem 19.19 A linear operator on F is an umbral operator if and only if its adjoint is an automorphism of the umbral algebra < . Moreover, if is an umbral operator, then

496

Advanced Linear Algebra

d ²!³ ~ ² ²!³³ for all ²!³ < . In particular, d ²!³ ~ !. Proof. We have already shown that the adjoint of is an automorphism satisfying (19.9). For the converse, suppose that d is an automorphism of < . Since d is surjective, there is a unique series ²!³ for which d ²!³ ~ !. Moreover, Theorem 19.18 implies that ²!³ is a delta series. Thus, [Á ~ º! % » ~ ºd ²!³ % » ~ º ²!³ % » which shows that % is the associated sequence for ²!³ and hence that is an umbral operator. Theorem 19.19 allows us to fill in one of the boxes on the right side of Figure 19.1. Let us see how we might use Theorem 19.19 to advantage in the study of associated sequences. We have seen that the isomorphism ª d maps the set K of umbral operators on F onto the set aut²< ³ of automorphisms of < ~ F i . But aut²< ³ is a group under composition. So if ¢ % ¦ ²%³ and ¢ % ¦ ²%³ are umbral operators, then since ² k ³d ~ d k d is an automorphism of < , it follows that the composition k is an umbral operator. In fact, since ² k ³d ²²!³³ ~ d k d ²²!³³ ~ d ²!³ ~ ! we deduce that k ~ k . Also, since k ~ k ~ ! ~ we have c ~ . Thus, the set K of umbral operators is a group under composition with k ~ k and c ~ Let us see how this plays out with respect to associated sequences. If the

The Umbral Calculus

497

associated sequence for ²!³ is

²%³ ~ Á % ~

then ¢ % ¦ ²%³ and so k ~ k is the umbral operator for the associated sequence

² k ³% ~ ²%³ ~ Á % ~ Á ²%³ ~

~

This sequence, denoted by

² ²%³³ ~ Á ²%³

(19.10)

~

is called the umbral composition of ²%³ with ²%³. The umbral operator ~ c is the umbral operator for the associated sequence ²%³ ~ 'Á % where c % ~ ²%³ and so

% ~ Á ²%³ ~

Let us summarize. Theorem 19.20 1) The set K of umbral operators on F is a group under composition, with k ~ k

and

c ~

2) The set of associated sequences forms a group under umbral composition

² ²%³³ ~ Á ²%³ ~

In particular, the umbral composition ² ²%³³ is the associated sequence for the composition k , that is, k ¢ % ¦ ² ²%³³ The identity is the sequence % and the inverse of ²%³ is the associated sequence for the compositional inverse ²!³.

498

Advanced Linear Algebra

3) Let K and ²!³ < . Then as operators, ²!³ ~ c ²!³ 4) Let K and ²!³ < . Then ² ²!³³ ~ ²!³ Proof. We prove 3) as follows. For any ²!³ < and ²%³ 7 , º²!³ d ²!³²%³» ~ º´d ²!³µ²!³ ²%³» ~ ºd ´²!³²c ³d ²!³µ ²%³» ~ º²!³²c ³d ²!³ ²%³» ~ º²c ³d ²!³ ²!³²%³» ~ º²!³ c ²!³²%³» which gives the desired result. Part 4) follows immediately from part 3) since is composition by .

Sheffer Operators If

²%³

is Sheffer for ²Á ³, then the linear operator Á defined by Á ²% ³ ~

²%³

is called a Sheffer operator. Sheffer operators are closely related to umbral operators, since if ²%³ is associated with ²!³, then ²%³

~ c ²!³ ²%³ ~ c ²!³ %

and so Á ~ c ²!³ It follows that the Sheffer operators form a group with composition Á k Á ~ c ²!³ c ²!³ ~ c ²!³c ² ²!³³ ~ ´²!³² ²!³³µc k ~ h²k ³Ák and inverse c Á ~ c ² ³Á

From this, we deduce that the umbral composition of Sheffer sequences is a Sheffer sequence. In particular, if ²%³ is Sheffer for ²Á ³ and ! ²%³ ~ '!Á % is Sheffer for ²Á ³, then

The Umbral Calculus

499

Á k Á ²% ³ ~ !Á Á % ~

~ !Á ²%³ ~

~ ! ² ²%³³ is Sheffer for ² h ² k ³Á k ³.

Umbral Shifts and Derivations of the Umbral Algebra We have seen that an operator on F is an umbral operator if and only if its adjoint is an automorphism of < . Now suppose that B²F ³ is the umbral shift for the associated sequence ²%³, associated with the delta series ²!³ < . Then º d ²!³ ²%³» ~ º ²!³ ²%³» ~ º ²!³ b ²%³» ~ ² b 1)[bÁ ~ ² c ³[Ác ~ º ²!³c ²%³» and so d ²!³ ~ ²!³c

(19.11)

d ´ ²!³ ²!³ µ ~ d ´ ²!³ µ ²!³ b ²!³ d´ ²!³µ

(19.12)

This implies that

and further, by continuity, that d ´²!³²!³µ ~ ´ d ²!³µ²!³ b ²!³´ d²!³µ

(19.13)

Let us pause for a definition. Definition Let 7 be an algebra. A linear operator C on 7 is a derivation if C²³ ~ ²C³ b Cb for all Á 7. Thus, we have shown that the adjoint of an umbral shift is a derivation of the umbral algebra < . Moreover, the expansion theorem and (19.11) show that d is surjective. This characterizes umbral shifts. First we need a preliminary result on surjective derivations.

500

Advanced Linear Algebra

Theorem 19.21 Let C be a surjective derivation on the umbral algebra < . Then C c ~ for any constant < and ²C f²!³³ ~ ² ²!³³ c , if ² ²!³³ . In particular, C is continuous. Proof. We begin by noting that C ~ C ~ C b C ~ C and so C ~ C ~ for all constants < . Since C is surjective, there must exist an ²!³ < for which C²!³ ~ Writing ²!³ ~ b ! ²!³, we have ~ C´ b ! ²!³µ ~ ²C!³ ²!³ b !C ²!³ which implies that ²C!³ ~ . Finally, if ²²!³³ ~ , then ²!³ ~ ! ²!³, where ² ²!³³ ~ and so ´C²!³µ ~ ´C! ²!³µ ~ ´! C²!³ b !c ²!³C!µ ~ c

Theorem 19.22 A linear operator on F is an umbral shift if and only if its adjoint is a surjective derivation of the umbral algebra < . Moreover, if is an umbral shift, then d ~ C is derivation with respect to ²!³, that is, d ²!³ ~ ²!³c for all . In particular, d ²!³ ~ . Proof. We have already seen that d is derivation with respect to ²!³. For the converse, suppose that d is a surjective derivation. Theorem 19.21 implies that there is a delta functional ²!³ such that d ²!³ ~ . If ²%³ is the associated sequence for ²!³, then º ²!³ ²%³» ~ º d ²!³ ²%³» ~ º ²!³c d ²!³ ²%³» ~ º ²!³c ²%³» ~ ² b 1)[bÁ ~ º ²!³ b ²%³» Hence, ²%³ ~ b ²%³, that is, ~ is the umbral shift for ²%³. We have seen that the fact that the set of all automorphisms on < is a group under composition shows that the set of all associated sequences is a group under umbral composition. The set of all surjective derivations on < does not form a group. However, we do have the chain rule for derivations[

The Umbral Calculus

501

Theorem 19.23 (The chain rule) Let C and C be surjective derivations on < . Then C ~ ²C ²!³³C Proof. This follows from C ²!³ ~ ²!³c C ²!³ ~ ²C ²!³³C ²!³ and so continuity implies the result. The chain rule leads to the following umbral result. Theorem 19.24 If and are umbral shifts, then ~ k C ²!³ Proof. Taking adjoints in the chain rule gives ~ k ²C ²!³³d ~ k C ²!³

We leave it as an exercise to show that C ²!³ ~ ´C ²!³µc . Now, by taking ²!³ ~ ! in Theorem 19.24 and observing that ! % ~ %b and so ! is multiplication by %, we get ~ %C ! ~ %´C! ²!³µc ~ %´ Z ²!³µc Applying this to the associated sequence ²%³ for ²!³ gives the following important recurrence relation for ²%³. Theorem 19.25 (The recurrence formula) Let ²%³ be the associated sequence for ²!³. Then 1) b ²%³ ~ %´ Z ²!³µc ²%³ 2) b ²%³ ~ % ´ ²!³µZ % Proof. The first part is proved. As to the second, using Theorem 19.20 we have b ²%³ ~ %´ Z ²!³µc ²%³ ~ %´ Z ²!³µc %

~ % ´ Z ² ²!³³µc % ~ % ´ ²!³µZ % Example 19.9 The recurrence relation can be used to find the associated sequence for the forward difference functional ²!³ ~ ! c . Since Z ²!³ ~ ! , the recurrence relation is b ²%³ ~ %c! ²%³ ~ % ²% c 1)

502

Advanced Linear Algebra

Using the fact that ²%³ ~ , we have ²%³ ~ %Á ²%³ ~ %²% c ³Á 3 ²%³ ~ %²% c ³²% c ³ and so on, leading easily to the lower factorial polynomials ²%³ ~ %²% c ³Ä²% c b ³ ~ ²%³

Example 19.10 Consider the delta functional ²!³ ~ log² b !³ Since ²!³ ~ ! c is the forward difference functional, Theorem 19.20 implies that the associated sequence ²%³ for ²!³ is the inverse, under umbral composition, of the lower factorial polynomials. Thus, if we write

²%³ ~ :²Á ³% ~

then

% ~ :²Á ³²%³ ~

The coefficients :²Á ³ in this equation are known as the Stirling numbers of the second kind and have great combinatorial significance. In fact, :²Á ³ is the number of partitions of a set of size into blocks. The polynomials ²%³ are called the exponential polynomials. The recurrence relation for the exponential polynomials is b ²%³ ~ %² b !³ ²%³ ~ %² ²%³ b Z ²%³³ Equating coefficients of % on both sides of this gives the well-known formula for the Stirling numbers :² b Á ³ ~ :²Á c ³ b :²Á ³ Many other properties of the Stirling numbers can be derived by umbral means. Now we have the analog of part 3) of Theorem 19.20. Theorem 19.26 Let be an umbral shift. Then d ²!³ ~ ²!³ c ²!³

The Umbral Calculus

503

Proof. We have º ²!³ d ²!³ ²%³» ~ º´ d ²!³µ ²!³ ²%³» ~ º d ´²!³ ²!³µ c ²!³ d ²!³ ²%³» ~ º d ´²!³ ²!³µ ²%³» c º²!³ c ²!³ ²%³» ~ º ²!³ ²!³ ²%³» c º c ²!³ ²!³ ²%³» ~ º ²!³ ²!³ ²%³» c º d ²!³ ²!³²%³» ~ º ²!³ ²!³ ²%³» c º ²!³ ²!³²%³» from which the result follows. If ²!³ ~ !, then is multiplication by % and d is the derivative with respect to ! and so the previous result becomes Z ²!³ ~ ²!³% c %²!³ as operators on F . The right side of this is called the Pincherle derivative of the operator ²!³. (See [104].)

Sheffer Shifts Recall that the linear map Á ´ ²%³µ ~

b ²%³

where ²%³ is Sheffer for ²²!³Á ²!³³ is called a Sheffer shift. If ²%³ is associated with ²!³, then ²!³ ²%³ ~ ²%³ and so c ²!³b ²%³ ~ Á ´c ²!³ ²%³µ and so Á ~ c ²!³ ²!³ From Theorem 19.26, the recurrence formula and the chain rule, we have Á ~ c ²!³ ²!³ ~ c ²!³´²!³ c d ²!³µ ~ c c ²!³C ²!³ ~ c c ²!³C ²!³ ~ c c ²!³C !C! ²!³ ~ %´ Z ²!³µc c c ²!³´ Z ²!³µc Z ²!³ Z ²!³ ~ >% c ? ²!³ Z ²!³ We have proved the following.

504

Advanced Linear Algebra

Theorem 19.27 Let Á be a Sheffer shift. Then 1) Á ~ <% c 2)

b ²%³

Z ²!³ ²!³ = Z ²!³ Z ²!³ ²!³ = Z ²!³ ²%³

~ <% c

The Transfer Formulas We conclude with a pair of formulas for the computation of associated sequences. Theorem 19.28 (The transfer formulas) Let ²%³ be the associated sequence for ²!³. Then cc

1) ²%³ ~ Z ²!³4 ²!³ ! 5 c

2) ²%³ ~ %4 ²!³ ! 5

%

%c

Proof. First we show that 1) and 2) are equivalent. Write ²!³ ~ ²!³°!. Then Z ²!³²!³cc % ~ ´!²!³µZ ²!³cc % ~ ²!³c % b !Z ²!³²!³cc % ~ ²!³c % b Z ²!³²!³cc %c ~ ²!³c % b ´²!³c µZ %c ~ ²!³c % c ´²!³c % c %²!³c µ%c ~ %²!³c %c To prove 1), we verify the operation conditions for an associated sequence for the sequence ²%³ ~ Z ²!³²!³cc % . First, when the fourth equality above gives º! ²%³» ~ º! Z ²!³²!³cc % » ~ º! ²!³c % c ´²!³c µZ %c » ~ º²!³c % » c º´²!³c µZ %c » ~ º²!³c % » c º²!³c % » ~ If ~ , then º! ²%³» ~ , and so in general, we have º! ²%³» ~ Á as required. For the second required condition, ²!³ ²%³ ~ ²!³ Z ²!³²!³cc % ~ !²!³ Z ²!³²!³cc % ~ Z ²!³²!³cc %c ~ c ²%³ Thus, ²%³ is the associated sequence for ²!³.

The Umbral Calculus

505

A Final Remark Unfortunately, space does not permit a detailed discussion of examples of Sheffer sequences nor the application of the umbral calculus to various classical problems. In [105], one can find a discussion of the following polynomial sequences: The lower factorial polynomials and Stirling numbers The exponential polynomials and Dobinski's formula The Gould polynomials The central factorial polynomials The Abel polynomials The Mittag-Leffler polynomials The Bessel polynomials The Bell polynomials The Hermite polynomials The Bernoulli polynomials and the Euler–MacLaurin expansion The Euler polynomials The Laguerre polynomials The Bernoulli polynomials of the second kind The Poisson–Charlier polynomials The actuarial polynomials The Meixner polynomials of the first and second kinds The Pidduck polynomials The Narumi polynomials The Boole polynomials The Peters polynomials The squared Hermite polynomials The Stirling polynomials The Mahler polynomials The Mott polynomials and more. In [105], we also find a discussion of how the umbral calculus can be used to approach the following types of problems: The connection constants problem Duplication formulas The Lagrange inversion formula Cross sequences Steffensen sequences Operational formulas Inverse relations Sheffer sequence solutions to recurrence relations Binomial convolution

506

Advanced Linear Algebra

Finally, it is possible to generalize the classical umbral calculus that we have described in this chapter to provide a context for studying polynomial sequences such as those of the names Gegenbauer, Chebyshev and Jacobi. Also, there is a q-version of the umbral calculus that involves the q-binomial coefficients (also known as the Gaussian coefficients) ² c ³Ä² c ³ 4 5 ~ ² c ³Ä² c ³² c ³Ä² c c ³ in place of the binomial coefficients. There is also a logarithmic version of the umbral calculus, which studies the harmonic logarithms and sequences of logarithmic type. For more on these topics, please see [103], [106] and [107].

Exercises 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18.

Prove that ² ³ ~ ² ³ b ²³, for any Á < . Prove that ² b ³ min¸² ³Á ²³¹, for any Á < . Show that any delta series has a compositional inverse. Show that for any delta series , the sequence is a pseudobasis. Prove that C! is a derivation. Show that < is a delta functional if and only if º » ~ and º %» £ . Show that < is invertible if and only if º » £ . Show that º ²!³ ²%³» ~ º ²!³ ²%³» for any a d, < and F. Show that º!e! ²%³» ~ Z ²a³ for any polynomial ²%³ F . Show that ~ in < if and only if ~ as linear functionals, which holds if and only if ~ as linear operators. Prove that if ²%³ is Sheffer for ²²!³Á ²!³³, then ²!³ ²%³ ~ c ²%³. Hint: Apply the functionals ²!³ ²!³ to both sides. Verify that the Abel polynomials form the associated sequence for the Abel functional. Show that a sequence ²%³ is the Appell sequence for ²!³ if and only if c ²%³ ~ ²!³ % . If is a delta series, show that the adjoint d of the umbral operator is a vector space isomorphism of < . Prove that if ; is an automorphism of the umbral algebra, then ; preserves order, that is, ²; ²!³³ ~ ² ²!³³. In particular, ; is continuous. Show that an umbral operator maps associated sequences to associated sequences. Let ²%³ and ²%³ be associated sequences. Define a linear operator by ¢ ²%³ ¦ ²%³. Show that is an umbral operator. Prove that if C and C are surjective derivations on < , then C ²!³ ~ ´C ²!³µc .

References

General References [1] [2]

Jacobson, N., Basic Algebra I, second edition, W.H. Freeman, 1985. Snapper, E. and Troyer, R., Metric Affine Geometry, Dover Publications, 1971.

General Linear Algebra [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19]

Akivis, M., Goldberg, V., An Introduction To Linear Algebra and Tensors, Dover, 1977. Blyth, T., Robertson, E., Further Linear Algebra, Springer, 2002. Brualdi, R., Friedland, S., Klee, V., Combinatorial and GraphTheoretical Problems in Linear Algebra, Springer, 1993. Curtis, M., Place, P., Abstract Linear Algebra, Springer, 1990. Fuhrmann, P., A Polynomial Approach to Linear Algebra, Springer, 1996. Gel'fand, I. M., Lectures On Linear Algebra, Dover, 1989. Greub, W., Linear Algebra, Springer, 1995. Halmos, P. R., Linear Algebra Problem Book, Mathematical Association of America, 1995. Halmos, P. R., Finite-Dimensional Vector Spaces, Springer, 1974. Hamilton, A. G., Linear Algebra, Cambridge University Press, 1990. Jacobson, N., Lectures in Abstract Algebra II: Linear Algebra, Springer, 1953. Jänich, K., Linear Algebra, Springer, 1994. Kaplansky, I., Linear Algebra and Geometry: A Second Course, Dover, 2003. Kaye, R., Wilson, R., Linear Algebra, Oxford University Press, 1998. Kostrikin, A. and Manin, Y., Linear Algebra and Geometry, Gordon and Breach Science Publishers, 1997. Lax, P., Linear Algebra, John Wiley, 1996. Lewis, J. G., Proceedings of the 5th SIAM Conference On Applied Linear Algebra, SIAM, 1994.

508 [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30]

Advanced Linear Algebra Marcus, M., Minc, H., Introduction to Linear Algebra, Dover, 1988. Mirsky, L., An Introduction to Linear Algebra, Dover, 1990. Nef, W., Linear Algebra, Dover, 1988. Nering, E. D., Linear Algebra and Matrix Theory, John Wiley, 1976. Pettofrezzo, A., Matrices and Transformations, Dover, 1978. Porter, G., Hill, D., Interactive Linear Algebra: A Laboratory Course Using Mathcad, Springer, 1996. Schneider, H., Barker, G., Matrices and Linear Algebra, Dover, 1989. Schwartz, J., Introduction to Matrices and Vectors, Dover, 2001. Shapiro, H., A survey of canonical forms and invariants for unitary similarity, Linear Algebra and Its Applications 147:101–167 (1991). Shilov, G., Linear Algebra, Dover, 1977. Wilkinson, J., The Algebraic Eigenvalue Problem, Oxford University Press, 1988.

Matrix Theory [31] Antosik, P., Swartz, C., Matrix Methods in Analysis, Springer, 1985. [32] Bapat, R., Raghavan, T., Nonnegative Matrices and Applications, Cambridge University Press, 1997. [33] Barnett, S., Matrices, Oxford University Press, 1990. [34] Bellman, R., Introduction to Matrix Analysis, SIAM, 1997. [35] Berman, A., Plemmons, R., Non-negative Matrices in the Mathematical Sciences, SIAM, 1994. [36] Bhatia, R., Matrix Analysis, Springer, 1996. [37] Bowers, J., Matrices and Quadratic Forms, Oxford University Press, 2000. [38] Boyd, S., El Ghaoui, L., Feron, E.; Balakrishnan, V., Linear Matrix Inequalities in System and Control Theory, SIAM, 1994. [39] Chatelin, F., Eigenvalues of Matrices, John Wiley, 1993. [40] Ghaoui, L., Advances in Linear Matrix Inequality Methods in Control, SIAM, 1999. [41] Coleman, T., Van Loan, C., Handbook for Matrix Computations, SIAM, 1988. [42] Duff, I., Erisman, A., Reid, J., Direct Methods for Sparse Matrices, Oxford University Press, 1989. [43] Eves, H., Elementary Matrix Theory, Dover, 1966. [44] Franklin, J., Matrix Theory, Dover, 2000. [45] Gantmacher, F.R., Matrix Theory I, American Mathematical Society, 2000. [46] Gantmacher, F.R., Matrix Theory II, American Mathematical Society, 2000. [47] Gohberg, I., Lancaster, P., Rodman, L., Invariant Subspaces of Matrices with Applications, John Wiley, 1986. [48] Horn, R. and Johnson, C., Matrix Analysis, Cambridge University Press, 1985.

References

509

[49] Horn, R. and Johnson, C., Topics in Matrix Analysis, Cambridge University Press, 1991. [50] Jennings, A., McKeown, J. J., Matrix Computation, John Wiley, 1992. [51] Joshi, A. W., Matrices and Tensors in Physics, John Wiley, 1995. [52] Laub, A., Matrix Analysis for Scientists and Engineers, SIAM, 2004. [53] Lütkepohl, H., Handbook of Matrices, John Wiley, 1996. [54] Marcus, M., Minc, H., A Survey of Matrix Theory and Matrix Inequalities, Dover, 1964. [55] Meyer, C., Matrix Analysis and Applied Linear Algebra, SIAM, 2000. [56] Muir, T., A Treatise on the Theory of Determinants, Dover, 2003. [57] Perlis, S., Theory of Matrices, Dover, 1991. [58] Serre, D., Matrices: Theory and Applications, Springer, 2002. [59] Stewart, G., Matrix Algorithms, SIAM, 1998. [60] Stewart, G., Matrix Algorithms Volume II: Eigensystems, SIAM, 2001. [61] Watkins, D., Fundamentals of Matrix Computations, John Wiley, 1991.

Multilinear Algebra [62] Marcus, M., Finite Dimensional Multilinear Algebra, Part I, Marcel Dekker, 1971. [63] Marcus, M., Finite Dimensional Multilinear Algebra, Part II, Marcel Dekker, 1975. [64] Merris, R., Multilinear Algebra, Gordon & Breach, 1997. [65] Northcott, D. G., Multilinear Algebra, Cambridge University Press, 1984.

Applied and Numerical Linear Algebra [66] Anderson, E., LAPACK User's Guide, SIAM, 1995. [67] Axelsson, O., Iterative Solution Methods, Cambridge University Press, 1994. [68] Bai, Z., Templates for the Solution of Algebraic Eigenvalue Problems: A Practical Guide, SIAM, 2000. [69] Banchoff, T., Wermer, J., Linear Algebra Through Geometry, Springer, 1992. [70] Blackford, L., ScaLAPACK User's Guide, SIAM, 1997. [71] Ciarlet, P. G., Introduction to Numerical Linear Algebra and Optimization, Cambridge University Press, 1989. [72] Campbell, S., Meyer, C., Generalized Inverses of Linear Transformations, Dover, 1991. [73] Datta, B., Johnson, C., Kaashoek, M., Plemmons, R., Sontag, E., Linear Algebra in Signals, Systems and Control, SIAM, 1988. [74] Demmel, J., Applied Numerical Linear Algebra, SIAM, 1997. [75] Dongarra, J., Bunch, J. R., Moler, C. B., Stewart, G. W., Linpack Users' Guide, SIAM, 1979. [76] Dongarra, J., Numerical Linear Algebra for High-Performance Computers, SIAM, 1998.

510

Advanced Linear Algebra

[77] Dongarra, J., Templates for the Solution of Linear Systems: Building Blocks For Iterative Methods, SIAM, 1993. [78] Faddeeva, V. N., Computational Methods of Linear Algebra, Dover, [79] Frazier, M., An Introduction to Wavelets Through Linear Algebra, Springer, 1999. [80] George, A., Gilbert, J., Liu, J., Graph Theory and Sparse Matrix Computation, Springer, 1993. [81] Golub, G., Van Dooren, P., Numerical Linear Algebra, Digital Signal Processing and Parallel Algorithms, Springer, 1991. [82] Granville S., Computational Methods of Linear Algebra, 2nd edition, John Wiley, 2005. [83] Greenbaum, A., Iterative Methods for Solving Linear Systems, SIAM, 1997. [84] Gustafson, K., Rao, D., Numerical Range: The Field of Values of Linear Operators and Matrices, Springer, 1996. [85] Hackbusch, W., Iterative Solution of Large Sparse Systems of Equations, Springer, 1993. [86] Jacob, B., Linear Functions and Matrix Theory, Springer, 1995. [87] Kuijper, M., First-Order Representations of Linear Systems, Birkhäuser, 1994. [88] Meyer, C., Plemmons, R., Linear Algebra, Markov Chains, and Queueing Models, Springer, 1993. [89] Neumaier, A., Interval Methods for Systems of Equations, Cambridge University Press, 1991. [90] Nevanlinna, O., Convergence of Iterations for Linear Equations, Birkhäuser, 1993. [91] Olshevsky, V., Fast Algorithms for Structured Matrices: Theory and Applications, SIAM, 2003. [92] Plemmon, R.J., Gallivan, K.A., Sameh, A.H., Parallel Algorithms for Matrix Computations, SIAM, 1990. [93] Rao, K. N., Linear Algebra and Group Theory for Physicists, John Wiley, 1996. [94] Reichel, L., Ruttan, A., Varga, R., Numerical Linear Algebra, Walter de Gruyter, 1993. [95] Saad, Y., Iterative Methods for Sparse Linear Systems, SIAM, 2003. [96] Scharlau, W., Quadratic and Hermitian Forms, Springer, 1985. [97] Snapper, E., Troyer, R., Metric Affine Geometry, Dover, [98] Spedicato, E., Computer Algorithms for Solving Linear Algebraic Equations, Springer, 1991. [99] Trefethen, L., Bau, D., Numerical Linear Algebra, SIAM, 1997. [100] Van Dooren, P., Wyman, B., Linear Algebra for Control Theory, Springer, 1994. [101] Vorst, H., Iterative Krylov Methods for Large Linear Systems, Cambridge University Press, 2003. [102] Young, D., Iterative Solution of Large Linear Systems, Dover, 2003.

References

511

The Umbral Calculus [103] Loeb, D. and Rota, G.-C., Formal Power Series of Logarithmic Type, Advances in Mathematics, Vol. 75, No. 1, (May 1989) 1–118. [104] Pincherle, S. "Operatori lineari e coefficienti di fattoriali." Alti Accad. Naz. Lincei, Rend. Cl. Fis. Mat. Nat. (6) 18, 417–519, 1933. [105] Roman, S., The Umbral Calculus, Pure and Applied Mathematics vol. 111, Academic Press, 1984. [106] Roman, S., The logarithmic binomial formula, American Mathematical Monthly 99 (1992) 641–648. [107] Roman, S., The harmonic logarithms and the binomial formula, Journal of Combinatorial Theory, series A, 63 (1992) 143–163.

Index of Symbols

*´²%³µ: the companion matrix of ²%³ ²%³: characteristic polynomial of crk²(³: column rank of ( cs²(³: column space of ( diag²( Á Ã Á ( ³: a block diagonal matrix with ( 's on the block diagonal ElemDiv² ³: the multiset of elementary divisors InvFact² ³: the multiset of invariant factors of @ ² Á Á ³: Jordan block ²%³: minimal polynomial of null² ³: the nullity of : : canonical projection modulo : 9 : Riesz vector for = i rk² ³: the rank of rrk²(³: row rank of ( rs²(³: row space of ( (Á) : projection onto ( along ) ( : the multiplication by ( operator supp² ³: the support of a function = : the - -vector space/- ´%µ-module where ²%³# ~ ² ³# = d : the complexification of = : assignment, for example, " º:» means that " stands for º:» : subspace or submodule : proper subspace or proper submodule º:»: subspace/ideal spanned by : ºº:»»: submodule spanned by : Æ ¢ an embedding that is an isomorphism when all is finite-dimensional. : similarity of matrices or operators, associate in a ring. d : cartesian product p : orthogonal direct sum ` : external direct product ^ : external direct sum l : internal direct sum :% & means º%Á &» ~ º&Á %»

514

Index of Symbols

w : wedge product n : tensor product n : -fold tensor product d : -fold cartesion product ²Á ³ ~ : and are relatively prime : affine combination

Index

Abel functional, 477, 489 Abel operator, 479 Abel polynomials, 489 abelian, 17 absolutely convergent, 330 accumulation point, 306 adjoint, 227, 231 affine basis, 435 affine closed, 428 affine combination, 428 affine geometry, 427 affine group, 436 affine hull, 430 affine hyperplane, 416 affine map, 435 affine span, 430 affine subspace, 57 affine transformation, 435 affine, 424 affinely independent, 433 affinity, 435 algebra homomorphism, 455 algebra, 31, 451 algebraic, 100, 458 algebraic closure, 30 algebraic dual space, 94 algebraic multiplicity, 189 algebraic numbers, 460 algebraically closed, 30 algebraically reflexive, 101 algorithm, 217 almost upper triangular, 194 along, 73 alternate, 260, 262, 391 alternating, 260, 391 ancestor, 14 anisotropic, 265 annihilator, 102, 115, 459 antisymmetric, 259, 390, 395

antisymmetric tensor algebra, 398 antisymmetric tensor space, 395, 400 antisymmetry, 10 Apollonius identity, 223 Appell sequence, 481 approximation problem, 331 as measured by, 357 ascending chain condition, 26, 133 associate classes, 27 associated sequence, 481 associates, 26 automorphism, 60 barycentric coordinates, 435 base ring, 110 base, 427 basis, 47, 116 Bernoulli numbers, 477 Bernstein theorem, 13 Bessel's identity, 221 Bessel's inequality, 220, 337, 338, 345 best approximation, 219, 332 bijection, 6 bijective, 6 bilinear form, 259, 360 bilinear, 206, 360 binomial identity, 486 binomial type, 486 block diagonal matrix, 3 block matrix, 3 blocks, 7 bottom, 10 bounded, 321, 349 canonical form, 8 canonical injections, 359 canonical map, 100 canonical projection, 89 Cantor's theorem, 13

516

Index

cardinal number, 13 cardinality, 12, 13 cartesian product, 14 Cauchy sequence, 311 Cauchy–Schwarz inequality, 208, 303, 325 Cayley-Hamilton theorem, 170 center, 452 central, 452 centralizer, 464, 469 chain, 11 chain rule, 501 change of basis matrix, 65 change of basis operator, 65 change of coordinates operator, 65 characteristic, 30 characteristic equation, 186 characteristic polynomial, 170 characteristic value, 185 characteristic vector, 186 Cholsky decomposition, 255 circulant matrices, 457 class equation, 464 classification problem, 276 closed ball, 304 closed half-spaces, 417 closed interval, 143 closed, 304, 414 closure, 306 codimension, 93 coefficients, 36 column equivalent, 9 column rank, 52 column space, 52 common eigenvector, 202 commutative, 17, 19, 451 commutativity, 15, 35, 384 commuting family, 201 compact, 414 companion matrix, 173 complement, 42, 120 complemented, 120 complete, 40, 311 complete invariant, 8 complete system of invariants, 8 completion, 316 complex operator, 59

complex vector space, 36 complexification, 53, 54, 82 complexification map, 54 composition, 472 cone, 265, 414 congruence classes, 262 congruence relation, 88 congruent modulo, 21, 87 congruent, 9, 262 conjugacy class, 463 conjugate isomorphism, 222 conjugate linear, 206, 221 conjugate linearity, 206 conjugate representation, 483 conjugate space, 350 conjugate symmetry, 205 connected, 281 continuity, 340 continuous, 310, 492 continuous dual space, 350 continuum, 16 contraction, 389 contravariant tensors, 386 contravariant type, 386 converge, 339, 210, 305, 330 convex combination, 414 convex, 332, 414 convex hull, 415 coordinate map, 51 coordinate matrix, 52, 368 correspondence theorem, 90, 118 coset, 22, 87, 118 coset representative, 22, 87 countable, 13 countably infinite, 13 covariant tensors, 386 covariant type, 386 cycle, 391 cyclic basis, 166 cyclic decomposition, 149, 168 cyclic group generated by, 18 cyclic group of order, 18 cyclic submodule, 113 cyclotomic polynomial, 465 decomposable, 362

Index

degenerate, 266 degree, 5 deleted absolute row sum, 203 delta functional, 475 delta operator, 478 delta series, 472 dense, 308 derivation, 499 descendants, 13 determinant, 292, 405 diagonal, 4 diagonalizable, 196 diagonally dominant, 203 diameter, 321 dimension, 50, 427 direct product, 41, 408 direct sum, 41, 73, 119 direct summand, 42, 120 discrete metric, 302 discriminant, 263 distance, 209, 322 divides, 5, 26 division algebra, 462 division algorithm, 5 domain, 6 dot product, 206 double, 100 dual basis, 96 dual space, 59, 100 eigenspace, 186 eigenvalue, 185, 186, 461 eigenvector, 186 elementary divisor basis, 169 elementary divisor form, 176 elementary divisor version, 177 elementary divisors, 155, 167, 168 elementary divisors and dimensions, 168 elementary matrix, 3 elementary symmetric functions, 189 embedding, 59, 117 endomorphism, 59, 117 epimorphism, 59, 117 equivalence class, 7 equivalence relation, 7 equivalent, 9, 69

517

essentially unique, 45 Euclidean metric, 302 Euclidean space, 206 evaluation at, 96, 100 evaluation functional, 474, 476 even permutation, 391 even weight subspace, 38 exponential polynomials, 502 exponential, 482 extension by, 103 extension, 6, 273 exterior algebra, 398 exterior product, 393 exterior product space, 395, 400 external direct sum, 40, 41, 119 factored through, 355, 357 factorization, 217 faithful, 457 Farkas's lemma, 423 field of quotients, 24 field, 19, 29 finite support, 41 finite, 1, 12, 18 finite-dimensional, 50, 451 finitely generated, 113 first isomorphism theorem, 92, 118, 469 flat representative, 427 flat, 427 form, 299, 382 forward difference functional, 477 forward difference operator, 479 Fourier coefficient, 219 Fourier expansion, 219, 338, 345 free, 116 Frobenius norm, 450, 466 functional calculus, 248 functional, 94 Gaussian coefficients, 57, 506 generating function, 482, 483 geometric multiplicity, 189 Geršgorin region, 203 Geršgorin row disk, 203 Geršgorin row region, 203 graded algebra, 392

518

Index

Gram-Schmidt augmentation, 213 Gram-Schmidt orthogonalization process, 214 greatest common divisor, 5 greatest lower bound, 11 group algebra, 453 group, 17 Hamel basis, 218 Hamming distance function, 321 Hermite polynomials, 224, 489 Hermitian, 238 Hilbert basis theorem, 136 Hilbert basis, 218, 335 Hilbert dimension, 347 Hilbert space adjoint, 230 Hilbert space, 315, 327 Hölder's inequality, 303 homogeneous, 392 homomorphism, 59, 117 Householder transformation, 244 hyperbolic basis, 273 hyperbolic extension, 274 hyperbolic pair, 272 hyperbolic plane, 272 hyperbolic space, 272 hyperplane, 416, 427 ideal generated by, 21, 455 ideal, 20, 455 idempotent, 74, 125 identity, 17 image, 6, 61 imaginary part, 54 indecomposable, 158 index of nilpotence, 200 induced, 305 inertia, 288 infinite, 13 infinite-dimensional, 50 injection, 6, 117 inner product, 205, 260 inner product space, 205, 260 integral domain, 23 invariant, 8, 73, 83, 165 invariant factor basis, 179 invariant factor decomposition, 157

invariant factor form, 178 invariant factor version, 179 invariant factors, 157, 167, 168 invariant factor decomposition theorem, 157 invariant ideals, 157 invariant under, 73 inverses, 17 invertible functional, 475 involution, 199 irreducible, 5, 26, 83 isometric isomorphism, 211, 326 isometric, 271, 316 isometrically isomorphic, 211, 326 isometry, 211, 271, 315, 326 isomorphic, 59, 62, 117 isotropic, 265 join, 40 Jordan basis, 191 Jordan block, 191 Jordan canonical form, 191 kernel, 61 Kronecker delta function, 96 Kronecker product, 408 Lagrange interpolation formula, 248 Laguerre polynomials, 490 largest, 10 lattice, 39, 40 leading coefficient, 5 leading entry, 3 least, 10 least squares solution, 448 least upper bound, 11 left inverse, 122, 470 left regular matrix representation, 457 left regular representation, 457 left singular vectors, 445 left zero divisor, 460 left-invertible, 470 Legendre polynomials, 215 length, 208 limit, 306 limit point, 306 line, 427, 429

Index

linear code, 38 linear combination, 36, 112 linear function, 382 linear functional, 59, 94 linear hyperplane, 416 linear least squares, 448 linear operator, 59 linear transformation, 59 linearity, 340 linearity in the first coordinate, 205 linearly dependent, 45, 114 linearly independent, 45, 114 linearly ordered set, 11 lower bound, 11 lower factorial numbers, 471 lower factorial polynomials, 488 lower triangular, 4 main diagonal, 2 matrix, 64 matrix of, 66 matrix of the form, 261 maximal element, 10 maximal ideal, 23 maximal orthonormal set, 218 maximum, 10 measuring family, 357 measuring functions, 357 mediating morphism map, 367 mediating morphism, 357, 362, 383 meet, 40 metric, 210, 301 metric space, 210, 301 metric vector space, 260 mimimum, 10 minimal element, 11 minimal polynomial, 165, 166, 459 Minkowski space, 260 Minkowski's inequality, 37, 303 mixed tensors, 386 modular law, 56 module, 109, 133, 167 modulo, 22, 87, 118 monic, 5 monomorphism, 59, 117 Moore-Penrose generalized inverse, 446

Moore-Penrose pseudoinverse, 446 MP inverse, 447 multilinear, 382 multilinear form, 382 multiplicity, 1 multiset, 1 natural map, 100 natural projection, 89 natural topology, 80, 82 negative, 17 net definition, 339 nilpotent, 198, 200 Noetherian, 133 nondegenerate, 266 nonderogatory, 171 nonisotropic, 265 nonnegative orthant, 225, 411 nonnegative, 225, 411 nonsingular, 266 nonsingular completion, 273 nonsingular extension theorem, 274 nontrivial, 36 norm, 208, 209, 303, 349 normal equations, 449 normal, 234 normalizing, 213 normed linear space, 209, 224 null, 265 nullity, 61 odd permutation, 391 one-sided inverses, 122, 470 one-to-one, 6 onto, 6 open ball, 304 open half-spaces, 417 open neighborhood, 304 open rectangles, 79 open sets, 305 operator adjoint, 104 operator characterization, 484 order, 18, 101, 139, 471 order ideals, 115 ordered basis, 51 order-reversing, 102

519

520

Index

orthogonal complement, 212, 265 orthogonal direct sum, 212, 269 orthogonal geometry, 260 orthogonal group, 271 orthogonal resolution of the identity, 232 orthogonal set, 212 orthogonal similarity classes, 242 orthogonal spectral resolution, 237 orthogonal transformation, 271 orthogonal, 75, 212, 231, 238, 265 orthogonality conditions, 480 orthogonally diagonalizable, 233 orthogonally equivalent, 242 orthogonally similar, 242 orthonormal basis, 218 orthonormal set, 212 parallel, 427 parallelogram law, 208, 325 parity, 391 Parseval's identity, 221, 346 partial order, 10 partially ordered set, 10 partition, 7 permutation, 391 Pincherle derivative, 503 plane, 427 point, 427 polar decomposition, 253 polarization identities, 209 posets, 10 positive definite, 205, 250, 301 positive square root, 251 power of the continuum, 16 power set, 13 primary, 147 primary cyclic decomposition theorem, 153, 168 primary decomposition theorem, 147 primary decomposition, 147, 168 prime subfield, 97 prime, 26 primitive, 465 principal ideal domain, 24 principal ideal, 24 product, 15 projection modulo, 89

projection theorem, 220, 334 projection, 73 projective dimension, 438 projective geometry, 438 projective line, 438 projective plane, 438 projective point, 438 proper subspace, 37 properly divides, 27 pseudobasis, 472 pure in, 161 q-binomial coefficients, 506 quadratic form, 239, 264 quaternions, 463 quotient algebra, 455 quotient field, 24 quotient module, 118 quotient ring, 22 quotient space, 87, 89 radical, 266 range, 6 rank, 53, 61, 129, 369 rank plus nullity theorem, 63 rational canonical form, 176–179 real operator, 59 real part, 54 real vector space, 36 real version, 53 recurrence formula, 501 reduce, 169 reduced row echelon form, 3, 4 reflection, 244, 292 reflexivity, 7, 10 relatively prime, 5, 27 representation, 457 resolution of the identity, 76 restriction, 6 retract, 122 retraction map, 122 Riesz map, 222 Riesz representation theorem, 222, 268, 351 Riesz vector, 222 right inverse, 122, 470 right singular vectors, 445

Index

right zero divisor, 460 right-invertible, 470 ring, 18 ring homomorphism, 19 ring with identity, 19 roots of unity, 464 rotation, 292 row equivalent, 4 row rank, 52 row space, 52 scalar multiplication, 31, 35, 451 scalars, 2, 35, 109 Schröder, 13 Schur's theorem, 192, 195 second isomorphism theorem, 93, 119 self-adjoint, 238 separable, 308 sesquilinear, 206 Sheffer for, 481 Sheffer identity, 486 Sheffer operator, 491, 498 Sheffer sequence, 481 Sheffer shift, 491 sign, 391 signature, 288 similar, 9, 70, 71 similarity classes, 70, 71 simple, 138, 455 simultaneously diagonalizable, 202 singular, 266 singular values, 444, 445 singular-value decomposition, 445 skew self-adjoint, 238 skew-Hermitian, 238 skew-symmetric, 2, 238, 259, 390 smallest, 10 span, 45, 112 spectral mapping theorem, 187, 461 spectral theorem for normal operators, 236, 237 spectral resolution, 197 spectrum, 186, 461 sphere, 304 split, 5 square summable, 207 square summable functions, 347

521

standard basis, 47, 62, 131 standard inner product, 206 standard topology, 79 standard vector, 47 Stirling numbers of the second kind, 502 strictly diagonally dominant, 203 strictly positive orthant, 411 strictly positive, 225, 411 strictly separated, 417 strongly positive orthant, 225, 411 strongly positive, 56, 225, 411 strongly separated, 417 structure constants, 453 structure theorem for normal matrices, 247 structure theorem for normal operators, 245 subalgebra, 454 subfield, 57 subgroup, 18 submatrix, 2 submodule, 111 subring, 19 subspace spanned, 44 subspace, 37, 260, 304 sup metric, 302 support, 6, 41 surjection, 6 surjective, 6 Sylvester's law of inertia, 287 symmetric, 2, 238, 259, 390, 395 symmetric group, 391 symmetric tensor algebra, 398 symmetric tensor space, 395, 400 symmetrization map, 402 symplectic basis, 273 symplectic geometry, 260 symplectic group, 271 symplectic transformation, 271 symplectic transvection, 280 tensor algebra, 390 tensor map, 362, 383 tensor product, 362, 383, 408 tensors of type, 386 tensors, 362 theorem of the alternative, 413 third isomorphism theorem, 94, 119

522

Index

top, 10 topological space, 305 topological vector space, 79 topology, 305 torsion element, 115 torsion module, 115 torsion-free, 115 total subset, 336 totally degenerate, 266 totally isotropic, 265 totally ordered set, 11 totally singular, 266 trace, 188 transfer formulas, 504 transitivity, 7, 10 translate, 427 translation operator, 479 translation, 436 transpose, 2 transposition, 391 triangle inequality, 208, 210, 301, 325 trivial, 36 two-affine closed, 428 two-sided inverse, 122, 470 umbral algebra, 474 umbral composition, 497 umbral operator, 491 umbral shift, 491 uncountable, 13 underlying set, 1 unipotent, 300 unique factorization domain, 28 unit vector, 208 unit, 26 unital algebras, 451 unitarily diagonalizable, 233 unitarily equivalent, 242 unitarily similar, 242 unitarily upper triangularizable, 196 unitary, 238 unitary metric, 302 unitary similarity classes, 242 unitary space, 206 universal, 289 universal for bilinearity, 362

universal for multilinearity, 382 universal pair, 357 universal property, 357 upper bound, 11 upper triangular, 4 upper triangularizable, 192 Vandermonde convolution formula, 489 Vector Space, 167 vector space, 35 vectors, 35 Wedderburn's Theorem, 465, 466 wedge product, 393 weight, 38 well ordering, 12 Well-ordering principle, 12 with respect to the bases, 66 Witt index, 296 Witt's cancellation theorem, 279, 294 Witt's extension theorem, 279, 295 zero divisor, 23 zero element, 17 zero subspace, 37 Zorn's lemma, 12

Graduate Texts in Mathematics (continued from page ii) 76 IITAKA. Algebraic Geometry. 77 HECKE. Lectures on the Theory of Algebraic Numbers. 78 BURRIS/SANKAPPANAVAR. A Course in Universal Algebra. 79 WALTERS. An Introduction to Ergodic Theory. 80 ROBINSON. A Course in the Theory of Groups. 2nd ed. 81 FORSTER. Lectures on Riemann Surfaces. 82 BOTT/TU. Differential Forms in Algebraic Topology. 83 WASHINGTON. Introduction to Cyclotomic Fields. 2nd ed. 84 IRELAND/ROSEN. A Classical Introduction to Modern Number Theory. 2nd ed. 85 EDWARDS. Fourier Series. Vol. II. 2nd ed. 86 VAN LINT. Introduction to Coding Theory. 2nd ed. 87 BROWN. Cohomology of Groups. 88 PIERCE. Associative Algebras. 89 LANG. Introduction to Algebraic and Abelian Functions. 2nd ed. 90 BRØNDSTED. An Introduction to Convex Polytopes. 91 BEARDON. On the Geometry of Discrete Groups. 92 DIESTEL. Sequences and Series in Banach Spaces. 93 DUBROVIN/FOMENKO/NOVIKOV. Modern Geometry—Methods and Applications. Part I. 2nd ed. 94 WARNER. Foundations of Differentiable Manifolds and Lie Groups. 95 SHIRYAEV. Probability. 2nd ed. 96 CONWAY. A Course in Functional Analysis. 2nd ed. 97 KOBLITZ. Introduction to Elliptic Curves and Modular Forms. 2nd ed. 98 BRÖCKER/TOM DIECK. Representations of Compact Lie Groups. 99 GROVE/BENSON. Finite Reflection Groups. 2nd ed. 100 BERG/CHRISTENSEN/RESSEL. Harmonic Analysis on Semigroups: Theory of Positive Definite and Related Functions. 101 EDWARDS. Galois Theory. 102 VARADARAJAN. Lie Groups, Lie Algebras and Their Representations. 103 LANG. Complex Analysis. 3rd ed. 104 DUBROVIN/FOMENKO/NOVIKOV. Modern Geometry—Methods and Applications. Part II. 105 LANG. S L 2 (R). 106 SILVERMAN. The Arithmetic of Elliptic Curves.

107 OLVER. Applications of Lie Groups to Differential Equations. 2nd ed. 108 RANGE. Holomorphic Functions and Integral Representations in Several Complex Variables. 109 LEHTO. Univalent Functions and Teichmüller Spaces. 110 LANG. Algebraic Number Theory. 111 HUSEMÖLLER. Elliptic Curves. 2nd ed. 112 LANG. Elliptic Functions. 113 KARATZAS/SHREVE. Brownian Motion and Stochastic Calculus. 2nd ed. 114 KOBLITZ. A Course in Number Theory and Cryptography. 2nd ed. 115 BERGER/GOSTIAUX. Differential Geometry: Manifolds, Curves, and Surfaces. 116 KELLEY/SRINIVASAN. Measure and Integral. Vol. I. 117 J.-P. SERRE. Algebraic Groups and Class Fields. 118 PEDERSEN. Analysis Now. 119 ROTMAN. An Introduction to Algebraic Topology. 120 ZIEMER. Weakly Differentiable Functions: Sobolev Spaces and Functions of Bounded Variation. 121 LANG. Cyclotomic Fields I and II. Combined 2nd ed. 122 REMMERT. Theory of Complex Functions. Readings in Mathematics 123 EBBINGHAUS/HERMES et al. Numbers. Readings in Mathematics 124 DUBROVIN/FOMENKO/NOVIKOV. Modern Geometry—Methods and Applications Part III. 125 BERENSTEIN/GAY. Complex Variables: An Introduction. 126 BOREL. Linear Algebraic Groups. 2nd ed. 127 MASSEY. A Basic Course in Algebraic Topology. 128 RAUCH. Partial Differential Equations. 129 FULTON/HARRIS. Representation Theory: A First Course. Readings in Mathematics 130 DODSON/POSTON. Tensor Geometry. 131 LAM. A First Course in Noncommutative Rings. 2nd ed. 132 BEARDON. Iteration of Rational Functions. 133 HARRIS. Algebraic Geometry: A First Course. 134 ROMAN. Coding and Information Theory. 135 ROMAN. Advanced Linear Algebra. 3rd ed. 136 ADKINS/WEINTRAUB. Algebra: An Approach via Module Theory. 137 AXLER/BOURDON/RAMEY. Harmonic Function Theory. 2nd ed.

138 COHEN. A Course in Computational Algebraic Number Theory. 139 BREDON. Topology and Geometry. 140 AUBIN. Optima and Equilibria. An Introduction to Nonlinear Analysis. 141 BECKER/WEISPFENNING/KREDEL. Gröbner Bases. A Computational Approach to Commutative Algebra. 142 LANG. Real and Functional Analysis. 3rd ed. 143 DOOB. Measure Theory. 144 DENNIS/FARB. Noncommutative Algebra. 145 VICK. Homology Theory. An Introduction to Algebraic Topology. 2nd ed. 146 BRIDGES. Computability: A Mathematical Sketchbook. 147 ROSENBERG. Algebraic K-Theory and Its Applications. 148 ROTMAN. An Introduction to the Theory of Groups. 4th ed. 149 RATCLIFFE. Foundations of Hyperbolic Manifolds. 2nd ed. 150 EISENBUD. Commutative Algebra with a View Toward Algebraic Geometry. 151 SILVERMAN. Advanced Topics in the Arithmetic of Elliptic Curves. 152 ZIEGLER. Lectures on Polytopes. 153 FULTON. Algebraic Topology: A First Course. 154 BROWN/PEARCY. An Introduction to Analysis. 155 KASSEL. Quantum Groups. 156 KECHRIS. Classical Descriptive Set Theory. 157 MALLIAVIN. Integration and Probability. 158 ROMAN. Field Theory. 159 CONWAY. Functions of One Complex Variable II. 160 LANG. Differential and Riemannian Manifolds. 161 BORWEIN/ERDÉLYI. Polynomials and Polynomial Inequalities. 162 ALPERIN/BELL. Groups and Representations. 163 DIXON/MORTIMER. Permutation Groups. 164 NATHANSON. Additive Number Theory: The Classical Bases. 165 NATHANSON. Additive Number Theory: Inverse Problems and the Geometry of Sumsets. 166 SHARPE. Differential Geometry: Cartan’s Generalization of Klein’s Erlangen Program. 167 MORANDI. Field and Galois Theory. 168 EWALD. Combinatorial Convexity and Algebraic Geometry. 169 BHATIA. Matrix Analysis. 170 BREDON. Sheaf Theory. 2nd ed. 171 PETERSEN. Riemannian Geometry. 2nd ed. 172 REMMERT. Classical Topics in Complex Function Theory. 173 DIESTEL. Graph Theory. 2nd ed. 174 BRIDGES. Foundations of Real and Abstract Analysis.

175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213

LICKORISH. An Introduction to Knot Theory. LEE. Riemannian Manifolds. NEWMAN. Analytic Number Theory. CLARKE/LEDYAEV/STERN/WOLENSKI. Nonsmooth Analysis and Control Theory. DOUGLAS. Banach Algebra Techniques in Operator Theory. 2nd ed. SRIVASTAVA. A Course on Borel Sets. KRESS. Numerical Analysis. WALTER. Ordinary Differential Equations. MEGGINSON. An Introduction to Banach Space Theory. BOLLOBAS. Modern Graph Theory. COX/LITTLE/O’SHEA. Using Algebraic Geometry. 2nd ed. RAMAKRISHNAN/VALENZA. Fourier Analysis on Number Fields. HARRIS/MORRISON. Moduli of Curves. GOLDBLATT. Lectures on the Hyperreals: An Introduction to Nonstandard Analysis. LAM. Lectures on Modules and Rings. ESMONDE/MURTY. Problems in Algebraic Number Theory. 2nd ed. LANG. Fundamentals of Differential Geometry. HIRSCH/LACOMBE. Elements of Functional Analysis. COHEN. Advanced Topics in Computational Number Theory. ENGEL/NAGEL. One-Parameter Semigroups for Linear Evolution Equations. NATHANSON. Elementary Methods in Number Theory. OSBORNE. Basic Homological Algebra. EISENBUD/HARRIS. The Geometry of Schemes. ROBERT. A Course in p-adic Analysis. HEDENMALM/KORENBLUM/ZHU. Theory of Bergman Spaces. BAO/CHERN/SHEN. An Introduction to Riemann–Finsler Geometry. HINDRY/SILVERMAN. Diophantine Geometry: An Introduction. LEE. Introduction to Topological Manifolds. SAGAN. The Symmetric Group: Representations, Combinatorial Algorithms, and Symmetric Functions. ESCOFIER. Galois Theory. FÉLIX/HALPERIN/THOMAS. Rational Homotopy Theory. 3rd ed. MURTY. Problems in Analytic Number Theory. Readings in Mathematics GODSIL/ROYLE. Algebraic Graph Theory. CHENEY. Analysis for Applied Mathematics. ARVESON. A Short Course on Spectral Theory. ROSEN. Number Theory in Function Fields. LANG. Algebra. Revised 3rd ed. MATOUŠEK. Lectures on Discrete Geometry. FRITZSCHE/GRAUERT. From Holomorphic Functions to Complex Manifolds.

214 JOST. Partial Differential Equations. 2nd ed. 215 GOLDSCHMIDT. Algebraic Functions and Projective Curves. 216 D. SERRE. Matrices: Theory and Applications. 217 MARKER. Model Theory: An Introduction. 218 LEE. Introduction to Smooth Manifolds. 219 MACLACHLAN/REID. The Arithmetic of Hyperbolic 3-Manifolds. 220 NESTRUEV. Smooth Manifolds and Observables. 221 GRÜNBAUM. Convex Polytopes. 2nd ed. 222 HALL. Lie Groups, Lie Algebras, and Representations: An Elementary Introduction. 223 VRETBLAD. Fourier Analysis and Its Applications. 224 WALSCHAP. Metric Structures in Differential Geometry. 225 BUMP. Lie Groups. 226 ZHU. Spaces of Holomorphic Functions in the Unit Ball. 227 MILLER/STURMFELS. Combinatorial Commutative Algebra. 228 DIAMOND/SHURMAN. A First Course in Modular Forms. 229 EISENBUD. The Geometry of Syzygies.

230 STROOCK. An Introduction to Markov Processes. 231 BJÖRNER/BRENTI. Combinatorics of Coxeter Groups. 232 EVEREST/WARD. An Introduction to Number Theory. 233 ALBIAC/KALTON. Topics in Banach Space Theory. 234 JORGENSON. Analysis and Probability. 235 SEPANSKI. Compact Lie Groups. 236 GARNETT. Bounded Analytic Functions. 237 MARTÍNEZ-AVENDAÑO/ROSENTHAL. An Introduction to Operators on the Hardy-Hilbert Space. 238 AIGNER, A Course in Enumeration. 239 COHEN, Number Theory, Vol. I. 240 COHEN, Number Theory, Vol. II. 241 SILVERMAN. The Arithmetic of Dynamical Systems. 242 GRILLET. Abstract Algebra. 2nd ed. 243 GEOGHEGAN. Topological Methods in Group Theory. 244 BONDY/MURTY. Graph Theory. 245 GILMAN/KRA/RODRIGUEZ. Complex Analysis. 246 KANIUTH. A Course in Commutative Banach Algebras.