SYMMETRIES AND LAPLACIANS introduction to Harmonic Analysis, Group Representations and Applications
NORTH-HOLLAND MATHEMATICS STUDIES 174 (Continuation of the Notas de Matematica)
Editor: Leopoldo NACHBIN Centro Brasileiro de Pesquisas Fisicas Rio de Janeiro, Brazil and University of Rochester New York, U S A .
NORTH-HOLLAND - AMSTERDAM
LONDON
NEW YORK
TOKYO
SYMMETRIES AND LAPLACIANS Introduction to Harmonic Analysis, Group Representations and Applications
David GURARIE Department of Mathematics and StaYjstics Case Western University Cleveland, OH, U.S.A.
1992
NORTH-HOLLAND - AMSTERDAM
LONDON
NEW YORK
TOKYO
ELSEVIER SCIENCE PUBLISHERS B.V. Sara Burgerhartstraat 25 P.O. Box 211, 1000 AE Amsterdam, The Netherlands
ISBN: 0 444 88612 5
0 1992 ELSEVIER SCIENCE PUBLISHERS B.V. All rights reserved. No part of this publication may be reproduced, stored in a retrieval system or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior written permission of the publisher, Elsevier Science Publishers B.V., Copyright & Permissions Department, RO.Box 521, 1000 A M Amsterdam, The Netherlands. Special regulations for readers in the U.S.A. - This publication has been registered with the Copyright Clearance Center Inc. (CCC), Salem, Massachusetts. Information can be obtained from the CCC about conditions under which photocopies of parts of this publication may be made in the U S A . All other copyright questions, including photocopying outside of the U.S.A., should be referred to the publisher, Elsevier Science Publishers B.V. No responsibility is assumed by the publisher for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions or ideas contained in the material herein. This book is printed on acid-free paper. Printed in The Netherlands
In the memory of my parents, whom I owe so much to Valentina
Eli and Mark
This Page Intentionally Left Blank
TABLE OF CONTENTS
Introduction .................................................................................................................. 1 Chapter 1. Basics of representation theory
51.1. Groups and group actions ............................................................................. $1.2. Regular and induced representations; Haar measure and convolution algebras .................................................................................... 51.3. Irreducibility and decomposition .................................................................. 51.4. Lie groups and algebras; the infinitesimal method ........................................
13 24 37 46
Chapter 2. Commutative Harmonic analysis
$2.1. Fourier transform: inversion and Plancherel formula .................................... 62 52.2* Fourier transform on function-spaces ........................................................... 71 52.3. Some applications of Fourier analysis ........................................................... 83 52.4. Laplacian and related differential equations .................................................. 92 52.5* The Radon transform ................................................................................. 120 Chapter 3 . Representations of compact and finite groups
53.1. The Peter-Weyl theory ............................................................................... 53.2. Induced representations and Frobenius reciprocity ..................................... 53.3* Semidirect products ...................................................................................
125 137 147
Chapter 4 . Lie groups SU(2) and Sq3) 54.1. Lie groups and SU(2) and Sq? and their Lie algebras ................................ 161 54.2. Irreducible representations of U(2)............................................................................. 164 &3* Matrix entries and characters of irreducible representations: Legendre and Jacobi polynomials .............................................................. 171 54.4. Representations of SO(3): angular momentum and spherical harmonics ..... 174 $4.5* Laplacian on the n-sphere .......................................................................... 184 Chapter 5 . Classical compact Lie groups and algebras
$5.1. Simple and semisimple Lie algebras; Weyl “unitary trick” .......................... 55.2. Cartan subalgebra, root system, Weyl group .............................................. $5.3. Highest weight representations ................................................................... 95.4* Tensors and Young tableaux ...................................................................... $5.5. Haar measure on compact semisimple Lie groups ....................................... $5.6. The Weyl character formulae ..................................................................... $5.7* Laplacians on symmetric spaces .................................................................
191 197 206 217 227 231 241
Chapter 6. The Heisenberg group and semidirect products
$6.1. Induced representations and the Mackey’s group extension theory .............. 257
...
Vlll
56.2. The Heisenberg group and the oscillator representation .............................. $6.3* The Kirillov orbit method .......................................................................... Chapter 7. Representations of SL,
274 290
.
305 57.1. Principal. complementary and discrete series .............................................. 57.2. Characters of irreducible representations .................................................... 313 57.3. The Plancherel formula on SL. (W) .............................................................. 317 67.4. Infinitesimal representations of SL. spherical functions and characters ....... 328 §7.5* Selberg trace formula ................................................................................. 334 $7.6* Laplacians on hyperbolic surfaces W/r....................................................... 348 57.7* SL. (C) and the Lorentz group .................................................................... 361 Chapter 8. Lie groups and hamiltonian mechanics 58.1. Minimal action principle; Euler-Lagrange equations; canonical formalism ... 369 58.2. Noether Theorem, conservation laws and Marsden-Weinstein reduction ..... 378 385 $8.3. Classical examples ...................................................................................... 58.4, Integrable systems related to classical Lie algebras..................................... 393 $8.5* The Kepler Problem and the Hydrogen atom ............................................. 408 Appendices: A: Spectral decomposition of selfadjoint operators .............................................. 423 B: Integral operators .......................................................................................... 427 C: A primer on Riemannian geometry: geodesics. connection. curvature ............. 430 Rderences ................................................................................................................. List of frequently used notations ................................................................................ Index .........................................................................................................................
439 447 449
Introduction Throughout this book word Group will be synonymous with transformations and symmetries. Groups are omnipresent in the mathematical and (possibly) physical universe. In Mathematics symmetries play several important functions. On the one hand they allow to reduce the “number of variables”, and often render problem soluble (like reducing the order, or separating variables in differential equations). On the other hand they allow to analyze and synthesize complex objects in terms of simple (elemental) blocks/constituents (eigenvalues of matrices and operators, expansion in Fourier modes, harmonic analysis in general). The prime example came from the very onset of the subject, the Galois theory of algebraic equations. The thrust of this early work was precisely to relate the symmetry (Galois) group of the equation to its solvability’ in radicals. Then Sophus Lie set up the task to develop the “Galois theory” for differential equations. He naturally came up with the concept of a continuous Lie group and its infinitesimal versions Lie algebra. Since its inception in the mid XIX-century the Lie theory rapidly grew and spread across the broad range of subjects, to occupy its present central position at the crossroads of Geometry, Analysis, Differential equations, Classical mechanics and Physics in general. Symmetries in physical systems2 are inextricably linked to conservation laws (via Noether’s Theorem), and the Galois chain repeats: “symmetries of hamiltonians”
+ “conserved integrals” + “solvability (via symmetry-reductions)”.
Our book sets more limited goals. The main objective is to introduce the reader to a wide range of concepts, ideas, results and techniques, that evolve around
. 1-
[symmetry-groups\, Irepresentations], and
We strived to stress the diversity
and versatility of the subject, but at the same time to unravel its common roots. More specifically, our main interest lies in geometrical objects and structures {X}, discrete or continuous, that possess sufficiently large symmetry-group G, like regular graphs (Platonic solids); lattices; symmetric Riemannian manifolds, and other. All such objects have a natural Laplacian A, a linear operator on functions over X, invariant under the group action. There are many problems associated with Laplacians on X. Typically one is interested in certain “continuous or discrete-time” evolutions, on
X, random walks, diffusion processes, wave-propagation. All these problems require to ‘Thence came the terminology of solvable groups (abelian ext,ensions of abelian groups), as the abelian ones were directly associated with solvability. ‘Symmetries acquire yet greater significance in the murky world of subatomic physics, where nothing could be observed, measured, sensed and compared to the everyday experiences of the macroworld. Here symmetries and conservation laws provide the only guiding light, and Lie groups enter in the very formulation of fundamental models of particles, fields and forces.
2
introduction
compute certain “functions of A”, L = f ( A ) , like powers { A ” } ; the semigroup, generated by A {e-“}; function {cos(tJ-6)}, etc. To explain and motivate the problems and the methods, based on symmetries, we shall start with a few simple prototypical examples. Fibonacci sequence is defined by a 2-term recurrence relation: Ixn+l = x, (or more general, xnS1 = a x ,
+b
+ .z,-~]
~ ~ - and ~ ) the , two initial values {zo;xl}.One is asked
to find all {z”}. To solve the problem, one first converts the 2-term numerical recurrence to a one-term recurrence for vectors: y, = Frobenius-type matrix L =
and we need to compute all iterates {L“}. The latter could be accomplished by diagonalizing matrix L , i.e. solving the eigenvalue problem: finding spectrum {A,$,}, and the matrix of eigenvectors U = {&;$,}.
The latter serves to diagonalize L,
L = U A U - I , where A=diag{A,;X,}. Once L is brought into the diagonal form, its iterates {Ln= UA”U-’} come right away, as well as more general “function of L”, f(L)= U f ( A ) U - ’ . So we immediately get an explicit solution
’.
Of course, no group-analysis was needed here, as the “eigenvalue problem” had an easy direct solution. However, similar “finite” problems could quickly become intractable, as we shall demonstrate in our next example.
Random w&
on graphs
I‘. The process consists of random jumps from any site
(vertex) of the graph to any one of its nearest neighbors with equal probabilities. Starting with the initial position at {xo}, or the initial probability distribution Po(,) (ZE
r),one is asked to find the probability distribution after the R time-steps, { P n ( x ) } .
Once again vector {P,(z)} is completely determined by the initial state Po and the transition-matrix L , which could be called the “Laplacian” of the graph, that sends any vectorffunction $(x) on
I‘ into L!b(x) = ?&b(Y),
sum over all m nearest vertices {y} of
Y
2.
Since the jumps are statistically independent
(the process has no memory of the past, when deciding to go from {x} to {y}), and
stationary in time (the “jump-rules” remains the same for all times n = 0; 1;...), we get a sequence of iterates, 3The classical Fibonacci numbers are given by
=c
I+&”
+
]-fin
, the generalized ones, z, = clX1” where coefficients, {cl;ez} depend on the initial data {zo;zl}. 2,
, ( ~ ) c2(-7_!
+ C,X,~;
3
Introduction
P, = Ln[P,],at time n. To compute powers
{L”}we
(2)
need once again to diagonalize
L, i.e. to solve the
eigenvalue problem. But this time the difficulty increases drastically with the size of
r,
if one tries a direct approach (imagine diagonalizing a 20 x 20 “dodecahedral” matrix!). There is not much to do in the general case, so we turn to a special, but important, class of regular graphs (like polygons, Platonic solids, a “Knight on a periodic
r are characterized by the abundance of symmetries. One can associate with any such r group G of vertex-transformations, that preserve all links (the incidence relations) between vertices. For regular r, group G takes any vertex x into any other4, so r becomes a homogeneow space o f G, r = H\G, where H denotes the stabilizer subgroup of some xo E r, H = {g:xog = xo}. For instance, the chessboard”, etc.). Regular graphs
regular m-gone is itself a finite abelian group Z , = {integers modulo m} (its symmetries are made of all finite rotations and reflections), whereas Platonic solids (cube, octahedron, dodecahedron, etc.) are homogeneous spaces H\G
of some well-known
symmetric or alternating groups (see 31.1 and 33.2). Given a homogeneous space X, group G acts on X by translations, g:x-+xg,and this action gives rise to the regular representation R = R X of G on the function-space
L = L(X ) over X , R g f ( 4= f ( 4 .
(3)
One easily verifies that map g+Rg, from G into operators on L, takes group multiplication into the product of operators, in other words defines a group-
L commutes with all operators {Rg},group G leaves L invariant, and we have to analyze the resulting “reduced
representation. Since Laplacian eigenspaces { E x } of
L
in
representations” of G on eigenspaces { E x } . Among all representations of the group one distinguishes the minimal ones, called irreducible, the ones that do not allow further reduction to smaller invariant subspaces (they play the role of “joint eigenvalues” for a g E G}).The typical problems in representation theory include: family of operators {Rg: (i) characterization all ineducible {n’s}for a given group G (analysis-problem); and
(ii) decomposition of an arbitrary (natural) representation R into the direct sum (or integral, to be explained below) of irreducibles (synthesis problem). The harmonic analysis on X suggests a possible approach to the spectral problem 41n fact, in many cases G takes any connected pair of vertices {qy} to any other such pair. Such graphs T’s could be called 2-point symmetric spaces, by analogy with their continuous counterparts.
4
Introduction
for operator L. Namely, we need
(I) to study the structure of the symmetry-group G; (11) to determine its elemental (irreducible) representations
{T};
(111) to interpret Laplacian L in terms of group-elements of G, to be able to assign operators { T ~ to} it for various irreducible { T } (IV) finally, to decompose the regular representation into the direct sum of irreducible components: RX = @ d .
i
The latter means that space L ( X ) is split into the direct sum of subspaces @ Ei, 3
invariant under all operators { R,:g E G}, irreducible relative to the G-action. Such decomposition, generalizes the notion of spectral/resolution (diagonalization) of a single operator. Steps (I-IV), when successfully accomplished, would lead to the desired spectral resolution of L. Namely, the reduced Laplacian L I E j r L becomes scalar5, N
So the “eigenwalue spectrum {X,(L)]” could be identified the “representation spectrum {d}of R”, the “eigenspaces” being “irreducible components” of R. Once L is diagonalized we immediately obtain all iterates {L”), and more general functions {f(L)}, needed to solve (3). { T = ~ X,(L);
for each irreducible
T}.
The details of the scheme could be found in 52.3 (Finite Fourier transform) for polygons, and in 53.3 for Laplacians on Platonic solids. After a brief excursion through the finite/discrete cases (graphs) we shall turn now to a completely different geometric setup: smooth Riemannian manifolds, Lie groups and their quotients, symmetric spaces X . All these objects will be introduced and studied at length in the due course (chapters 1,4,5,6,7).
Here for the sake of
introduction we consider 3 familiar examples: the Euclidian space R”, the n-sphere
S” c Rn+’, and the hyperbolic (Poincare-Lobachevski) plane‘ W2 = { z = z+iy:y 2 0). To motivate the continuous case we shall mention 3 typical problems in the integral geometry with some physical context. The setup for all 3 is basically the same. We are given a solid body D of constant (or variable) density p ( z ) . In the first case one considers all lines {y}, passing through
D,and computes integrals along
y: ?(y)
=
I,
pds. In the
’There is a general reason, why Laplacian L goes into the scalar operator nL under any irreducible A, known as Schur’s lemma. ‘the latter could be also interpreted, as a continuous limit of properly scaled i n space and time random walks on trees.
5
Introduction
second solid D (assumed to be convex, symmetric) is illuminated from all possible directions, and we measure the area of the shadows, cast by D. In the third case (somewhat simpler than the second) we take all cross-sectional areas of D by the family of planes, passing through its center. In all 3 cases one would like to recover either density p (case l), or the geometric shape of D (case 2 and 3) from such data. The first case is essentially the celebrated X-ray transform, used in tomography. The 2-nd and 3-rd give somewhat different versions of the Radon transform on the 2-sphere. Indeed, a (convex) body D can be represented as a graph of function f(z) on the 2-sphere, S2 = { I z I = 1). The cross-sectional areas of D are obtained by integrating :fz along all
great circles {y} in S2. So in all cases we are dealing with certain integral transforms of functions, either on the Euclidian space, or the sphere.
W have the natural Riemannian metric (§1.1), and possess G of distance-preserving transformations (isometries). The
All 3 spaces, R", S", large symmetry-group
Euclidian isometries consists of rigid motions: group En, generated by all translations and rotations in R"; the n-sphere S" has the orthogonal symmetry-group SO(n+l), made of all (n+l) x ( n + l ) - orthogonal matrices in Rn+' 3 S"; symmetries of
W coincide
with the ubiquitous unimodular group SL,(R) - all 2 x 2 real matrices of detg = 1, acting az+b by the fractional-linear transformations, g: z-+Cz+d. In fact, all 3 examples represent the of group G modulo the maximal compact
so called symmetric spaces, quotients K\G
subgroup K , that fixes a particular point in X , K = {g:z,g examples:
Ii' = S q n ) for Rn and S",
while the hyperbolic
' A
= z,}.
In the above
= SO(2).
The analog of the discrete-time random-walk on space X is played either by
spherical means:
(average value of function
f over the sphere of radius r , centered at
{z}), or by more
general integral kernel,
I
u(z)+K[u]= Ii'(z;y)u(y)dy. Transition-operator7 K takes any distribution p,(y) on X into pl(z) = li'[po], at a single time-step. As before we are interested in iterates
{I{"},
and once again have to
face the "diagonalization problem" for Ii'. In general, the problem is hardly tractable, so we turn to a special class of integral kernels, K ( z ; y ) ,that commute with the G-action, 7Strictly speaking, density K ( z ;y)dy represents transitional probabilities t o j u m p from point {y} to {z}, provided jK(z;y)dy = 1.
6
Introduction
(3) on X . Any G-invariant operator has kernel K , that depends only on the distance’ T = d(z;y), between z and y, 1-(z;y) = K [ T ) .Spherical means give one such example, the Radon and X-ray transforms also obey this condition, but the foremost of all Ginvariant operators is the natural Laplacian’ A (the Laplace-Beltrami operator) on X . Associated to A there are various evolution processes on X (continuous analogs of the “discrete-time random walks”), described by partial differential equations (pde). One of them is the standard heatdiffusion problem, ut
- AU= 0; ~ ( 0 =) f;
(4)
whose formal solution of (4) is given by the heat-semigroup of A, u = et”[fl.
Another important model is the wave equation: utt - Au = ...,whose solutions can also be represented by “functions of A”, {u = e *it-;
or cos t a ; sin tJ--Zij}.
Since A commutes with the G-action on X , all its functions { f ( A ) }do the same, including the heat and wave-propagators (another names for the “transition-matrices”, I( = etA; e i t f i ) . But any such I( has a radially symmetric kernel (I = K ( r ) . So a pdeproblem for I(, is often reduced to an ordinary differential equation (see chapters 2;4;5). Those in many cases, including R”, S”, W, could be solved explicitly in terms of the well-known special functions. Once again we are lead to analyze the G-action on X , and the resulting regular representation R,f(z) = f(zg),on suitable function-spaces over X.
So spectral decomposition of any radial operator K is reduced to the harmonic analysis of representation R. The latter involves among other a good understanding of the symmetry-group G itself, its ‘elemental (irreducible) representations” nature of the direct sum and/or integral decomposition,
R=
@ x j ; or 3
xSdp(s).
{T},
and the
(5)
To give some feeling of the issues involved, let us take the Euclidian space Rn, as an abelian group with translations only. Irreducible representations of R” are all 1‘This result is a direct consequence of the “double-transitive” action of G on rank-one symmetric spaces X,i.e. any pair {z;y} C X , is taken any other equidistant pair. ’In fact, any spherically symmetric kernel K on X (in either one of 3 examples) is given by a “function of the Laplacian”, K = f(A). This result, far from obvious, comes from the harmonic analysis of rank-one symmetric spaces (chapters 4-5). It has to do with the “multiplicity-free spectrum” of regular representation R, equivalently, with the fact that the commutator of R (all operators, that commute with it) forms a commutative algebra. “Functions of operators” {f(L)}will be introduced and discussed in chapter 2 and Appendix A via spectral resolution of operator L.
7
Introduction
dimensional (characters!), they consist of the family of exponentials {eit. €:[ E R”}. Decomposition (5), amounts to the Fourier-transform of f(z), expansion in the Fourierseries or integral. In chapter 2 we explore in detail various aspects of the commutative Fourier analysis and apply it to differential equations. The relevant analysis on the n-sphere, however, has a very different flavor. Here one has to start with “irreducible representations”
S q n t l ) , that would play the role of exponentials
{T}
of the orthogonal group
{ e i ” * < } on
R”. This, by itself a
challenging task, requires to develop a fair amount of the linear and multilinear (tensorial) algebra, as well as the relevant structure theory of Lie groups and algebras (chapters 4-5). Irreducibles
{T}
of S q n t l ) turned out to be finite dimensional, and the
decomposition of R has purely “discrete spectrum”lo,
R=
co @ rk.
k=O The corresponding irreducible subspaces { 36k} of L2(Sn), called spherical harmonics, have many remarkable features. Aside of being eigenspaces of the spherical
Laplacian, they proved to be closely connected to harmonic polynomials on Fin+’, and the classical Legendre and Jacobi functions. Our last example
W2 reveals even more striking difference. SL, is a
simple, not
compact group (in fact, the “smallest” of them). Unlike R” or S q n ) , its irreducible
unitary representations are infinite-dimensional, they make up certain continuous and discrete families {#:s E R} and {rrn:mE Z}. The decomposition problem for regular representations { R X } on homogeneous spaces, strongly depends on the “geometric nature” of quotient
X
continuous (direct integral), R = discrete subgroup of
+
= H\G. In some cases (Poincare half-plane W) it becomes purely
G) -
TSdp(s);in other cases (compact quotients r\G,
purely discrete (direct sum) R =
%0 r S j .It
r-
could also be a
combination of both, like the regular representation R on the entire group
G! The
relevant harmonic analysis becomes quite involved (see chapter 7). Its most spectacular application to the spectral theory of Laplacians appears in the context of quotients X = r\G, modulo discrete subgroup r c G. It turns out that “spectrum of such X” (either Laplacian A,, or representation Rx) is intimately connected with some fine “arithmetic properties” of via the celebrated Selberg-trace Theorem (a noncommutative version of the “Poisson summation” on R”). As an off-short we
r,
-Both results are general and hold for arbitrary compact (finite in size) groups (chapter 3). But specific examples, like S q n ) will carry a great deal more structure and information about {m} and
R.
8
Introduction
establish some interesting links between the “eigenvaluespectrum {A,}
of A,”, and the
“geometric length-spectrum of X” (length of all closed geodesics in X ) . We elaborate some aspects of spectral theory on compact quotients r \ G in 57.6-7.7. Our discussion of the “continuous cases” vs. “discrete cases”, although very different in technical terms, clearly demonstrates similarities in the basic procedures. Namely,
I) investigation of the group-structure of G, and its “elemental” irreducible representations { T } ;
11) study of G-invariant objects “Laplacians” on G, or on its quotients X = H\G; 111) decomposition of natural G-actions, regular representations R X , RG; its L on G and X ;
connection to Laplacians
IV) application of parts (1)-(111), particularly, the correspondence between “spectrum of L” and “spectrum of R” to various “junctions of L”,like iterates {L”}, semigroups { e-“; e - t f i } , and solutions of the associated differential equations. The development in the book largely will largely follow the general scheme (IIV), although not always step by step, as in the model examples. We may or may not start with a simple model problem. Often a significant effort has to be spent first to study the relevant groups, and their representation theory: problems of irreducibility and decomposition (cf. chapters 5-7). As we mentioned our main interest lies in the natural G-actions and representations, regular RG; RX,and their generalizations called
induced representations. The latter extend G-actions on scalar function { f(s)}, to vector-functions and sections of vector-bundles over X (so G-action combines translations in the X-space with “twisting” in fiber-spaces). The decomposition of the regular and induced representations makes up the content of the so called Plancherel Theorem. Although we follow steps (I-IV) throughout the book their relative “size” and “weight”, as well as the amount and depth of applications, varies from chapter to chapter. Thus chapter 2 is largely devoted to applications of the commutative Fourier transform, while chapters 5 or 7, deal mostly with the analysis of groups and represent at ions. An introductory chapter 1 brings in the principal players: basic examples of
9
Introduction
groups and geometric structures (51.1). Then we proceed to develop some fundamental concepts and results of the representation theory (§§1.2-3), and the Lie theory ($1.4). In chapters 2-7 we take on the main themes of the book with many facets of the general scheme (I-IV). Each chapter has its own LLfavorite” Laplacians and symmetry-groups: 0
standard Laplacian A on commutative groups Wn;Un and certain domains
D c Rn, in chapter 2 0
polyhedral Laplacians and the relevant finite groups in chapter 3
0
the spherical Laplacian A on S2 and S”, whose theory is based on Lie groups
SU(2);SO(3), in chapter 4 0Laplacians on more general compact Lie groups and symmetric spaces in chapter 5 0
the harmonic oscillator H = - A
+ I x I ’,and the Heisenberg group in chapter 6
0Laplacians on the Poincare plane
W, SL,,
and compact Riemann surfaces r \ H
in chapter 7. Chapter 8 stands somewhat aside from the mainstream, as it takes on the subject of “symmetries in nonlinear(!) hamiltonian systems”, and their role in integrability. But the very end of chapter 8 (58.5) brings us back to the main issues in grouprepresentations. We discuss the quantization problem of classical hamiltonians, particularly, the “quantized Kepler problem” - the hydrogen atom, one of the best studied models in quantum mechanics (see [LL]). The “Laplacian” pops in here in a quite interesting and unexpected form. A remarkable feature of the Kepler (hydrogen) in @, is that the “discrete hamiltonian, Schrodinger operator H = - A -1 Izl
(negative) part of H” is equivalent to the inverse Laplacian A - ’ on the 3-sphere. Along with the main themes (I-IV) a number of side issues and problems come to the discussion in different places. The reader will learn, for instance, why the geodesic flow is integrable on Sz (chapter 8), and ergodic on compact negative-curved Riemann surfaces X (chapter 7); what group-representations have to do with a topological problem of “counting linearly independent vector fields on spheres” (chapter 3); how Lie groups and representations explain some peculiar properties and rela.t,ionsof special functions (Legendre, Jacobi, Hermite, etc.). The book was designed as an introduction to harmonic analysis and group representations for graduate students in Mathematics and applications, or anybody
10
Introduction
(non-expert), interested in the subject, who would like to gain a broad perspective, but also learn some basic techniques and ideas. In the words of H. Weyl
“... it
is primarily
meant f o r the humble, who want to learn as new the things set forth therein, rather than f o r the proud and learned who are already familiar with the subject and merely look for
quick and ezact information...” The material of the book is based on the lectures and seminars, given by the author over the passed few years at UC Irvine, Caltech, and CWRU. Student comments and suggestions were helpful in preparing the manuscript. Our goal was to cover a wide range of topics, rather than to delve deeply into any particular one. The exposition is largely based on examples and applications, which either precede or follow the general theory. Some are important on their own, others serve to elucidate and motivate general concepts and statements. Of course, if the general approach seemed conceptually easy and directly leads to the main point, we do not hesitate to bring it forth (like the “Peter-Weyl theory” of chapter 3). But we never engage in “abstract” studies for their own sake. In order to keep the minimal prerequisites, and to shorten the background preparations, the book often appeals to intuition, example and analogy, rather than formal derivations. So the reader versed to some degree in the basic Riemannian geometry, functional analysis and operator theory, should be able to go through most of the topics without difficulty. The less prepared reader would be granted a paragraph (or two) of a footnote/appendix style explanations, to enable him to grasp quickly a new concept (or idea), and to follow the rest. Certain parts of the book (sections/paragraphs) are addressed to a better prepared reader, without detracting from the main themes.
As we wanted not to rely heavily on the standard (4-8 semester) staple of real analysis, algebra, topology, the book provides a number of ‘shortcuts’. So some basic concepts in algebra, geometry, topology are introduces “on fly”, as the need arises. We were somewhat more patient and systematic with the analysis of operators and differential equations: chapter 2 could serve as a brief introduction to PDE’s, mostly from the classical standpoint (cf. [CHI; [WW]). We also provided 3 appendices: on spectral decomposition of self-adjoint operators (A); on integral operators (B), and on basic Riemannian geometry (C). Finally, to cover a sizable material in a moderate-size volume required a departure from certain standards of mathematical exposition. We found it impossible
11
Introduction
(and undesirable) to try to maintain a uniform level of rigor and detail throughout the text. So some results are provided with fairly complete arguments, others are only outlined, relegated to problems, or just stated. The role of formal proofs is in general downplayed. The book contains sufficient material for a 1 or 2-semester course in the Harmonic analysis and Group representations. The instructor could make several choices, and follow different path in selecting topics. Aside of chapter 1, that provides a general core and background, all other parts are relatively independent. So the reader could start at practically any place in the book, going back and forth, as deemed necessary. The only exception are chapter 4 and 5, which should proceed in their natural order. In each chapter we marked with
* more
advanced topics, that could be
omitted in the first reading. In writing an introductory text on a fairly broad subject, one inevitably has to make certain choices, and put aside some important topics. Our selection and style reflected largely on the author’s personal experience and prejudices, rather than anything else. Among a few important topics, left outside the scope, let us mention the representation theory of infinite-dimensional Lie groups and algebras (Kac-Moody), which was actively pursued over the passed 20 years, and recently came to the focus in connection with the String theory (monographs [Kac] and [GSW] present the mathematical and the physical view on the subject). Another important topic is related to symmetries of differential equations, dynamical systems, integrable hamiltonians. Although we do touch upon integrability in chapter 8, our analysis is limited to finitedimensional systems. The exciting developments in the field of “infinite integrable hamiltonians” over the past 30 years were also left out (see [Per]; [Olv] for references). The book was composed on the IEXP-2 word processing system, with an additional help of the Microsoft Windows PBRUSH-graphics, responsible for the figures (the author takes entire responsibility for errors, misprints, omissions). After a somewhat bumpy initiation to the world of the modern information technology, the author had a highly rewarding experience working with both programs. Both became the most indispensable tools in the arduous enterprise of writing and organizing the manuscript.
He thought he saw a Garden-Door That opened with a key He looked again, and found it was A Double Rule of Three: ‘And all its mystery”, he said, “Is clear as day to m e !” He thought he saw an Argument That proved he was the Pope He look again, and found it was A Bar of Mottled Soap. “A fact so dread, ” he faintly said, “Eztinguishes all hope !” Lewis Carroll, “The Mad Gardener’s Song”
Chapter 1. Basics of representation theory. 51.1. Groups and group actions. We introduce basic examples of discrete and continuous transformation groups; classical matrix Lie groups; isornetries of symmetric spaces; rigid motions of the Euclidian, spherical and hyperbolic geometry, as well as symmetries of regular polyhedra.
1.1. Geometric transformations. Groups, discrete and continuous, typically arise as symmetries of “geometric” structures of different kinds, which includes both discrete objects (graphs, polyhedra), and the continuous, like manifolds, symmetric spaces. Important examples of discrete groups include:
i) permutations of a finite sets A = {l;...n}, G(A) = W,; ii) isomorphisms of graphs, lattices, regular polyhedra;
iii) matrix groups over finite fields. Some of them will be described at the end of the section (example 1.1). Throughout this book we shall be mostly interested in the continuous (Lie) groups. The latter typically arise as transformations (linear or nonlinear) of vector spaces,
or
manifolds
A, equipped
with
certain
geometric
structure.
Such
transformations q5:A+A preserve the structure (or transform it in a prescribed manner), in other words they represent symmetries of A. For instance, classical (matrix) Lie groups over vector spaces R”, c”, or quaternionic1 Q”, consists of linear transformations, that preserve certain bilinear/quadratic forms, like the general linear
group made of all non-singular n x n - matrices, Gf, = {A:detA# 0 } , its subgroup
Sl, = { A :detA = l}, called special linear (or unimodular) group, as well as their numerous subgroups (orthogonal, unitary, symplectic etc.). Lie groups also arise naturally as isometries of certain Riemannian (pseudoRiemannian) manifolds, or more general conformal transformations2. The foremost cases are symmetric spaces, Riemannian manifolds which possess large isometry groups. Here we shall briefly discuss 3 basic examples of symmetric spaces (flat, spherical and hyperbolic), and the related symmetry groups. a) Euclidian space
Rn, respectively Minkowski Mn,equipped with either positive-
+ + +
‘Quaternions Q = {( = a b i c j dk a, b, c,d € R} form a noncommutative field (division algebra), generated by 3 imaginary units:i2 = j 2 = k2 = -1; i j = - - j k k ; j k = -kj=i; t i = -ik = j . They can be also represented by complex 2-vectors: ( = z wk ( z , w E C), where k2 = -1, with the multiplication rule: kzu = ii~ t. Quat,ernions along with R and C are known to form the only 3 possible division algebras over reals (see J3.1 for further details).
+
§l,l.Groups and group actions.
14
definite product: z . y = Czjyj, or indefinite product3: (z I y) = zoyo-
n-1
zcjyj. The
isometries of the Euclidian space form a group E, of rigid motions of W", generated by
+ a}, and rotations (orthogonal matrices) {U:'UU = I } . So each the form 4(z) = Uz + b. The proof is outlined in problem 1. The tricky part is to show that any rigid motion q5 is affine, d(z) = A z + b, with some
all translations {a:z+z transformation
4 E En is of
matrix A. Then orthogonality of A follows fairly straightforward. Thus group becomes a subgroup of a larger affine group,
En
+
A f f , = {4(z) = AX b: A E G L ; b E W"}. The En-linear factors { A } are either general orthogonal matrices, A E q n ) , or special orthogonal, A E S q n ) = {U:detU = l}, if 4 preserves the orientation.
Mn the role of the orthogonal affine factors is taken by the
In Minkowski space
Lorentz (pseudoorthogonaZ) group Sql;n - l), which c nsists of matrices B, that preserve the indefinite product (z I y) = Jz y, where J = - matrix of Minkowski form,
41-11)
{ A : ( A zI A d = (z I Y)). So symmetries of the Minkowski space form the Poincare group P, of Special Relativity. The latter is generated by all translations and all Lorentz transformations, 4(z) = Az
+ b; b E Rn;A E S q 1 ; n ) .
Let us notice that in all three cases (affine, Euclidian, Poincare) elements
4 are
2We recall that a Riemannian manifold Ab carries a positive definite metric g = Cgijdtidzj, i.e. an inner product (( I tJg = Cgij
O, on tangent spaces of Ab {( E Tz}.Pseudo-riemannian refers t o the indefinite metric g, usually of the type (+; - ;- ;...). Any diffeomorphism (change of variables) z = #(y), on a Riemannian (pseudo-Riemannian) manifold Ab transforms the metric: g+T = g# = #'*(go #) #', the new entries being N
gkm(Y)
= C (9; j 4) 'I' '
azjwhere
aykaYm9
4' = (%) denotes the Jacobian (matrix) of #. aY
Map # defines an isometry of Ab, if the transformed metric g4 is equal to g. Such maps # obviously preserve the length of any path y = { y ( t ) : 0 5 1 5 T } ,
IY['
T
=
{ Cgij(~(t))i.ii.jdt,
hence the distance between points, d(+;y) = rnin{L[y]: y(0) = 2 ; y ( T ) = y}. Conversely, for any distance preserving diffeomorphism #:Ab-+Jb, the differential 42 (Jacobian matrix of 4), considered as the map of tangent spaces, #':T,+Ty (y = d(z)), preserves the metric (norm) on T,, (#:([) I#:(()) = ((I (), for any tangent vector (. Isometries of Ab clearly form a group. Conformal maps # do not preserve metric, but multiply it with a scalar (conformal) factor p(z), i.e. g4(z) = p(z)g. So they form a larger symmetry group of Ab. 31n Special Relativity R4 = {zo;
...;z3} represents a simplified version of space-time,
serves as time variable, while (zl; z2;z3)represent space coordinates.
where
I,,
51.1. Groups and group actions
15
identified with pairs (A; b), and the group multiplication takes the form
( A ;b) * (A';b') = (AA';Ab' t b).
(1.1)
It is easy to check that translations {b} form a normal subgroup
H
N
Rn of G,
while linear factors { A } , or { U } , form a subgroup I(
N GL,, or S q n ) ; S q 1 ; n ) . So group G is decomposed into a semidirect product4, G = H D K . The representation theory of
semidirect products will be analyzed in chapters 3 and 6. Two other examples of geometric symmetries arise on the sphere and the hyperbolic (Poincare-Lobachevski) space, the prototypes of the spherical and hyperbolic geometries. They also serve as the simplest prototypes of symmetric spaces of the compact and non-compact (hyperbolic) type. b) Sphere: Sn-' = {3::113:11~= l} in Rn with the natural (Euclidian) metric has the isometry group G = S q n ) - all orthogonal transformations (rotations) in R". In fact,
Sn-' can be identified
with the quotient (coset) space of group G = S q n ) , modulo the
stabilizer (isotropy subgroup)'
K = S q n - 1) of
fixed point zo in
a
Sn-' (the
North
Pole!), Sn N G / K . In other words (definition of stabilizer), I( = { U : U ( z 0 )= x0}. c) Hyperbolic space W, can be realized either as Poincare-Lobachevski complex half-plane { z = 3:
+ iy: y > 0 } , with metric
or as complex disk D = {IZI
< 1 ) with metric, ds2 =
the Mobius transformation,
One can easily check that
0
rn. dzdf
Both spaces are related by
takes W onto D, and transforms their Poincare
metrics, one into the other. The isometry group of 04 in both realizations is made of the fractional-linear transformations: az+c @z+w = bz+d'
U:Z-+X",
4A semidirect product G = H D U , of groups H and U , where U acts by automorphisms on H , (zE H ; u E U ) , consists of all pairs g = (z,u) witbthe multiplication rule (2,u)* (y, w) = (2 * y" ;uw). It is easily seen that H = {(z;e)} forms a normal subgroup of G, while K = {(e;u)} a
subgroup, so that H
nK = { e } , and the whole group G = H . K .
'Stabilizer (isotropy subgroup) of point c in a G-space X , consist of all elements {g E C}, that leave z fixed, K , = {g: zg = z}.
31.1. Groups and group actions.
16
5
1
In the half-plane case matrices {A = fi]:ad- bc = 1 belong to SL,(R) - the real unimodular group. In the disk realization they are given by complex matrices
detA=iaiZ-IPlZ=l A that preserve the indefinite hermitian form: J ( z ) = I z1 1' - / z 2 on C2 = {(q; 2,)). This group is called conformal and denoted SU(1;l), by analogy with the standard unitary group SU(2), that preserves the positive
in
other
words
matrices
I,,
hermitian product, l z l z = 1z11' +1zzl2. Once again the hyperbolic space coincides with the quotient of G = SL, (or SU(l;l)),modulo stabilizer I< of a fixed point z,, = i E W. The stabilizer I< is easily seen to coincide with an orthogonal group in R2,
I< = {A:
= i} =
SO(^), so w N SL,/SO(~).
Remark: The hyperbolic nature of the Poincare-Lobachevski geometry stems from its close connection to a hyperboloid r = { z 2 - (z2+y2) = 1) in R3. The former represents an orbit of the Lorentz group Sq2;1), acting by linear transformations of @ N h.03. So the Lorentz metric ds2 = ds2+dyZ - dz2, restricted on I?, remains invariant under Sq2;1). Furthermore, its restriction, ds2 I r becomes positive-definite (since all rtangent vectors [ = (z;y;z) are space-Me, 2 + y 2 - z 2 > O!). So r turns into a Riemannian manifold with a large symmetry group (symmetric space). In fact,
r N Sq2;1)/S0(2), where Sq2) acts as a stabilizer of the vertex (0;O;l)of I?. The connection between I' and D is established via a stereographic map @ (see fig.1). We parametrize
r and D by polar coordinates in the sy-plane, { ( r ; 8 ) : z= 2 1 +r2)for l?,and { ( p , 8 ) ) for D.
Then @:p+r = 2'2P So 1-P
the metric on dr' dsz _-_ -$
is taken into the Poincare metric in D,
r:
l+r2
ds2 = 4
dp2
r2dO2,
+ p2d@
(1 - p y
.
Fig. 1. Sfereographic m a p 9 takes a I': z2 - (z2 + yz) = 1, in hyperboloid R3 -M3, i n f o f h e unit disk D p2 = x 2 + y z < 1, in the xy-plane, and transforms the nafural ( S O ( 2 ; 1)-invariant metric on r into the Poincare metric on D.
51.1. Groups and group actions
17
The relation between I?, D and W suggests that their groups of isometries should be identical. Indeed, we shall see (chapters 5,7) that SL,@) 21 Swl;1) makes a two-fold cover of SO(2;1)(problem 5). Thus we have exhibited 3 classes of manifolds with rich symmetry groups: Euclidian/Minkowski spaces, spheres and the hyperbolic plane. In all 3 cases group G acts transitively on A (each point is moved to any other by an element of G). Hence space A is identified with the quotient G / K , where K denotes a stabilizer of a point xo E A, namely Rfl
21 E,/SO(n);
(I = S q n ) - stabilizer of {O};
Mn 21 P,/SO(l;n-l) Sn-'
N
- stabilizer of {0} in P,
SO(n)/So(n - 1) - stabilizer of xo = (1; 0; ... )
W IISL,(R)/So(2);
or D = SU(1;1)/SO(2); K = So(2)-stabilizer of {i} (or (0))
In fact in all three cases group G acts in a stronger double-transitive manner on A, meaning that any pair of points (2; y } can be transformed into any other equidistant pair {x';y'}, d(e; y ) = d ( d ; y'). This fact has important implications for the analysis on such manifolds. Let us remark that all the Riemannian examples above (except Minkowski) belong to a wide class of symmetric spaces. These manifolds can be generally described as quotients of Lie groups modulo maximal compact subgroups, A N G / K , and we
shall see many more examples in subsequent sections.
1.2. Finite groups. The easiest to describe are commutative finite group. The complete list includes cyclic groups: Z, 21 Z/nZ, which could also be written in the - the n-th primitive root of unity), and complex form: {eiEw:O 5 k 5 n - l}, (w = their direct sums: G = Zfll x ... x Z ,.
2) Symmetries of regular polyhedra and finite subgroups of SO(3). We shall start with regular polygons in the plane.
i)
Symmetries of the regular n-gon consist of rotations by angles EZ , - a finite subgroup of all planar rotations SO(2), or the larger dihedral group, D, = Z, Z, c O(2) - a semidirect product of rotations and a reflection
{gk:k = 0; l...}
about any symmetry axis. D, can be thought of as the symmetry group of the dihedral A,, a solid in R3, built of 2 pyramids based on a regular polygon R, in W2 and a pair of opposite vertices that project onto the center of R, (see figure ). Then Z, implements
18
$1.1. Groups and group actions.
axisymmetric rotations of the dihedral in the plane, while a generator of
Z, flips 2
opposite (spatial) vertices. In 3-D one also has 5 Platonic perfect solids, each one with a symmetry group of orthogonal transformations in
W3 that
preserve its vertices:
ii) tetrahedral symmetries: A, - alternating group (even permutations) of order 4. The generators of A, are cyclic permutations of 3 vertices in each tetrahedral face, that leave the opposite vertex fixed. iii) cubo-octahedral symmetries: the symmetric (permutation) group W,. Cube contains two opposite regular inscribed tetrahedra (fig.2), so its group contains an A, plus an element u that transposes both tetrahedra. Hence Gcube= A,
UaA, = W,.
Fig.:! shows cube with 2 opposit inscribed tetrahedra.
A, - alternating group of order 5 [Cox]. Indeed, a dodecahedron contains 5 inscribed tetrahedra {TI;...;T5}(the set of 20 vertices is evenly split into 5 quadruples). A rotation of order 3 about any pair of opposit verteces leaves a pair of tetrahedra (say T,;T,) fixed and cyclicly permutes the remaining triple (2'5 T,;T,). Obviously, even permutations, {(123);(234);(345);...} generate As! iv) Icoso-dodecahedral group:
Fig.3 demonstrates one of 5 regular tetrahedrons, inscribed inside the dodecahedron. Any pair of opposit vertices ( A ;A') selects a pair of tetrahedra, and any rotation of order 3 about the AA'-aixis, leaves the pair fixed, and cyclicly permutes the remaining triple.
We shall see now that (i-iv) completely describe all finite subgroups of the orthogonal group S q 3 ) . Classification Theorem 1: A n y finite subgroup G of S q 3 ) coincides with one of the
above polyhedral groups (i)-(iv).
19
51.1. Groups and group actions The proof involves several steps. 1) We observe that any rotation U E S q 3 ) has two fixed points on the unit sphere
S2 = {It t 11 = l}, indeed the eigenvalues of an orthogonal U in R2 are X = e
* ip; l!
2) Take set X of all G-fixed points on S2. Set X is G-invariant, hence splits it into the union of G-orbits:
X = w1
G, = { U : U z = z}, and call
u...Uwm.
For each point z E X we consider its stabilizer
I G, I = n2 - the
degree of
(2).
same orbit have equal degrees (stabilizers are conjugate!),
80
Obviously, all points on the nz
= n(w 1.) = n3., for I E w j .
3) Next we count the number of pairs { ( z ; U ) :z E X ; U E Gz} (the “total degree of C”) in two ways: #{fixed points of all U E G\{e}} = “sum of degrees of all {z}”.This yields,
c
2(IGI-1)=
c rn
(IG,l
-l)=
lwjl(nj-1).
(1.3)
+EX 1 4) Dividing both sides of (1.3) by I G I we derive the equation relating the order of G t o degrees of its fixed points:
I
I
The rest of analysis closely resembles a classification of Platonic solids. Namely, one can show that m in (1.4) could take on two values m = 2;3 only! Each case is analyzed separately. 5) In case
m,one easily shows n1 = n2 = n = I G I. So G has two 1-point orbits of
degree n, which implies G = Zn- cyclic!
6) In case1 Case
there are 3 possible subcases:
PI = n2 = 2; n3 = nl. Here G has two “n-point” orbits of deg = 2 and a “2-point”
orbit of deg = n, which implies G = Dn- the dihedral group!
-1
Case
yields 3 possibilities for
n3
3; I G I = 12; Gtetr ( 3 orbits of degree: 2;3;3) 4; I G I = 24; GCube(3 orbits of degree: 2;3;4) 5; I G I = 60; Gdodec(3 orbits of degree:2;3;5)
In a similar vein one can describe symmetries of regular polytopes in higher dimensions [Cox].
3. Other interesting classes arise as automorphisms of finite groups, and regular graphs. Among them we shall mention finite groups of Lie type: G = GLn(F), and
SLn(IF),made of n x n-matrices with entries in
a
finite field IF. There many similarities
51.1. Groups and group actions.
20
in the analysis and representation theory of classical and finite-type Lie groups, but our attention will be focused mostly on the continuous case.
1.3. Compact groups. These are topological groups with compact space G. The E [O;l]}ci Rn/Zn; the classical most important examples include t o w 8" = {(tl;.,.tn):tj compact Lie groups, like orthogonal - S q n ) ; unitary - SU(n); symplectic - Sp(n), et al. The general theory of the compact and finite groups will be developed in chapter 3. The classical compact Lie theory will be covered in detail in chapters 4-5. Other interesting examples arise as matrix groups over p-odic numbers (pprime): Q,, or integers Z,, i.e. the closure of rationals Q, or integers Z, in the padic norm: II a llP = p-". Here n denotes the largest power (positive or negative) of prime p in fraction a = p"a'.
The corresponding p - a d i c Lie groups: GLn(Qp); SLn(Qp), and their compact subgroups
GL,(Z,); SL,(Z,), consists of all Qp(Zp)-valued n x n matrices (respectively matrices of det
= 1). These groups find applications in the number theory and algebraic geometry,
but we shall not venture into the subject of the padic analysis (see [GGP]; [JL]).
1.4. Lie groups form the most important and interesting class, which plays the fundamental role in many areas of Mathematics and Physics: analysis, differential equations, geometry; classical, quantum, statistical mechanics. In general, Lie groups are defined as manifolds with smooth (differential) group structure: multiplication and inversion operations. A brief introduction to the Lie theory is provided in $1.4. The main bulk of the book will deal with the analysis and representations of Lie groups, emphasizing both general aspects of the Lie group theory as well as many specific examples and applications. Lie groups could be divided into two large and distinct classes: solvable and nilpotent (a subclass of solvable); simple and semisimple. The definitions of both classes will be given in 51.4, and the detailed analysis conducted in chapter 6 (nilpotent and solvable groups), and chapters 4,5,7 (simple and semisimple groups). Although both classes differ substantially in their structure, there are abundant connections, and parallels in the harmonic analysis and representation theory. The first class is exemplified by
+
b}- all affine transformations of W. i) 1-D affine group: A f f = {da,*:x-iax This group can be also realized by 2 x 2 matrices of the form { A = b E W;a E R*}. ii) the celebrated Heisenberg group 3 x 3 matrices
W, (and its higher-D
(" i);
cousins Hn), realized by
21
31.1. Groups and group actions
iii) The group B, of upper/lower triangular matrices in GL, (called often the
Bore1 subgroup),
{
B,= A =
r
A, b
1
Xj;a,b,cER,orC
Groups Aff, and B, are solvable, whereas
.
W, - nilpotent.
In $1.4 we shall see that any Lie group can be decomposed into a semidirect
product B D H , of the solvable normal subgroup B, and a semisimple subgroup H . Any semisimple group H in turn can be decomposed into the direct product of sample groups. The latter were completely classified in celebrated works of E. Cartan. They comprise 4 series of classical Lie groups, listed below, and a few exceptional groups. Many classical Lie groups arise as linear transformations of vector spaces over 02, C or quaternions Q, that preserve certain bilinear/quadratic forms, in other words as subgroups of GL, or SL,. These include,
i) Orthogonal groups: q n ) , S q n ) , preserve Euclidian inner product x .y , q n ) = {U:Ux . Uy = z . y } , and S q n ) = {U E q n ) :det U = 1). There are also indefinite orthogonal groups q p ; q ) and S q p ;q), that preserve indefinite products: P P+Q (z y ) = f: z j y j - C z j y j in P+l
I
so that
RPtQ,
or
CP+q,
( U z I Uy) = (z I y ) , for all z,y. The principal difference between the definite-type and indefinite-type groups in the real case is that the former are compact, while the latter are not, as exemplified by the S q n + l ) and the Lorentz S q l ; n ) , mentioned earlier. But in the complex case the difference disappears, so S q p ;q ) N Sqp+q), for p , q. ii) Unitary groups: U(n), U(k;m), preserve definite (or indefinite) hermitian inner product in C“, z . F , or k
( z I w) = ?zjwj
- c zjwj; in cktrn. ktm
ktl
A particular example is the conformal group SU(1;1). iii) SympIectic grou s: Sp(n) consist of all 2n x 2n matrices that preserves a skewsymmetric form J =
[-I IF (z
I z’) = a:.y’-z’.y
= J z .z‘; z = ( 2 , y ) .
$1.1. Groups and group actions.
22
In all cases the corresponding group consists of matrices, that satisfy
U*JU = J , where U* denotes the transpose (or hermitian adjoint) of U , and
J is either identity (for
the orthogonal and unitary groups), or the symmetric form Xkm with k pluses and m minuses on the main (for the indefinite orthogonal and unitary groups), or the skew symmetric form J in the symplectic case. The above list contains most of the classical examples, but it does not represent their classification scheme (see chapter 5 ) , our emphasis was mainly on construction. The reader will find (problems 5,6), that 3 series overlap, particularly in low dimensions. More examples of this nature will come in $1.4 and subsequent chapters. 1.5. Discrete groups. Those typically arise as discrete subgroups of continuous
(Lie) groups, like lattices Z' 2 Z" in W", or their noncommutative counterparts, lattices r c Gf,; SL,, and other Lie groups. Important examples of the noncommutative lattices are i) the unimodular group I' = SL,(Z), and some related subgroups of SL, ii) discrete Heisenberg group: matrices (1.5) with integer a , b,c E Z.
iii) discrete subgroups
r of the Euclidian motion group En.
The latter are called crystallographic gToups, as they describe all possible crystalline arrangements in Euclidian spaces. They were thoroughly investigated, and classified in dimensions 2 and 3. As E, itself such groups are decomposed into a
r =A
U , where A 21 Zn forms a lattice in R", while a finite group U c Gf, acts by automorphisms (linear transformations) of A. semidirect product,
In 2-Dthe complete list contains 17 groups, whose U-components could take only 10 possible values: {Zm;Dm:m = 1;2;3;4;6}. The complete list in 3-D includes 219 nonisomorphic and 230 nonconjugate crystallographic groups. It was worked out by Fedorov, Schoenfliess and Barlow at the turn of the last century. As a step towards classification one needs all finite orthogonal symmetries, derived in Theorem 1. For further details and relates issues we refer to [BH]; [HC]; [Sch].
$1.1. Groups and group actions
23
Problems and Exercises: 1. Show that any rigid motion on R" (Euclidian or Minkowski) is linear, d(z) = A z for some a E R", and matrix A. Follow steps:
i) The Jacobian q5'(z) is an orthogonal matrix, i.e.
+ a,
aid. ajq5 = 6ij
ii) Differentiate the orthogonality relation (i) in the k-th variable, ak(...), then dot,. Show the resulting 4-tensor tLJ = (8,,,+4 Bid) to be symmetric multiply with 84 in both row-indexes: tik = tki; tm3= tJm; and antisymmetric in column-indexes (changing i - j or k-m changes its sign). iii) Show that any such t must be at once symmetric and antisymmetric in the pair of indexes {ij}. Hence t = 0, for all quadruples ijkm, which implies 8'4 = 0, i.e. q5 is linear! 2. The isometry group of S"-' consist of rotations {V E Sqn)}. Hint: any isometry of s n - 1 extends to an isometry of R"\{O} N S"-'xR+. But the latter are linear by problem l!. 3. Show that fractional-linear transformations (1.2) of SL, on H, or SU(1;I) on D, are isometries. 4. Show that all finite subgroups of S q 3 ) are either polyhedral, or polygonal groups, described in Example 1.1.
5. Establish and find explicit form of the isomorphism SL,(R)+SU(l;I), using their fractional-linear actions on H and D, and the Caley transformation 4: z - + G , from M z+z
to D. Notice that both are subgroups of the larger complex unimodular group SL,(C), with a Caley element u E SL,(C). Find and can be obtained by conjugation, U+U-'UU, U!
6. Show that the symplectic group Sp(1) coincides with SL,. 7. Show that complex groups S O ( p ; q ) and SO(p+q) are isomorphic (Hint: the indefinite product ( z 1 w ) in Cp+* is equivalent to the definite product z . w by conjugation with a complex diagonal matrix).
$1.2. Regular and induced representation
24
51.2. Regular and induced representations; Haar measure and Convolution algebras. We introduce two basic concepts of regular and induced representations, discuss continuity and unitarity and develop some basic algebraic constructions: direct sum, direct integral, tensor product. In the process we introduce the (invariant) Haar measure on groups and homogeneous (coset) spaces H\G, define convolution (group) algebras L’(G), and find the links between representations of the group and those of its groupalgebra.
2.1. Regular representations: In the last section we have described some examples of transformation groups acting on various geometric structures (graphs, manifolds, symmetric spaces), including the action of group on itself by the right/left multiplication (translation), g: x-+xg; or x+g-’x. With any such action of group G on space X (denoted by x - + x g ) , we can associate a
linear
G-action on functions over
homomorphism of G into linear operators over
X, {f(x)} = e = e ( X ) , i.e. a
e,
In particular, the right/left translations, x-+xg; x+g-lx,
on G give
IR,f(x) = f ( g - ’ z ) (left), or f ( x g ) (right), on e(G)l Homomorphisms g+T,, vector space
Y,are
(2.2)
of group G into linear transformations/matrices on a
called representations, and formulae (2.1)-(2.2) provide important
examples of regular representations of G on X, or on itself. In the analysis of group actions and associated representations on various function-spaces, it is often necessary to “integrate” over G or X . So we need some measures on X and G . 2.2. The Haar measure. It turns out that all discrete and continuous (locally compact) groups G , as well as large classes of homogeneous spaces6 X = H\G have an invariant measure, dp(xg) = dp(x), for all x E X, g E G, called the Haar measure. The general proof will be outlined below. More important, however, will be to compute the Haar measure in a suitably chosen coordinate system on group G or space X. Here we shall list a few examples of Haar measures on groups and homogeneous spaces (many more will appear throughout the book). ‘homogeneous (quotient) space X = H\G, or G / H , consists of all right/left cosets {z = Hgo:gOE C } , or {z goH:go E C}. Group G acts on X by right/left translations, g:z+zg = ( H g o ) g , or 2-9(z).So X could be viewed as a G-space, with a transitive G-action, and subgroup H coincides with the stabilizer of a fixed point, H = {g:rog= zo}.
$1.2. Regular and induced representation
a) finite/discrete group G has dp(z) =
25
C C ~ ~sum ( Xof) &-functions over
all
Y
y E G.
b) for commutative groups Wn;Tn, dp coincides with the standard Lebesgue measure (volume element), dz = dxldz, ...dx,. We have shown in $1.1that Rn could be regarded as a homogeneous space of the Euclidian motion group G = En. The Lebesgue measure is clearly invariant under all translations and rotations on
W", hence it also
forms a G-invariant measure on homogeneous space W" = E n / S q n ) . c) on compact Lie groups S q n ) , SU(n), the Haar measure can be explicitly calculated in terms of suitably chosen coordinates, like Euler angles (chapters 3-4). The same could be done on non-compact (Lorentz, conformal, etc.) groups, like S q 1 ; n ) or
SU(1;n ) , by a combination of spherical and hyperbolic angles.
W is a homogeneous space of dz d y measure d p = -is G-invariant.
d) the hyperbolic (Poincare-Lobachevski) half plane
G = SL,(R). One can check (problem l), that
Y2
e) GLn(W) has a natural set of coordinates, matrix entries {zjk}'&l
of g. One can
show (problem 6) the invariant measure on GL,
dPk)=
&=
ndXjk. Jk
For more examples see problems 5-8.
Existence Theorem: i) On any compact (and more general locally compact) group
there ezists a Bore1 measure dp(x), positive on all open subsets, finite on all compact subsets and invariant under all right (or left) translations: EX) = p ( E ) , for all subsets E c G and elements x E G. ii) A right and left Haar measures: d,x; d,x are unique up to a constant factor. iii) On a compact group G the left and right Haar measures are equal, furthermore dp is invariant under the group inversion and conjugation,
dp(x-') = dp(x), dp(g-'sg) = dp(z). Let us briefly outline the proof.
Existence: To construct a right-invariant measure dp, we pick a small neighborhood U of the identity e E G to serve as a gauge. Given a (open, closed) subset E optimal (least) covers of E and G by translates of U , m n E C u U z j , G = U U yj*
j=1
j=l
C G, we choose
51.3. Regular and induced representation
26 The ratio
represents an approximate relative size of E in G in the ”U-gauge”.
m ( E ; U ) as U-.{e}, (by standard arguments such limit always exists!), one Taking limit n(C;U ) gets an honest Borel measure on G. The limiting measure, dp, thus constructed is easily seen to be right-invariant. Also p ( E ) > 0 for any open E, and p(G)
< 00 for a compact
G , since G is covered by finitely many translates of E. Uniqueness: If dji is another right-invariant measure, then for small U ,
m ( E , U ) ef-i ( E ) while n ( G ; U )e m(U)’
subsets E.
fi(U)’
whence p ( E ) = lim
U-lel .
$ = Const f i ( E ) , for
all
I
iii) T o show that the right and the left-invariant Haar measures on a compact G are equal, we take a left translate, d j i ( z ) = dp(az), of a right-invariant measure d p ( z ) . Of course, dji is also right-invariant. Hence dji = p ( o ) d p , by the uniqueness of dp. The map a+p(a) is a homomorphism of
G into the multiplicative group R+. But there are no such
nontrivial continuous homomorphisms on a compact group (continuous functions on compact sets are always bounded, whereas nontrivial homomorphisms p: G-R,
are
unbounded!). Hence, p = 1, and d p is also right-invariant. Two other invariance properties of d p easily follow now. Indeed, change of variable 2-z-l
takes a right-invariant measure into a left-invariant measure, the conjugate-
invariance immediately follows from the bi-invariance: d p ( a r b ) = d p ( z ) for all a, b. This completes the proof. R e m a r k In many cases (e.g. Lie groups) the right and left invariant measures are dlb) absolutely continuous one relative t o the other. The corresponding density -= A(g) is 4(9) a character (homomorphism into the numbers) of G with positive real values
A:G+R,; A(gh) = A(g)A(h), called the modular function. Groups with equal right and left Haar measures, A(g) = 1, are called unimodular. Large classes of groups are known t o be unimodular, for instance, all compact groups, semisimple and nilpotent Lie groups. Nonunimodular examples include: a f i n e groups and more general Borel groups of all upper/lower triangular matrices (problem 8).
The invariant (Haar) measure, topology/metric, and the differential structure on group G allow to int,roduce a variety of function-spaces. The most important among those are
C(G) - continuous (bounded) functions {f(z)}on G, with norm
II f II 00 = S U P I f(z) I ZEG
7
$1.2. Regular and induced representation and its subspaces functions; 0
e,, -
LP-spaces on
27
functions vanishing at {cm},and C, - compactly supported
G with respect
to the Haar measure dz,
in particular Hilbert space LZ(G)with the standard inner product
0 For Lie groups G (smooth manifold) one can consider a variety of differentiable function-spaces: Y ( G ) - Um-smoothfunctions”, C”(G) - infinitely smooth functions; their subspaces of functions vanishing at {m}, or compactly supported ep, with norm
er;
sum over all
llfllrn = C S U P Jaaf(41, XEG “partial derivatives” of order I a I 5 m (the meaning
: f l * f z ) ( z= )
J
fl(zy-’) f z ~ d )y = G
J
of “derivatives”) on
~ I ( Y )f z ( ~ - ’ z d5 )
(2.3)
G
integration with respect to the Haar measure d y on
G. On
finite/discrete groups
integration in (2.3) is replaced by the sum,
f*h =
c
f(zy-’)h(y).
YEG
The meaning of convolution, as an extension of the group multiplication to functions, becomes transparent here, as each group element {z} is identified with the delta-function 6,, so that
31.2. Regular and induced representation
28
6,* 6, = SZy;for all x,y E G. For continuous groups
G convolution-integral (2.3) is well defined on compactly
supported functions (ec;ep),then it extends to larger classes by the standard density arguments7. Spaces closed under convolution forms convolutzon/group algebras. One important example is space L'(G) of all integrable functions on G. Here,
Ilf*hli
(2.4)
A similar estimate holds for a convolution of L' and LP-functions,
1 f* h Il p 5 (1 f 1 IIh IIp ; for all pairs f E L', h E LP.
(2.5)
In the standard terminology L' is a Banach algebra, while other LP are modules over L'. Estimates (2.4)-(2.5) follows from the well known Minkowski inequality,
applied t o function F = f(zy-')h(y).
The inequality (2.6) has an obvious interpretation,
any function F ( z ; y ) (z E X ; y E Y) defines a map F:Y-.LP(dz), then (2.6) claims:
1 JF(y)dyll<
lIIF(y)Ildy, where
(1 11 means the LP(dz)-norm. Other examples of
convolution-algebras include spaces of continuous and differentiable functions:
em,
e, em,
as well as bounded (Borel) measures on G, A ( G ) 3 L'(G). Condition (2.5) can be
stated as boundedness of the convolution (bilinear) map, L'*LP-LP,
or L'*LP
c LP, 1 5 p 5 03.
It is also easy t o see that L2*Lz C Loo, and Ilf*hll, 5 llfl1211 hllz (Cauchy-Schwartz). Such results can be extended to many other triples: LP*Lq
c L', and other "scales of
function-spaces" by interpolation discussed in chapter 2 (J2.2).
2.4. Continuity and unitarity. The study of group representations combines two
aspects: algebraic, which reflects purely algebraic properties of groups and operators, and analytic (topological), which involves the geometric structure of manifolds A (differentiable, complex, etc.), as well as topological properties of the relevant vector (function) spaces and operators over A. In the topological context (locally compact/Lie groups G, Banach or Hilbert spaces V) one needs to impose some continuity assumptions on 'Spaces
ec and ep are dense in all Lp, 1 5 p < co!
T.One possibility would
$1.2. Regular and induced representation
29
be to require continuity of the map g+Tg in the operator (norm) topologfl on space %(V), of all bounded operators on T. This notion, however, turns out to be too restrictive, (see problem 2, and the author's paper ', JFA, 35, 1980).
A weaker notion of continuity asks vector functions {TgfiG+T}, or the matrix t E T*-dual space} to be continuous. It is called entries of T , {tcv(g)= (T,J I ~ ) E: T;Q strong continuity. Strong continuity holds in most natural examples, including regular representations (2.1), (2.2), and (2.8) below, in different function-spaces: e; em;LP(G), etc. (both on G and on homogeneous spaces X=H\G, G modulo a closed subgroup H ) . Let us remark that quotient-spaces X of topological groups often inherit basic structures of G, like metric, differential structure, and in some cases G-invariant (Haar) measure. An important class of group representations consist of unitary representations in Hilbert (inner product) spaces:
The natural example is furnished by the regular representation R on squareintegrable functions V = L 2 ( X ) ,provided space X has a G-invariant Haar measure, d ( x g ) = dx, for all g E G. The Haar measure was shown to exist on any locally compact group G, but quotient spaces X = H\G may or may not have it, as we shall demonstrate. The invariant measure always exists, if subgroup H is compact, e.g. on symmetric spaces. However, in any case a homogeneous space (manifold) X comes equipped with some measure (volume element) dx. Group elements ( g E G}, acting on X transform dx into another measure dxg, which will differ from dx by a density factor: In case of the smooth (Riemannian) manifolds and G-actions (e.g. a(x;g) symmetric space X ) , a represents the Jacobian determinant of the map 4: x - d ,
=%.
a ( x ; g )= d e t # ( x ) .
(2.7)
We can still construct a unitary representation of G on L * ( X ; d z )by combining the group (translational) action on X with multiplication by fi,
Function a ( x ; g )is easily seen to obey the so called cocycle condition, 'We remind the reader that the space of all bounded linear operators in a normed vector space 11(11< l}, or sup{(A( 7):all ( ; q } in the Hilbert space setup.
Y is equipped with an operator-norm: 11 All = sup{IIA(Il: all
'We have proved that any norm-continuous irreducible T must be finite dimensional, which automatically excludes large and important classes of m-D representations.
51.2. Regular and induced representation
30
4";9192)= +;g1)4z91;gz); all gI;gZ E G , which yields the representation property for operators T , (2.8),
(2.9)
T g l g 2= Tg1Tg2, for all gr;g2in G . Obviously operators {T,} preserve the L2-norm of
f, so we get once again a
unitarity representation of G on L 2 ( X ) !
2.5. Induced representations. Regular representations (2.2)-(2.1) on G and quotients H\G, as well as more general representation (2.8) are obtained by the process called induction. In general, induction allows to construct representations of group G , starting from representations of its subgroup H. An alternative definition is based on the notion of cocycle u(z;g)on the homogeneous space X = H\G, i.e. a function u: X
x G + {numbers} or {operators},
that satisfies the cocycle condition (2.9)
a(z;glg2)= a(z;gl)u(xgl;gz);for all gl;g2 E G , x E X . The induced representation, T = ind(a I H ; G ) , acts on the space of (Y-valued) functionslO on X,
e = C(x;Y)
= {f(z) : X + Y },
and is given by
We leave to the reader, as an easy exercise, to check that (2.10) does give a representation of G. Furthermore, if u is continuousfunitary and X has a Haar measure
d x , then T , also becomes a continuousfunitary representation'' of G on space L 2 ( X ; Y ) , of square-integrable Y-valued functions on X (problem 9). Let us remark that the cocycle condition (2.9) for elements of stabilizer H = G, = {g: xg = z}, yields a representation ~ ( h=)a ( z ; h )of H on
Y,u ( h l h 2 )= u(hl)u(hz).So one could start with a
representation u of H and construct a cocycle u(z;g)!This alternative way to define 2nd will be discussed in $3.2. Regular representations R on G and quotients X = H\G, are induced by the trivial cocycle u,
R = ind(1 I H;G). '"In the topological (Lie) setup one takes a suitable function-space on X, e.g. continuous, differentiable, L2-functions, etc. "In the absence of invariant measure on X we can modify operators T, (2.10) by a scalar (2.7) to make T unitary. Obviously, the product of two cocycles (I and makes factor another cocycle!
fi
$1.2. Regular and induced representation
31
We shall see the induction procedure to appear throughout this book in various places and contexts, in the representation theory of finite, compact, simple, nilpotent, solvable groups alike. The role of induction will be twofold: on the one hand it extends the notion of regular representations on G and X to vector functions and sections of
vector bundles12, on the other hand many irreducible representations of groups will be constructed via induction. Here we shall illustrate the induced action by two examples of homogeneous spaces H\G, without G-invariant measure (see problem 4), and construct the relevant regularlinduced unitary representations.
+
1) Affine group G = {g = (a;b):z-iaz b = zg} of $1.1 acts transitively on R z H\G, where H denotes the stabilizer of {0}, H = {(a;O)} R* (multiplicative group of reds). The “az+b” action on R has no G-invariant measure. Indeed, the translational invariance ({b}-part) requires dp to be Lebesgue, but multiplication with the {a}-factors dilate it, dp-‘ 1 a 1 d p . However, we get the cocycle a ( rr g) --@dz =lab
(2.11)
hence a unitary action of G on L2(R),
The representation theory of affine groups will be discussed in chapter 6 (56.1). 2) Group G = SL,(R) and its subgroup of upper-triangular matrices
H
=
{(R akl)}
yield the quotient-space X N R, with G acting by the fractional linear transformations, g: 5
ac+b
3 cz+d9
for
(2.12)
=
Indeed, cosets z E H\G can be parametrized by matrices
{rz = (i
coset representatives, via the Gauss factorization: (almost) any element g = factors into the product,
(: 3 = (d-l
$(b/ld
l):z E R}
-
(ii), ( d # 0 )
1) =
To find the cocycle we take the product yzg, and factor out the h-term, y z . g = ( az+b a cr+d ) = h . y ,
where
vector bundle TI consists of the base manifold X , stuffed with fibers (linear vector spaces) {YZ} associated to points of z E X . The natural example is the tangent bundle, made of all tangent spaces {Tz: z E X } of a smooth base-manifold.
A-
51.2. Regular and induced representation
32
The density factor a is computed from the transformation law (2.12) to be a(t;g) = 1 ct+d
I - 2.
(2.13)
So a unitary representation of SL, on space L2(R) takes the form,
More general representations of SL,(W) on Lz(W) are obtained by taking complex powers of the cocycle,
(2.14) These are so called principal series representations of SL,, that will be studied in detail in chapter 7.
2.6. Representations of group algebras. Any group representation T, gives rise to a representation of a suitable group-algebra L = {f}, of functions on G, by integrating T against f , (2.15) It is easy to check that map f+Tf satisfies
Tf*,, = TfTh, for any pair f , h E L . Conversely, any algebra-representation f -tTf, subject to a minor technical assumption (span{T,('V):f E L} is dense in 'V), can be shown to come from a group representation g-tT, via (2.15). As usually, in the topological context one has to give a meaning to integration (2.15) and to find a proper class of functions { f } on G. If T is continuous (which is always the case), and f - compactly supported, integral obviously exists and yields a bounded operator, )IT,II 5 Jlf(~)IIITllld..
Tf
(2.16)
Estimate (2.16) also indicates the class of functions, for which T , can be extended, as a bounded operator. The latter depends on the behavior of function w(z) = IIT,II on G. Clearly T f extends to the space (algebra) of functions, integrable
{f:/lf(z)lw(s)ds < m}. In particular, unitary (or more general bounded) representations, I(T,(I5 C, yield bounded operators {T,} for all f in the group algebra with weight w,
L = L'(G), indeed, I)T,(II Cllflll. If T is a left regular representation (2.2) on G, then operators (2.15) are nothing but left convolutions with f,
51.2. Regular and induced representation
33
R f [ h ]= f*h. The main advantage of passing from the group representation T , to its integrated form (2.15) has to do with nice (smoothing/regularizing) properties of operators { T I } . To wit the “unitary/bounded group operators” {T,} could yield a better class of “algebra operators” { Tf}:compact, Hilbert-Schmidt, or trace-class. The latter often depends on the regularity properties of
L = em(G);P‘(G), the “smoother11
{ f ’ s } one takes the “better” { T I }results. A simple illustration is furnished by convolutions with f E Cm(G) (or Coo). It is easy t o check (problem l l ) , that any such operator
R,: ZP-.Cm
(P), i.e. Cm*ZP C em,for all p, m.
In other words, convolutions with m-smooth
{f} transform ZP-singular functions { h } into
m-smooth functions. Smoothing properties of convolution operators with nice (smooth) functions are well known and frequently utilized techniques in harmonic analysis and differential equations, that go under the name of regularization methods. We shall use them extensively in subsequent sections, but here we just give a simple illustration of smoothing by convolutions in problem 10.
2.7. Algebraic constructions. We shall conclude this section with a few algebraic constructions of representations. Direct sum of {T’; Tz;...}lacting in spaces Yl;Y,...is defined in the obvious way by taking the direct sum of spaces Y,@ Y2@ ... , and setting,
(T’ @ T2@
...)g((1;(2;...)
= TL(1@Ti(, @ ...; (j E Yj.
(2.17)
There is a continuous version of (2.17) called the direct integral decomposition. Given a family of representations {TX:XE A}, in spaces {X’} (assumed to be regularly embedded in the embient space X), and a measure dp on A, we define direct integral X’dp(X), to consist of all L2-vector valued functions on A, that for each X E A, take on values in 36,: F = f(X) E X’. We then define direct-integral operators,
to map vector-functions f(X)
+ Tx[f(X)].
Tensor (Kronecker) product T’ &I T 2 is defined in the tensor-product space’
Y1@ Y,: (T’ @ T2),({&I 17) = Tb{ &I Tiq.
81.2. Regular and induced representation
34
An analytic version requires a proper topological product
Y,@ Y2.
Thus for
regular representations R3 ( j = 1,2) in spaces YJ = L 2 ( X j ) , the product space Y = Y 1 @ Y 2 L~ ' ( X , x X , ) , a n d G a c t s o n Y b y
Rgf(";Y) = f(& Y,). The contragredient (dual) representation Pg = Ti-,, in the dual space Y* to Y.Here T* denotes the dual (adjoint) operator to T. 0
The adjoint representation: Ad = AdT in the space L(Y)-all linear transformations on Y, or its subspaces, Ad,(A) = Ti'AT,, A E L(Y). Space L(Y) is naturally identified with the tensor product of Y 8 Y' (Y by its dual Y'), where tensor (@e+"rank-one operator (el ...)?. Then one can show that AdT= T@F - tensor product of T and its contragredient. A natural example of Ad is the action of GL, by conjugations on the space of all matrices, g:A+g-'Ag; g E GL,, A E Mat,,. More generally, the adjoint action appears in the context of Lie groups acting on their Lie algebras, and will be explained in 81.4. 0
-Tensor product of 2 vector spaces W = Y,@ Yz is spanned by linear combinations of {( 8 q:( E Yl;q E Y2}, subject to the natural bilinear rules: elemental products ( ( + q ) 8 ( = < @ ( + q @ { , and similarly for ( @ ( q + ( ) . The easiest way to think of W is in terms of function-spaces: Y,N e ( X ) , Y2 N e ( Y ) (for finite-D spaces Y,sets X and Y are finite!). Then W is also a function-space e ( X x Y ) , where elemental tensors {( 8 q ) correspond t o product-functions {<(z)q(y)). Such correspondence extends to some infinite function-spaces, e.g. L 2 ( X )8 P ( Y ) u Z2(X x Y ) .
35
51.2. Regular and induced representation
Problems and Exercises: 1. Show that the fractional-linear action (2.12) of group G = SL,(R) on the Poincare, and that d p is the Lobachevski half-plane H leaves measure dp = ~ - ~ d z d yinvariant only G-invariant measure of H. 2. Show that the regular representation of G = R" (1")is strongly continuous on various function spaces over R" (1"): (i) (ii)
e, = {all continuous functions vanishing at co} with
norm 11 f 11 = ma4 f(z)I;
e r = {m-times
continuously differentiable functions, vanishing a t co), with norm IIfIIm= C la1 <mmazl'of(z)I; (iii) C F = Q with a sequence of seminorms (11 f m=O; 1; ...)
er;
Ilm:
(iv) Lebesgue Lp-spaces, 1 5 p
< 00, with l l f l l p = (llf(z)rd~)'/~.
Hint: Verify continuity of the map a + R , f , for compactly supported functions f in e, or (uniform continuity!), then utilize the standard functional analysis: density of compactly supported f ' s in e,, and LP, and the uniform boundedness of the family of translations { R,: Q E R"}, R , = I!
er
1 1
The above argument applies t o other groups G and homogeneous spaces X, and their regular representations R on e ( X ) ;Zp(X), etc. as long as transformations g: z+zg, are continuous and measure preserving, i.e. X has a G-invariant Haar measure. (v) The regular representation R on G or X = H\G is not uniformly continuous, as
1 R , - 111= 2, for all g in the vicinity of {e}.
3. Check that any cocycle a ( z ; g ) defines a representation T = T f fvia formula (2.8). Show that the product, a@, and a power 19of cocycles, are also cocycles, hence T" of (2.14) is also a representation. 4. A cocycle (I is trivial, if a(z,g)= @(z8) for some function p(x).
4.)
(i) Show that cocycles (2.11) and (2.13) are nontrivial. (ii) A trivial cocycle yields a Haar measure on X. Thus in both examples (affine and SL,) the quotient H \ G has no Haar measure.
5. Show that the Haar measure on a semidirect product G = H D K, with compact group K (e.g. Euclidian motions En = R" D Sqn)), is equal t o dg = dhdk - the product of Haar measures (Hint: Haar measure is invariant under any compact group of automorphisms).
6. Show that the Haar measure on SL,(R) = {g = (5
!) detg = l},
Use the Haar measure on GL, = SL, x R', write any g E GL,
is given by
as
and compute the Jacobian a(a; b; c; d ) y; 2;A)
' ( 2 ;
=I
I
Xz'
.'7 Show that the Haar measure on SL,(R) parametrized by its entries (excluding zll!)) is equal t o
d
g
=
1
{g
= (zjk):j + t > 2
36
51.2. Regular and induced representation where M , , denotes the 1,l-minor of g. Formulae of problems 5-6 look elegant, but found little practical use in harmonic analysis on SL,. The latter exploits a different set of coordinate, based on certain factorizations of SL, (Cartan, Iwasawa), described in chapter 7.
8. a ) Find the left and right Haar measures on the atline group G = {az and compute the modular factor: A(a; b) = l/a.
+ b } over R, or C,
b) Do the same problem for the Bore1 group of all upper-triangular matrices
9. Show that formula (2.10) defines a representation T of G. The induced representation is continuous/unitary, provided u is continuous/unitary and X has a G-invariant measure. Another simple (functorial) property of ind, ind(ul @ u z )= ind(ul) @ ind(u2).
10. Consider convolutions on the unit circle G = 1, 2rr
~ , p=] f * h = J f ( t - e) q e ) de, on L2-space ( h E L*), and show
0
i) L'-functions f yield compact operators R , ii) L2-functions f yield Hilbert-Schmidt operators R , iii) Cm-smooth functions (rn 2 2) yield trace-class operators. Hint: use Fourier expansion of f = C a m e i m t ; and some basic facts of Fourier analysis on 1: (a) f E L'*{a,}-O, as k-m; ( b ) f E L 2 e { a , } E e2; and (c) f e e m 3 a, = O(k-m), as k-m. We remind the reader that compactness for diagonalizable (self-adjoint) operators means Hilberl-Schmidt means C I A, I < 00, and trace-class is that their eigenvalues A,+O, equivalent to C I A, I < 00 (see Appendix B). 11. Show that convolutions f * h with f E em on groups R", T", take any LP-space back into Cm (interchange integration J ...d y , and differentiation 8:). The same argument works on any Lie group and its homogeneous spaces, when integration and differentiation are G-invariant!
$1.3. Irreducibility and decomposition
37
51.3. Irreducibility and decoinp~tion. Section 1.3 studies irreducibility and decomposition of representations. We establish basic irreducibility tests: Schur’s Lemma and Burnside’s Theorem, discuss direct sum/integral decompositions, equivalence, intertwining spaces, and outline some basic problems of the representation theory.
3.1. Irreducibility. One of the main goals in the representation theory is to
analyze the structure of any representation T in terms of its simple constituents, called
irreducible representations. In this section we shall discuss some general features of irreducibility and decomposition.
A representations T of grouplalgebra in space Y is called irreducible, if operators { T g }have no joint nontrivial invariant s~bspaces’~:Y~ c Y; T , I Yo c Yo, for all g E G,+ Yo = {0} or Y. An
invariant subspace Y 0 c Y reduces
T , so one can talk about a
subrepresentation (restriction) T“ = T I Yo, and the factor-representations on the quotient-spaces Y/Yo.Irreducible representations play the role of elementary building blocks, in the decomposition of an arbitrary T . In this respect they resemble eigenvalues of linear operators, or characters (joint eigenvalues) of commuting families of operators. Two general results provide useful irreducibility tests: Schur ’s Lemma and Bumside (von Neumann) Theorem. Each has a purely “algebraic” (finite-dimensional) version, and a “topological” (unitary) version. Since the arguments are very similar in both cases, we shall discuss them in parallel. Given a representation T , we introduce its
commutator algebra:
Com(T)={A:AT, = T,A, for all g E G} - all operators that commute with T . Schur’s Lemma: (i) An irreducible finite-D representation T in complex space Y has scalar commutator algebra, Com(T)= {XI}. (ii) A unitary representation T is irreducible, iff Com(T) is scalar.
Proof: (i) If Com(T) has a nonscalar operator Q, then any (nontrivial) of Q, is obviously invariant under T , which means
eigensubspace E , = {(:Qt = A(} reducibility.
(ii) The unitary version requires some basic facts of the spectral theory of unitary and selfadjoint operators, described in Appendix A. Let us remark that algebra -In the topological context (Banach/Hilbert spaces subspaces {Yo}.
Y) one usually considers closed invariant
81.3. Irreducibility and decomposition
38
Com(T) is closed under conjugation Q+Q* (adjoint operator to Q ) . So a nontrivial algebra Com(T)# {XI} has a nonscalar selfadjoint operator Q. The role of the eigenspaces in part (i) will be played now by the spectral
subspaces15 {E(A)} of Q. Each subspace E ( A ) is known to be invariant under the commutator algebra of Q , since the corresponding spectral projections { P ( A ) } are approximated by “functions (polynomials) of Q . The argument is easily completed now. A nontrivial Q = Q* in Com(T) would yield a nontrivial T-invariant spectral subspace E(A). So irreducibility of T implies Com(T) = {XI}. Conversely, if T has an invariant subspace Yo, then the orthogonal projection P: T-+Y,,
gives a nonscalar
element of Com(T),QED. Let us remark that for general (nonunitary) T , Schur’s Lemma gives only a necessary condition of irreducibility. It is easy to find examples of reducible T with scalar commutator, for instance, all upper-triangular matrices in Cn (problem 3). An immediate corollary of Schur’s Lemma is Corollary: A n irreducible finite-D or unitary representation of a commutative
group/algebra is one-dimensional,
T, = x ( g ) - a character of G.
Remark: There is a version of Schur’s Lemma for representations in real spaces Y. Here irreducibility of T does not always imply Com(T) to be scalar. But the commutator algebra must contain nonsingular matrices only, hence it makes a division
algebra. There are only 3 such beasts: reals W (the corresponding T a r e called real-type); complex numbers 43 (complez-type T), and quaternions Q (puaternionic-type T). All 3 cases can be characterized in terms of complexification of T , i.e. the family of operators {TC(E+iq)= T[+i(Tq):(,qE V}. Then “real type” yields an irreducible representation TC= T @ C; “complex-type” - a pair of inequivalent (conjugate) representations: TCN T I @ T while quaternionic-type results in a pair of equivalent representations:
TCN T $ T . A simple illustration of this statement is given by 3 algebras, Iw;C;$, themselves, of a selfadjoint operator Q in Hilbert space 36, consists of { A E R: so that (AZ-Q) is not boundedly invertible). Spec(Q) is always a c l w d subset of R, and operators with 1-pointed spectrum, spec(&) = {A), are known t o be scalar, Q = XI. Furthermore, any closed subset A C R has an associated spectral subspace E ( A ) and the corresponding spectral projection, P ( A ) : 36-+E(A).Spaces { E ( A ) }are invariant under Q, and have spec(Q I E ( A ) )c A n Z . For operators with discrete spectrum { A k } , space E ( A ) = @ E(Aj),consists of all eigensubspaces with A j E A . An easy way to obtain { E ( A ) ) is to use the canonical model of a selfadjoint operator Q,namely, a multiplication: f(A)-,Af(A), on the space of square-integrable (scalar or vector) functions on C = spec(&), f E Z*(Z;dp). Then E ( A ) consists of all L2-functions vanishing outside A (see Appendix A). -Spectrum
$1.3. Irreducibility and decomposition when realized by matrices (problem ll),
R = { [ a ] ; a E W}; 43 =
{;[
] a , b E R}; and
Q=
39
{b
: ] u , b E C}.
Burnside (von Neumann) Theorem: (i) A finite-D representation T in space
(3.1)
Y is
irreducible if and only if t he algebra A = A(T), spanned by T, coincides with the full operator/matriz algebra L(Y) N Matd, where d = d ( T ) = d i m Y , called the degree of T. (ii) A unitary (infinite-D) representation T in Hilbert space 36 is irreducible, i f and only af the closure of A ( T ) in the weak operator topology16 coincides with the algebra of all bounded operators %&).
Sufficiency follows from the irreducibility of the full matrix/operator algebra Mat,/%(%). The proof of necessity: “irreducibility” ==+ ‘@ll-operator algebra”, exploits a
L‘bootstrap’’argument (see problems 7,8). In the m-D case one could not expect the purely algebraic (Burnside’s) version of the Theorem to hold, as there are many transitive operator-algebras, different from %(%), like all finite-rank operators 5,or compact operators R in 36. But the weak (operator) closure of both coincides with
%(36)! A Theorem of von Neumann extends the algebraic result to symmetric transitive subalgebras A c %(%), by combining LLbootstrap”with some standard functional analysis, to show that A must be weakly dense in ’3(36).
3.2. Decomposition. If a finite-D representation T on space 36 is reducible, one X0,so that T I Mo becomes irreducible.
can always find a minimal invariant subspace
Hence, any such representation can be brought into a ‘quasi-triangular form” with irreducible diagonal blocks. But for unitary T each invariant subspace 36, has also an Tinvariant orthogonal complement 36, = 36 8 Xo. Therefore, quasi-triangular reduction for unitary T amounts to LLquasi-digonalizationnor complete decomposition. Proposition 1: A finite-dimensional unitary representation can be decomposed into
the direct orthogonal s u m of irreducible components: T = 9 Ti, i.e. 36 = @ Mi3 3 direct s u m of invariant subspaces, and T3 = T I Xj. In other words a unitary finite-D representation is completely reducible. Thus irreducible representations provide the “building blocks” of which any T is made. In this sense Proposition 1 extends the well-known diagonalization of a unitary/self-adjoint
16Weakly dense algebra A means that any operator B E %(%) can be approximated by r) E %}, equivalently, by all traceoperators { A E A } , relative to a sequence of seminorms { ( A t I r)):t, class operators K , identified with functionals on %(%), via the natural pairing K : A-.tr(KA). In other words the distance between operators { A ; B } is measured by n-tuples of vectors {ti;qi: 1 5 j 5 m}, or trace-class operators { K } : { ( ( A- B ) t j I qi)} or {tr K ( A - B ) } .
51.3. irreducibility and decomposition
40
operator (Appendix A). There exists a continuous version of Proposition 1, called the direct integral decomposition of a unitary representation in m-D case. We briefly introduced it in the previous section and will discussed later in 56.1. Let us remark that the decomposition of Proposition 1 is not unique in general, due to possible multiplicities of irreducible components of T , the same way as multiple eigenvectors of a matrix are not unique. We shall describe a canonical (unique!) decomposition of T into the direct sum of multiples of irreducibles, ?r@m,where m rn(?r;T)denotes the multiplicity of
?r
in
T.
Theorem 2 (primary decomposition): A unitary finite-D representation T in space 36 is uniquely decomposed into the direct sum of multiples of irreducible
representations, i.e. % = @36, - direct sum of invariant subspaces, so that k T/%, N x k @ mk - a multiple of an irreducible ?rk, and
(3.2) with mk = m ( x k ; T )- multiplicity of 'a in T . Multiples of irreducibles Irk}are often called primary components of T , spaces { % k } are primary subspaces, and
(3.2) is referred to as a primary decomposition of T .
The main difference between Theorem 2 and Proposition 1 is that the primary subspaces Xj = %(d) are uniquely determined, whereas within each Nj there are many ways to break T
I Xj into the sum of irreducibles. So the primary subspaces represent
noncommutative analog of the eigenspaces of a linear operator. There are different ways to establish the primary decomposition. Some depend on specific structures of groups in question. For instance, for finite and compact groups G (3.2) could be derived via irreducible characters of G and the associated primary projections (see
chapters 3-5). Another method is based on central projections of the commutator algebra
Com(T). Here we shall exploits matrix entries of T: given a pair of vectors ( , q E 36, we def ( 9 ) = (Tg( I q ) defines a function on G. For each vector ( E 36
define the matrix entry t
€v
we introduce a cyclic subspace, Y (() = Span{Tg(: g E G}. Obviously, under T. Any vector q E 36 defines a linear map,
Y(() is invariant
9:, (-+(Tg(I q ) , that takes V(() into a
subspace '3 of functions on G (matrix-entries), invariant under right translations, f ( z g ) E GJ, for any f E '3. In fact, Qq intertwines the subrepresentation T regular representation R 1'3,
RgQq = QqTg,for all g E G.
I Y(()
and the
a
51.3. Irreducibility and decomposition
41
This shows, in particular, that any irreducible T can be embedded into the regular representation on a suitable function space. We say that vector (E 36 transforms according t o an irreducible
A,
if all matrix-entries { t
span T ( A ) of matrix-entries of
A,
(2): r)
€r)
a R-invariant subspace of L(G). An equivalent
statement: all irreducible constituents of T I Y(() are copies of claim that T I 36(()
E 36) belong to the linear
T , 80
T I Y(() u T 8 rn. We
u A has multiplicity 1. Indeed, by Schur’s Lemma one can show that
the only cyclic subrepresentations of
S = A 8 m are copies of
A.
Now we are able to
identify primary subspaces { % ( A ) } of Theorem 2: % ( A ) = {( E 36: transformed according T}.
Spaces % ( A ) are obviously T-invariant, their intersection % ( T ) ~ ) % ( A ’ ) = 0, if
# A‘,
and their sum, @ % ( A ) , spans 36. Thus we get a unique (canonical!) primary
to 7r
decomposition of T, QED.
Proposition 1 and Theorem 2 naturally lead to the following problems:
(I) Construct and classify all irreducible representations of the group/algebra G. (11) Decompose the given representation (regular, tensor-product, etc.) into the sum/integral of irreducibles. These are indeed two basic questions in representation theory, that will be addressed at length in chapters 3-7. We shall construct, classify and analyze irreducible representations for many different classes of groups: 0
finite and general compact groups (chapter 3);
0
classical compact Lie groups S q 3 ) ;Sl.42) (chapter 4), as well as their higher-D
analogs: SU(n);S q n ) ,... (chapter 5); 0
semidirect products, including the Heisenberg group:
W,
N
R D Cn, the
Euclidian motion groups: E I , = R” D S q n ) , and affine groups (chapter 6); 0
the celebrated SL, in its various incarnations: conformal SU(1; 1); Lorentz
Sql;2) and S q l ;3), and also the Poincare group: P,
N
R4D S q 1 ; 3 ) in chapter 7.
Once the list of irreducible components of G is found (usually denoted by
G,and
called the dual object of G ) , one can turn to the decomposition Problem (11) for specific classes of representations. The most interesting examples include regular and induced
representation R on G , or homogeneous spaces X = H\G. We shall see that in the compact case ( X or G), the decomposition is always “discrete”,
-l F CB
C 3 m ( T ) - direct sum of irreducibles.
For non-compact groups/spaces the decomposition is typically “continuous”
51.3.Irreducibility and decomposition
42 (direct integral),
or more generally a combination of the "discrete" and the "continuous" parts, as is the
case with SL,(W) (chapter 7). The resulting measure dp on G (whether discrete or continuous) is called the Plancherel measure of R. It will be one of principal objects of our study. The classification problem could easily become "unmanageable", if one insists on "all irreducible representations" of G, since any T can be deformed in infinitely many ways by conjugations with invertible operators,
Q : T ~ +T$ = Q - ~ T ~ Q . Such deformations, T-T', however, retain many essential features of T (e.g. spectrum), and could be regarded trivial. Thus we are lead to the notion of equivalence: a pair of representations, T and S in spaces 36 and W , are called equivalent, if they are conjugated (intertwined)by an invertible linear map Q: %+W,
The space of all intertwining operators is denoted by Int(T;S), its dimension (if finite) is called an intertwining (branching) number m(T;S).Operators that intertwine T with itself form the commutator algebra of T , I n t ( T ; T )= Com(T), and other spaces Int(S;T ) acquire the structure of {Com(T);Com(S)}- bimodule, i.e. Com(T) acts on Int(S;T )by the left multiplication while Com(S)by the right one. Problem (I) usually refers to classification of equivalence classes of irreducible representations, rather than individual {T's}. In some cases, however, equivalent representations T and S may be given in different realizations. Then one is interested in specific intertwining operators that implement the equivalence. Let us mention yet another useful application of Schur's Lemma. Proposition 3: For any pair of irreducible representations T and S, space Int(T;S ) is either 1-D,if T and S are equivalent, or (0). So the intertwining number m(T;S ) = 1 or 0. Indeed, for any
Q E Znt(T;S), its range %(Q)and kernel N(Q) are invariant subspaces
of S and 7'. So space Znt of an irreducible finite-D pair ( T; S ) consists of invertible operators Q. But QIQ;'ECom(S)
for any pair Q I Q z in Znl(T;S), so by Schur's
51.3. Irreducibility and decomposition
43
Lemma, QIQ;’ = XI, i.e. Q1 = XQ2. The unitary version requires a slight modification. We observe that two spaces Znt(T;S) and Znt(S;T)are naturally conjugate one to the other by taking the adjoint Q-Q*. X-’/*Q
By Schur’s Lemma Q*Q= XI, hence operator
is unitary and invertible, and the rest of the argument proceeds as above.
Our final result in this section gives irreducible unitary representations of the direct product group G = H x K in terms of its factors.
Theorem 4 Each irreducible representation
of G = H x K is decomposed into
?r
the tensor productl’l of two irreducible representations of H and K ,
?r
= u @ 7,
where u E 8, 7 E I?. Conversely, the product of two irreducibles u E and 7 E k , is an irreducible representation of G, T = u @ , ~G. E So the dual object of G = H x I( also factors into the product, A
1Proof:The second statement follows directly from the Burnside’s Theorem and the known facts of multilinear algebra about tensor-products of operator/matrix algebras (problem
1%
Mat,, 8 Matm N Mat,,;
(3.3)
or in terms of operator algebras
%(X)8 %(Y)
2: %(36
T o prove the first statement we take an irreducible first factor
A
8 r). A
(3.4)
E G in space 36, restrict it on the
I H, and write the primary decomposition of Theorem 2, r1HN
$Uk@mk; k
where ak E if, and mk- the multiplicity of uk in
A
IH.
so the representation space
36 N $36, - direct sum of primary subpaces of H. Since each 36, is invariant under the k
commutator Com(r I H), hence under K , it follows that 36 has only one primary component, 36 N Kk,and
A
I H N u 8 I,,,.
In other words, space 36 is a direct sum of { m }
copies of the “irreducible space” V of u,
36 N Y a3 ...$ Y N Y 8 cm.
x
Any operator A in 36 can be represented by a block-matrix A = (aij), aij E Mat(Yi;‘Tj). We are interested in operators A E Corn(* I H). By Proposition 3 all “matrix entries” {aij) of such A must be scalars: aij = yijZ; yijE C. When restricted on subgroup K,
matrices (yij(k)) clearly yield a representation of K. So space 36 is decomposed into the product
Y 8 C d , as well as operators {rg), g = (h;k) E G, rg = u,, 8 rk. Obviously, for A
17The reader should not confuse the tensor product u 8 y of Theorem 4 with the Kronecker product of 51.2. The latter is obtained, when the tensor product u@y of the product-group G x G is restricted on the diagonal subgroup ( ( 9 ;g)) N G C G x G .
51.3. Irreducibility and decomposition
44
t o be irreducible, y must also be irreducible, which completes the proof!
Problems and Exercises. GL, and u(n) on C" are irreducible.
1. Show that the natural actions of
2. Show that an eigenspace E A or a root subspace ET = {<: (T-Al)m( = 0) of a matrix T in C" are invariant under C o m ( T ) .
3. Show: the group B , of all upper/lower triangular matrices in R"; C" has trivial (scalar) commutator, though B, is clearly reducible (This exercise shows Schur's Lemma t o provide only a necessary condition for general non-unitary representations!).
X = {1,2, ...n} induces the regular representation R on the function-space: e ( X ) N C", R , f ( j ) = f( j ' ) .
4. A symmetric group W, (permutations of n objects) acting on
(i) Show that e ( X ) is the sum of two invariant subspaces: f(j ) = 0). constant function fo( j ) = 1, and 36, = 32f:
C
X0 = {Afo} - spanned by the
(ii) Show that subrepresentations ro = R 1 360-trivial, and r1 = R I M1-irreducible (Use Schur's Lemma for rl, represent any operator Q on e ( X ) by a kernel/matrix [ q j k ] ,
QfW =
qjkfk).
5*.(i) Show that the adjoint representation Ad,(A) = g-'Ag, of the group G = SL, on the space of all traceless matrices: Jb, = { A E Mat,,: t r A = 0) is irreducible (Hint: any Q E C o m ( A d ) is uniquely determined by its.values on the subspace '3 C A, of diagonal matrices, moreover Q:P+P, commutes with the natural action of W, on 9. Use problem 3!).
r''..]
(ii) Prove the same result for the special unitary group SU(n).
6. Show that a pair of matrices: A =
)t n
(diagonal), B =[1
1) (cyclic
permutation), generate the full matrix algebra in C", provided all {Aj} are different.
7. Let A be an operator-algebra in space Y, vector ( E Y. We denote by ) =I(() the annihilator of in A, ) = { A : A< = 0) - a left ideal A, ( A ) C I), and by Yo = Yo(<) - a cyclic subspace of (, Yo = {A<: A E A).
<
(i) Show that Yo is A-invariant, and the restriction of A on Yo is equivalent t o the regular representation of A on factor-space A / ) , R,(B) = A E ( m o d 8). (ii) If )(<) c )(9), then there exists a linear map Q: Yo(<)+Yo(q), that intertwines two representations of A on Yo(() and Yo(q). (Set: Q [ A < ]= A?, for all A E -4, and check that Q is well defined!).
8. Prove the Burnside's Theorem by the "bootstrap argument", as outlined below. It is convenient t o use the following terminology an operator-algebra A is transitive (irreducible!), if for any vectors ( # 0 and q there exists A E A, that takes A: <+q; algebra A is double-transitive, if for any pair (t1;t2) of linearly independent vectors and any pair (q1;q2) there exists A E A, so that At l = q l , A t Z = q2. Similarly, a k-transitive A takes any linearly independent k-tuple k-tuple (ql; ...qk), i.e. A t i = q3., j = 1; .A.
(F1;...(k) into any
81.3. Irreducibility and decomposition
45
Clearly, a n-transitive algebra in n-space coincides with the full matrix algebra! The idea of ”bootstrap” is to show that “ t r a n s i t i v i t y ” ~“2-lransilivity”* transitivity“.
...“k-
Step 1’: Show “transitivej2-transitive”. For any vector [ E Y denote by 3 = I ( [ ) the annihilalor of 5 in A (problem 6). Given an operator-algebra A and a left ideal 1, consider 1-cyclic subspaces Yo = Yo(q) = {Bv: BE 7 E Y. Each of them is A-invariant, A) c ). So for transitive (irreducible) A, each space Yo must be either 0 or the entire Y.
a},
Pick a linearly independent pair (tl;t2), and consider the ideal ) = )(tl) of t1,and the C )-cyclic subspace, Yo = Y0(t2). Use problem 6 to show: if Yo = {0}, i.e. extends to an intertwining operator Q E Com(A), impossible for then map Q: irreducible A (Schur’s Lemma)! Conclude, that Yo = Y,i.e. for any q E Y there exists A E A, Show that the latter implies double-transitivity of A.
50
that Atl = 0, A t 2 = q.
Step 2”. Demonstrate in a similar fashion that “k-transitive*(k+l)-transitive”, completes the argument of Burnside’s Theorem.
and
The same “bootstrap” applies to transitive algebras A in (oo-D) Hilbert spaces, to show that A is k-transitive for all k! Then fairly standard functional analysis implies that A is weakly closed in %(36) (von Neumann). 9.
Prove: (i) the commutator algebra of a primary representation T N ~ 8 m , Com(T) N Mat,,,, and any Q E Com(T) factors into the product I, 8 Q ,in the tensorproduct decomposition of space YT N Y, 8 Cm (rn = multiplicity of A in T!). (ii) If representation T N A’ 8 mj - is completely reducible (direct sum of primary components), then the algeura spanned by T,A ( T )N @ Matd , @ I m( .d j = d e g d ; m3. = multiplicity of A’ in T), while the commutator algebja
’
Com(T)
2:
e,3 I d j 8 Mat,
’
..
1
(iii) The intertwining number of a pair of representations m(T;S) is defined as d i m I n t ( T ; S ) .Show, if T N @ ~ @ m ( r ; Tand ) , S2: @ n 8 m ( n ; S ) then ,
10. Show that the tensor product of matdx/operator algebras is equivalent to the full matrix algebra on the tensor-product space: Mat, 8 Mat,,, u Mat,. ,,,; and %(X) 8 %(Y) N %(36 8 Y) (Use rank-one operators and vectors). 11. Check formulae (3.1) for C and Q in terms of real/complex matrices. Show that complexification of C yields the direct sum of two conjugate representations: z + ( z ; 5 ), z E C; while complexified Q yields two copies of Q itself.
$1.4. Lie groups and algebras; The infinitesimal method
46
$1.4. Lie groups and algebras; The infinitesimal method. Here we shall introduce two principal players in the Lie theory: Lie algebras and Lie groups. Although different in appearance both are intimately linked through the ezponential map, and its inverse log. Two maps allow to express the local group structure (multiplication) through the Lie bracket (Campbell-Hausdorff formula). We discuss the basic classes of Lie algebras: solvable; nilpotent; simple, semisimple, and state the decompition result (Levi-Malcev). The section is concluded with an infinitesimal method, that allows to reduce group representations to those of its Lie algebra, and thus provides a powerful tool in their analysis.
4.1. Lie groups and Lie algebras. We remind the reader that Lie groups are
smooth manifolds G with differential, (sometimes analytic, or algebraic) group structure. This means that the group operations: ( ~ , y ) + z y and z-u-’, are differentiable (analytic or algebraic). The most common examples of Lie groups are subgroups of matrices, G c Mat,,, like the general and special linear groups: GL,, and SL,; orthogonal, unitary, symplectic, and other groups described in 51.1. Let us remark, however, that not all Lie groups can be realized by finite-D matrices. One such example is a simply connected cower of Sf,, G = SL, (examples in 4.2 below). The class of Lie groups admits all natural operations, performed with groups in general, like taking closed subgroups, factor-groups, direct and semidirect products, etc. All those result in new Lie groups. N
Lie algebras can be viewed as “linearizations” of Lie groups, although formal definition does not allude to an underlying group structure. Formally, Lie algebra 0 is defined as a vector space equipped with the Lie bracket (product) [X;Y], a bilinear map 0 x 0 + 0, that satisfies: (i) skew symmetry: [X;Y] = - [Y;X] for all X,Y.
(ii)Jacobi identity: [X;[Y;Z]] + [Y;[Z;X]] + [Z;[X,Y]]= 0, for all triples X, Y,Z. We can rephrase (i-ii) is in terms of the so called adjoint action of Lie algebra on itself, adx(Y) = [X;Y]; adX:@+@. Then, (i) adX(Y) =
.-
ady(X), for all X, Y
(ii)adlX;yl = [ad,; ady] = adxady - adyadx,
(4.1)
51.4. Lie groups and algebras; The infinitesimal method
47
In particular, the Jacobi identity means that map, X+ad,, is a representation of the Lie algebra 6 by linear transformations over (5. It can also be written as ad,[Y;Z] = [adx(Y);Z] -t- [Y; adx(Z)], for all Y,Z, (4.2) which means that operators {adx} are derivations of the Lie algebra. The examples of Lie algebras are abundant. Any associative algebra (e.g. matriz/operator algebra Mat,) turns into the Lie algebra with respect to the commutator bracket: [x,y ] = xy - yx. Considered as a Lie algebra it is usually denoted A n ) . a
Any vector space A with the trivial bracket, [x;y]= 0, gives a commutative Lie
algebra.
*Important examples of Lie algebras are given by vector fields on a manifold {X=Cui(x)aj} (in local coordinates { x l ;...z,}), with the standard Lie bracket of vector fields, [X;Y] = X Y - Y X =
x ( x u j i 3 j b j - b j a j a i ) a i = C ( a x [ b i ] - d y [ a j ] ) a i ; (4.3) i
i
j
Here aX;ay denote directional (Lie) derivatives along fields X,Y. Those include the algebra 9(A)of all vector fields on A, or its subalgebras of fields, preserving certain structures on A: volume, symplectic (chapter 8), etc. Many examples arise as Lie subalgebras of Mat,, and will be examined below. There is a close relationship between Lie groups and algebras.
Theorem 1: (i) Any Lie group G has an associated Lie algebra (5, which consists of infinitesimal generators of all one-parameter subgroups of G. (ii) Conversely, given a Lie algebra with algebra (5.
(5
one can construct a ‘’local” Lie group18 G
(iii) A local Lie group can be extended to a global group in many different ways. Namely, there exists a unique simply-connected extension G, and any other Lie group G with Lie algebra (5 is isomorphic to a factor-group G/r, G modulo a discrete subgroup of the center
r c Z(G).
Theorem 1 summarizes the basic facts of the Lie theory. We shall not provide complete details of the proof, but rather demonstrate different statements in several 18By local Lie group we mean an open neighborhood of the identity {e} in a Lie group, i.e. a manifold, where the group operations: z.y, and z-*,are defined only locally.
$1.4. Lie groups and algebras; The infinitesimal method
48
important cases. The easiest case to consider are matriz groups and algebras, i.e. subgroups and subalgebras of Mat,. We shall start with infinitesimal generators. It is well known that any (continuous) one-parameter subgroup { U , } of Mat, is generated by a matrix A, in the sense that
U , = etA; where A = dt Such A is naturally to call a generator of subgroup of Mat,.
Lo.
(4.4)
{U,}.Let now G be a closed (Lie)
Proposition 2: (i) The set of infinitesimal generators { A } of G coincides with the
tangent space Te(G)at the identity.
(ii) The commutator of two generator A , B E Te(G) is also a generator, [A;B ] E Te(G),so tangent space forms a Lie subalgebra (5 c Mat,. Proof: (i) Obviously, generators {A} C T,, and the latter forms a subspace of Mat,. To show that {A} is a subspace in T, we pick two subgroups of G { U , = etA], {V, = efB}, generated by A , B E T, according to (4.4), then a subgroup generated by A
+ B, can be
approximated by an “amalgamation” of U , V,
= hence { e t ( A t B ) }
liyi(ezp(;A)ezp(;B))n,
(4.5)
(I G (problem 2). So A+B E T,, and clearly XA E T, for any scalar A.
Conversely, if matrix A is tangent to G, i.e. tangent t o a curve g ( t ) c G, A = g’(O), then we can easily construct a I-parameter subgroup tangent to A, by setting,
u, = g~&(t/4)n.
(4.6)
The generator of U , is A (problem 2), thus we show that generators {A} span the entire space T,! (ii) It suffices to exhibit [ A ; B ]as a tangent vector of some curve in G. Once again we take a pair of subgroups U t = e t A , V, = e t B , and consider the group commutator g ( t ) = cI;’V;’U,V,.
Expanding U and V in the Taylor series, multiplying and collecting
terms we get g(t) = I
+ t2[A;B ] + fl (t3).
Neglecting the cubic and higher order terms the commutator becomes the tangent vector a t { e = I},
[ A ; B ]= lim t-+O
g(,) - I ; QED. t2
We have shown that the Lie algebra of a matrix group G can be identified with the “set of infinitesimal generators’’ = “tangent space of G at { e } ” . A similar description
51.4. Lie groups and algebras; The infinitesimal method
49
could be given for all (non-matrix) Lie groups, namely 6 N T,(G). Lie algebra 6 can also be characterized, as the set (subalgebra) of left/right invariant vector fields on G. Group G acts on itself by left/right translations g: z+g-lz left/right invariant vector fields, {X =Caj(z)aj
- in
(or 2-zg),
gives rise to
local coordinates (zl;...z,) on G}.
Lefl-invariance of X means that the transformed field”:
XO(z)
= g‘[X]og-’
= X,
where g’ denotes the Jacobian map of the left translate g: z+gz. Left-invariant vector fields form a subalgebra
(s
in the algebra of all vector fields 9(G)with the Lie bracket
X can be identified with its value at { e } , a tangent for any point g E G. So 19 N T,2: {left/right invariant fields
(4.3). Any left-invariant vector field
vector
t E T,,X(g)
= g’[(],
on g}.
4.2. The exp and log maps. The correspondence between Lie groups and Lie
algebras is given by two maps: exponential and tog, exp: 6+G; log: G+@.
Map exp: 6 NT,+G, sends a tangent vector/line { t t } into a one-parameter subgroup { e x p t t } of G, tangent to
<
at {e}20. Map d(<)=exp(t)
is a local
4;
= I!), and its inverse is called log. So exp takes a neighborhood of {0} in 6 onto a neighborhood of identity { e } in G, while log does vice diffeomorphism (since its Jacobian
versa. The easiest way to introduce exp and log is for matrix groups and algebras. Here two maps are given by their Taylor-expansions,
Exponential map on the full matrix algebra Mat,, = gl,, takes it onto the full linear
group
GL, = { z E Mat,,:
det(z) # 0).
Clearly, exp on
is a local
Mat,
diffeomorphism (near X = 0), but globally it may not be 1-1. For instance, all complex matrices { X } with purely imaginary integer eigenvalues {Aj = 27rim}, are taken to the identity by exp. We shall list now a few basic and familiar properties of exp and log. W -e remind the reader that a diffeomorphism (coordinate change) y = 4(z) on a manifold transforms vector field X = C a .a. into a field Y = X4 = (#[XI)o d-’, with components J Z bi = C4:jaj, where 4’ denotes a Jacobian (matrix) of 6. 20Such notion of exp arises in differential geometry and is closely connected with geodesics and parallel transport (Appendix C). On Lie groups the parallel transport is furnished by the left/right group multiplication, z-g-lz, g E G. The latter gives rise a notion of parallel field (along a curve ~ ( t ) )the , geodesic path y (whose tangent field y‘(t) is parallel), and exp:t< E Ta-y (t) - a geodesic initiated at {a} in the direction (!
<
51.4. Lie groups and algebras; The infinitesimal method
50
Lemma 3 (i) If matrices A and B commute, AB = BA, then particular {e"A}t is a one-parameter subgroup of Mat,.
= eAeB, in
(ii) det(eA) = elrA. (iii) exp preserues complez conjugation, transposition and taking adjoint matriz: exp(2) = 3, exp(=A)= TexpA; exp(A*) = (expA)*.
The proof is fairly straightforward (problem 1). Using Lemma 3 we are able to characterize Lie algebras of many classical Lie groups (orthogonal, unitary, symplectic, etc.), listed in $1.1. For instance, the expmap takes a Lie algebra of n x n hermitian antisymmetric matrices, denoted u(n), or real antisymmetric matrices, denoted o(n), into the unitary and orthogonal groups: U(n) and q n ) . Special (traceless) Lie algebras { X :trX = 0) are taken into special { detg = 1) Lie groups, like 0
d ( n )= {A:trA = 0) (all traceless matrices) goes into SL,
0
so(.)
0
su(n) = {A* = - A; t r A = 0 ) (unitary traceless) goes into SU(n).
= {TA = - A; tr A = 0 ) (orthogonal traceless) goes into S q n )
In fact, in all cases the exp map is locally onto, since by Lemma 3 generators of groups SL,; U(n); q n ) , etc., must traceless, unitary, orthogonal. We have already shown that Lie algebras are obtained as tangent spaces at {e} of Lie groups G, or as infinitesimal generators of one-parameter subgroups of G. Going in the opposite direction one would like to construct Lie group G, starting from Lie algebra (5. As we stated earlier (Theorem 1) the correspondence C k G is not 1-1, as many Lie groups could have the same Lie algebra. Although all such groups are locally isomorphic, their global structure may be different. The local group-structure, however, is completely determined by the Lie algebra formula bracket, via the so called Campbell-Haudorff-Dynkin
Here the inner sum extends over all tuples of integers { p l , q l ;... p,,q,) with the property: pj+qi 1 1, and [AP1[Bql ...Bqm]] denotes the iterated commutator [A[A...[B...[B...]]]]. The Campbell-Hausdorff series can be shown to converge for all Lie algebra elements (matrices) of norm 11 Alk llBll< and thus defines a local group
a,
31.4. Lie groups and algebras; The infinitesimal method
51
structure in a neighborhood of {0} in Q. Formula (4.8) explains, why the exponential sends any Lie subalgebra 0 c Mat, into a Lie group. Indeed, e x p A e x p B = e x p ( A t B + i [ A ; B ] t ...), so e x p o
is closed under the matrix multiplication.
The local Lie group uniquely extends to a global simply connected Lie group G (not necessarily matrix group) with the Lie algebra (5. Any other Lie group G with algebra (5 is then isomorphic to G/I', where is a discrete subgroup of the center of G, identified with the fundamental group21 of G . So all locally isomorphic Lie groups {G} with the given Lie algebra are in 1-1 correspondence with discrete subgroups r c Z(G ).
r
N
N
We shall illustrate the relations between Lie algebras Q and groups G; G and with a few simple examples.
r
1) The easiest of the sort is the abelian Lie algebra 8 = W", which gives rise to several Lie groups: G=Rn;or G=T"=Wn/Z"orG=Bm$W"-m,for a n y m = 1 , 2 , ... corresponding to various discrete subgroups N Z", or Zm, of G.
r
2) Another example is the unitary group SU(2) (simply connected), which forms a 2-fold cover of S q 3 ) (chapter 3). All other (higher-D) orthogonal groups S q n ) also have 2-fold simply connected covers, called spinor groups Spin(n), S q n ) N Spin(n)/Z2. 3) Group G = SL,(W) is not simply connected, and its simply connected cover cannot be realized as a matrix group!
N
G
Indeed, group G is topologically equivalent to the product R2 x 1 ' (this follows from realization of the Poincare-Lobachevski half-plane H as quotient SL2/So(2), and could also be established by the Iwasawa decomposition, as explained in $7.1, chapter 7). So the
"We remind the reader that a fundamental group of manifold W consists of the hornotopy equivalence classes of closed path (loops) in W . Two loops 7 1 ; 7 2 in W are equivalent, if rl can be continuously deformed into yz inside W . Any pair {yl;y2)of closed path (loops) passing through a fixed point z,,E W can be formally multiplied by combining them into a single path ~ 1 ~ ("71'' 7 2 followed by "y2"), and any 7 can be inverted, by traversing it in the opposite direction. Such multiplication and inversion are easily verified to respect the (homotopy) equivalence on the space of loops passing through a fixed point. So equivalence classes of (7) acquire the group structure, the identity being made of all path contractible to {zo}. The resulting fundamental group of A, = al(W), is a discrete group, that carries some important topological information a b u t the manifold. It acts freely (by the so called deck Iran_sformations) on a simply connected cover W of W , so the latter is isomorphic to a quotient-space &IT. The fundamental group of a E e group G is c G , hence belongs to the center of G (problem a), identified-with a normal discrete subgroup and G rz G
/r.
r
r
51.4. Lie groups and algebras; The infinitesimal method
52
fundamental group
r of SL, is isomorphic to Z,and its simply connected cover G has
Z. If SL, were realized by matrices we would get a finite-D faithful (one-te one) representation of SL, in C”.The detailed analysis of finite-D representations of SL, in chapter 4 shows that any such T has { d e t T , = 1; for all g E C}, and the same holds for representations of G (in other words T m a p S L p S L , ! ) . Furthermore, any center Z N
representation T of G or
2:
is completely reducible ( u “direct sum of irreducibles a”).
Since center of any group is represented by scalars { X I } under an irreducible (Schur’s Lemma), and for
SL,
i~
EG
these scalars must be m-th roots of the identity { w :
wm = 1 ) (due to d e t r , = I), it follows that center Z(G) is mapped by T into a finite
group of matrices with diagonal entries { w } . But Z N Z is infinite, so some g to { I } , which means T can not be exact. In other words,
Z must go
Sf, can not be realized
as
matrix group!
Let us make a few comments concerning the explog correspondence: (5-G. It clearly, preserves all general (functorial) properties of groups and algebras. So a
closed (connected) subgroups of G u subalgebras of
(5,
normal subgroups H c G u ideals @ c (5 - subalgebras invariance under the adjoint action of (5: [@;@I c @, or adx(@) c 8, for all X E (5).
Z(G), commutant G, = [G;G] (a subgroup, generated by all group commutators {g-lh-lgh: g , h E G}),etc., have proper counterparts in the Lie algebra: central ideal 3 = 3((5), commutant a other “natural mbgroups”, like center Z =
(5,
= [(5;(5] = S p a n { [ X Y ] X : , Y E (5}, etc.
Also any decomposition of group in the direct, semidirect product, yields a similar decomposition of its Lie algebra. 4.3. The adjoint action. Lie group acts on its Lie algebra by the so called adjoint action Ad,. First let us observe that G acts on itself by the group conjugation g : h6+g - ‘hg. The lakter can be lifted to the Lie algebra via exponential map:
g
-1
(expA)g = exp ( g - ‘Ag), for A E 6.
(4.9)
We call the resulting action (4.9) of G on (5 the adjoint representation Ad,. Of course, for matrix groups G,the adjoint action is nothing but a conjugation with g, Adg(A) = g - ‘ Ag . In general, adjoint action represents G by automorphisms of its Lie algebra,
g1.4. Lie groups and algebras; The infinitesimal method
53
Ad,[A;B] = [Ad&A);Ad,(B)]. We shall see that the adjoint action of group G by automorphisms of its Lie algebra corresponds to the adjoint action (4.1) of 6 on itself by derivations, adA(B)= [A;B]. In fact derivations {adA} represent infinitesimal generators of oneparameter subgroups of automorphisms {AdexptA},as will be explained in Theorem 5 below. 4.4. Basic classification of Lie groups and algebras; Examples. On the algebraic level the realm of Lie algebras (it is easier to start with them, as one avoids many topological complication) contains two large and distinct classes:
I) solvable and nilpotent algebras (the latter make up a subclass of solvables); 11) simple and semisimple algebras. We shall briefly describe both of them. The definition of the former is given in terms of the upper derived series of repeated commutators,
8 3 8'= [0;0] 3 8" = [@';@'I Here each subsequent term X , Y E d k - l ) } of the preceding term
3 ... 3
dk) = [(5@-1); d k - 9
dU is derived dk-').
3 ...
(4.10)
via commutators, Span{[ X ;Y ] ;
Algebra 8 is solvable (of step n), if its derived series terminates at {0}, (similar notion applies to groups with group-commutators { g - ' h - ' g h } in place of Lie brackets [ X ; Y ] )A. commutator 8' of Lie algebra is the smallest ideal 8 c 8,with the commutative factor-algebra 8/@, and one can check that all are also ideals of 8. So solvable algebras/groups are obtained by subsequent abelian extensions of abelian algebras22. Nilpotent algebras form a subclass of solvable, they are defined in a similar manner with a (smaller) lower derived series,
dn+') = (0)
{dk)}
8 3 8,= [@;@I 3 8, = [O;O,] 3 ... 3
[8;(5J 3
...
(4.11)
So each subsequent term is the commutator of the entire algebra 8 with the preceding term. Once again the series is required to terminate at {0}, On+' = [@;On] = (0) (step-n nilpotent). In particular, 8, belongs to the center of 8, and the entire algebra is obtained by a sequence of central extensions. The natural examples of solvable and nilpotent algebras are made of upper22The terminology came from the Galois theory, where solvability of the Galois group of an algebraic equation implies that the equation can be solved in radicals. Radicals correspond to cyclic subgroups of quotients dk)/dk-').
$1.4. Lie groups and algebras; The infinitesimal method
54
1
triangular matrices, respectively matrices with (0) on the main diagonal (problem 5 ) . Other examples incude t e affine group Affl, matrices of the form{ and
[
realized by the upper-triangular
(a step-one solvable group), and the Heisenberg groups HI
W, (step-one nilpo ent g oups) (see $1.1and chapter 6).
Next we turn to semisimple and simple algebras. The former are characterized by the property that the commutator [0;(5] = (5 (in other they have no commutative quotients), the latter (simple) have no ideals at all (hence no nontrivial
quotient^)^^.
These properties have important implications for the structure and analysis of simple and semisimple groups and algebras (see chapters 4-6). At this point we shall mention only two basic results:
i) any semisimple algebra is a direct sum of simple ones, 0 21 @I Gj;in other 3 words the adjoint action (4.1) on (5 is completely reducible; ii) all simple Lie algebras can be completely classified; they consist of the classical
series: unimodular, orthogonal, symplectic, and a few exceptional ezamples. Here we shall list the basic classical simple Lie algebras, which correspond to
classical Lie groups of $1.l. Examples: 1) Lie algebra gqn) = Mat,, of the general linear group GL,; and its subalgebra sqn) = {A:t r A = 0 } , which corresponds to a special linear group SL,. Other classical groups (11.1) preserve certain quadratic/bilinear forms J on Wn;
Cn, so that 'gJg = J ; or g*Jg = J ; for group elements ( 9 ) .
(4.12)
The corresponding condition for the Lie algebra elements {X} becomes
'XJ + J X = O ; o r X * J + J X = O , which easily derived from (4.12) by differentiating g = exptX, at t = 0. Thus we get 2) so(n)
- algebra of all n x n real/complex antisymmetric traceless matrices
{'X = - X ; tr X = 0}, which corresponds to the orthogonal group S q n ) . In the real case "traceless" condition holds automatically for all skew-symmetric matrices.
3) u(n) and su(n) - algebras of all n x n complex (hermitian) antisymmetric matrices, { X * = - X}, and traceless antisymmetric matrices { t r X = 0}, correspond to to the unitary, or special unitary groups, U(n) and SU(n). In a similar fashion one can of ad,
23We can phrase both definitions in terms of the adjoint action: simplicity means irreduci6ility on 8 , while semisimple implies that operators {adx(@):X E 8 ) span entire space 8.
51.4. Lie groups and algebras; The infinitesimal method
55
describe Lie algebras so(p;q); su(p;q) of ( p ;q)-indefinite orthogonal and unitary groups S q p ;q) and S y p ;q), symplectic algebra sdn). For instance, matrices X E s o ( p ; q ) have block structure d E sa(q), and c = Tb E Matp
4.) (problem 6). So
{X=(i
:)}with
a€+);
and similar representations could be found for ur(p; q),
vector space 4 p ; q ) is isomorphic to 4 p ) x so(q) x Mat,
Q, which
yields the dimension of Lie group S q p ; q).
We shall see more examples in subsequent chapters 4-6. Two classes: solvable and semisimple, provide the building blocks for arbitrary Lie algebras S!. The following general result will be stated here without proof. Levi-Malcev decomposition Theorem: A n y finite-dimensional Lie algebra I? has the maximal solvable ideal '32, called radical, and a semisimple subalgebra 6,so that S! is decomposed into a semidirect product: S!='[email protected] h e radical 8 is a characteristic ideal of I? (invariant under all automorphisms of S!), and obviously unique, while a n y two semisimple factors 6;6' differ by a n automorphism of !2, 6'= a(@),a E AutI?. Let us recall that a semidirect product S! = ?R@ s6, is defined for a pair of Lie algebras {%;@}, where 6 acts by derivations on '32,
6 3 z+a = az E Der('32), where a ( [ a ; b ] = ) [ a ( a ) ; b+ ] [ a ; a ( b ) ) for ; all a,b E '32. By definition, I? consists of all pairs { ( a ; z ) }with the Lie bracket (problem 6), [(a;2);(4Y)1 = ( a -t
4)); 2 -t Y ) .
(4.13)
The natural examples of semidirect products include the Euclidian motion and the Poincare algebras: En = W" @so(.) and Sp, = M" @ 4 1 ;n - 1). In both cases the radical '32 = W" (or M") is commutative, and a (simple) factor 6 II 4.) or 41;n-1) (orthogonal/pseudo-orthogonal)acts on '32 by linear transformations. Another important example, discussed in chapter 6, is a semidirect product of the Heisenberg and symplectic Lie algebras: 8, @ 4.). Further details and references on the general Lie theory could be found in [Che]; [Jac]; [Kir]; [Ser]; [Hell. 4.5. The infinitesimal method. The correspondence between Lie groups and Lie
algebras extends to their representations. It is often convenient to replace Lie group representations by those of its Lie algebra, as the latter are much easier to analyze (study irreducibility; decomposition, etc.). We shall conclude this section with a general
56
81.4. Lie groups and algebras; The infinitesimal method
result that serves the basis of the infinitesimal method in representation theory. It will require a notion of the generator of a 1-parameter group of unitary (bounded) operators U t in an oo-D space 36. Proposition 4 (i) A one-parameter strongly continuous group {U,} of matrices, or bounded operators in space 36 has a generator A, defined on a dense subspace of “smooth vectors” {t}(i.e. vectors which yield smooth vector-functiow t(t)= U t t ) ,
Conversely, any selfadjoint (bounded and unbounded) operator A ( o r more general A, subject to some technical constraints) generates a 1-parameter group, denoted by Ut( = e t A t .
(ii) Generators of orthogonal/unitary groups {Ut} are skew-symmetIic, A* = - A . The algebraic (finite-D) version exploits the well known feature of any {U,}: differentiability t o any order (even real analyticity)! The same holds for a uniformly continuous group U , in Hilbert/Banach spaces. In both cases, U , = etA, with a bounded operator24 A (problem 4). The strongly continuous (unitary) case is more subtle, as not all vectors [ E 36 have differentiable “tails” { [ ( t ) = U t [ } , in general. However, any U has smooth vectors { < : [ ( t )E el} (or
em;Coo; real
analytic), which form dense subspaces in
36: 36 3 36, 3 ...36, 3 ...X6, 3 A. Indeed, smooth vectors can be obtained by mollifying (smoothing) operators of J1.2,
It is easy to check that operators {U,} send 36 into the m-smooth subspace 36,. The density of the embedding 36, supported functions {f E
c 36 is verified by observing that m-smooth compactly
er}approximate any L’-function
(in the weak sense!). If f,4, then images U ,
on R, as well as distribution 6
[I
are easily seen t o approach any [ E 36. d Thus we get a densely defined (closed) generator: A = &Y, on 36, c 36. For unitary { U } operator A is selfadjoint, A*=A,
90
n
its exponential eitA (as well as any other function
=f(A)”) can be defined via spectral resolution (Appendix A), eitA =
I
eiXtdE(X) = U,.
Smooth vectors form a domain 9 ( A ) = 36, of a closed, skewadjoint (in the unitary case) -This fact is a generalization of differentiability of continuous one-parameter groups of numbers, or characters (~(1): t E R) c R or C. Such x is shown to be differentiable, by smoothing i t 1 where c = J f ( s ) x ( s ) d s# 0, and the out via convolution with nice f E e:(R), i.e. ~ ( t=) i;(x*f)(t), RKS is obviously as smooth as f! Now it follows immediately, that ~ ( t=) eat, for a E R or C.
51.4. Lie groups and algebras; The infinitesimal method
57
unbounded operator A . Space 9 ( A ) is closed under a stronger (graph) norm,
11 ( ]I1 = 11 ( 11 + 11 A(
1
or norm ( I +I A
b( 1 where I A I = (A*A)’/’- the modulus of A.
In fact, any such A defines a scale of densely embedded (“Sobolev”) spaces 36 3 36,3
where 36, = ( I
+ I A I )-“36
intersection 36, =
n
36,-
... 3 36,3 ... = 36,;
=domain
of A’” = {rn-smooth vectors in 36}. The
consists of oo-smooth vectors, and is also dense in 36
m>O
(moreover, 36, contGns yet another dense subspace A of analytic vectors!). All subspaces are invariant under the group U , , and of course, U,136,
is strongly continuous in the rn-
norm.
Let us illustrate the general concepts of smooth vectors; closedness; domains; “strong” vs. “uniform” continuity of operator-groups with a few simple examples. 4.6. Examples: 1) Multiplication operator, A f ( z ) = zf(z), on L2-spaces: L2(W); ~ ~ ( bl); [ a ;or more general
L’(w; dp), generates a unitary group: U,f = e i t t f ( z ) , also
acting by multiplications. i) On finite intervals [a;b],or for measures dm of “finite support”, the group is uniformly (norm) continuous, and generator A is bounded, 11 A 11 5 C ,if supp I c [ - C;C ] . The two conditions (norm-continuity and boundedness) always come together. Furthermore, all vectors {f} in L2 are “00-smooth” (even real analytic), so
36 = 36, = A = L*.
ii) On infinite intervals (e.g. W), A is no more bounded, and U , is only strongly continuous: 1 U ,f - fIP0, as t+O, for any f E L2. Not all vectors are smooth, however, the space of “1-smooth vectors”: 361 = P(A) = {f: similarly “m-smooth vectors”:
I f 111 = I (I+ I 2 I 1f(z)llL2
36, = 9(Arn) = { f :llfllrn =Il(ltI z I
)“f(.)IIL2
< CQ}’ < 00}, etc.
Vector-function ei‘”f(z) could be differentiated at t = 0, only for 1-smooth vectors { f } , to get Af = zf!
2) The translation-group W: U,f(z) = f ( z - t ), acing on L2 (or other LP-spaces) is strongly continuous. Its generator A = d is an unbounded operator with a dense domain made of “1-smooth” vectors/functions, f E L2,so that f’ is also L2,
{
1 1
36, = f:1 f 1 1 = 1 f 1 L~ t f’
L2}
- the so called first-Sobolev space.
$1.4. Lie groups and algebras; The infinitesimal method
58
Similarly, LLm-smoothvectors” coincide with um-smooth (Sobolev) functions” (see
36, =
{f : l l f l l m
= IlfllL2
-+ ... +I1 f(m)IIL2} - m-th Sobolev space25.
The last example explains the terminology “smooth vectors”. It could be extended to more general one-parameter group actions, e.g. flows on
manifolds,
4,:
z+y(z;t) on A. Any such flow is generated by a vector field
as a solution of an ordinary differential system:
X = Cajaj,
+ = X(y); y(0) = z (so we could write:
dt = exptX). Associated to any flow dt is a 1-parameter group of linear transformation on function-spaces over A, e.g. L’(A),
U , f ( z ) = ~ ( I # J , ( z )or ) ; more general V,f(z) = a ( z ; t ) f od,(z), where
(I
satisfies the cocycle condition: a(+;t + s ) = a(+;t ) a ( d , ( z ) ; s ) The . generator of
X considered as a 1-st order differential operator C ajaj on (it plays the role of D = in example 2). Smoofh vectors consist of dz
group U,, is the vector-field functions over A
functions differentiable along the flow. In the second case, the generator is the 1-st order operator of the form: A = X + q , where q is a multiplication operator with function q(z) = crt(z;O)(at- derivative of
(I
at t = 0). If function q is sufficiently smooth, one can
check that Um-smoofh vectors” of A are the same as for X.
Once the notion of generator of one-parameter groups was made clear, we can proceed to more general Lie group representations.
Theorem 5: Any representation g+T, of Lie group G in space 36 gives rise to a representation X +Tx of its Lie algebra (li: (4.14)
In the 00-D c u e generators { T x } are unbounded operators defined a joint dense core of “smooth vectors” 36, = {.$(Tg(I q ) E C?(G);for all q E 36). Furthermore, for unitary T operators { T x } are skew-symmetric. Conversely, a representation of Lie algebra 0 by (closed, unbounded) generators { T x } can be ezponentiated (lifted) to a representation of its simply connected Lie group,
T,
= exp T x .
(4.15)
Proof: On the purely algebraic level (finite-D, uniformly continuous case) the result becomes a simple consequence of the exp and log maps (defined via Taylor expansions)
1 + I D I )“f
-One can show (52.2) that the m-th Sobolev norm is equivalent to (1 I D I = ( D * D ) 1 / 2 is the modulus of the differentiation operator D = L. dz
lL2;where
51.4. Lie groups and algebras; The infinitesimal method
59
and the Campbell-Hausdorff formula (4.2). On the functional-analytic (unitary) level some technical complications appear, as generators { TX} could be unbounded (densely defined) operators, each with its own domain !D(TX).However, all of them have a joint dense core of I-smooth uecfors 36,. As in the one-parameter case, 36, contains the range
{( = T f ( q ) :q E 36) of any mollifier T with f E e,!(G). In fact, there exists the whole
f
scale of
3 6 , 3
m-smooth
spaces: 36 3 36, 3 ...36,
3
...36,
3A
(analytic vectors), with
{ T f ( q ) :E f ey(G)}. All of them are dense and invariant under the group action
{Tg}(for analytic vectors this result is due t o L.G$rding). On each space 36, one can define the m-fold products and powers of generators {TX...T X :XjE a}, furthermore 00
1
,
on analytic vectors {( E A } the Taylor series:F$TF(E), converges. Thus we get the group action on analytic vectors { T gI d:g = e x p X } , which then extends rough the whole space 36. An alternative way t o obtain the group representation in the unitary case (skewsymmetric generators) is via spectral resolution:
T X = JXdE(X), E(X)-spectral measure, whence the corresponding unitary group: exp(t T
~= T~~~~~ ) = Je’”dE(X),
as in Proposition 4. This completes the proof.
Let us remark that many natural properties of representations, including irreducibility, decompositiona6, etc. can be transferred from Lie algebra 8 to Lie group
G. However, in the absence of simple connectivity of G local representation { T e X p x } given by (4.15) may not be extendible through the whole group. Indeed, only “half’ of irreducible representations of simply connected group SU(2) extend through the representations of Sq3)u SU(2)/{i I} (chapter 4), and a similar pattern holds for all other non-simply-connected groups. Clearly, the dual object of any factor-group G / H consists of all i~ E G, that annihilate H , i~ I H = 1.
26with some obvious qualification: for any invariant subspace of the Lie algebra generators So irreducibility of { T x } means any {Tx}-invariant subspace 36 is dense in 36.
ITx} we take its closure in the %norm.
60
51.4. Lie groups and algebras; The infinitesimal method
Problems and Exercises: 1. Prove Lemma 3 (Hint: some statements, like (ii), are easy to check first for diagonalizable mairices A = U-'DU (D - diagonal, U E GL,), and then to use density of diagonalizable {A) in Mat,!). 2. (i) Prove formula (4.5) for matrix subgroups, using the Taylor expansion of { e x p t A } . It is known as Trotter product formula (4.5), and has numerous extensions and applications to groups (and semigroups) of operators in spaces of finite and infinite dimensions. (ii) Show that formula (4.6) defines a 1-parameter group with generator A. 3. Show that subspaces d,; so(n) and
4.)of g/, = Mat,
form Lie algebras.
4. A (strongly continuous) one-parameter group of unitary Hilbert-space operators { U t } has a selfadjoint generator A = A*, i.e. U , = exp(iAt). Show U , is uniformly (norm)
continuous iff operator A is bounded. Steps: i) Use spectral decomposition of a selfadjoint (not necessarily bounded!) operator: A is unitarily equivalent to a multiplication, r : f + r f ( z ) , in LZ(R;dp) or in the direct sum
e,3 L2(R;dpj).
ii) Show: A is bounded iff s u p p t d p } or U s u p p { d p j } is compact in R. iii) Unitary group { U t : f-e""f(z)} compact.
in L2(dp) is norm-continuous iff s u p p { d p } is
5. Show that all upper triangular matrices over R or C form a solvable Lie algebra 23,. Find the derived series for B,, and the corresponding Lie group. Do the same for the the algebra U, of all upper triangular matrices with 0 on the main diagonal (show that U is nilpotent, find its lower derived series, and its Lie group).
6. Find block-structure (as explained in example 3) of Lie algebras: 4 n ) , su(n), s o ( p ; q ) , w ( p ; q ) , sp(n), and use it to compute dimensions of the corresponding Lie groups.
7. Show that (4.13) defines Lie bracket on the semidirect product % @ 8 of Lie algebras.
r
8. Show that a normal discrete subgroup of a Lie group G belongs in the center Z ( G ) (Hint: conjugate each y E r by elements g E G, y+g-'-yg, and observe that all {g} in a vicinity of {e} fix y).
Chapter 2. Commutative Harmonic Analysis. Schur's Lemma implies that irreducible representations of a commutative group G consists of I-D characters, i.e. homomorphisms { x:G+C* numbers)}, or {x:G-+T
(multiplicative complex
= {e"} - unitary characters}. We shall mostly deal with the
latter. The characters can be multiplied and inverted: (x1;xz)
+
x1x2;
x
x-l =
-+
1 Xb)' -
so they also form a commutative group, called the dual group
e. Dual group of a locally
compact G , is itself locally compact [Po,], hence carriea an invariant (Haar) measure. Three commutative groups lie a t the heart of the classical harmonic analysis: R";
Z" C R" and 1"N R"/Z". The characters of all three consist of exponentials X,(z)
= e i t . '.
In the case of R", parameter ( runs over R",
80
the group becomes isomorphic t o its dual.
For Z", identified with an integral lattice {m = (ml; ...mn)] in R", characters
x = xO(m)= e " . m ,
...
are labeled by the dual torus, 0 = (Ol; On), 0 5 O 5 2u, so
A
Z" u T". The characters of torus T" = {(el;...On) : 0 5 Oj 5 2 x 1 , or more general (0 5 Oj 5 Tj], are in turn labeled by the lattice points,
x = Xm(4
eim .O.
, m E Z",
so 1"N Z". Let remark that the mutual duality of groups 1" and
Z" exemplifies the
general Pontrjagin Duality principle for locally compact Abelian groups [Pon]: the second dual group (characters of G) is isomorphic t o G itself, via the map: u-G(x) = x(a), a E G (elements a E G define characters on G). So the classification problem (I) of
chapter 1 for commutative groups G amounts t o characterization of
6,
while the
decomposition problem (11) for the regular representation R will lead to study the Fourier transform 9 on L2-spaces and other spaces. This marks the starting point of the Fourier analysis. Chapter 2 gives the basic overview of the Commutative Fourier analysis, and some of its applications, the main emphasis being on differential equations.
$2.1. Inversion and Plancherel Formula.
62
$2.1.Fourier transform: Inversion and Plancherel Formula. In this section we shall introduce the Fourier transform/series expansion on R"; 2"; T", discuss their basic properties (with respect to translation, differentiation, dilation, rotations. Then we establish the key result of the Fourier analysis, Plancherel/Inversion formula, in 2 different ways. One exploits Gaussians and Schwartz functions, another Poisson summation.
1.1. Fourier transform 4 on group G = R";T" is defined by integrating functions {f} against characters {x = e i t e Z } ,
while on
Znintegration becomes summation,
Sometimes we shall indicate the
"I
to
transform by TZjt.
From the
definition one could easily check a few basic properties of 4. 1) Convolution-+product, 4:f*g---+?(t)G(t) 2) Translation-tmultiplication with character, 4:R, f = f ( z - a)-+eia't?(t)
3) Differentiationjmultiplication, 4:8,f +itj?(t); hence 4:aaf+(it)"Y Here
{aj} and (8%= ( a l...; a,,)}denote partial differentiations in I . The latter
relation is easily verified via integration by parts:
/(a"f)g = / f ( - a ) a g ;
with g = e T i Z . t .
We introduce an involution on functions { f ( s ) }on G,as a linear map,
-
f(-.),
*: f(.)-f*(.) = with the obvious properties: ( f * g ) * = g**f*; (f*)* = f . Clearly, 4) Involution-tcomplex conjugation, 9:
f*(z)+?(t).
5)A linear change of variable, A:z+Az
( A E GL,)
on Rn,
defines a
transformation of functions,
f + f A ( z )= f o A = ~ ( A I ) . Fourier transform takes such linear change into its inverse transpose, precisely 4 : f AdetA + ~ ~('AA-'). o
$2.1. Inversion and Plancherel Formula
63
Finally, relation (3) differentiation-+multiplication, can be reversed into
6) m ultiplication-tdifferentiation, 9:xjf(x)+$3,,f^([); As we mentioned earlier characters
+~~f-(ia,)~?
{x} describe all irreducible representations of
the commutative group G. On the other hand they can be interpreted as joint eigenfunctions of all translations { R,:a E G } , respectively, of their infinitesimal
generators {aj} (on 53"; U"), as well as convolution operators: R f [ g ]= ( f * g ) ( x ) . Formulae (1-3) can be rephrased now in the language of the representation theory by saying: map 9 diagonalizes regular representation R of G,
9R9-
N
"direct sum/integral of characters".
Precisely, any function f (in a suitable space) can be expanded into the direct sum/integral of characters
x = eix
*
€,
C
1
f(x) = bEeix* t; or f(z) = b( [) eiz ' '%(, (1.3) with "Fourier coefficients/transform" b(() = ?([), and the regular representation turns into multiplication on the space of "Fourier coefficients",
Later on (chapters 3-7) we shall establish various noncommutative versions of this result.
1.2. Plancherel and Inversion formula. The fundamental property of exponentials { e i < ' + } is their orthogonality and completeness. In case of compact G (e.g. U"), orthogonality can be understood literally',
as all
{x,}
are L2-functions on G. Completeness means that any f can be
approximated, hence uniquely expanded into the Fourier series. An elementary proof for
T"exploits the
Cezaro means (problem 3).
More general explanation has to do with compactness of convolution-operators, R f:u (z )d f*u (+) , in L2 (LP)-spaces on 1" (Appendix B). Compact self-adjoint (f = f*)
operators, are well known to have a complete orthogonal system of eigenvectors, and exponentials {xm = ei+ * m} are precisely the eigenvectors of R
f , f*eiz
*
= F(m)eic ' €.
'The same holds on an arbitrary compact group G (see [HR]; [Loo]),namely,
(X, I xa) = I&',&d+
= I G I 6km;
I G I = Ud(G);6km-
Kronecker 6.
62.1. Inversion and Plancherel Formula.
64
Similar arguments apply to all compact groups (chapter 3).
For non-compact groups W"; Z" orthogonaZity should be understood in a generalized (distributional) sense,
l(eit.z I ,ir)
.z)=
c
,i(t-r))
*
zdz
(1.5)
In other words, the role of Kronecker 6,k = S(m - k), being played by the Dirac 6-function. The integral (1.5) is strictly speaking divergent, but it makes sense as a distribution. Namely, for any nice (testing) function f(q) on R",
We shall see that in such form orthogonality is equivalent to the
Inversion/Plancherel formula for 9. Namely the transform,
4':h(q)-tL(x) = ( h I eWiV.') = Ih(q)eiV*'dq, takes
f^
(1.6)
back into Const x f. Therefore, (1.6) defines the inverse Fourier
transform of h. The precise result is usually stated for a suitable class of
functions on G. We shall treat separately compact (T";Z")-cases and noncompact Rn- case.
1.3. Compact case. A suitable class of functions on torus, called Fourier algebra, consists of absolutely convergent Fourier series, A = {f(x) = CakeiZ * k: Clearly, algebra Ace(U"),
c I ak I < m}.
is made of continuous functions on
U", and
contains sufficiently many smooth functions, A 3 em,for m > n. The latter is easy verified using differentiation formula 3). Indeed, Fourier coefficients of a partial derivative
aaf = C(ik)aakeiz'k , by (iii). So if f E em, then its
Fourier coefficients ak = o( I k I,),-
which
guarantees convergence of
c l%I. Uniqueness and Plancherel Theorem on T": (i) Any function f E A(Tn) is uniquely determined by its Fourier coefficients, {ak}*f(z)
(ii) For any function f =
=
Cakeiz' '.
CakeiZ " E L',
In other words 9:L2(T")+12(Z"), is a unitary map. Both results are fairly
52.1. Inversion and Plancherel Formula obvious corollaries of the completeness and orthogonality' (problem 3).
65 of {eiz * k: k E Z"}
1.4. Continuous case W". Here a suitable class consists of all functions { f},
integrable along with their 9-transforms,
{f E ~
T E L I } = L' nA.
1 :
As above, A denotes the Fourier algebra {y(z): f E L'}, of absolutely convergent Fourier integrals. As with T" algebra A is shown to consist of continuous functions, decaying at {oo}, a consequence of the Riemann-Lebesgue p ' - compactly supported mTheorem (problem 1 of 52.2). Furthermore, A 3 C smooth functions3, so
C'y c A c C',,
for m > n.
Inversion/Plaucherel Theorem on R": (I) A n y function f E A n L', can be uniquely
recovered from
1 b y the Inverse Fourier transform
(11) transformation 4,normalized b y factor L2(Rn) onto L"), i.e.
1becomes a unitary map from (27r)"/2 '
The Fourier transform of an Lz-function in (1.9) should be understood in the "square-mean" sense, i.e. by approximating f with nice (L'; compactly supported) functions. Let us make a few comments: 0
the general inversion formula (1.8) is equivalent to a special case z=O (and is
often stated in that form), G
multiplication of ?(() by e i z * < corresponds to a shift f(O)+f(z), under the inverse transform 9':
?+f.
'Indeed, for any complete orthogonal system (basis) { $ J ~in } Hilbert space 36, any f E 36, is uniquely decomposed into a series f = C akdlr,and 11 f 11 = C 1 ak I Gk (Parceval identity). Clearly, factor ( 2 ~ ) "in (1.9) represents L2-norms of Fourier harmonics {gk=
1) 1 '
3The inclusion
er c A
j(<= ) U( I < I -"'), Slf(f)ld<.
whe_nce
follows from the differentiation formula (3): 9[13"f] = (i<)"j for anyf E, :e which yields absolute convergence of the integral
52.1. inversion and Piancherel Formula.
66 0
Plancherel formula (1.9) follows from the Inversion (1.8) applied to a function:
g(4=
If(. + Y)f(y)dY
Indeed, from (i)-(ii) it follows that
5 via (1.8), 0
?(()
=
f*f**
(1.10)
=I ?(() f, while g(0) = Ilfll'.
So inverting
we get the Plancherel formula (1.9) for f .
The result holds for any locally compact Abelian group
4 transforms functions { f(z)} on G into defines a unitary map: L'(G)+L2(8).
{ j ( x ) } on
G ([Loo];[HR]). Here
e, takes AnL'(G)+AnL'(G), and
We shall present two arguments based on very different, but equally important ideas. One of them exploits the Gaussian and Schwartz functions, another a "discrete (lattice) approximation" of
W", and the so called Poisson summation formula.
The Gaussian method. Obviously, it suffices to prove (1.9) for a dense class of functions in AnL' or L '.
One possible choice are infinitely smooth {f},rapidly decaying a t {co}
with all their derivatives
I z I -"aaf(z)-+0,
as z - o q for all m and a,
equipped with a suitable system of seminorms, e.g.
IIf llaP = "$2 I ."f%) These are
90
1
called Schwarlr functions on R". The class of Schwartz functions and
distributions is discussed in detail in 52.2. Here we shall only remark that the Schwartz class is sufficiently rich. It contains, all C3-
compactly supported
{f}; the Gaussian
.
function: G = G,(z) = e-tz2 (z2 = z z = I z I 2), and all products: G x "polynomials".
T h e latter class, 9 = Szf = p ( z ) e - z 2 / 2 : polynomial p ( z ) } , will be used in our proof of inversion Theorem. Let us remark that space 9 is dense in C,(R"),
by the Weierstrass
"uniform approximation" principle, as well as in other spaces: Lz, ZP, A n L ' ,
Y, with
appropriate norms. The inversion for the Gaussian itself is verified directly (problem 5):
9:Gt-Gt = (2nt)"/2e-t2/4', i.e. Gaussian is Fourier-transformed into another (differently scaled) Gaussian, whence follows G(z) = T-'[C]. Then it remains to apply the differentiation and multiplication rules 3)-6), to get Fourier-inversion for any product-function, f = p , ( z ) G in 9. The proof is completed by the standard density argument: once Plancherel formula is established for a dense subspace 9 C Lz, it follows for all f E Lz.
1.5. Poisson Summation Formula: Let functions f and
1
be integrable and
52.1. Inversion and Plancherel Formula
67
continuow on Rn (e.g. f E A n L'). Then for any period T
> 0, (1.11)
Similar result holds for an arbitrary lattice: r = { m = Emitj}, spanned b y vectors { t , ;...tn} c R", and its dual lattice, r1 = (7= E kjaj}, where a;. t . = 2x6. .. Namely, 3 83 (1.12)
where J L = R"/r
21
T" denotes the fundamental domain of
r.
The proof is fairly easy and straightforward. We define a new f-periodic function:
Series obviously converges and gives a continuous function on torus R " / f
N
Tn. Function
F is expanded in the Fourier series,
with coefficients {a,) on the dual lattice f':
Next the torus inversion formula applies,
and the result follows.
Remark Formula (1.11) has many different meanings, interpretations, and S(x - rn) on Rn, "6applications. One of them is in terms of distribution: 6, = function of r".The LHS of (1.12) pairs b r to a nice test-fun%sxf(z), while the RHS defines the Fourier transform of 4 = S f . By definition, ($If^) = ($If), for any distribution 4 and a test-function f. So (1.11)-(1.12)reads,
c
in other words LL6-function of
r is Fourier-transformed into -volX1
"6-function of
rl".
$2.1. Inversion and Plancherel Formula.
68
In such s form Poisson summation turns into a special case of the general result for arbitrary Abelian locally compact groups: if
r
is a closed subgroup of G, and
r'
denotes
r in G, r'={ x : x I r = l}, fhen 6-funcfion of r (understood as (6rIf ) = Sffdy,dy-Haar measure on r), is Fourier fransformed into the 6function of Z", T:Cr-Constbrf. The key idea is that averaging over r, sends functions the annihilafor of integral
f(x) on G into functions on the factor-group G / r ,
f ( . ) + f ( z + r )= Jrf(z+V)dY = 6 r * f . The dual group l o G / r is identified with
r', and
the Fourier transform takes the A
A
projection map: f --df *f (from G to G / r ) ,into the restriction map: f -f
1 r'.
Poisson summation immediately yields the inversion/Plancherel formula on R". We let T-co in (1.11.), and remember that f(z)-+O at {m}. The LHS of (1.11) clearly goes to f(O), while the RHS converges to
&/?(O&, so f(0) =
(2x1
?(t)d<, QED.
Further applications of the Poisson summation will appear in $2.4 and chapter 7.
52.1. Inversion
69
and Plancherel Formula
Problems and Exercises: 1. Show that any continuous character on R (a homomorphism x:R-C') is of the form x ( a ) = eZa, for some z E C (prove it first for rational a E Q, then use continuity of x ! ) . Deduce the character formulae (1.0) for R"; Z"; 1". 2. i) If a sequence of functions {Fn]on R"; 1" approximate &function: (F, I h) -h(O), as n-co, for any continuous h, then the corresponding convolution kernels: F,*f(z)-f(z), as n+m, for all {f} in C and Lp, 1 5 p < cm (such sequences are called approximate unities).
ii) Any sequence of positive functions {Fn] with the properties: (a) J F , d z = 1; (b)
1
1x1
F,dz
-.)
0, as n+m, for any
6
0,
56
forms an approximate unity. 3. Completeness of characters
eim * =} on 1":any f E e, or Lz is approximated by
{x,=
trigonometric polynomials,
c
PN=
lml I N
bm
eim.I
Hint: I t suffices to check the result for continuous {f] with the standard 11 1 1 ~ . Note, that partial sums, may not converge t o f E e uniformly, in general. However, certain means of s, do, like Ces a Po, i) Show that both sums s, and cn are given byconvolutions, s,(f) = S , * f ; c , ( f ) = F,* with kernels, which in
sn(e) =
f,
1-Dcase take the form, sin'? ( l - Ln+ )e1 i k e = - , sin':
sin (n+,$)~
called the Fejer kernel.
sin^
; andF,(B)=
Ikl 5"
ii) Show that the Fejer kernel forms an approximate unity (problem 2), hence F,*f(z)+f(z), uniformly in e, as well as in any LP-norm, 1 5 p < 00. 4. Establish the decay rates of Fourier transform of m-smooth functions
1"and
emand
36,
on
R": estimates (1.11); (1.0); (1.0).
5. Fourier integrals on R are often evaluated by the techniques of complex analysis: calculus of residues, branch-cuts (if necessary), changing path of integration. Use complex analysis t o evaluate the following integrals:
-t2
i) Gaussians: G, = T(e-t22/z n- D).
sin? iii) T(-) = 2n sinh bz cos@ + ~ b
= G, (in 1-D); and (znt)-"/ze-t2/zt (in
on R 0 < a < b (calculus of residues).
0 ~ 5 '
b
Integrals (ii-iii) are evaluated by the residue calculus. They represent the Poisson kernels (j2.2) in the half-plane and a strip.
70
$2.1.Inversion and
Plancherel Formula.
The following 3 Fourier integrals appear in the harmonic analysis on group SL,(R) celebrated Selberg-trace formula ($7.6 of chapter 7):
00
vi)
xis
d 1 - 2Xcos 0 +
X2
dX = -
rsinhs(O+r) sin 0 sinhns
' -'<
- the
< O'
6 . Show that the Fourier transform o f the radial function f = f(lz I) in R" is also radial, find the convolution of two radial functions f(r), g(r) in polar coordinates. 7. Show that the Gaussian G,(z) and the Poisson kernel Pi form convolution-semigroups:
G,*G, = Gt+,, and P,*P, = Use the Yproducl,+convolutionn property of 4, and explicit form of
and
p).
52.2. Fourier transform on function-spaces
71
52.2.Fourier transform on function-spaces. In this section we shall briefly discuss the basic properties of Fourier transform on various function-spaces, including LPspaces, Sobolev spaces, distributions and analytic functions. The material will provide some background for a subsequent study of differential equations in 52.4 and other parts of the book.
2.1. L p - spaces and interpolation. Plancherel formula implies unitarity of the normalized Fourier transform '3':f+(27r)-"/23 on L2, On the other hand Riemann-Lebesgue Lemma means that 5:L'+Lm
boundedly,
Such estimates can be extended to other LP-spaces by the Riesz interpolation Theorem. Namely, S:L*-+Lq, where q = L,is the dual Holder index, furthermore P-1
1-
(Hamdorff-Young inequality).
The exact constant in the Hausdorff-Young was calculated by Beckner in ~ O ' S ,
2
and was shown to be attained on the Gaussian function f =e-" 12. Interpolation provides a powerful tool in estimating linear (bilinear, multi-linear) operators in { L p } and other scales of spaces. The simplest linear version states,
Riesz-Thorin Theorem: i f operator T maps boundedly Lpl(X)+Lql(Y), and L P 2 ( X)+LQ2(Y ) , IITf IIQ' I ClIlf then for any pair of indices:
IIP1'
1-l-s Bp1
andllTf
s +--' P2'
I Q2
1-14
I C2Il.fIIP2'
&
B -T i+ operator T is bounded from L p ( X ) to Lq(Y), and its
I T IILPLQ I c:-sc;. So function
4($f) = ZoglITllLpLqis
convez in the plane of reciprocal parameters
{++I. The Hausdorff-Young inequality arises via interpolation between pairs of Holder indices ( 1 ; ~ )and (2;2) for '3. Of course, interpolation does not yield the best constant
$2.2. Fourier transform on function-spaces
72
( 2 . 1 ) , as C , = C2= l : + C ( p ; q ) = 1 , for all p . The interpolation also applies to convolutions of Lp-spaces. This time we deal with triples of indices so that Lp*Lq c L', in other words, (f;g)--+f*g,maps
(f;$i),
boundedly Lp x Lq4.L'. The general principle for bilinear T remains the same, as for
and LP2x LQ2+L'2, with
linear operators. So if T maps boundedly LplxLql+L'l, bounds C,;C,, then for any interpolating triple:
1-L
8-1
1-
8
1 - A s-1 T-rrl+F;
8-1
q-q+%; T maps Lp x L b L ' , with the norm estimate p-pl+pz;
To apply ( 2 . 2 ) to convolutions of LP-spaces we observe that L'*Lp words operators { R : f E L'} are bounded in L p by
f
llfl/l).
c L p (in other
On the other hand
L2*L2c L". Interpolating between the resulting triples ( 1 ; h fand ) ($;;;O), we get the Young's inequality,
Young's inequality holds for all groups G = R"; T";Z". Moreover, for compact groups (like T") the range of (2.3) can be extended to { ( p , q , ~ ) : 2 $ 1 - l } , due to the
+
obvious inclusions: L"
c LQ,for all 1 I q < p I 00.
The proof of the Riesz interpolation Theorem is based on a version of the maximum principle for holomorphic functions in the strip, called 3-line Lemma (see problems 2,3,4). Let us remark, that so far we didn't use any special features of R"; 1" or Z", so all results about Lp-spaces, which involve interpolation (like Young and Hausdorff-Young
inequalities) extend to arbitrary locally compact Abelian groups [HR].
2.2. Smooth functions and Sobolev spaces. Riemann-Lebesgue Theorem (problem
1) implies that continuous functions on T" (or continuous, compactly supported
R") have
{f} on
bounded, decaying at {m} Fourier transform. Thence one could easily show
(problem 5) that m.-smooth functions f E Cm(Tn), or er(R"), have their Fourier transform/coefficients {T(k)}; {f^([)} decay at a polynomial rate:
Conversely, if Fourier transform
T([)
decays as
I [ I -p,
with p
> m+n,
then f(z)
is at least m-smooth. Both results are easily verified via the differentiation formula:
( 8 f j= (it)'?([),
and the obvious relation:
f^ = O( I ,$ I -"-')+?
E L'. Hence, by
52.2. Fourier transform on function-spaces Riemann-Lebesgue Theorem,
73
f = S-l(f^) E C,.
Thus Fourier transform relates the "degree of smoothness of f" to the "decay rate of f^ at (00)". More precise results, however, could be stated in terms of L2-norms, rather than uniform (L"). The main disadvantage of smooth functions from the standpoint of harmonic analysis, is that erndoes not allow a simple characterization in terms of Fourier transform/coefficients, which could done for Lz-functions. Thus we are naturally lead to the notion of Sobolev spaces.
Sobolev space X,, all their derivatives
{ti(@)
of order rn, consists of functions {u},which belong to L2 with E Lz: I (I I 5 rn}. It is equipped with the norm
The Plancherel Theorem and the differentiation rule of 52.1 give an explicit characterization of 36, in terms of Fourier transform/coefficients {ti(()} of u,
- sum of weighted L2-norms with polynomial weights w(()= (a. It is easy to check that the sum of monomials {ka} in (2.6) (respectively partials
{aa} in (2.5)), could be replaced by powers of (1+ I k I 2)m, respectively (l-A),,
where
A = V . V = Ca: is the Laplacian on Wn. So norm (2.6) is equivalent to IIuI?mz J ( 1 t
~E~2)m~~(()~z=~~(l-A)m~2~~~~.
In other words 4 takes Sobolev space 36, onto the weighted Lz-space, Lz(wmd() (or L2(wm)in the Tn-case), with polynomial weight turn(() = (1
+ I(I
2)m,
and
This allows to extend the notion of Sobolev n o m and derivatives (powers of
Laplacian) from integers to all real (fractional; negative) values of m. By definition,
I
(l-A)-% = T 1 ( 1 -t I [ I z)-s/24u = I(l+I ( 2)-s/2G([)e""'td[.
(2.7)
Fractional Laplacian (2.7) are given by convolution kernels, Bs*u, with the so called Bessel potential -s/2
B s = ~ ~ , ( I ~ I ) = ~ - "It1 ( 1 t) I, where K,(r)-modified Bessel function (3-rd kind) of order v = 2
Thus we get a scale of Sobolev spaces {X,:s 2 0 } ,
(problem 3).
62.2. Fourier transform on function-spaces
74
36, = (1-A)-"/2[L2] = B,*L2 = 9-'[L2(w,d()], where w, = (1+ I
t I ')'.
The above definition also extends for negative s by duality. Namely, we define space 36-, as the dual space (of distributions) to 36,, via the pairing,
( F I f) = J f ( z ) F W s =
J 3 ( 0 f i ( O d t , f E 36,;
F E 36-s.
(2.8)
We shall list a. few basic properties of Sobolev spaces. Theorem 1: i) Spaces (36,) form a continuous scale of embedded Hilbert spaces:
with 36-, = "dual to 36, via the natural pain'ng (2.8)".
ii) O n t o w 8" ( o r more general compact manifolds, domains) the embedding 36, c 36, is compact f o r all t < s. iii) Sobolev embedding: spaces (36,) are related to
{em}in the following way:
Xm c e, for m > n/2, more generally,
ern+'c 36, c em-n/2,for any m I s < m+1, Proof: (i-ii) Each 36, is mapped onto 36, by a fractional Laplacian (l-A)-m, which is Fourier-equivalent to a multiplication of ['-sequences
m = d-t 2 '
of coefficients by a
decaying sequence { a k = (1+ I k I 2)-m/2}, clearly a compact operator. (iii) Multiplying and dividing a sequence of Fourier coefficients of f E 36, by a polynomial weight (1+ I k
c
lakl
I 2)m/2,
and applying the Cauchy-Schwartz inequality, we get
=~~(1+~kI2)m/2a~J(1+I~I2)-n~2~~~{(1+I~I2)m'*~~}~~e2~(l+IkI*)-m
which shows that the Fourier series of u(z) =
C akeik
'
is absolutely convergent, whence
follows continuity.
Remark 1: The last statement of the Theorem extends to embedding of Sobolev spaces 36, into the LP-scale. The general result (Sobolev-Hardy-LittlewoodTheorem) claims: space
36,
c LP - continuously
embedded, i.e.
II f llLP I I ( 1 - 4 , h for all
;1,
(2.9)
$ = f - 3 (on R"), and $ 2 f - 3 (on 1").The argument exploits interpolation for
Bessel potentials B , E Lq, with q = y . Convolutions, u+B.*u, with such B, shifts any LP-space into Lr, where
$2.2. Fourier transform on function-spaces
Consequently, for negative s, a,,,,,
75
2 36-, 2 '3,,,-n,2. Therefore,
as claimed in (i).
Remark 2: Sobolev spaces provide a natural framework to analyze solutions of differential equations, as outlined in $2.3. This often requires Sobolev spaces of both integral and fractional order, as well as Sobolev spaces on manifolds and domains
RcW". Let us also notice that spaces 36, constitute "m-smooth vectors" of regular representation R on L2(Wn), or L2(T"), in the terminology of $1.4. 2.3. Distributions { F } are defined as continuous linear functionals on suitable spaces of test-junctions {u}, like em, em, etc. One usually writes them as formal integrals
/
( F I u ) = F(z)u(s)dz.
(2.10)
So for the class of bounded continuous functions, f E e ( X ) , all functionals are given by finite Borel measures p E A ( X ) , ( p I u ) = u(z)dp(z).Test-spaces en, em we equipped with a sequence of seminorms,
I
{Ilullm=
c
maz I a"u(z) I ;m = 1;2;...},
therefore the corresponding distribution spaces, called gm,goo,consist of functionals satisfying
m
I ( F I U ) I ICCIIuIIj.
(2.11)
j=O
Using (2.11) one can show Proposition: Any distribution F E bDm (m = 0; 1;..LO),is equal to a fznite sum of
derivatives of Borel measures, (2.12)
Integer m is called the (differential) order of F . Derivatives of measures in (2.12) are defined via pairing to smooth functions:
I f) =
/ aaf(.)dP(.).
The proof of (2.12) is fairly straightforward. Indeed, the m-jet map
f ( z ) 4 ) ( 4 :I a I I ml7 takes P ( R ) into the direct sum:
e, =
C(R") @ la1 5 m
- N-component vector functions on R";
N = (n+m m 1-
$2.2. Fourier transform on function-spaces
76
Distribution F defines a bounded linear functional (in the supnorm of
em c eN,which can be extended
through the entire space
vector-valued continuous functions, Idpa: I a
eN, is
e,)
on the image
e. But any linear functional on
given by an N-tuple of Bore1 measures
I I m}, QED.
Let remark that all natural linear operations on test-functions can be transferred to distributions by duality. For instance, differentiation is introduced via
( P F I u ) = ( F I (-a)%), by analogy with the standard "integration by parts" formula for test-functions. Clearly, daF E am, for any F E am, in fact aa:am+gm+I a I . Thus we get the following chain of test-function spaces along with their distribution-spaces
1-
00
...
ln7eEC ... C e m c ... c e c L ' c ~ = P o c ca,c
co
...c a = ua, (2.13) 0
Examples of distributions include (i) &functions and their derivatives: 6 z ( x ) = S(x-z), or (6(");u)= ( - 1) I a I u(a)(O), as well as their linear combinations, e.g. hA = cb(x-m) (A- lattice in Rn); (ii) &functions of hyper-surfaces and submanifolds: (bC I u) = J,udS, where dS is the natural surface area element on C, like the spherical distribution SS = (S = S,- sphere of radius r in Rn), that appears in the wave propagator solution of the wave equation); (iii) Certain singular integrals, like Halbert transform,
I
(HI u ) = P.V. +x
(principal-value integral) = lim
WX,
as well as its higher-D analogs, called Calderon-Zygmund integrals, can be interpreted as distributions ([Zy];[St];[SW]). (iv) Distributions often arise as limits of families of regular functions4, e.g. lim 1_ = H i~(x). ,4J z+ae
+
Another linear operation on functions, the change of variable, f +f o 4, can also 4Such approximating families {fc(z)}, called regularizations, exist for any F E 3.The easiest way to construct a regularization is to take a em- smooth approximate unity {hr+6}, e.g. h, = c-"h($), with h E ( h 2 0, and J h = l), and to convolve h,*F = f, (problem 1).
cr
$2.2. Fourier transform on function-spaces
77
be transferred to distributions (problem 2), def
( F 0 4 I 21) = ( F I
'&'4-7 jF ( l ) ( h ) 4-Y+. =
0
0
Distribution spaces on torus can be characterized by their Fourier coefficients, the same way as smooth functions. For any F E !Dm we define
bk = ( F I e-iz * k =
j
F(z)e-ik*z.
T"
Fourier coefficients of continuous functions, u(x) = x a k e i k ' 2 , are bounded and go to 0, as k-too, those of cm-functions decay at a polynomial rate, ak = o( I k I -"). Since, ( F I 21) = x a k b k (an extension of Plancherel's Theorem to pairs 'functiondistributions"), it is natural to expect {ak} to increase polynomially. Indeed, if F € 9,
I l b r l = l ( F l e i k * z ) l
(2.14)
for any F of order m. We shall illustrate (2.14) by the example of &function and its derivatives,
6=
C eik
*
I , so all bk= 1;
6(")=C(ik)a eik", so bk - (ik)a* The latter extends to all (integer and fractional) powers of the Laplacian:
(-A)sb=
C 1k12seik'z,or(1-4)'6=
~(1+Ik12)Seik*z.
Hence both belong to space gm,for m > 2s. Remark 3 On R" test-function and distribution spaces are defined similar to the compact (torus) case, but additional constraints are imposed at {co}. If no such constraints are
imposed (spaces
em and P)one gets
compactly supported C-spaces 2(';:
The R"-Sobolev spaces 36,
compactly supported distributions '3. Conversely,
Cr) give all distributions (with unlimited support).
satisfy all the properties (i-iii), except compactness of the
embedding 36, c 36,. To show for instance, the embedding: 36, c C,, for m
> n/2, we
replace Fourier sums by Fourier integrals,
If
3 E L',
then f =
the rate of decay of
fF'(3)E C,. 1 at
As in the torus case smoothness of f can be linked to
00,
f E cy *?(c) = O( I < I Conversely, sufficient decay rate O( I C I -"')
or
f E 36, *(I+ I < I m ) j E ~
2 .
off. o f f guarantees (m-g)-smmthness 2
52.2. Fourier transform on function-spaces
78
2.4. Schwartz functions and tempered distributions. The most convenient classes
of functions and distributions combine both features (smoothness and the decay rate ay {co}). These are so called Schwartz finctiom 9 and tempered distributions Y'. Space I consists of €?-functions, that decay faster than any polynomial, Y = {f:z*f (z)+O,as
(a)
z+co}, along with all their derivatives. It is equipped with a sequence of semi-norms,
I(f 1I.p
=
1 z*@f I Lm
; or equivalently
1 f llpq
1
= (1t I 5 I ')P(l-A)qf
/IL=-.
(2.15)
Let us that the choice of supnorm in (2.15) is not essential, it could be replaced with any other Lp-norm. Also the order of two operations (differentiation and multiplication) could be interchanged (problem 4). So space 9 can be characterized by semi-norms,
lp-A)p(l+ Space
Y
I I 2)Pf11L2 <
is sufficiently rich, it contains
(2.16)
{ep}; the
Gaussian { e - t z 2 } ;
{polynomials}x { e - t z 2 } , etc. The important features of the Schwartz spaces are due to the fact that T:Y+lf, represents an isomorphism of 3, i.e. semi-norms (2.15) or (2.16) of f and are equivalent. Schwartz (tempered) distributions Y' are linear functionals on Y. So distribution F E Y", if it satisfies,
7
I ( F I u)I 5 C
~ I 1lpq; I Ufor any Schwartz u.
This class includes all examples we encountered so far, like the lattice &function,
8,
6, = c 6 ( z - k ) ; its transform;
= z e i k z ; &functions of
hyper-surfaces, or
submanifolds in Rn; their derivatives, etc. We shall list a few other examples of Schwartz distributions and their transforms (see problem 7,8): i) Dirac-6 is transformed into constant 1, hence distribution 6 E KS, for s > n / C . ii) &function of subspace E
c Wn, 6E+(27r)n-k6E1;
where E' is the orthogonal
complement of E , k = dimE, iii) F = bS (&function of the r-sphere S) + P(() = cn(r I ( I )-"l2+lJ n/2-1('
I€ I)
- the Bessel function of order (;-I);
iv) fractional derivatives of Dirac 6: (-A)-",
(Riesz potential), and (1+ I ( I 2)-s/2
or (1-A)-s6
I ( I -2s
(Bessel) potentials (problem 9).
v) &function of the light cone C = { z = c
m
} in
W3 is Fourier-transformed
into 8&;0;C)
turn into
1 =r n H ( c C-
m),
52.2. Fourier transform on function-spaces
79
where H is the Heaviside function. So the Fourier transform of 6, is supported in the dual cone C ' = {(:(.z 2 0). isomorphically.
Y the Fourier transform maps Y'+Y'
As with
2.5. Analytic functions and the Paley-Wiener Theorem. After the Fourier analysis of LP-spaces and smooth functions
decaying functions LP, = { f :f (z)e"
{ern;Xm}, we turn
our attention to exponentially
I I E Lp}, and yet smaller spaces of compactly
supported functions. Clearly, in both cases the Fourier transform ?(() extends from the c Cn. In the former case the domain of
3
real space W" to a complex domain {( = (+iq}
is a tubular neighborhood of the real space R", O,= {(+if): 171
imaginary iq, furthermore,
1
I,r,z(d()
5 eprlllf
(2.17)
llL2'
The necessity of (2.17) is fairly straightforward. The proof of sufficiency is more involved. T o give the reader a flavor of the arguments we shall briefly outline the 1-D case. The proof exploits the so called LZ-Hardy spaces, made of analytic functions { F ( z ) } in the half-
>
plane H = {Sz 0}, square-integrable along lines {% = yo}, parallel t o the real axis, such that all L2-norms are uniformly bounded, m
]' I F(z+iy) I ' d z 5 C; for all y > 0.
-m
Any Hardy function F ( z ) is obtained by Fourier-transforming an Lz-function f ( i ) , supported on the positive half-line [O; m) (problem 6): 00
F ( z ) = f(t)e""' tdt,
z
E H.
(2.18)
0
A similar result holds for functions { F ( z ) } , holomorphic in the lower half-plane, the corresponding {f(i)} being supported on the negative half-line
(-00;
increasing functions { F ( z ) } (2.17) the support of {f(t)} gets shifted by
01. For exponentially p to the left/right, an
easy consequence of the Fourier-translation formula of J2.1:
eizpF(z) = 9 t - , J f ( i + ~ ) I , and the fact that {eazpF(z)}belongs to the Hardy Lz-class. Therefore, analytic function F ( z ) , exponentially bounded in both the upper and the lower half-planes, has its inverse 9-transform inside the interval [-p;p], QED.
62.2. Fourier transform on function-spaces
80
Problems and Exercises: 1. Riernann-Lebesgue Theorem: Fourier 1,ansfot-m ?(<) of any L'-function f on R" is continuous and goes lo 0, as [+m, i.e. f 6 e., Hint: Check the result directly for indicator functions {xe} of rectangular boxes: B=[al; b,] x ...x [a,;b J. Then use the standard density argument: any f E L' is approximated by linear combinations of xw 2. Three line Lemma: If function F ( z ) is holomorphic in a complex stn'p 0 < 9?ez has maximal values on both sides of the strip:
< 1, and
M , = auP(F(it)\ M1=sup I F(l+it) 1, t then F is bounded inside the strip'and /F(F(s+it) I
5 MA-";,
for all z=s+it.
1'. Truncate F ( r ) by a family of exponentials {Mb-'M;e('2-')~n}, 2
(i) product G,(z) = MA-zMfe(z -"/"F(z),
to ensure that
goes to 0, as z=s+it+m, and
(ii) maximal values of G, on both sides of the strip I1.
2 . Use the maximum principle for holomorphic functions to show IG,(z) Observe, G , ( z ) + F ( z ) , as n-m!
I 5 1, for all
z.
3. Prove the Riesz interpolation Theorem,
I s ( T f )(z)s(z) dz I IMA-":
IIf llp(s)II9 ll*'(s);
P'
=
gig
for a dense subset of stepfunctions functions {f;g}, i.e. linear combinations of the indicator functions {xi = ,yo ,} of subsets {Q C X).
1 . Represent: f =
Cajx3;
ck bkXk; with coefficients
3
g=
j
ie . a . = r .e 3; b, = pkei4,. 3
3
2 . It is convenient to use reciprocal Hijlder indices:
a =:;; p = I. 9' a 0- Po'
a,=+ 90
al, pl; a(.) = (1-z)ao+zal,
etc.
Introduce two families of functions (one-parameter deformations of f ; g) depending on complex z , - Cr+z)/aeiejxj; gz = c p j ,1-@(z)l/(1-@)ei4kxk, z -
3
where a,p in the denominators of powers of {r .}, { p , ) mean a ( s ) and p ( s ) for some real 0 < s < 1 (without loss of generality assume 11 = 1; 11 g 11 l-p = 1).
f]la
3 . Consider a holomorphic function F(z) = of the strip: 0 < 32er < 1. For
s T(f,)g,dz,
and estimate it on both sides
express F ( i t ) = $T(f)Gdz,in terms of new functions f = Crjao/aez(ej+...) x j ; G = C P k(l-@O)/(l-@)ei(~k+"')xk.
Apply the (ao;C)o)-estimate to the pair { j ; j } ,to show
Is
m
GdzII...
I~ollflll~aollGll'~@o=~o.
1-
Similarly, on the other side of the strip
IF(l+it)I
=I S T ( f , + i t ) S l + i t d ~ ( I M lIlf Ilo119111-p=M1.
Conclude the proof by the Three Line Lemma (problem 2).
52.2. Fourier transform on function -spaces 4. Outline the proof of the Interpolation Theorem (specifically the construction of holomorphic function F ( z ) ) , for a bilinear map T.
5. Compute LP-norm of the Gaussian and check the Beckner estimate of the YoungHausdorff constant (2.1). 6. Hardy classes on the half-plane [Hofl are defined by (2.18) with an L2-norm (or, more
generally, LP-norm, 1 < p < m). i) Hardy functions are known t o have the (limiting) boundary values in Lz (Lp):
1
F(z+ic)-F(z), as (-0, and F ( z )
1 5 C.
This holds for yet more general harmonic functions u(z) in {Sr> 0) (solutions of the u = 0, while F ( r ) solves only the Cauchy-Riemann Laplaces equation: Au = equation: 4 F = O!), as a consequence of the Poisson representation of u (§2.2),
I Py(z-r)u(T)dr, Y =r(22+y2)'
u(z+iy) = (Py*u)(z)=
with the Poisson kernel P&Z)
The latter forms an approximate unity {Py}, as y-0, i.e. P , 4 , and Py*u+u, for all u E LP (see problems 2,5,7 of 52.1 and problem 2 of $2.2). ii) Furthermore, analytic (Hardy) functions F ( z ) are represented through their boundary value by the Cauchy integral:
a convolution of the Cauchy kernel K , ( z ) = 2 4 2 parts are
WK = py(z) =
-+@(z) 2r(zZ+y2)
SK =Hy(z)=
2r(z2+y2)
+ iy)'
whose real and imaginary
- Poisson kernel;
+m;
where H denotes a distribution P.V.(i), called HilberZ transform (see 32.2, and [SW]). iii) The FouGer tran_sforms of both distributions, Hilbert H and Dirac 6, on R are easy t o compute: 6 = 1; H ( t ) = sgn t (52.2). iv) Use (i-iii) t o show that, any Hardy function F(z) in H is given by Fourier transforming an Lz-function f ( t ) on the half-line (2.18). The latter provides a crucial step in the proof of the Paley-Wiener Theorem.
7. Smooth regularization of distributions. Take a smooth compactly supported function h ( z ) , consider a family of dilations: h,(z) = r - " h ( z / f ) , and show that convolutions {h,*F} yield a family of regular functions, that converge t o F: ( f c 1 u)-(F I u), for any test-function u.
8. a ) Show that (6 o 4) = y
d ( z - y ) ,
E ToI 4 " V ) I
for the Dirac delta on R, sum over all zeros
ro= {y} of functiond. b) If$: R"+R, then (6 o 4) =
dS
4(z)= 0vgl' I
where d S is the Euclidian surface
element on the level set {z:d(z) = O}. 9. Bessel and Riesr potentials are formally _defined as Fourier transforms of tempered distributions: B , ( ( ) = (1 1 ( I z)-s/2, and R , ( ( ) = I ( I -'on R".
+
81
52.2. Fourier transform on function-spaces
82
i) Use polar coordinates and the Bessel-function representation of is = Jv(r ...) (on the r-sphere) to compute B,( I z I ) in terms of the modified Beasel function K,,. A Hankeltype integral relates Bessel functions J and K of the 1-st and 3-rd kind:
In our case v = n- 1; p = S- I. 2 2 ii) Check that Fourier transform takes homogeneous functions/distribuGons of degree s: f(tz) = t-'f(z), t > 0, into homogeneous functions of degree (s-n): f(t() = t""f((). Apply it to the Riesz potential to show, R,( I z I ) = C(s) I z I Compute constant
'-".
C(S).
10. Show that the Lm-norm in the definition of Schwartz functions Y can be replaced with any other Lp-norm. Show also that operations (1+ I z I 2, and (1 - A ) can be interchanged in the definition of 9. (Hint: apply a version of the Hardy-LittlewoodSobolev inequality (2.17) for LP-Sobolev spaces: 36: = (l-A)-'I2LP, to show that 36; is embedded in LQ, for = - & and
II f 11LQII (1-A)"'2f
IILP'
$2.3.Applications o f Fourier analysis
83
v.3.Some applications of Fourier anal+. Here we picked up a few selected topics and applications of the commutative Fourier analysis: Central limit Theorem of probability; the Heisenberg uncertainty principle; Finite Fourier transforms, Bochner’s Theorem and the Mellin transform. The latter has application to special functions and differential equations in R“, and also in the harmonic analysis on SL, (chapter 7).
3.1. The uncertainty principle. The Heisenberg uncertainty principle of the Quantum mechanics will be discussed in 56.2 of chapter 6. It has a simple Fourier-
for a n y function $ in
Iz I
and
I< I
Rn, such that $ and
4 are both square-integrable
with weights
respectively.
Rephrasing (3.1) in terms of the derivative (gradient) a$, we get
In this form it becomes an easy consequence of the standard integration by parts formula, /z$# =
/z(i I $ I
2)1
=
-$I 14 I ’,
(1-D); or/V$.z$ = -
I31
$ I ’(Vex) =
-51
I $ I 2, (n-D),
and the Cauchy-Schwartz inequality applied to the LHS,
I/z$.v+(/
lv$,~yz~
Izdi~)li2(l
Relation (3.2) implies that one can simultaneously localize functions $ and
4 at
I $ 11L2
(0) maintaining the Lz-norm = 1 (LHS), in other words, localizing supp($) in zspace will ‘‘stretch” its Fourier transform $, hence increase gradient, and vice versa. In quantum mechanics operators z: $+$, and &$+a$, represent the “position and momentum” of the quantum particle, while integrals
(1 I4
V d x ) and (IlW 1 2 4 ) estimate errors in their measurements. So the Heisenberg principle prevents
simultaneous precise determination of both operators to any degree of accuracy.
3.2. Finite Fourier transform. Here we shall briefly discuss the Fourier transform on finite commutative groups. The simplest of them is a cyclic group Z,. Its characters
52.3. Applications of Fourier analysis
84
coincide with the n-th roots of unity ( x Q $ j ) = w J } , where w = e z p ( F ) . So the dual group Z, N if,, and t,he Fourier transform on Z, becomes
c5
s:f(j)+?(e) =
f(j)w
3e.
(3.3) Characters { x Q= (wQJ)} form an orthogonal system of eigenvectors of the o53
n-1
translation operator R:f(j)+f( jtl),on space L*(Z,)
N
C"
, with norms
(x I x) = IIXIIZ = n. The inverse transform takes the form,
c
+i(j)= A
4-1:qe)
o
while the Plancherel formula,
a(e)w~l,
5 Q 5 n-1
llfllZ= ~lf(dlZ=AE I?(e,12=11?112, both being obvious consequences of the orthogonality of {xl}.
":I
We shall study now the czrculunt mat& A, that appears in many applications of finite groups (in linear algebra, probability, etc.),
[ :; :
... A = ... ... ... ... ...On-1 an a2
1 . -
each subsequent row of A is obtained by a cyclic permutation of the n-tuple (ul; ..a,,). Notice that matrix A is a convolution with function u(k) on Z,,
A f = a* f. So Fourier transform (3.3) applies to diagonalize A (find its eigenvalues and eigenvectors), compute inverse, determinant, etc. A specific model of such A is given by the
random walk on
Z,, where
az =a,
=f
(probabilities to j u m p from point (1) t o its neighbors (2) and {n) in a unit time), and the rest aJ = 0. More generally, we consider a stationary stochastic matriz':
C a , = 1 (aJ measures probability
aJ
2 0;
to j u m p from cite (1) t o { j ) ) .
Characters { A @ } form a complete set of eigenfunctions of A with eigenvalues,
c~,~J~,
A, = q e ) = - the l-th Fourier coefficient of function { ~ ( j ) l} ,= 0;1;...n-1. Diagonalizing A, via 4, 'Stationarity means that all rows of A are permutations of a single row, so the process is independent of the starting point: probability of "i+j"-jump depends only on the difference i - j E 2,: a . . = a(i-j). 13
92.3. Applications of Fourier analysis
85
allows to compute d e t A ; A-', all iterates A"', and yet more general "functions of A", B = f ( A ) . Indeed, any such B is itself a convolution with function b ( j ) , whose Fourier transform, b = f i?, so A
0
b(j)=T'[foi?] =
Sc e
foZ(e)wje.
As an application to random walk on Z,, we immediately find the probability distribution of the process after m steps (units of time), i.e. find entries of the m-th iterate Am. If p$:)+k denotes the probability of jump from j to j+k after m steps, then p$:)+k = s-'[xT].
Hence, pi:) =
ax
and P$:3)+k = 4
cosm (T); 2=e
e
c
cosm(@) cos(@).
e
Remark: More general finite groups G are known t o decompose into the direct product of cyclic groups, G cz
n Z,.
Hence, G =
nZ,
u G, and the Fourier transform reduces to
the Fourier transforms (1.3) on cyclic components of G. One interesting application of such finite Fourier transform is the random walk on the n-cube (problem 9).
3.3. Central Limit Theorem. Another interesting application of the Fourier Analysis is to r a n d o m walks on groups W , Z,or more generally, to sums of independent r a n d o m variables, S , = X,+X,+ ...+X,. Here we shall establish the Central Limit Theorem, that describes asymptotic distribution of S,, as n + w .
Any real random variable X defines a probability measure (distribution), d p X on R. Independence of two random variables X , Y means that their joint distribution d p x , y ( z , y ) on RZ is the product d p x ( z ) d p r ( y ) . So the sum of two independent random variables X + Y has distribution,
+x+,
= dPX*dClY (3.4) W. Another simple transformation of X ,
- convolution of two probability measures' on
scaling X + o X , results in the dilation of the measure dPx+dPx(3
(3.5)
We consider the sum of n independent identically distributed r a n d o m variables,
s,
x,,
= XI+ ...+
with equal probability distributions, d P x . = 4% 3
and ask about the asymptotic behavior of sums {S,} at large n. 'Convolution of L'-functions f*g on R" (or any group G) can be extended t o bounded measures {dm}, and other (compactly supported) distributions/generalized functions (see 92.2), via pairing to continuous test-functions
If).
$2.3.Applications of Fourier analysis
86
A typical example will be random walk on the lattice Z. Here variable X takes on values f 1 (jumps from
k t o k f 1) with equal probabilities
f, so dp = 36,+6-,).
Then the sum
of n variables, S,, measures the position of the walk after n time-steps, relative to the initial point.
It turns out that the limiting (large n ) behavior of sums {S,} carries some universal features, independent of a particular distribution dp. In what follows, we always talk about convergence of random variables {Y,} in the distributional (weak) s e m e , namely,
Yn+Y, if
I
fdpy
n-I
f d p y , for any continuous f(z)on W.
Theorem 1: L e t { X n } r be a sequence of independent identically distributed r a n d o m variables with mean: E = xdp(x), and the dispersion
I
0=
J
I(x-E)'dp = z'dp - E2 > 0.
Then n
(i) S, = EXi has n o limiting distribution, as n+w. j=1
.
(ii) $ , 4 , n
(iii) L E ( X I . - E ) + N ( E ; U ) = 0.
normal distribution with dispersion
J;;1
Rephrasing Theorem 1 we can say that the mean value of sum: E(S,) increases linearly with n, while its dispersion o(Sn-nE) u f i , grows as f i .
-
= nE,
The proof exploits the notion of the characteristic function of variable X ,
the Fourier transform of measure p. The characteristic function has the following properties:
A ) d(() is continuous and positive-definite 7, for all tuples
{t1;... tn} in R and {al; ... an] in C; also d(0) = 1; Id(€) I 51.
B) if measure dp has all moments to order k, i.e. 'The converse result is also true (see Bochner's Theorem below). is also true. Namely (problem 12): any positive-definite continuous function on group R; R", or more general commutative G , is the Fourier transform of a positive (probability) measure.
52.3.Applications of Fourier analysis then function
I
I z J I dp < co, j=O;
d(<) if k-smooth, and its derivatives:
87
1; .A,
#(O)
= iE; d"(0) = -o-E2.
C ) For any sequence of random variable { Y k } measures dp,
iff q5~n(<)+&(<),
n
uniformly on compact sets of R. The first two properties follow directly from the definition of 3 and positivity of measures dp. The last one relies on the fact that families of exponential {e'"(} with ( varying over compact sets D, approximate all continuous functions {f} on compacts in {z}. So the convergence of characteristic functions is equivalent to the weak convergence of measures. We consider characteristic functions
{dn} of S,.
By (3.4)
d"(0 = %*P*...*P) where
d(<) = b(() -
= d(O",
the characteristic function of a single variable distribution dp.
4 1, i.e. p = 6. This proves (i). Next we apply (3.4) and (3.5) to the =ergodic mean" AS,,, Property (A) shows that
d(<)" does not
converge to a continuous function, unless
ds,/,(€) = d(4". Taking the Taylor expansion of
d(() at
{0}, we get
{ 1 + iETi € + O(n-')}"+e'tE= '5(15,),as n+co. This implies by (C), dps /,,+bE,
(*r,
as claimed in (ii). Finally, Fourier transforming
of variable Y , = L ( S ,
distribution-measures dpy
- nE),
we get by (3.4)-(3.5),
J;; where 4 denotes thne shifted measure d p ( z + E ) , with the zero mean,
I
z d p ( z + E ) = 0.
Taking the Taylor expansion of 1 €2
(1 -lo,+O(n
-3
d(<) at 0 we get by (B)
11n+,-oE
2/2
(e-z2/2u)
=9
- the Gaussian.
Sequence of measures {dp, } converges, as above, to the normal distribution, QED. n
We shall illustrate the foregoing argument with an example of the random walk on the lattice aZ (problem 8). Then measure d p = f ( ~ 5 ~ + 6 - has ~ ) , characteristic function $6 = corn( = (1
jxdp = 0,
with mean: and dispersion: Hence, the limit:
-$[ "...},
0
= jx2dp = 2.
(1 - 2n a2t2
+ ...) +e -a2(2/2*
88
52.3.Applications of Fourier analysis 3.4. Positivedefinite functions and Bochner’s Theorem. In the previous part we
I
encountered Fourier transforms of positive (probability) measures q5(€) = e-jZ ’ (dp(z), which are positive-definite on R (or W”) in the sense of definition (6). It turns out that all positive-definite functions are obtained in this way. Bochner’s Theorem: Any continuous positive-definite function q5(()
on
W” is the
Fourier transform of a positive (probability) measure,
q5(() =
j
ei= * % p ( z ) .
Two direct arguments, outlined in problem 5, utilize spectral theory of selfadjoint operators, and regularization of distributions. There is yet another fairly general argument, based on the convezity and eztreme points. It applies to all locally compact groups (and even more positive functionals on Banach algebras). We observe that the set of all continuous positive-definite functions {q5} is a convex cone 36 in space e(W”), aq5 b$ is positive-definite, along with {q5;$}, for any
+
pair of coefficients a , b > 0. When normalized at {0}, subset 36, = {q5:q5(0) = 1 ) becomes a convex compact set in the Banach space
e.
c 36,
Any convex compact I< in a
+
Banach space has the set of eztreme points, e z ( K ) = {$:$ # adl ( l - a ) & } . By the general Krein-Milman Theorem, any point inside K can be represented by a “convex linear combinations” of extereme points. In other words, for any q5 E I<, there exists a probability measure d,u&$), supported on ez(K), so that q5 is a baricenter of dp,
(3.7) Applying this result to positive-definite functions one first shows that ex(%,) consists precisely of all characters {$,(()
= eiz‘t} (“extreme directions” in 36 must be
translation-invariant!). Hence, by the Krein-Milman, any q5(() = je-jr
*
(d,u(~),QED.
Remark The Bochner’s Theorem has a far reaching extension to noncommutative groups, known as Gelfand-Naimark-Segal Theorem. The latter characterizes positive-definite functions on G in terms of matrix-entries of unitary representations. The extreme points of % are then identified with irreducible representations, and formula (7) gives a decomposition of a unitary representation Td, associated to 4, into the direct sum/integral of irreducibles.
3.5. Mellin transform. Mellin transform of function f(t) on the half-line t E [O;m)
defined by integrating f against a family of multiplicative characters {X,(t) = t S ; sE C} with respect to the multiplicative-invariant measure d p = $ on R+;
is
$2.3.Applications of Fourier
89
analysis
t,7 E R+,
d p(t7) = d p ( t ) , for all
m
-m:j(t)+3(s) = p ( t ) t s - ' d t .
(3.8)
0
So A can be viewed as Fourier transform on the multiplicative group R+. Its inverse 31t-l is obtained by integrating T ( s ) over the imaginary line: {Res = b } i.e. { s = b iu},
+
F+f(t) = &Jmf(.)t-Sds.
(3.9)
b-im
A few important examples of the Mellin transform include 1) r-function:
r(s)=
T
e-tts-ldt
= ~ t L [ e - ~and ];
0
7
2) ~ - s r ( s= ) e-"ts-'dt
=~ [ e - ' ~ ]
0
the latter can be also viewed as the Laplace transform, L:f(t) = ts-'+F(X) = ~ ( S ) X - ~ ; W
3) r ( s ) r ( l - s) = /&t% 0
=
-+&I
4) Bessel function of the 3-rd kind (McDonald):
I%-$('
W
K,(r) =
+ l/t)"tu-ldt = At+v[e-i(t
+ l/l)z]
(3.10)
0
5) Riemann Zeta-function:
Formulae (1) and (2) are essentially the definition of r; function I{, of (4) is directly verified to satisfy the modzjied Bessel equation (problem 10); two remaining relations (3), (5), however, would require further explanation. The former (3) is based on the so called B-function,
for special values p = s, q = 1-s. The latter exploits formula (2) and the definition of Riemann zeta-function.
Let us remark that all examples (1)-(5) of Mellin transforms involve integrals of the form
$2.3. Applications of Fourier analysis
90
m
Z(s) =
F(t) tS-'dt, 0
with certain meromorphic/rational function F. Such integrals can be evaluated by the Residue Theorem, combined with the branch-cut along the half-line R+, Z(S)
=2 1-e2"'8 %
Res,j(F)z;-';
sum over all residues (poles) of F.
Fig.1: Branch cut and residues of e-z/2
F(z) = 2 sinh 212 in the Zeta-function transform. Here poles: zk = 2 r i k ; k = f 1; f 2; ...
Thus we compute the right hand side of (3) and (5), and find (3.12) ) s - l ( z b ) " z . - ](5) &[ A 2 sinh t / 2
- sin(?)
As the result we derive the functional equations for I' and Z-functions, which relate their values at {s} to those at {-s}, (3.13)
The Mellin transform, in general, and functional equations (3.13), in particular, play extremely important role in the analysis of both functions, numerous applications (see
r
and 2, and their
[Tit];[Ven];[Ter];[HST]). We shall see m in the next section,
and later in chapter 7 (representations and harmonic analysis of SL,). The standard notion of Zeta-function (11) can be extended to Zeta-functions of differential operators L (e.g. Laplacian on a manifold),
Z L ( s )= tr(L+) =
C Ale-';
{ Ak}-eigenvalues of L.
Analytic continuation (regularization) of Z L ( s ) , particularly, its zeros, poles and residues in C, contain some interesting spectral and geometric information about operator L (57.6). Among other things it allows to define a seemingly absurd (but highly useful) entity, the "determinant of L", det(L) = e-' one finds from (3.13), Z(0) = -f; Z'(0) = -$ln(27r).
I
('1. For the standard Riemann Z
52.3.Applications of Fourier analysis
91
Problems and Exercises: 1. Work out an example of the random walk on Z with uneven probabilities: p (jump to the right) and p = 1-p (jump to the left). What is the most likely position after n (large) steps? What is the probability t o find oneself in the range [ ( p - q ) n + a f i ; ( p - q)n P f i ] after n steps?
+
2. Apply the finite Fourier transform to the random walk on the n-cube
11
I, x
... x Z,
x
to
find P$:$~. 3. Verify that function K = K , ( z ) of (10) satisfies the modified Bessel's equation in variable I, a K " + $ K ' + ( l + Sz) K = 0. 4. Evaluate integrals (12) and establish functional equations (13) for
r and Z.
5. Prove Bochner's Theorem in two possible ways.
i) Using spectral theory of self-adjoint operators (Appendix A). Hint: show that the quadratic form, determined by 4,
(S,?
I 3)= J I d(t-~)3(O%dtdq > 0,
is positive for all functions f E 2'
nA, hence defines an inner product on A n L', II f Ib = (043I 3).
A
,
.
The regular representation, R,f = f(<-u), is unitary in Ailbert space Xd = 11 lidclosure of A n L ' . Furthermore, space %+ has cyclic vectors: any f, whose Fourier transform f(z)# 0, for all 2. Observe, that Bochner's Theorem is equivalent to the spectral decomposition (existence of spectral measure dp) for a family of commuting unitary operators {R,) on Xd, or for their infinitesimal generators (Appendix A). ii) Direct prpof of Bochner's Theorem on groups R"; Z", exploits a regularization of distribution d(z) by an approximate unity (problem 2): pJ2)
= €-"p(f),
p(z) 2 0; J p d z = 1.
Set Jc = pcq& and show that {$<} defines a family of positive linear functionals on continuous functions e(R") (or T"),
I(i,lf)l 5d(O)llfIlm; ( & l I ) Z O , Conclude that
forf20.
4 =€lim4 04,is a finite positive measure d p on R" of 11 p 11 = d(0).
52.4. Laplacian and related differential equations.
92
52.4. Laplacian and related differential equations. The Fourier transform provides an efficient tool in the study of differential/difference equations on R”; T”, due to its diagonalizing properties. Solving a differential equation is then reduced t o an algebraic problem, followed by the Fourier inversion. The utility and power of Fourier analysis will be shown by the basic examples of differential equations (heat, wave, Schriidinger), associated t o the Laplacian - the most symmetric of all differential operators on R”. In all case8 we shall construct explicit fundamental solutions (Green’s functions) and analyze their properties. After the freespace Green’s functions are constructed and analyzed in via the Fourier techniques in part (I), we turn t o the boundary-value problems in special regions (part (11)). Our analysis will emphasize groupsymmetries of the problem (discrete and continuous). One such techniques exploits the reductions (separation) of variables t o bring the problem t o exactly solvable ODE. Another one, known as the “method of images”, gives solution of the boundary-value problem expressed through the free-space Green’s function, subjected t o a finite/discrete group of transformations. Our third example, the ball problem, exploits yet another conformal symmetry of the Laplacian.
4.1. The role of Fourier transform in differential equations is that 9 diagonalizes all constant coefficient differential operators: L = a , < a(,Da, ( D = +a), on R”; I Tn.Namely,
TL[u]= a(<)G((),for all
{u},
or TLT-’ = a(<).
So 4 takes L into a multiplication with a polynomial function a(() = C a a t a , called symbol of L. This clearly follows from the eigenvalue property of exponentials {eic‘”} for any such L. In fact, {eiz.(} form joint eigenfunctions of all such operators8, with symbol a(() serving as “joint eigenvalue”. In many applications operator L is the Laplacian - A , or more general elliptic operato? L = - v ‘ p v q, with a positive
+
definite coefficient-matrix p = (pjk). A few basic examples of differential equations associated to operator L include:
I. Elliptic: -(L
+ X)u = f; (called Helmholtz, or “reduced wave” equation)
8 0 n torus 1” characters {erm*’} are honest L2-eigenfunctions, while on R” they represent the generalized eigenfunctions (bounded, but not L2). ’We keep a convention of writing elliptic operators with sign (-), to make L a positive operator: (Lf I f ) =
J C p i j ( a i f ) ( a j f ) d z 2 0, for all {f}.
93
$2.4. Laplacian and related differential equations
11. Parabolic (heat-difbion): a,u
+ L[u]= f; for u = u(s;t )
111. Elliptic equation in n+l variables: (a2 - L)u = f (called Poisson); IV. Hyperbolic (wave): utt + t [ u ]= f(s;t) V. Schrodinger: atu = iL[u]. Each equation must be supplemented by an appropriate boundary and initial
conditions. On regular boundaries (hyper-surfaces C c R”) one usually takes one of the 3 basic conditions: u I = ... (Dirichlet); d,u I r:= ... (Neurnann); or more general
+
I
( a ,LM,)u = .... (mized). The condition at (00) (“singular point”) typically require u to be bounded, or to have a prescribed asymptotics (incoming/outgoing “radiation
condition”). The initial conditions depend on the type of equation. One has u 1 t=O = uo (for elliptic/parabolic problems 11-111; V), and Ilt=O uo; (for hyperbolic IV).
{;
I
t=o - u1 We recall the general definition of the Green’s function, for a boundary value problem t
M[u] = F; in D . (4.14) B[u] = h; on aD’ where M could be any differentiafoperator in a space (or space-time) region D (e.g.
M = - V . pV+q; or
at2- V . pV+q),
and E a suitable boundary/initial conditions on
a D (e.g. B = a + pa,). The Green’s function (fundamental solution) of (4.14) is a n integral kernel K(z;y), or K(z,t;y,s) (possibly distributional) on D x D , that satisfies the differential equation M in z-variables with the 6-source at {y}, (also the formal adjoint equation in y-variables with the 6-source a t {z]), and the zero boundary conditions,
BJK] = 0
M,*[K] = ~ ( z - Y ) By[K] = O *
For instance, heat-problem (11) has operators M = , = 8,
+ L,,
M*,, ,=
(4.15)
- a, + L,,
and
the Green’s kernel K(z,y;l-s). Function K of either one of problems (4.15) allows t o solve a nonhomogeneous system (4.14) with an arbitrary RHS { F ; h ] ,
+
u(z) =
K(z;y)F(y)dy P(z;y)h(y)dS(y); dS-surface area element on aD. J D So the Green’s term K gives a contribution of the “continuously distributed sources”
F I D, while the so called Poisson term P(z;y) on D x a D , those of the “boundary sourcesn h
I
The Poisson kernel P(z;y) is not independent of K, in fact, K is
constructed explicitly in terms of K and the boundary operator B. For instance, the Dirichlet condition for
L = -V .pV + q, yields
P(z;y) = p(y)a, K(z;y) (normal derivative of K in y o n the boundary), Y
while the Neumann condition gives
52.4. Laplacian and related differential equations.
94
P(z;y) = p K ( z ;y), for z E D;y E aD. Similar constructions, based on Green’s identities could be given for arbitrary M and boundary operator B.
I. Formal solutions and frespace Green’s functions. We shall construct fundamental solutions (Green’s functions) for each of problems (I-V) in
Wn, by first writing its formal solution, then applying the
Fourier analysis. Formal solutions of problems (I-V) can be written in the form of certain ‘functions” of L”. Thus we get
u = (A
+
u = etL‘[uo] and similarly,
+ L)-’[f], for the elliptic problem (I);
1:
e(t-s)L[f(...;s)]ds, for the heat-problem (11),
u = eitL[uo];for the Schrodinger problem (V). Here uo represents the initial value of u,and f - its RHS, the “heat-sources”. Elliptic L‘n+l”- problem (111), is solved by
*
u=e 4 [ u 0 ] .
< 00 (unbounded range), one takes the negative “forward-stable” exponential {e-‘...}, while finite time-intervals, a 5 t 5 b,
So in the half-line case: 0 5 t
with 2-point boundary condition require a suitable combination of 4 exponentials. The wave problem (IV) is solved in either of 2 forms,
The choice depends on the type of the LLboundaryconditions” in t-variable, either “initial” at t=O, or “asymptotic” at t = f00. In all cases (I-V) fundamental solution { d ( L ) }is given by an integral kernel,
K ( z ;y), called the Green’s function of the problem. For constant-coefficient L these are convolution-integrals K = K(z-y), due to translational invariance of operator L. The Fourier transform yields explicit form of kernels K ( z ) , z = z-y. By definition of K we have to find a distribution
K ( z ) , z=z-y,
that satisfies the equation,
luuFunctions of L” for a self-adjoint operator, could be given precise meaning via diagonalization (spectral resolution) of L. Given a diagonal (multiplication) operator D:d(A)+Ad(A), on Lz(dp)-space of functions/sequencea {d(A)), function f ( D ) is also a multiplication operator, f ( D ) :d+f(A)d(A). But any self-adjoint L is unitary equivalent to a diagonal operator: L = U D U - ’ , then f(L)= U f ( D ) U - (Appendix A).
’
$2.4. Laplacian and related differential equations
L [ K ]= 6(z-y).
95
(4.16)
Applying ’3 to both sides of (4.16) we get
where aL(<) is the symbol of L. So we immediately get the transformed Green’s function 1 R(t) = -
(4.17)
ad€)’
in other words “symbol of K” is inverse to “symbol of L”. Then it remains to Fourier-invert l / a L ( ( ) to recover K. The Fourier-inversion, however, involves the techniques of complez analysis, distribution theory, and also
reduction of variables via group-symmetries, as we shall demonstrate below. 4.2. Laplacian is the most symmetric of all differential on
Wn as it
commute
all translations and rotations (Euclidian motions). In fact, A generates all such operators, in the sense that any operator M on a suitable functionthat commutes with En is a convolution with a radiallyspace (LZ;Lf’;ew), symmetric function M ( I z I ),
WuI=
pc I
5
- Y I MY)dY.
(4.18)
The Fourier-transform of (4.18) gives a multiplication-operator with another radially symmetric function
% = F( I t 1 ),
u ^ - q I t I )u^, where F ( p ) is related to M ( r ) through the reduced Fourier-Bessel (Hankel)
transform (52.1-2.3). Since, p2 is the symbol of - A , we can call M a “function of A”, M = F ( m ) .The proof of (4.18) is an easy application of the Fourier transform. It contains 2 statements. Theorem 1: i) A n y translation-invariant operator M on a suitable function-class
(e.9. Schwartz) is a convolution: MR,u = R a M u , for all a E Rn, and all u E Y,+ M [ u ]= f * u , f o r some distribution f .
ii) Rotationally-symmetric convolutions, M T , = T,M, g E Sqn), are given by radial functions, f = f( I z I ). Indeed, any such M is taken by 9 into an an operator
% = 9M9-’,that
commutes
with multiplications by characters { e i b * t},
--
f i e i b * = eib ’ €I@, for all b E R”. A-
Hence
A
fi commutes with all multiplications, M h = h M , h in the Fourier-algebra A. But
52.4. Laplacian and related differential equations.
96
algebra A contains sufficiently many functions, so any operator, commuting with A must
= F ( ( ) , so Mu = k*u. The 2-nd statement is obvious, as 5
itself be a multiplication,
takes radial functions into radial.
Our discussion will be limited to the Laplacian, L = -A, although many results and techniques extend to more general (2-nd and higher-order) operators. The corresponding Green's functions then become inverse '%transforms of certain symbols. i) resolvent kernel: RX(z)= (X+L)-
4(, ii) heat-semigroup (Gaussian): G,(z) = etL has 9[Gt] = G ( t ; t )= e-t I ( I 2.
has Fourier transform P(t;t )
iii) half-space Poisson kernel: Pl(z)= e-td, = e-,
(a semigroup, generated by
a)
I.
JL JL'
sin I
iv) cos- and sin-type 'waue propagators": U , = c o s t a ; V , =*-
U ( z ; t )= ( 2 ~ ) - ~ j c o Is [( 1t )e'c'zd(. sintI(1
.
V ( z ; t )= ( 2 ~ ) - " 1 ~ e ' ( ' ~ d ( . or the unitary group { e
generated by
fi.
v) Schrodinger propagator W, (a unitary group generated by L),
W, = (21)-n/ei(t I t I 2+t.z)dt. In some cases the Fourier-inversion is easy and straightforward like for the Gaussian and Schr6dinger propagators (problem 5 of 52.1)'
L
I
I
I
The latter is obtained by analytically continuing G,to imaginary time, t-it. 4.3. The resolvent kernel R(z;X)is expressed through the Bessel functions of the 1-st and 3-rd kind [Leb];[Erd];[WW]of order Y = - 1,
5
(4.20)
In the limiting case, eigenvalue X = 0, the Green's functions (-A)-' with the classical Newton's potentials:
coincides
$2.4. Laplacian and related differential equations
97
To derive (4.20) we notice, that all kernels (i-v) are radial functions of z , an easy consequence of the underlying rotational symmetry of A. Writing integral R(z;A) in polar coordinates
I z I = p;
I € 1 = r, we get (4.21) The inner integral gives the
80
called Poisson representation of the Bessel function J,, u =
5- 1,
Hence we get in (4.21)
The latter integral can be evaluated by complex integration in the r-plane of holomorphically extended Bessel function J,, whence comes K , ( z ) = J,(iz). We shall illustrate the foregoing argument in R3. The inner integral r r p cos
2sin rp (4.22) sinode = -. rP ' since J 1 , 2 ( z ) is an elementary function &. We substitute the RHS of (4.22) in the outer je.
0
Jnt
4.4. The half-space Poisson kernels. Next we turn to the Poisson kernel Iin the half-space W?+'. The 1-D Poisson kernel is computed directly
- e-tl t-
(problem 5 of 52.1),
The multidimensional case is, however, more subtle, and could be derived via the so called subordination principle, which allows to express the Poisson-semigroup, generated by
fi,in
terms of the heat-semigroup, generated by L, (4.23) The derivation of (4.23) is fairly straightforward, via 1-D Fourier transform (problem 1). In the case of Laplacian L = -A on Rn, we get the Poisson kernel P ( z ;t) in RT++' in terms of the Gaussian { G t } . By (4.24) and an explicit form of the Gaussian (4.19) we find,
98
$2.4. Laplacian and related differential equations.
P, = tn(1 +
Cn Cnt ' I I 2 / p ) ( n + 1)/2 - ( p + I I 2 ) ( n + 1)/2
(4.25)
(4.26) is then given by the convolution integral: u ( z ; t )= P l * f , with kernel (4.25). As an application of (4.19), (4.25) one can show
Theorem 2 Solution u ( z ; t ) of the elliptic Dirichlet problem (I.%), and the initialvalue heat-problem: ut = Au;u I t=O = f, converges to the boundary value f(z)as t+O.
The proof is outlined in problem 2. After the Dirichlet Poisson problem is solved we can proceed to the Neumann problem for the half-space Laplacian, (4.27) The formal solution of (4.27)
-f=3-t'%
21 = = &t[fl, is given by the conjugate Poisson erne1 related to P , via
co
&(...;t ) = / P (...;3) ds, and P = $. d
(4.28)
t "Let us remark that Green's functions (i-v) are related one t o the other through various transforms. For instance, the resolvent RA is obtained by Laplace-transforming the heat-semigroup G t , (A
+ L)-'
do
=
1e-"-Ltdt, 0
while (4.27) gives a similar representation for P,. Formula (4.27) exemplifies such relations. It represents a special case of the so called Sonin-Mehler integral representation of modified Bessel functions of the 3-rd kind (McDonald functions) OD
K J r ) = i J e t p [ - ( t + t ) i ] t-'-'dt. 0
Precisely, (4.28) correspond t o order u = elementary, K 1 / J r ) = &e-'.
i, the
only case, when Bessel-function becomes
99
52.4. Laplacian and related differential equations
00
as a consequence of an identity, L-le-tL = Se-sLds. Substituting (4.28) in (4.25) we t find
So the Neumann problem (4.27) is solved by the convolution kernel ti(";
t ) = &(...;t)*f.
4.5. Wavepropagatom. We shall use Poisson kernels P and Q, found in the previous part to find the wave-propagators, U ( z ; t )= c o s c t a , V =
9
= 0,
i.e. fundamental solutions of the propagation) with 2 different-type initial conditions:
( c = speed of
u I t=O = 0 ; for V . "tIt=o = 0' "1 I 1=0 = As in the Gaussian case this will be accomplished via analytic continuation of
I t=o = *. for u, or
*
the time parameter through the imaginary axis, t-it. The Fourier transforms of both propagators: n
U ( &t ) = cos ct
sin ct I < I I ( I , V((;t= ) 1c1 ' n
*-
as well as Poisson P , are easily inverted in 1-D. Thus,
turns into a Heavi.de jump-function, truncated at { fc t } . Similarly, the cos-propagator,
cos(ct I c I ) = ;(eict I c I + ,-ict I € I )
* u = ;{*(Z-ct) + *(z+ct)},
which is the sum of the right and left-traveling *-pulses, the classical d'Alambert solution. One has to compute two limiting distributions:
In dimension n = 2 both are given by certain densities,
$2.4. Laplacian and related differential equations.
100
{I
z-y I 5 c t } . So for each fixed t > 0 support of U(..;t); V( ...;t ) belongs to a disk of radius R = ct, centered at the source12 {y} (fig.2). In 3-D space we get supported in the light-cone
1
1
2ECt
I z I 2 + (c+ict)Z - I z I 2 t (r-ictlz} = lim (c2t2- I I 2 + E 2 ) 2 + qE2
2 t 244.29)
The RHS of (4.29) represents a limiting distribution of the family
{wd=*} I ’.
with a regular function f = fr(.z) = cZtZ- I z
From the basic distribution theory of
$2.2, we find,
where dS means the natural surface element on the level-set {f = 0). In other words, the limiting distribution bf is made of the “zero-level” &function of f , divided by a continuous density IVf(z) 1 . In our case (4.29) we get
Vt(Z) = k b ( c t 2 V,(z) = a,v = + ( 4*(ct)
12 1 2 )
=(6’47rct
12
I z I - C t ) t &a’(
I -Ct) I2 I -C t )
So the 3-D wave-propagator { V ; V }can be thought of either as distributions in supported on the forward light-cone C = ((2;t): I z I = c t } c R4, or as a t-parameter family of spherical &functions ( 6 * of radius R = c t } in R3. The SR* corresponding solution of the initial value problem: u I = f ; ut I = g, in 2-D and 3-D (2;t)-variables
are given by convolution kernels,
Remark: Similar expressions hold for higher-D wave-problems. Namely, the Green’s functions V ( z ;t); V ( z ;t ) are given by distributions supported either inside the light cone
C,{ 1 z I 5 ct} (even dimensions), or on the surface of the cone { I z I = ct}, as illustrated in fig.:! below (problem 4). The exacts solution of the n-D wave problem are given in terms of spherical means of function f,
1 M,f(z) = ,n-l wn-l
s
lyl =P
f(Z-Y)dS(Y),
-Green functions are often called point-sources, aa they describe solutions produced by sharp (&type) initial impulses.
$2.4. laplacian and related differential equations
101
where w , - ~ is the area of the unit sphere in R". Then the initial data u I, = 0; ut10 = f(z), yields [Hell u(2; t ) = 2 (n-2)! -0-2
1
[
~ { M 8 f ) ( x ) s ( t z -n-3 s2)~ds
(4.31)
In special cases n=2;3 one recovers solutions (4.30). Formula (4.31) demonstrates the
Huygens principles for solutions of the wave equation: any initial disturbance propagates inside the light-cone based on supp{f) with a finite speed c. Furthermore, in odd dimensions disturbance of any finite extent vanishes after a signal passes (sharp Huygens
principle), whereas in odd dimensions once a particular space-point is reached by the disturbance will remain forever, dying out exponentially fast with time. Indeed, derivatives of integral (4.31) in odd dimensions take the form
c
Pk(t)a:[M'f
1,
with some polynomial coefficients { P k ( t ) } .
Fig. 2. The forward and backward light-cones for the wave equation. The left figure (forward) shows a point-source solution (Greens function) at 3 differenl lime cross-sections. At each 1 the disturbance is localized either on the surface (sphere S R ( x ) of radius R = ct, centered at the source {t}), or inside the ball BR(z) = { yx I 5 ct}. The right figure shows the domain of dependence: any solution u(z;t) depends on the initial data inside the ball { I y-+ I 5 ct). There are different way to derive (4.31), all exploiting to some degree the underlying group symmetry of the problem. We shall mention 2 of them: [He121 (chapter 1) is baaed on the Radon transform on R" (and rotational symmetries), whole R. Howe [How] presents an elegant derivation baaed on the representation theory of the Heisenberg group (cf. $6.2).
We have summarized the main results of the section in the table at the end of the section. It. Symmetries and differential equations in regions of R".
After we obtained the free-space fundamental solutions in Wn, let us turn to boundary-value problems. The analysis below will emphasize group-symmetries (discrete
102
52.4. Laplacian and related differential equations.
and continuous), and whenever possible provide explicit formulae and constructions of Green's functions. Of course, our discussion will be limited to special classes of symmetric regions and differential equations.
(I = f(L) L; region Q), which
4.6. Symmetry reduction and the sou~cecondition. Green's function
clearly inherits all original symmetries of the problem13 (operator
allows to reduce the number of variables in the Green's function K . For instance, hermitian symmetry,
L = L*,implies symmetry of integral kernel, K ( z ,y) = It'*(y, z)
complex conjugate. Time-independence of L means that K = I'(t-s;
5,y)
-
are time-
convolutions kernels. Similarly, constant-coefficient, i.e. translational-invariance of L in R" yield translational-invariant (convolution) kernels K = K(z-y). Larger symmetry groups play the role for special operators L, like orthogonal rotations S q n ) for the Laplacian/Helmholtz operators: A , A + m 2 on Rn, and hyperbolic rotations 5 q l ; n ) for the d'Alambertian (wave operator) 0 = 6': - A , or the
Klein-Gordon (KG) operator: a: - A f m2.The corresponding Green's functions, It'(z) or K ( t ; z ) ,z = z-y, retain the symmetries of L, which allows to reduce Zt' to a single variable function I'(r), of the Euclidian or hyperbolic radius: T = I z I , or r = In both cases the radial function I' solves the reduced ODE,
d q .
which is singular at the end-points r = O;m, hence requires an appropriate boundary condition at both. The "source condition" refers to a proper choice of the boundary condition at a "singular (source) point". We shall demonstrate with 2 familiar examples. 4.7. Elliptic Caae. Laplacian/Helmholtz-type operator: L = -A
+ q.
The condition at (0)
represents the so called Upoinl-source",
,.
p - 1 ~ 1
()Ir=O==
1
(4.33)
2rnI2 - surface area of the unit sphere in R". To derive (4.33) we take any where Cn-l =---
r("/a
function u(y) on R", write L[u]= F(y), apply kernel K(z,y) to the equation, and integrate by parts (Green identity) over the complement R = R"\B,(r), of a small ball B , ( z ) centered at {z}. On one hand,
W -e shall not pursue any systematic study of symmetries in differential equations here, but refer to the recent book [Olv], which explores the subject in great depth (see also [Ovs];[Mil]). Our goals will be limited to exploiting some apparent and well know (geometric) symmetries, rather than investigate all of them.
82.4. Laplacian and related differential equations
103
but
First integral in the RHS vanishes, since L,[K] = 0 outside of the source. Remembering that
K = K ( Iz-y
I ), we get the second integral
f while the third
4
... = Cn-l F 1 K ’ ( e )(u(z)+ o ( E ) ) , Iz-yl
=€
... FT t ” K ( c )Au(z) +O,
as € 4 0 .
12-yI = c
Thus
+ {c” K(c)}Au(z),
u(z) = lim {C,-,r”-’K’(c))u(z) C+O
for any function u, which yields the source condition (4.33). Equation (4.32) for q = 0 takes the form
+V
K,,
K r = 0,
it has 2 solutions (regular and singular): { l;logr} in n
= 2, and
{1;r2-,} for n 2 3. So
K = c1 + c2rZ--n(or logr). The boundary condition for K at {m} eliminates the constant term, while the source condition
(4.33) yields c2, whence the familiar form of the Newton/Gauss potential, the Green’s function of the Laplacian,
A similar treatment applies to the case q # 0. But this time (4.32) becomes the equation of Bessel-type, y”+qyrfXy=O,
with X = q .
It is easily reduced to the standard (I-st or 3-nd kind) Bessel equation 2
Y ” + k Y ’ f ( l - E ) Y = O , oforder v=&, 2 r2
(4.34)
by the change, y(r)
= r-’Y,(Jir).
Thus solution of (4.32) with positive rn2 becomes a combination of the 1-st and 2-nd kind Bessel functions: K = r-”{clJ,(mr)
+ cZYv(mr)},of order
v=
9. The source-condition at (0) is
furnished by a non-vanishing coeficient c2. The condition at {co} should be chosen either as incoming, or outgoing “radiation” condition
The latter with a proper normalization at (0) uniquely determines K as the MacDonald function, M, = Const ( J , f iY,), so
52.4. Laplacian and related differential equations.
104
~
Similarly, for A-m’,
the reduced equation changes into the modified Bessel,
Krr + n-1 7 K r - m2K = 0, whose solutions are
I,
-
1-n r 2 emr, and K,
-
1-n
rTeWmr.
We obviously choose the exponentially decaying Kelvin function K,, which could be normalized
to satisfy the source condition a t (0). 4.8. Hyperbolic b. The wave and KGequations:
= utt - cZAu = 6;and (0f m2)u= 6, have all the symmetries of R” +
space-time, translations as well as hyperbolic rotations
So(1;n). The corresponding Green’s functions K = K(1,z) are reduced as above t o a single variable, the hyperbolic radius r =
d-.
They satisfy the Bessel-type equation
Kr,+:K,fm2K
= 0.
(4.35)
But unlike the elliptic case it is not easy to write down the proper %wee condition” a t r = 0. In the last section we used the Fourier analysis t o derive the wave-propagators, which yields for dimensions n = 1,2,3,
(4.36) In higher dimensions K becomes yet more’singular distribution, supported in the light cone
{ I z I 5 ct) or its surface { I z I = ct} (4.31), i.e. K = d ( r ) - a distribution in r = ct - I z 1, or hyperbolic r =
d m .T h e KGGreen’s functions satisfy the Bessel-type (4.35) equation
in hyperbolic r, and we can use solutions (4.35), as proper “source-conditions”. This yields
in spatial dimensions n = 1,2,3.
4.9. Discrete symmetriea and the method of images. The method of images
(reflected sources) applies to differential equations in regions of Rn, and other symmetric spaces (see 57.6 of chapter 7), obtained by “discrete symmetries” (reflections, translations, conformal mappings), like the half-space, quadrants, strips/ slabs, rectangles, spheres, etc. The Green’s function of the corresponding problem is obtained from the free-space Green’s function KO by subjecting KO to a (discrete) set of symmetry transformations, or by subtracting from
KO a suitable regular solution.
32.4. Laplacian and related differential equations
105
Namely, K Osatisfies the source (6 -function) condition, Lz[KO] = 6 ( x - y) in R, but it fails to vanish on the boundary, B[K,] # 0. So we want to find a family of regular functions {u(z;y)}, depending on the position of the source {y}, so that
Lz[u(...;y)] = 0 in 0; and Bz[u]I r = Ko(z;y).
(4.37)
4.10. Renectiona and half-upam problems. We consider the half-space R?+’ t
> 0},
and assume operator Z to be reflection-invariant: (z;t)+(z;
= { ( z , t ) : z E Rn;
-t),
for instance
L = 8: + A - the half-space Laplacian. The corresponding Green’s function depends on the type of boundary condition. For Dirichlet Problems the Green’s function KD
= K ( z ; t , s )= K,(r;t - s) - I f o ( z ; t + s),
represents the difference between the “sourcesolution” K , ( y ; s) (e.g. Newton’s potential) and the
“reflected source” K , ( y ;
- s)
(fig.3). The reflected source gives the requisite correction u in
(4.37). So the half-space Dirichlet Green’s function represents a convolution kernel in 2-variable, K ( z - y ; t , s) of the form
If = &{( I z 1 2
+ ( t - s)2)
n-1 -2
-( Iz
12
+ ( t +.)2) -F}.
This yields, in particular, the half-space Poisson kernel P ( z - y ; t ) , derived earlier in by the Fourier-transform methods
(4.38)
Let us observe, that solution of the Dirichlet problem: Au = F ; u I r: = f; in any region R with boundary C = 80, is represented by the Green’s identity, as
4.) = where
an denotes
/
nK(z;Y)F(Y)dY
+ / za,K(z;Y)f(Y)dS(Y),
the normal derivative of the Dirichlet Green’s function in y E C. So the
Green’s function K picks up the “continuously distributed sources” F over R, while its normal derivative P = 8°K - the Poisson kernel gives the contribution of the “boundary sourcesn.
Fig.%: Free-space Green’s functions at a source {y] and the reflected source {b} cancel each other on the boundary r, separating two half-spaces (Dirichlet problem). In the Neumann problem 2 “sources” add up, so the normal derivative vanishes on the boundary.
106
$2.4. Laplacian and related differential equations. For Neumann Problem the difference of two source8 is replaced with the sum
K N = Ko(z;t-s) +Ko(z;t+s), (so that normal derivative8 would cancel each other). As application we obtain the conjugate
Poisson kernel Q in the half-plane14,
found earlier by the Fourier transform methods. In a similar vein one constructs Green’s functions for other problems, e.g. the half-space Dirichlet/Neumann “heat problem” has K(z,y; t ) = G(2 - y;t)
- the difference/sum {y’
C(z - y’; t )
of two Gaussians with “sources” a t {y = (yl; ...y,))
and the reflected point:
= ( - yl; ...yn)}. Quadrants and other product-type regions are treated in a similar fashion.
Thus the quadrant {(z1;z2) > 0) Green’s function consists of 4 reflected sources: K(z,Y ...) = Ko(z - y) - K o ( z - y’)
- Ko(2 - Y”)
+ KO(Z + y);
where y’; y” represent reflections of {y) relative to the 1-st, respectively 2-nd coordinate axis. 4.11. Green’s functions in finite regions. Green’s functions of elliptic differential operators in
bounded regions S2 (compact manifolds) can always be expanded in terms of eigenfunctions {$k} of L, subject to a proper boundary condition. We leave t o the reader as as exercise t o write expansions for the inverse operator L-’, the heat semigroup e-tL; Poisson kernel exp(-t&); wave-propagators, etc. Let us remark that the Green’s function of any regular S-L problem on interval [O;T]can be represented in two different ways: via the eigenfunction expansion 15 K(z;y) = Cfdk(Z)iJ,(Y), (4.39) k or in terms of a fundamental pair of solutions: L[ul,,] = 0, one of which satisfies the zero
boundary condition on the left (ul I t=O = 0)’ and the other on the right (uz I T = 0). Given a fundamental pair {ul; uz}, the Green’s function is constructed as, (4.40)
For constant coefficient operators L =
- 8 + q,
two representations are related by another
version of the Mtthod of images. Namely, we take source y E [O;TJ,and “reflect it” infinitely 141ndeed, solution of the Neumann problem: Au = F; a,u identity, as
I = f; is represented
via Green’s
where K = K N is the Neumann Green’s function, so Q(z;y) = K ( z ;y), for z E 0 , y E E. 151f one of eigenvalues A, = 0, then kernel K (4.5) (sum over nonzeros A’s defines the so called modified Green’s ,function, i.e. inverse of L on the subspace spanned by nonzeros eigenfunctions {dk},so the product of operators L K = b(z - y) - v,bo(z)$o(y).
52.4. Laplacian and related differential equations
107
many times by the lattice Z = {mT}. So the free-space Green's function of Z on R.
is shifted and summed over the
corresponding Green's function
K on [O;T]has different representations depending on the type of boundary condition. Namely, the periodic bounda y condition (torus), gives ~ ( z , y= )
C~
+
~- y ( mT) z
= +C'ezp(i%$z - y)); (4.41) k with eigenvalues: {Ak = ( r k / T ) ,+ q: k E Z}, while for the Dirichlet/Neumann condition the m
source a t {y} is reflected to {-y},
and then the process is repeated periodically with period 2T.
So one has
The reader should recognize relations (4.41) and (4.42), as special cases of the Poisson summation for function f ( t ) = KO(I z - y I
+ t ) on the lattice Z = {mT}. Similar considerations
apply to the heat, wave, KG and other equations, and to equations on multi-dimensional tori. Two such examples are
4.12. Dirichlet Laplacians on triangles.
Rightisoseeles triangle makes up
i of the square. The (rx r)-square eigenmodes are made of
products of the even and odd Fourier modes { c k = cos(k +:)z}, and
y).
{ C k ( t ) s,(y);
So
consist
of
products
{srn= sin2mz) (in variables z
{ c k ( ~ ) c , ( ~ ) ) , { s k ( ~ ) s , o ) and
s k ( z ) c , ( y ) } . The former two (cc- and ss-types) correspond to different families
+i)2+
+
(m +$)'}, and {4(k2 m')}, so they never mix. The latter ( c s - and types) produce eigenmodes Gkrn of the form Ck(z)6rn(Y) 2 L,(')'k(Y), that vanish on
of eigenvalues {(k
sc-
{4krn}
diagonals z = f y, hence yield the triangle Dirichlet eigenfunctions. Equilateral triangle T makes lattice
r c C, spanned by
i of the diamond,
the fundamental region of the hexagonal
{ l ;a = e i S / , } (fig.4). Its Dirichlet eigenfunctions fall into 3 families,
according to the natural action of dihedral symmetry group D, = Z, D Z, on T. The latter has 3 irreducible representations: trivial
xo = 1, x1 = f 1 (trivial on rotations { 1;u;CJ:u= eizXl3}
and
equal -1 on any reflection), and a 2-D representation u, uu:((;q)+(u(;CJq) (for rotations {I;wp}),
and u,((;q) = ( q , ( ) , for reflection r about the z-axis (see 53.3). So family (I) consists
of eigenfunctions16, symmetric relative to the D,-action
, family
(11) are rotation-invariant, but
"Dirichlet eigenfunctions are uniquely determined by the values of their normal derivatives on , on all 3 sides of T. the boundary, so the description here refers to the values of a+
108
$2.4. Laplacian and related differential equations. mapped t o -(1, by reflections, finally (111) gives a pair of conjugate values { @ ; a }{-l;-l},{Sp} , on each of 3 sides of T (see fig.5~). Our goal is to link eigenvalues of A, to the well known eigenvalues of the f-periodic A on torus
c/r, A,
= I i k + a m I = k2 +m2- km.
Figure 4 demonstrates how each eigenfunction of family (I), (11) and (111) extends t o a fperiodic eigenfunction of A: for (I)-(11) is equal to the size of the triangle (the plane is cut into the union of copies of T, and 11, is taken into all images {T‘} of T with signs f , depending on the orientation of T‘. That same reflection procedure applied to a 111-type eigenfunction requires to triple the period. Since each periodic eigenfunction is a combination of Fourier modes {exp[i(kz+m%(E::y)]:k,m E Z}, we obtain the eigenvalue-spectrum of A,,
t o be made of 2
sequences:
1 Atm
= k2 + m2 - km (for 1-11) A’km = g(k2 + rn2 - km) (for 111)
I
Fig.6 below demonstrates a 11-type eigenfunction as a combination of rotated and reflected exponents, (1, = XI - x2
+ x3 - x4 + x 5 - xs. Fig.4 shows cubic roots {w,G = w z } of 1, along with their negatives { a = - w-; E = - w } . Numbers { 1;w;G } give 3 irreducible characters of subgroup Z, c Q,, while reflection (complex conjugation) defines a 2-D representation of D,. Fig.5 illustrates a periodic extension of eigenfunctions of types I, 11, and 111. A Dirichlet eigenfunction (1, in T is determined by its normal derivatives {fi =a,$, on the i-th side Si, i = 1,2,3}.
We schematically represent each I f i } by a curve along Si; the portion of the curve inside the triangle corresponds t o a positive derivative a,f, while the outside portion t o the negative derivative. For type-I all functions { f i } are equal and symmetric with respect t o the center of the side.
So given a type-I (1, in a “+”-triangle, we reflect it t o a neighboring “”triangle, changing the sign to+-$, and repeat the process. The resulting fperiodic function on C is clearly an eigenfunction of A. For type-I1 11, (b) boundary function f is odd on
109
$2.4. Laplacian and related differential equations
each side Si and repeats itself under rotational symmetries with { 1 ; q Z }. Once again reflections +-$, from T + t o T - across {S,;S,;S,} produces a r-periodic eigenfunction of A in C. The type-111 (c) is somewhat more complicated. Here we pick a pair of complex conjugate functions {+;$}. Then one ca? easily verify that the boundary functions {fi;f i} have absolute values I f l I = I f z I = I f JI = p (an even positive function), and differ one from the other by a_ factor w , f 2 = wfl; f, = 5 f l (and similarly for f .). If we choose fl.(left side) to have phase a= then f 2 (horizontal) have phase - 1 , while f, = E p. The reflection $+-4 about side S, takes triple { a ; - l ; Z } into { u = - E ; + l ; G = a } , and the process has to continued 3-times along the (horizontal) line t o get the original triple { a ; - ; E } . So type-I11 $ yields a 3r-periodic eigenfunction of the Laplacian in C.
&,
Fig.6 shows a type-I11 eigenfunction constructed a combination of 6 exponentials: 4 = x1 - XZ x3 - x4 x 5 - x6.
as
+
+
4.13. Conformal symmetry: Green's function and Poisson kernel in the ball. The group-symmetries used so far in our discussion involved space-translations (continuous and discrete); rotations (spherical and hyperbolic) and reflections. Our last example will exploit yet another transformation, the conformal symmetry of the Laplacian, to study harmonic functions in the balls B , = {
12
I 5 R } c W".
The Green's function
A'
and
Poisson kernel P of the unit disk D c R2 = 43 are well known from the complex analysis, and can be derived in many different ways. We just state the result, referring to problem 6 for details, (4.43)
where z = reit; w = pei6 are complex points in D ( p = 1 for P). Our goal is to extend
(4.43) to higher dimensions. Namely, the Poisson kernel in the n-ball is given by
(4.44)
52.4. Laplacian and related differential equations.
110
2~r 4 2 - volume of the where 12 1 = r; I y I = 1; e = angle between z and y, and w ~ =- r(n/2) unit sphere in Wn. The derivation of (4.44) will utilizes a particular conformal transformation in R", the inversion, u: z+z*
=A *
(4.45)
IZl2'
We recall that name conformal refers to diffeomorphisms {$} of R" (or any Riemannian manifold) that preserve angles between vectors, but not necessarily norms. T h e Jacobian of such m a p $, A = $', is a conformal matriz, product 'scalar x orthogonal", i.e. t A A = AZ (problem 8). Conformal m a p u of (4.45) takes interior of the unit ball in R" into the exterior, and vice versa,
I I = 1). T o construct the Green's function of A on B (Newton's potential) K O = C, I z - y I - " (for n > 2),
and acts identically on the unit sphere { we take the free-space Green's function and &ln
Iz-y I
in R2. It satisfies the differential equation: A,K = 6(z - y), but not the
boundary condition, Ko(z - y) function u&z) in
I an= h,(z) > 0, 80
n, that satisfies A,[u,]
we need t o correct KO by a regular harmonic
I
= 0; ).(,it
an= h , ( z ) ,
K ( r ; y )= Ko(z - Y) - )(., Such uy can be constructed by inversion of KO.We shall use the following identity valid for any function u, A(
I2
I 2-nu(+)) IZI
= I z I -"(Au)
(-1,
(4.46)
IZI
The proof of invariance-relation17 (4.46) is outlined in problem 8. Its immediate corollary is
Proposition: Function u is harmonic iffv(x) = Applying Proposition t o the Newton potential
12
I "-'u(z/ I z I ')
is harmonic.
KO = C I z - y I - " we get the requisite
harmonic correction: u
Cn
(2)=
l(z/Izl-
IzlY)l"-2'
Obviously, uy is harmonic and coincides with K o ( r - y) on the unit sphere. Thus the Green's function of the unit ball (4.47)
Finally, t o compute the Poisson kernel we take the normal derivative 8°K = ny.V,K,and observe that for y on the boundary ( I y I = l), ny = y. Hence,
It remains to note that both denominators in (4.48) are equal (see fig.7),
50
we get Poisson
17The R"-Laplacian is not invariant under the conformal map u, but comes 'very near" t o it. !), Indeed, (4.30) means that the unitary operator U , : f ( z ) + I z I -"f(z*); ( I z I -" = ,/intertwines operators A I z I and A, (A I z I ' ) U , = U,A.
32.4. Laplacian and related differential equations
111
kernel (4.44),
From the Poisson kernel on the unit ball we can easily derive P for any radius p,
(4.49) Fig.?
demonstrates equality of two denominators of (4.49), that measure the distances between pairs {z;y} and
I./
I2
I ; I 2 I Yl.
Formula (4.49) has many applications in the theory of harmonic functions. It follows from (4.49)that any harmonic u(x) is real analytic, since P ( q ...) is real-analytic in x for each y on the boundary. In fact, one could estimate the radius of convergence of the Taylor series ~ ~ ~ ( ~ ) ( ofz harmonic - - x ~ )function ~ u ( z ) at each interior point {xo}
C. Furthermore, (4.49) gives the minimal and maximal value of the Poisson kernel for each 0 < r < p, they correspond to in terms of the distance from {xo} to the boundary
case=
fl,
(4.50) As a consequence of (4.50)we get an important result in the theory of harmonic
functions.
Harnack inequality: The ratio between the maximal and minimal values of a harmonic positive function u(x) 2 0 an the ball of radius r is estimated by max {u(x):I z I = r } min {u(x):I z I = r ) -
(4.51)
4.13. Green’s functions and Poiason kernels in arbitrary domains D are constructed, as in
solid sphere case, from the free-space Green’s function (Newton/Bessel potential) KO,and an auxiliary (harmonic) function u(x; y)
(2,y
E D), that satisfies the homogeneous
equation: A,u = 0; and the boundary condition: u( ...;y) or a,u
I aD = -a,Ko(x-y)
I,
= -Ko(z-y) (Dirichlet);
(Neumann). Then, the Green’s function of D,
62.4. Ladacian and related differential equations.
112
K(z;y) = Ko(z-y)--(z;y); while the Poisson kernel, P ( z ;y) = -B,K(z; y)(Dirichlet); and P(z;y) = -a,K(z;y) (Neumann). Solution of any inhomogeneous problem: Au = F; u I
=I
= f ; is given by
,my ) ~ ( y ) d y+ r aDp(z;y ) f ( y ) d w .
Another construction of P(z;y) in terms of the solid-angle form is outlined in problem 12.
Additional comments and results. Our discussion of differential operators was limited to specific examples and methods. Here we give a cursory introduct