Analysis in Positive Characteristic

This page intentionally left blank CAMBRIDGE TRACTS IN MATHEMATICS General Editors ´ S , W . FU LTO N , A . K AT O K ...

Author: Anatoly N. Kochubei

11 downloads 745 Views 1MB Size Report

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

Report copyright / DMCA form

DOWNLOAD PDF

This page intentionally left blank

CAMBRIDGE TRACTS IN MATHEMATICS General Editors ´ S , W . FU LTO N , A . K AT O K , F. K I RWA N , B . B O LLO B A P. S A RN A K , B . S I M O N , B . T O TA RO 178 Analysis in Positive Characteristic

Analysis in Positive Characteristic ANATOLY N. KOCHUBEI National Academy of Sciences of Ukraine

CAMBRIDGE UNIVERSITY PRESS

Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo Cambridge University Press The Edinburgh Building, Cambridge CB2 8RU, UK Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521509770 © A. Kochubei, 2009 This publication is in copyright. Subject to statutory exception and to the provision of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published in print format 2009

ISBN-13

978-0-511-51646-7

eBook (EBL)

ISBN-13

978-0-521-50977-0

hardback

Cambridge University Press has no responsibility for the persistence or accuracy of urls for external or third-party internet websites referred to in this publication, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.

Contents

Preface 1 Orthonormal systems and their applications 1.1 Basic notions 1.2 Additive Carlitz polynomials 1.3 Hyperdiﬀerentiations 1.4 The digit principle 1.5 Finite places of a global function ﬁeld 1.6 The Carlitz module 1.7 Canonical commutation relations 1.8 Comments

vii 1 1 13 17 22 32 36 40 44

2

Calculus 2.1 Fq -linear calculus 2.2 Umbral calculus 2.3 Locally analytic functions 2.4 General smooth functions 2.5 Entire functions 2.6 Measures and divided power series 2.7 Comments

47 47 60 75 82 90 99 102

3

Diﬀerential equations 3.1 Existence and uniqueness theorems 3.2 Strongly nonlinear equations 3.3 Regular singularity 3.4 Evolution equations 3.5 Comments

105 105 114 120 131 135

4

Special functions 4.1 Hypergeometric functions

137 137

v

vi

Contents 4.2 4.3 4.4 4.5 4.6

5

Analogs of the Bessel functions and Jacobi polynomials Polylogarithms K-binomial coeﬃcients Overconvergence properties Comments

The Carlitz rings 5.1 Algebraic preliminaries 5.2 The Carlitz rings 5.3 The ring A1 5.4 Quasi-holonomic modules 5.5 Comments Bibliography Index

142 145 154 159 166 167 167 173 178 185 201 203 209

Preface

One of the natural options in the development of mathematical analysis is to investigate analogs of classical objects in a new environment. To obtain the latter, it suﬃces to change the ground ﬁeld. For a nontrivial analysis, we need a nondiscrete topological ﬁeld. If we conﬁne ourselves to locally compact ﬁelds (which can be extended subsequently via algebraic closures and completions), then the spectrum of possibilities is not very wide – R, C, the ﬁelds Qp of p-adic numbers and their ﬁnite extensions, and non-Archimedean local ﬁelds of positive characteristic, that is the ﬁelds of formal Laurent series with coeﬃcients from ﬁnite ﬁelds Fq . Over R and C, the whole of classical mathematics is developed. The p-adic ﬁelds constitute the environment for p-adic analysis, a rapidly developing branch of contemporary mathematics embracing analogs of classical function theory, Fourier analysis, diﬀerential equations, theory of dynamical systems, probability, etc. Note that there are two kinds of p-adic analysis. One of them deals essentially with functions from Qp to R or C, like characters or probability densities, and its main results remain valid for local ﬁelds of positive characteristic (see [59]). The second treats functions with non-Archimedean arguments and values, like polynomials and power series, and here the extension to ﬁelds of positive characteristic is much more complicated. To see the essence of these diﬃculties, it suﬃces to notice that some of the simplest standard notions of analysis do not make sense in the case of characteristic p > 0. For example, n! = 0 in a ﬁeld of characteristic p, if d n n ≥ p. Similarly, dt (t ) = 0 if p divides n, that is the classical diﬀerential calculus cannot be used to investigate important classes of functions. For the same reason, the standard formulas do not work if one wants to vii

viii

Preface

deﬁne analogs of the most important special functions, beginning with the exponential function. The ﬁrst steps in developing analysis over local ﬁelds of positive characteristic were made in 1935 by Carlitz [22] who found appropriate positive characteristic counterparts for the factorial, exponential, logarithm, and classical orthogonal polynomials. Carlitz’s constructions were highly unusual for that time, and their algebraic meaning was understood only about 40 years later, within the theory of Drinfeld modules and their generalizations. After that they became important elements of a branch of number theory, function ﬁeld arithmetic; see the books by Goss [45], Rosen [94], and Thakur [111]. Though the initial paper by Carlitz contained important insights into analysis (like the diﬀerence operator, very close to a later notion of the Carlitz derivative), the active development of analysis in positive characteristic began much later, after the paper [43] by Goss (with an exposition of Carlitz’s work in more modern language) and the paper [109] by Thakur who introduced analogs of the hypergeometric function and the hypergeometric diﬀerential equation. That opened the way for a systematic development of the counterparts of the theory of Fourier series, the theory of diﬀerential equations, etc. The main achievements in these and related directions are summarized in this book. In contrast to the books by Goss [45] and Thakur [111], whose material has some common features with the present one, this book is intended for specialists in analysis, from the graduate student level. Therefore only the most basic knowledge of algebra is presumed; in any case, an explanation of any algebraic notion, not given explicitly in the text, can be found in Lang’s Algebra. In order to make the material more elementary, most notions are introduced on the local ﬁeld level (though in some cases their global ﬁeld interpretation is also outlined). The author’s aim was to reach analysts who do not usually read books or papers in algebraic number theory and algebraic geometry. Accordingly, the author did not try to touch on some nonelementary material covered exhaustively in [45, 111], like the zeta and gamma functions and their arithmetic applications (a new analog of the zeta function, appearing in Chapter 4, is still only an example of an application of diﬀerential equations with the Carlitz derivatives). Some results requiring involved algebraic techniques are given without proofs. Here is a short description of the contents. In Chapter 1, we introduce and study systems of functions on a local ﬁeld of positive characteristic,

Preface

ix

which are crucial for solving various speciﬁc problems of analysis – the additive Carlitz polynomials and hyperdiﬀerentiations. These functions are Fq -linear; a method of lifting a Fq -linear basis to a basis of a space of continuous functions was invented by Carlitz (1940) and found to be actually a very general construction (“the digit principle”) by Conrad (2000). We describe both the general results and some speciﬁc cases. Then we consider the ﬁrst special functions, the Carlitz exponential and logarithm, and the Carlitz module, and show how the above material appears in the positive characteristic analog of canonical commutation relations of quantum mechanics. Chapter 2 deals mainly with the representation of functions by their expansions in a series of Carlitz polynomials, the correct understanding of the smoothness and analyticity of functions and their characterizations in terms of Fourier–Carlitz coeﬃcients, the ﬁrst notions of diﬀerential and integral calculus, which agree (in contrast to the classical ones) with the positive characteristic framework. The diﬀerential calculus is then extended to a kind of Rota’s umbral calculus. Given the notion of a derivative, it is natural to study diﬀerential equations, and that is done in Chapter 3, including analogs of Cauchy’s existence and uniqueness theorems for analytic solutions, counterparts of regular singularity theory and partial diﬀerential equations of evolution type. Some special functions satisfying diﬀerential equations with Carlitz derivatives are discussed in Chapter 4. Finally, in Chapter 5 we introduce and study rings of diﬀerential operators of the above kind. These rings can be seen as analogs of the Weyl algebras, though some of their properties are quite diﬀerent – in fact, already the Carlitz derivative is, for instance, nonlinear (actually, Fq -linear). Nevertheless, there exists a notion of quasiholonomic modules over these rings, having some common features with holonomic modules in the sense of Bernstein, and connected to some special functions in the spirit of Zeilberger’s theory. The author is grateful to D. Goss, W. H. Schikhof and D. S. Thakur for their helpful consultations in non-Archimedean analysis and function ﬁeld arithmetic. Anatoly N. Kochubei

1 Orthonormal systems and their applications

In this chapter we introduce the main notion of analysis over a local ﬁeld K of positive characteristic, and study two most important orthonormal systems appearing in this theory, the Carlitz polynomials and hyperdiﬀerentiations. Their role, especially that of the Carlitz polynomials, will be crucial throughout this book. However the ﬁrst applications, to the Carlitz module expressing the main functional relation for the counterparts of the exponential and logarithm, and to representations over K of the canonical commutation relations of quantum mechanics, will be given in this chapter. We also describe the digit principle connecting the analysis of additive functions and that of general continuous functions on subsets of K.

1.1 Basic notions 1.1.1. Function ﬁelds. Factorials. It is well known [17, 38, 122] that any local ﬁeld (= nondiscrete locally compact topological ﬁeld) of positive characteristic p is isomorphic to the ﬁeld K of formal Laurent series with coeﬃcients from the Galois ﬁeld Fq of q elements, q = pυ , υ ∈ N. There is an absolute value | · | on K deﬁned as follows: |0| = 0; if z ∈ K, z=

∞

ζi xi ,

m ∈ Z, ζi ∈ Fq , ζm = 0,

i=m

then |z| = q −m . The absolute value has the following properties: |z1 z2 | = |z1 | · |z2 |,

|z1 + z2 | ≤ max(|z1 |, |z2 |),

1

z1 , z2 ∈ K.

2

Chapter 1

The second property, called the ultra-metric inequality (or the non-Archimedean property), implies that K is totally disconnected in the topology deﬁned by the metric |x − y|. Here we encounter all the peculiar properties of non-Archimedean spaces (see, for example, [98]) – standard subsets of K, like balls or spheres, are simultaneously open and closed (in fact, compact); two balls either do not intersect, or one of them is contained in the other, ∞ ak , ak ∈ K, converges if and only if ak → 0, etc. a series k=1

We will denote by K a ﬁxed algebraic closure of K. The absolute value | · | can be extended in a unique way onto K, and the completion K c is algebraically closed (see [98]). The ﬁrst “appropriate” analog of the factorial is the sequence Di = [i][i − 1]q . . . [1]q

i−1

i

[i] = xq − x (i ≥ 1), D0 = 1,

,

(1.1)

Another factorial-like sequence is Li = [i][i − 1] . . . [1] (i ≥ 1), L0 = 1.

(1.2)

It is easy to see that q i −1

|Di | = q − q−1 ,

|Li | = q −i .

(1.3)

As we will see, in some cases it is natural to consider Di as a counterpart of the factorial of the number q i . Then analogs of other values of the factorial are deﬁned as follows. Let us write any natural number j in the base q as j=

n−1

αi q i ,

αi ∈ Z, 0 ≤ αi < q.

i=0

Then we set Γj =

n−1

Diαi .

(1.4)

i=0

Lemma 1.1 (i) The elements Dm , Lm , and Γqm −1 are connected by the identity Γqm −1 Lm = Dm . (ii) For any i = 1, 2, . . ., the element [i] is the product of all monic irreducible polynomials m ∈ Fq [x], such that deg m divides i.

Orthonormal systems

3

(iii) For any i = 1, 2, . . ., the element Di is the product of all monic polynomials g ∈ Fq [x] with deg g = i. (iv) Li is the least common multiple of all polynomials from Fq of degree i. Proof. (i) By the deﬁnition, Γqm −1 = (D0 . . . Dm−1 )q−1 . We have Dm =

m

[i]q

m−i

, Lm =

i=1

m

[i],

i=1

D0 . . . Dm−1 =

j m−1

[i]q

j−i

=

j=1 i=1

=

m−1

m−1 m−1

[i]q

j−i

i=1 j=i

m−1

q

j−i

[i] j=i

=

i=1

m−1

[i]

q m−i −1 q−1

,

i=1

so that Γqm −1 Lm = [m]

m−1

[i]q

m−i

= Dm .

i=1

(ii) Let m ∈ Fq [x] be a monic irreducible polynomial. Then m divides i [i] = xq − x if and only if deg m divides i. Indeed, suppose that m divides [i]. Let α be a root of m lying in the i splitting ﬁeld of m. Then αq = α, so that α ∈ Fqi . Therefore the simple extension Fq (α) of the ﬁeld Fq is a subﬁeld of Fqi . Since m is irreducible, we have [Fq (α) : Fq ] = deg m, and it follows from the basic properties of ﬁnite ﬁelds [75, 76] that deg m divides i = [Fqi : Fq ]. Conversely, if the number l = deg m divides i, then [75, 76] Fql is a subﬁeld of Fqi . If α is a root of m in its splitting ﬁeld over Fq , then i [Fq (α) : Fq ] = l, whence Fq (α) = Fql . Therefore α ∈ Fqi , so that αq = α, i and α is a root of the polynomial xq −x. Since m is irreducible, it generates the principal ideal in Fq [x] consisting of polynomials for which α is a root. i This means that m divides xq − x. i d [i] = −1, so that [i] = xq − x has no Now it remains to note that dx multiple roots in its splitting ﬁeld over Fq . Therefore each monic irreducible polynomial appears exactly once in the canonical decomposition of [i] in the ring Fq [x]. (iii) Let π ∈ Fq [x] be an irreducible monic polynomial, deg π = d. Among

4

Chapter 1

all monic polynomials g ∈ Fq [x] of degree i ≥ d, q i−d polynomials are multiples of π, q i−2d (i ≥ 2d) polynomials are multiples of π 2 , etc. Thus the number of polynomials of degree i, whose canonical decompositions contain π with exactly the power 1, equals q i−d − q i−2d . We can perform similar evaluations with higher degrees of π, so that the product of all monic polynomials of the degree i contains π with the power equal to i−d − q i−2d + 2 q i−2d − q i−3d + · · · + int q

int( di ) i i−int( di )d q q i−νd = d ν=1

where int(s), s ∈ R+ , means the biggest integer not exceeding s. On the other hand, considering Di we see that π is contained once in the canonical decomposition of [j] exactly for j = νd, ν ≤ int di . Since i i−j Di = [j]q , we ﬁnd that π is contained in the canonical decomposition j=1 int( di )

of Di with the power

q i−νd , the same as the above one.

ν=1

(iv) Reasoning similar to (iii) shows that both elements to be proved equal contain π with the same power int di . 1.1.2. Fq -linear functions. A special feature of the ﬁeld K is the availability of many functions f with the property of Fq -linearity f (t1 + t2 ) = f (t1 ) + f (t2 ),

f (ξt) = ξf (t),

where t, t1 , t2 belong to a Fq -subspace of K, ξ ∈ Fq . Such are, for example, qk all power series ck t , ck ∈ K c , convergent on some region in K or K c . In particular, let us consider Fq -linear polynomials. Such a polynomial qk over K c (and over any inﬁnite ﬁeld) has the form ck t [45]. Obviously, the roots of an Fq -linear polynomial form an Fq -vector space. For separable polynomials, the converse is also true. Proposition 1.2 A separable polynomial P ∈ K c [t] is Fq -linear if and only if the set W = {w1 , . . . , wm } of all its roots forms an Fq -vector subspace of Kc Proof. The necessity is obvious, and we prove the suﬃciency. We have P (t) =

m i=1

(t − wi ).

Orthonormal systems

5

It is clear that P (t + w) = P (t) for any w ∈ W . Suppose that y ∈ K c . Set H(t) = P (t + y) − P (t) − P (y). Then deg H < deg P = m, and at the same time H(w) = 0 for all w ∈ W . This means that H(t) = 0 for all t, so that H is additive. Similarly, for α ∈ Fq set Hα (t) = P (αt) − αP (t). Taking a basis in the ﬁnite Fq -vector space W we see that m = q l for some l l ∈ Z+ . Then deg P = q l . Since αq = α, we conclude that deg Hα < q l . On the other hand, Hα (w) = 0 for w ∈ W . Therefore Hα (t) ≡ 0, as desired. A detailed exposition of algebraic properties of Fq -linear polynomials is given in [45]. Here we describe some properties of Fq -linear power series convergent on a neighborhood of the origin. ∞ k ak tq where ak ∈ K, Let RK be the set of all formal power series a = k=0 k

|ak | ≤ Aq , and A is a positive constant depending on a. In fact each series a = a(t) from RK converges on a neighborhood of the origin in K (and K c ). RK is a ring with respect to the termwise addition and the composition l

∞ ∞ l k qn an bl−n tq , b = b k tq , a◦b= l=0

n=0

k=0 k

as the operation of multiplication. Indeed, if |bk | ≤ B q , then, by the ultra-metric property of the absolute value, l l−n qn n l qn an bl−n ≤ max Aq B q ≤ Cq 0≤n≤l n=0

where C = B max(A, 1). The unit element in RK is a(t) = t. It is easy to check that RK has no zero divisors. ∞ k k If a ∈ RK , a = ak tq , is such that |a0 | ≤ 1 and |ak | ≤ Aq , |A| ≥ 1, k=0

for all k, then we may write k

|ak | ≤ Aq1 if we take A1 ≥ Aq

k

/(q k −1)

−1

,

k = 0, 1, 2, . . . ,

for all k ≥ 1. If also b =

∞ k=0

k

bk tq , |bk | ≤ B1q

k

−1

,

6

Chapter 1

B1 ≥ 1, then for a ◦ b =

∞

l

cl tq we have

l=0 i

|cl | ≤ max Aq1

−1

i+j=l

B1q

j

−1

q i

≤ C1q

l

−1

where C1 = max(A1 , B1 ). In particular, in this case the coeﬃcients of the series for an (the composition power) satisfy an estimate of this kind, with a constant independent of n. Proposition 1.3 The ring RK is a left Ore ring, thus it possesses a classical ring of fractions. Proof. By Ore’s theorem (see [50]) it suﬃces to show that for any elements a, b ∈ RK there exist elements a , b ∈ RK such that b = 0 and a ◦ b = b ◦ a.

(1.5)

We may assume that a = 0, ∞

a=

k

ak tq ,

b=

∞

k

b k tq ,

k=l

k=m

m, l ≥ 0, am = 0, bl = 0. Without restricting generality we may assume that l = m (if we prove m−l (1.5) for this case and if, for example, l < m, we set b1 = tq ◦ b, ﬁnd m−l a , b in such a way that a ◦ b1 = b ◦ a, and then set a = a ◦ tq ), and that al = bl = α, so that l

a = αtq +

∞

k

ak tq ,

l

b = αtq +

k=l+1

∞

k

b k tq ,

k=l+1

α = 0. We seek a , b in the form a =

∞ j=l

aj tq , j

b =

∞

bj tq . j

j=l

The coeﬃcients aj , bj can be deﬁned inductively. Set al = bl = 1. If aj , bj have been determined for l ≤ j ≤ k − 1, then ak , bk are determined from the equality of the (k + l)-th terms of the composition products: i i k k ai bjq = bk αq + bi aqj ak αq + i+j=k+l j=l

i+j=k+l j=l

Orthonormal systems

7

(the above sums do not contain nontrivial terms with ai , bi , i ≥ k, since aj = bj = 0 for j < l). In particular, we may set bk = 0,      

i i q q −q k ai bj − bi aj . ak = α    i+j=k+l  i
If this choice is made for each k ≥ l + 1, then we have bi = 0 for every i ≥ l + 1, so that i k ai bqj . (1.6) ak = α−q i+j=k+l i
Denote C1 = |α|−1 . We have |bj | ≤ C2q for all j. Denote, further, C3 = max(1, C1 , C2 ), C4 = C3q

l+2

. Let us prove that k

|ak | ≤ C4q . i

Suppose that |ai | ≤ C4q for all i, l ≤ i ≤ k − 1 (this is obvious for i = l, since al = 1). By (1.6), |ak | ≤ C1q

k

≤ C3q

k

i

max C4q C2q

i+j

i+j=k+l i
+q k+l+1 +q k+l

as desired. Thus a ∈ RK .

k

≤ C1q C4q

k−1

(1+q l +q l+1 )q k

= C3

C2q

k+l

l+2 qk k ≤ C3q = C4q ,

Every nonzero element of RK is invertible in the ring of fractions AK , which is actually a skew ﬁeld consisting of formal fractions c−1 d, c, d ∈ RK . Proposition 1.4 Each element a = c−1 d ∈ AK can be represented in the −m −m m form a = tq a where tq is the inverse of tq , a ∈ RK . Proof. It is suﬃcient to prove that any nonzero element c ∈ RK can be m written as c = c ◦ tq where c is invertible in RK . ∞ k k Let c = ck tq , cm = 0, |ck | ≤ C q . Then k=m

c = cm

t+

∞ l=1

q c−1 m cm+l t

l

◦ tq

m

8

Chapter 1 −1 l where cm cm+l ≤ C1q −1 for all l ≥ 1, if C1 is suﬃciently large. Denote w=

∞

q c−1 m cm+l t , l

c = cm (t + w).

l=1

The series (t + w)−1

=

∞

(−1)n wn converges in the standard

n=0

non-Archimedean topology of formal power series (see [83], Sect. 19.7) n because the formal power series for wn begins from the term with tq ; recall that wn is the composition and t is the unit element. Moreover, power, ∞ (n) (n) q j q j −1 n aj t where aj ≤ C1 for all j, with the same constant w = j=n

independent of n. Using the ultra-metric inequality we ﬁnd that the coef∞ j ﬁcients of the formal power series (t + w)−1 = aj tq (each of them is, up to a sign, a ﬁnite sum of the coeﬃcients Therefore (c )−1 ∈ RK .

j=0 (n) aj ) satisfy

the same estimate.

The skew ﬁeld of fractions AK can be imbedded into wider skew ﬁelds where operations are more explicit. Let Kperf be the perfect closure of the ﬁeld K. Denote by A∞ Kperf the composition ring of Fq -linear formal Laurent ∞ k series a = ak tq , m ∈ Z, ak ∈ Kperf , am = 0 (if a = 0). Since τ is an k=m

automorphism of Kperf , A∞ Kperf is a special case of the well-known ring of twisted Laurent series [83]. Therefore A∞ Kperf is a skew ﬁeld. ∞ Let AKperf be a subring of AKperf consisting of formal series with |ak | ≤ k

Aq for all k ≥ 0. Just as in the proof of Proposition 1.4, we show that AKperf is actually a skew ﬁeld. Its elements can be written in the form −m tq ◦ c where c is an invertible element of the ring RKperf ∈ AKperf of ∞ k ak tq . In contrast to the case of the skew ﬁeld formal power series k=0

−m

by c is indeed the composition of AK , in AKperf the multiplication of tq (locally deﬁned) functions, so that AKperf consists of fractional power series understood in the classical sense. Of course, AKperf can be extended further, by considering K or K c instead of Kperf . The above reasoning carries over to these cases (we can also consider the ring RK c of locally convergent Fq -linear power series as the initial ring). In each of them the presence of a fractional composition −m factor tq is an Fq -linear counterpart of a pole of order m. Thus we

Orthonormal systems

9

may see the above skew ﬁelds as consisting of functions of meromorphic type. 1.1.3. Orthonormal bases. In the sequel we will need some elementary notions from non-Archimedean functional analysis. The reader can consult [93, 100, 82] for further information on this rich and welldeveloped subject. Let E be a vector space over K (similarly one can deal with any nonArchimedean local ﬁeld). A norm on E is a map u → u from E to R, such that for u, v ∈ E, λ ∈ K, u ≥ 0,

λu = |λ| · u ,

u + v ≤ max( u , v ),

and u = 0 if and only if u = 0. If E is complete as a metric space with the metric (u, v) → u − v , then the space E is called a Banach space over K. In this book we will often deal with the Banach space C(O, K) of all Kvalued continuous functions on the ring of integers O = {z ∈ K : |z| ≤ 1}, equipped with the norm u = sup |u(s)|, s∈O

and its subspace C0 (O, K) consisting of Fq -linear functions. A sequence {f0 , f1 , f2 , . . .} in a Banach space E over K is called an orthonormal basis, if each u ∈ E can be represented as a convergent series u=

∞

cn fn

n=0

where cn ∈ K, cn → 0, and u = max |cn |. n≥0

The coeﬃcients in such a representation are determined in a unique way. Denote by P ⊂ K the prime ideal P = xO = {z ∈ K : |z| < 1}. Then the residue ﬁeld O/P is isomorphic to Fq . If E0 = {u ∈ E : u ≤ 1}, then the residual space (or the reduction) E = E0 /P E0 is a vector space over the residue ﬁeld O/P . In an obvious way, any element from E0 possesses a reduction from E.

10

Chapter 1

Denote by E and |K| the sets of values of the norm u , u ∈ E, and of the absolute value |z|, z ∈ K, respectively. Proposition 1.5 Suppose that E = |K|. A sequence {fn } ⊂ E is an orthonormal basis in E, if and only if all the vectors fn lie in E0 , and their reductions fn ∈ E form an Fq -basis of E in the algebraic sense, that is the sequence {fn } is linearly independent, and every element from E is a ﬁnite linear combination of elements of this sequence. Proof. The necessity is evident. Let us prove the suﬃciency. If u ∈ E0 , denote by u its image in E. Then u can be written as a ﬁnite linear combination, u = ξi f i , ξi ∈ Fq . Considering ξi as classes from O/P , we (1) (1) can take their representatives zi ∈ O. Then u − zi fi ∈ P E0 , so that (1) u= zi fi + xz (1) , z (1) ∈ E0 . Repeating the above procedure for z (1) and iterating we come to a representation u= zi fi , zi ∈ O, zi → 0. If u = 1, then necessarily sup |zi | = 1 (otherwise |zi | ≤ q −1 for all i, so i

that u < 1). Since E = |K|, for any u ∈ E we have u = q n , n ∈ Z, ∞ so that xn u = 1, xn u = xn zi fi , and we have seen that sup |xn zi | = 1, i=0

whence sup |zi | = q n = u . i

i

In fact, every separable Banach space E over the ﬁeld K, such that E = |K|, possesses an orthonormal basis (the Monna-Fleischer theorem, see [93, 90, 98]). In this book we are interested mostly in the explicit construction of orthonormal bases for some function spaces. Let E1 and E2 be Banach spaces over K with the norms · 1 and · 2 , ˆ 2 is the completion of respectively. The topological tensor product E1 ⊗E the algebraic tensor product the E1 ⊗K E2 with respect to the norm r vi ⊗ wi , vi ∈ E1 , wi ∈ E2 u ⊗ = inf max vi 1 · wi 2 : u = 1≤i≤r

i=1

where the inﬁmum is taken over all possible representations u =

r i=1

(1.7) vi ⊗wi ,

Orthonormal systems

11

vi ∈ E1 , wi ∈ E2 . See [100] for the proof of the fact that the expression (1.7) indeed deﬁnes a norm, and that v ⊗ w ⊗ = v 1 · w 2 . The topological tensor product has the usual functorial properties with respect to continuous linear mappings. A typical example of a topological tensor product is the Banach space C(O, V ) of continuous functions on O with values from a Banach space V (over the ﬁeld K), with the norm u = sup u(z) V . For this space, z∈O

ˆ (see [100]). If C0 (O, V ) is the subspace of Fq C(O, V ) = C(O, K)⊗V ˆ . This will linear continuous functions, then C0 (O, V ) = C0 (O, K)⊗V follow (see Section 1.2 below) from an explicit uniformly convergent expansion u(z) =

∞

cn fn (z),

cn ∈ V, z ∈ O,

n=0

of an arbitrary function u ∈ C0 (O, V ) in a system of K-valued normalized additive Carlitz polynomials. More generally, let E1 , E2 be Banach spaces over K. ˆ 2 is isoProposition 1.6 If E2 has an orthonormal basis {fn }, then E1 ⊗E metrically isomorphic to the space Z of all sequences Y = {yi ∈ E1 , yi 1 → 0, as i → ∞}, with the norm Y ∞ = sup yi 1 , Y = {yi }. The isomori

phism is implemented by the mapping Z Y →

∞

ˆ 2. yi ⊗ fi ∈ E1 ⊗E

(1.8)

i=0

ˆ 2 . Denote by π Proof. It is clear that the series in (1.8) converges in E1 ⊗E the mapping (1.8). We have π(Y ) ⊗ ≤ Y ∞ . Let t =

(1.9)

ξj ⊗ ηj ∈ E1 ⊗K E2 (thus the sum is ﬁnite). We can expand ηj :

j

ηj =

∞ i=0

cij fi ,

cij ∈ K, |cij | → 0 for each j,

12

Chapter 1

so that ηj 2 = max |cij |. Then i

t=

∞

 

i=0

 cij ξj  ⊗ fi ,

j

and we have a linear mapping ω : E1 ⊗K E2 → Z, ω(t) = cij ξj . j i Note that ω(t) does not depend on the representation t = ξj ⊗ ηj – if ∞

j

yi ⊗ fi = 0, where yi ∈ E1 , yi 1 → 0, then yi = 0 for all i (one can

i=0

apply the linear mapping λ ⊗ id where λ : E1 → K is an arbitrary linear functional, and use the basis property of {fi }). Now ω(t) ∞ = sup cij ξj sup |cij | · ξj 1 = sup ξj 1 · ηj 2 . ≤ sup i i j j j 1

The left-hand side does not depend on the representation of t, thus ω(t) ∞ ≤ inf sup ξj 1 · ηj 2 = t ⊗ . j

ˆ 2 , and This means that ω can be extended by continuity onto E1 ⊗E ω(t) ∞ ≤ t ⊗ .

(1.10)

It remains to note that ω and π are inverse to each other, so that the inequalities (1.9) and (1.10) mean that ω and π are both isometric and implement the required isomorphism. The ﬁeld K c is obviously a Banach space over K. For this case, Proposition 1.6 means, in particular, that an orthonormal basis {fn } in C(O, K) or C0 (O, K) is simultaneously an orthonormal basis in C(O, K c ) and C0 (O, K c ) respectively, that is any function u ∈ C(O, K c ) (or C0 (O, K c )) admits a uniformly convergent expansion u(t) =

∞

cn fn (t),

t ∈ O,

n=0

where cn ∈ K c , |cn | → 0, as n → ∞, and sup |u(t)| = max |cn |. t∈O

n≥0

Orthonormal systems

13

1.2 Additive Carlitz polynomials 1.2.1. A basis in C0 (O, K). The system of additive (in fact, Fq -linear) Carlitz polynomials is deﬁned as follows. Let e0 (t) = t, (t − m), i ≥ 1. (1.11) ei (t) = m∈Fq [x] deg m
By Proposition 1.2, the polynomials ei are Fq -linear. Proposition 1.7 (i) The polynomials ei satisfy the recursive relations q−1 ei−1 (t), ei (t) = eqi−1 (t) − Di−1

ei (xt) = xei (t) + [i]eqi−1 (t),

i ≥ 1; i ≥ 1.

(1.12) (1.13)

(ii) There is an explicit representation ei (t) =

i

i−j

(−1)

j=0

where

i qj t j

(1.14)

i Di = . j j Dj Lqi−j

(iii) The relations for special values ei (xi ) = Di , ei

1 x+1

=

(−1)i Di , (x + 1)l

l = 1 + q + · · · + qi ,

(1.15) (1.16)

hold for any i = 0, 1, 2, . . . Proof. First we prove (1.15). The class of polynomials xi − m, m ∈ Fq [x], deg m < i, is exactly the class of all monic polynomials of degree i. Therefore (1.15) follows from Lemma 1.1. In order to prove (1.12), note that both sides are monic polynomials of the same degree q i . Thus we need only to show that they have the same set of roots m ∈ Fq [x], deg m < i (the number of such polynomials is just q i ). It follows from the deﬁnition (1.11) that both sides of (1.12) vanish on t ∈ Fq [x], deg t < i − 1. If t = ζxi−1 , ζ ∈ Fq , then ei (t) = 0, while the

14

Chapter 1

q q−1 right-hand side of (1.12) equals ζ Di−1 − Di−1 · Di−1 = 0, which proves (1.12). The identities (1.14) and (1.16) are obvious for i = 0 and follow, for any i, by induction from (1.12). Using (1.14) we can check (1.13) by a direct computation. Denote by fi the normalized Carlitz polynomials: fi (t) =

ei (t) , Di

i = 0, 1, 2, . . .

Let us introduce also a sequence of diﬀerence operators which will play a crucial role throughout this book. Let (∆u)(t) = u(xt) − xu(t),

j−1 ∆(j) u (t) = ∆(j−1) u (xt) − xq u(t),

(1.17) j ≥ 1,

(1.18)

where ∆(0) = I (the identity operator), so that ∆(1) = ∆. It is easy to prove by induction that

∆(j) u (t) = u(xj t)   j−1 ij−k i1  u(xk t). + (−1)j−k  xq +···+q 0≤i1 <...
k=0

(1.19) The identity (1.13) can be written as (∆ei )(t) = [i]eqi−1 (t),

i ≥ 1,

(1.20)

whence q (t), (∆fi )(t) = fi−1

i ≥ 1.

(1.21)

Obviously, ∆e0 = ∆f0 = 0. More generally, we prove by induction that  j D   jk eqk−j , if k ≥ j, q ∆(j) ek = Dk−j (1.22)  0, if k < j, so that

(j)

∆

fk =

j

q , if k ≥ j, fk−j

0,

if k < j.

(1.23)

Orthonormal systems

15

Theorem 1.8 The sequence {fi }∞ 0 is an orthonormal basis of the Banach space C0 (O, K) of Fq -linear K-valued continuous functions on the ring of integers O ⊂ K. The coeﬃcients of the expansion u(t) =

∞

t ∈ O,

ci fi (t),

ci → 0,

(1.24)

i=0

of an arbitrary function u ∈ C0 (O, K) can be written explicitly: ci = ∆(i) u(1),

i = 0, 1, 2, . . .

(1.25)

Proof. It follows from (1.21) that q (xn ), fi (xn+1 ) = xfi (xn ) + fi−1

n ≥ 0.

By induction and Fq -linearity, we see that fi (t) ∈ Fq [x] for any t ∈ Fq [x]. Since Fq [x] is dense in O, we get fi ≤ 1 (the supremum norm), and the identity (1.15) implies the equalities fi = 1, i = 0, 1, 2, . . . If u is Fq -linear, we have u(0) = u(pt) = pu(t) = 0, and if u is also continuous, then u(xi ) → 0, as i → ∞. From (1.19) we ﬁnd for the coeﬃcients (1.25) the estimate |ci | ≤ max |u(xi )|, max q k−i |u(xk )| 0≤k≤i−1

≤ max |u(xi )|, u q −i/2 , max |u(xk )| −→ 0, k>i/2

as i → ∞. Therefore the series in (1.24), with the coeﬃcients given by (1.25), converges in C0 (O, K). Let us prove that it converges to u. Denote the right-hand side of (1.24) temporarily by v(t), so that v(t) =

∞

ci fi (t),

t ∈ O.

(1.26)

i=0

Applying a linear bounded operator ∆(k) to both sides of (1.26) we ﬁnd that ∞

qk ∆(k) v (t) = ci fi−k (t), i=k

so that ci = ∆(i) v(1), and by (1.25) ∆(i) (u − v)(1) = 0,

i = 0, 1, 2, . . .

16

Chapter 1

Using (1.19) we ﬁnd consequently that u(xn ) = v(xn ) for each n. Therefore, by Fq -linearity, u and v coincide on Fq [x] and, by continuity, u = v, and we have proved (1.24). This equality shows that u ≤ max |ci |. The opposite inequality follows from (1.25).

i

Remarks 1.1 (1) Theorem 1.8 and its proof remain valid for functions with values from a Banach space V over K. As mentioned in Section 1.1.3, ˆ . this proves the equality C0 (O, V ) = C0 (O, K)⊗V (2) The condition ci → 0 is also necessary for the convergence of the series (1.24) on O. Indeed, take t = (1 + x)−1 . By (1.16), 1 fi = 1, x+1 and the convergence of (1.24) at this value of t means that ci → 0, as i → ∞. !∞ m (3) If m is a ﬁxed positive integer, then fiq is an orthonormal i=0

basis in C0 (O, K). Actually, this is a very general fact, valid for bases in any spaces of K-valued continuous functions [53]. For the proof note m that the diﬀerence fiq − fi takes its values in the maximal ideal P of O. m Therefore fiq − fi ≤ q −1 for all values of i, and the result is a consequence of Proposition 1.5 – the reductions of both sequences (considered in Proposition 1.5) are the same. (4) It follows from Theorem 1.8 and the arguments used in its proof that the sequence {fi }∞ 0 ⊂ A[t], A = Fq [x], forms an A-basis (in the algebraic sense) of an A-module Int0 (A) of Fq -linear integer-valued polynomials from A[t] (a polynomial is called integer-valued if f (t) ∈ A for any t ∈ A). The expression (1.25) for the coeﬃcients of expansions in the Carlitz polynomials remains valid in this situation. 1.2.2. Strongly singular functions. The series (1.24) always makes sense for t ∈ Fq [x] – for each such t only a ﬁnite number of terms is diﬀerent from zero. A function u(t), t ∈ Fq [x], deﬁned by the series (1.24), is called strongly singular, if the series does not converge for any element t ∈ O \ Fq [x]. We will need the following property of the Carlitz polynomials. As we know, fi (xn ) = 0, if n < i. Let us consider the case where n ≥ i. Lemma 1.9 If n ≥ i, then |fi (xn )| = q i−n .

Orthonormal systems

17

Proof. For any ω ∈ Fq [x], deg ω < i, we have |xn − ω| = |ω|. Writing (t − ω) ei (t) = t 0=ω∈Fq [x] deg ω
we ﬁnd that

|ei (x )| = |x | n

n

ei (t) . |ω| = |x | · lim t→0 t n

deg ω
It follows from (1.14) that lim

t→0

ei (t) Di = (−1)i Li t

whence |fi (xn )| = as desired.

q −n = q i−n |Li |

Now we get a general suﬃcient condition for a function (1.24) to be strongly singular. Theorem 1.10 If |ci | ≥ ρ > 0 for all i ≥ i0 (where i0 is some natural number), then the function (1.24) is strongly singular. Proof. It is suﬃcient to ﬁnd, for any t ∈ O \ Fq [x], a sequence ik → ∞ such that |fik (t)| = 1, k = 1, 2, . . .. ∞ In fact, if t ∈ O \ Fq [x], then t = ξn xn , ξn ∈ Fq , with ξik = 0 for some sequence ik → ∞. Then fik (t) =

n=0

∞

ξn fik (xn )

n=ik

where fik (xik ) = 1 by (1.15), |fik (xn )| = q ik −n ≤ q −1 for n > ik (by Lemma 1.9), and |ξik | = 1. Therefore |fik (t)| = 1 for any k.

1.3 Hyperdiﬀerentiations 1.3.1. Deﬁnitions and main identities. Hyperdiﬀerentiations (or the

18

Chapter 1

Hasse derivatives) form a sequence of K-valued functions Dk (t), k ≥ 0, t ∈ O, deﬁned as follows. Set D0 (xn ) = xn , Dk (1) = 0 for k ≥ 1, n n−k n x , (1.27) Dk (x ) = k n = 0 for k > n. The binomial coeﬃcients are where it is assumed that k natural numbers making sense as elements of K. They can be deﬁned by the recurrence relation n−1 n−1 n = + (1.28) k k k−1 with appropriate boundary conditions; this deﬁnition does not involve divisions which are impossible in positive characteristic. Then Dk is extended onto Fq [x] by Fq -linearity: N

N n n ξn xn−k , ξn ∈ Fq , ξn0 = 0, = ξn x Dk k n=n n=max(n0 ,k)

0

n0 ∈ Z+ , so that

N N ξn xn ≤ q −(max(n0 ,k)−k) ≤ q k ξn xn . Dk n=n0

n=n0

Therefore each function Dk is continuous at the origin (thus on Fq [x], due to additivity), and can be extended to an Fq -linear continuous function on O. Note that each function Dk (k ≥ 1) is nowhere diﬀerentiable. Indeed, it is suﬃcient to check that Dk is not diﬀerentiable at the origin, that is to ﬁnd a sequence tn ∈ O such that tn → 0, as n → ∞, but t−1 k (tn ) does n D n x−k does not converge. Set tn = xn ; then tn → 0, but t−1 n Dk (tn ) = k not converge. The deﬁnition (1.27) shows that in characteristic zero one would have 1 dk Dk = , but this expression does not make sense in positive charack! dxk teristic. However the hyperdiﬀerentiations share some properties with the usual derivatives. One has the Leibnitz rule Dk (st) =

k j=0

Dj (s)Dk−j (t),

s, t ∈ O.

(1.29)

Orthonormal systems

19

Indeed, by continuity and Fq -linearity, the proof of (1.29) reduces to the case s = xm , t = xn , where the Leibnitz rule reduces to the Vandermonde formula k n m m+n (1.30) = k−j k k j=0 (see [88]). On the other hand, the form of the sequence Dk (that is, the expression (1.27)) follows from (1.29) and the assumptions D0 (x) = x, D1 (x) = 1, Dk (x) = 0 for k ≥ 2. The relations between hyperdiﬀerentiations and polynomial systems are given by the following results. Proposition 1.11 (i) For any t ∈ O, Dk (t) =

∞

Ank fn (t)

(1.31)

n=k

where fn are the An1 = (−1)n−1 Ln−1 ,

normalized

Ank = (−1)n+k Ln−1

0
additive

Carlitz

1 , [i ][i ] . . . [ik−1 ]
polynomials,

k > 1.

(1.32)

(ii) For any k ≥ 0, we have Dk = 1 (as before, · is the norm in C0 (O, K)). (iii) For any t ∈ O, n ∈ N, n

tq =

∞

[n]k Dk (t).

(1.33)

k=0

(iv) For any t ∈ O, fn (t) =

∞

Bkn Dk (t),

n = 0, 1, 2, . . . ,

k=n

where Bkn =

n (−1)n−i i

i=0

Di Lqn−i

[i]k .

(1.34)

20

Chapter 1

Proof. (i) By Theorem 1.8, we have to show that Ank = ∆(n) Dk (1). In fact, we will prove that ∆(n) Dk (t) =

k−1

An,k−i Di (t),

t ∈ O, k ≥ 1,

(1.35)

i=0

which gives the required identity for t = 1. We proceed by induction on n. It is checked directly, using (1.27) and (1.28), that ∆Dk = Dk−1 , k ≥ 1;

∆D0 = 0.

(1.36)

If (1.35) holds for some n, then, writing D−1 = 0, An0 = 0, we get from the recursive deﬁnition (1.18) that ∆(n+1) Dk (t) =

k−1

n

An,k−i Di (xt) − xq An,k−i Di (t)

!

i=0

=

k−1

n

An,k−i xDi (t) − An,k−i Di−1 (t) − xq An,k−i Di (t)

!

i=0

=

k−1

{−An,k−i [n] + An,k−i−1 } Di (t) =

i=0

k−1

An+1,k−i Di (t),

i=0

which completes the proof. " # (ii) It is clear that D0 = 1. For D1 we ﬁnd that D1 = max q −n+1 n≥1

= 1. Next, by (1.32), Ank ≤ q −n+1

max

0
q k−1 = q −n+k ,

so that |Ank | < 1, if k < n. As we know, |Ann | = 1. The required result follows from the fact that fn = 1 for any n. (iii) By (ii), the series in (1.33) is uniformly convergent, and it is suﬃcient to check the equality (1.33) for t = xm , m ∈ N. For this case, (1.33) follows easily from the Newton binomial formula. (iv) We ﬁnd from (1.14) and (1.33) that fn (t) =

∞

Bkn Dk (t),

t ∈ O, n = 0, 1, 2, . . . ,

(1.37)

k=0

where Bkn has the required form. In (1.37) we set successively t = 1, x, . . . , xn−1 , xn . Since fn (xn ) = 1 and fn (xj ) = 0 for j < n, we ﬁnd that Bkn = 0 for k < n. Thus, (1.37) gives the required identity (1.34).

Orthonormal systems

21

1.3.2. The basis property. The next result, the main one in this section, shows not only that {Dk } is an orthonormal basis, but also that the sequence of hyperdiﬀerentiations is closely connected with powers ∆n of the Carlitz diﬀerence operator, just as the sequence {fk } of normalized Carlitz polynomials is connected with higher diﬀerence operators ∆(n) . Theorem 1.12 The sequence {Dk }∞ 0 is an orthonormal basis of the Banach space C0 (O, K). The coeﬃcients of the expansion u(t) =

∞

ck Dk (t),

t ∈ O,

(1.38)

k=0

of an arbitrary function u ∈ C0 (O, K) can be written explicitly: ck = ∆k u(1) =

n

(−1)n−i u(xi )Di (xn ),

k = 0, 1, 2, . . .

(1.39)

i=0

Proof. Using the representation (1.31), in which Ann = 1 and |Ank | ≤ q −1 for k < n (see the Proof of Proposition 1.11) we ﬁnd that fk − Dk ≤ q −1 ,

k = 0, 1, 2, . . .

Since {fk } is an orthonormal basis of C0 (O, K), it follows from Proposition 1.5 that the sequence {Dk } possesses the same property. In order to prove the ﬁrst equality in (1.39), it suﬃces to apply the operator ∆ to the equality (1.38), to take into account the identity (1.36), to set t = 1, then to repeat the procedure, etc. To get the second equality, we use induction on n. For n = 0, the required identity is evident, since D0 (1) = 1 and c0 = u(1). Suppose it holds for k < n. Let us take t = xn in (1.38). As Dn (xn ) = 1, Dk (xn ) = 0 for k > n, we get n

u(x ) = cn +

n−1

ck Dk (xn ),

k=0

so that cn = u(x )Dn (x ) − n

n

n−1 k=0

ck Dk (xn ).

22

Chapter 1

The induction hypothesis allows us to rewrite this as   n−1 k  (−1)k−j u(xj )Dj (xk ) Dk (xn ) cn = u(xn )Dn (xn ) − k=0

= u(xn )Dn (xn ) −

n−1 j=0

= u(xn )Dn (xn ) −

n−1

   

j=0 n−1 k=j n−1

j=0

k=j

 (−1)k−j Dj (xk )Dk (xn ) u(xj )  n k xn−j  u(xj ). (−1)k−j k j

Using simple binomial identities [88] k n n n−j = , j k j k−j n−1

(−1)k−j

k=j

n−j k−j

(1.40)

= (−1)n−j+1 ,

we obtain that n

n

cn = u(x )Dn (x ) +

n−1

n−j

(−1)

j=0

and this is just the required formula.

n n−j x u(xj ), j

As we discussed in Section 1.1.2, typical classes of Fq -linear functions form rings with composition as the multiplication operation. In connection with this, it is interesting to note the following property of hyperdiﬀerentiations: k+l Dk+l (t), t ∈ O. Dk (Dl (t)) = l Its proof follows from the identity (1.40) and the deﬁnitions.

1.4 The digit principle 1.4.1. A general theorem. This section is devoted to the following problems. Given an orthonormal basis {ϕj }∞ 0 of the space C0 (O, K) of Fq -linear continuous functions, how do we construct an orthonormal basis {Φj }∞ 0 of the space C(O, K) of all continuous functions? How do we ﬁnd

Orthonormal systems

23

the coeﬃcients of the expansions in such bases? While the second problem is connected with special properties of each basis, the ﬁrst one admits a general solution. Let us write every integer i ≥ 0 in the base q as i = α0 + α1 q + · · · + αn−1 q n−1 ,

0 ≤ αj ≤ q − 1.

Set α

α1 n−1 0 Φi = ϕα 0 ϕ1 · · · ϕn−1

(1.41)

where it is assumed that ϕα j = 1 for α = 0 even if ϕj vanishes at some points. Note that ϕj = Φqj . The construction (1.41) is called the extension of the basis {ϕj } by digit expansions. Theorem 1.13 (“Digit principle”) The extension of an orthonormal basis of C0 (O, K) by digit expansions produces an orthonormal basis for C(O, K). Proof. Using Proposition 1.5 we reduce our problem to that of linear algebra over the ﬁnite ﬁeld Fq . The reduction of the space C0 (O, K) is $ HomFq (O, Fq ) = Fq ϕj where the reductions {ϕj }∞ 0 form an Fq -basis j≥0

of HomFq (O, Fq ). Here HomFq (O, Fq ) denotes the set of all continuous Fq linear maps (with respect to the discrete topology in Fq ). Similarly we will denote by C(O, Fq ) the set of all continuous functions O → Fq , the reduction of the Banach space C(O, K). n−1 % Let Hn = ker(ϕj ), n ≥ 1. Then Hn is a closed subspace of Fq j=0

codimension n in O, Hn+1 ⊂ Hn , and

∞ %

Hn = {0}. Therefore O ∼ =

n=1

lim O/Hn . ←−

Note that O/Hn is a ﬁnite set for each n.

Denoting by

Maps(X, Y ) the set of all mappings from a ﬁnite set X to a ﬁnite set Y , we see that C(O, Fq ) = lim Maps(O/Hn , Fq ). −→

Considering ϕ0 , . . . , ϕn−1 as functions on O/Hn we see that they form an Fq -basis of the Fq -dual space (O/Hn )∗ . Thus our problem is reduced to the following one. Let V be a ﬁnite-dimensional Fq -vector space, dim V = n, and let ψ0 , . . . , ψn−1 be a basis of V ∗ . Extend the functionals ψj to a system

24

Chapter 1

of polynomial functions on V by using digit expansions, that is for 0 ≤ i ≤ q n − 1 write i = α0 + α1 q + · · · + αn−1 q n−1 and set α

n−1 Ψi = ψ0α0 · · · ψn−1 ,

so that ψj = Ψqj and Ψ0 = 1. We need to show that the functions Ψi form a basis of Maps(V, Fq ). By a dimension count, it suﬃces to prove that the Ψi span Maps(V, Fq ). Let {v0 , . . . , vn−1 } be the dual basis to {ψ0 , . . . , ψn−1 }. For v ∈ V , write v = a0 v0 + a1 v1 + · · · + an−1 vn−1 where aj ∈ Fq . Consider functions hv : V → Fq of the form hv (w) =

n−1

n−1 1 − (ψj (w) − aj )q−1 = 1 − (ψj (w) − ψj (v))q−1 .

j=0

j=0

Since hv (w) = 1 for v = w and hv (w) = 0 when w = v (because at least one of the diﬀerences ψj (w) − ψj (v) is in this case a nonzero element of the ﬁnite ﬁeld Fq ), the Fq -span of all the hv is the whole set Maps(V, Fq ). Expanding the product deﬁning hv shows that hv is in the span of the Ψi since the exponents of the ψj in the product never exceed q − 1. This completes the proof. 1.4.2. The general Carlitz polynomials. Let us apply the construction (1.41) to the case where ϕj = fj , the normalized Carlitz polynomials. Denote by {Gi } the resulting polynomial system, so that G0 (t) = 1, Gi (t) =

n−1

fj (t)αj =

j=0

n−1 1 ej (t)αj Γi j=0

where the factorial-like sequence {Γi } is deﬁned by (1.4). By Theorem 1.13, {Gi } is an orthonormal basis of C(O, K). Let us study expansions with respect to this basis in detail. We will need an auxiliary polynomial system gi (t) =

n−1

gαj qj (t)

j=0

where

gαj qj (t) =

α

fj j (t),

if 0 ≤ αj < q − 1;

α fj j (t)

if αj = q − 1.

− 1,

Orthonormal systems

25

Proposition 1.14 The polynomial systems {Gi } and {gi } have the following properties: (i)

Gi (ξt) = ξ i Gi (t), ξ ∈ Fq ;

gi (ξt) = ξ i gi (t), 0 = ξ ∈ Fq .

(ii)

i Gj (t)Gl (s); Gi (t + s) = j

(1.42)

(1.43)

j+l=i

i i Gj (t)gl (s) = gj (t)Gl (s). gi (t + s) = j j

(1.44)

j+l=i

j+l=i

(iii) If 0 ≤ l < q ν , k ≥ 0, ν ∈ N, then

0,

if k + l = q ν − 1,

(−1)ν ,

if k + l = q ν − 1.

0,

if k + l = q ν − 1,

(−1)ν ,

if k + l = q ν − 1.

Gk (m)gl (m) =

m∈Fq [x] deg m<ν

(1.45)

(iv) If 0 ≤ k, l < q ν , then

Gk (m)gl (m) =

m monic deg m=ν

(1.46)

Proof. (i) The ﬁrst equality in (1.42) follows from the congruence n−1 j=0

αj ≡

n−1

(mod q − 1).

αj q j

j=0

Similarly we get the second equality if αj < q−1 for all j. If some αj = q−1, then for that j we get gαj qj (ξt) = gαj qj (t), if ξ = 0. Therefore we come to the required equality assuming that ξ = 0. (ii) To prove (1.43), note that αj

fj (t + s)

αj

= (fj (t) + fj (s))

=

αj αj l=0

l

fj (t)l fj (s)αj −l ,

26

Chapter 1

whence Gj (t + s) =

αj n−1 j=0 lj =0

αj fj (t)lj fj (s)αj −lj lj

α0 αn−1 ··· f0 (t)l0 · · · fn−1 (t)ln−1 = ··· l0 ln−1 l0 =0 ln−1 =0 × f0 (s)α0 −l0 · · · fn−1 (s)αn−1 −ln−1 α0

=

α0 l0 =0

αn−1

α0 αn−1 Gβ (t)Gi−β (s), ··· ··· l0 ln−1 αn−1

ln−1 =0

with β = l0 + l1 q + · · · + ln−1 q n−1 . It follows from Lucas’ theorem about binomial coeﬃcients modulo a prime number (see [46]) that αn−1 i α0 ··· (mod p), (1.47) ≡ l0 ln−1 β and it follows from (1.47) that i ≡ 0 (mod p) if β = l0 + l1 q + · · · + ln−1 q n−1 β with lj > αj for some value of j. This results in (1.43), as well as in (1.44) for the case where αj < q − 1 for every j. If αj = q − 1 for some j, then q − 1 fj (t)β fj (s)l − 1. gαj qj (t + s) = (fj (t) + fj (s))q−1 − 1 = β β+l=q−1

Considering separately the terms with β = 0 and β = q − 1, and repeating the above reasoning we come to (1.44). (iii) Writing k in the q-digit expansion, k = k0 + k1 q + · · · + kn−1 q n−1 , kn−1 = 0, we have Gk (t) = f0 (t)k0 f1 (t)k1 · · · fn−1 (t)kn−1 . If m ∈ Fq [x] with deg m < ν, then fj (m) = 0 for j ≥ ν. Therefore we need only prove (1.45) in the case where k < q ν . Thus, let k = k0 + k1 q + · · · + kν−1 q ν−1 , l = l0 + l1 q + · · · + lν−1 q ν−1 , with 0 ≤ ki ≤ q − 1, 0 ≤ li ≤ q − 1 for 0 ≤ i ≤ ν − 1. We have Gk (t) = f0 (t)k0 f1 (t)k1 · · · fν−1 (t)kν−1 ,

Orthonormal systems gl (t) = f0 (t)l0 − δ0 f1 (t)l1 − δ1 · · · fν−1 (t)lν−1 − δν−1 ,

27

where δj = δq−1,lj for 0 ≤ j ≤ ν − 1. Therefore Gk (t)gl (t) = f0 (t)k0 +l0 − δ0 f0 (t)k0 f1 (t)k1 +l1 − δ1 f1 (t)k1 · · · × fν−1 (t)kν−1 +lν−1 − δν−1 fν−1 (t)kν−1 . (1.48)

Denote His = fi (xi+s ), i, s ≥ 0. Let m ∈ Fq [x] with deg m < ν. Then m = cν−1 xν−1 + · · · + c1 x + c0 , cj ∈ Fq , so that

fi (m) = cν−1 Hiν−1−i + · · · + ci+1 Hi1 + ci Hi0 ,

0 ≤ i ≤ ν − 1.

Now, by (1.48),

Gk (m)gl (m)

deg m<ν

=

ν−1

ki +li cν−1 Hiν−1−i + · · · + ci+1 Hi1 + ci Hi0

c0 ,c1 ,...,cν−1 ∈Fq i=0

−

δi cν−1 Hiν−1−i

=

ν−1

+ ··· +

ci+1 Hi1

+

ki ci Hi0

ki +li cν−1 Hiν−1−i + · · · + ci+1 Hi1 + ci Hi0

c1 ,...,cν−1 ∈Fq i=1

− ×

δi cν−1 Hiν−1−i

+ ··· +

ci+1 Hi1

+

ki ci Hi0

k0 +l0 cν−1 H0ν−1 + · · · + c1 H01 + c0 H00

c0 ∈Fq

0 k0

− δ0 cν−1 H0ν−1 + · · · + c1 H01 + c0 H0

.

28

Chapter 1

Let B = cν−1 H0ν−1 + · · · + c1 H01 . Since H00 = 1, we get

k0 +l0 cν−1 H0ν−1 + · · · + c1 H01 + c0 H00

c0 ∈Fq

k0 − δ0 cν−1 H0ν−1 + · · · + c1 H01 + c0 H00 =

(B + c0 )k0 +l0 − δ0

c0 ∈Fq

def

(B + c0 )k0 = Σ1 − δ0 Σ2 .

c0 ∈Fq

We ﬁnd that Σ2 = B

k0

k0 k0 B k0 −j + j j=0

cj0

=

0=c0 ∈Fq

0,

if k0 = q − 1;

−1,

if k0 = q − 1,

since for 0 ≤ j ≤ k0 ≤ q − 1 we have

cj0

=

0=c0 ∈Fq

−1,

if j = 0, q − 1;

0,

otherwise.

Similarly, Σ1 = B k0 +l0 +

k 0 +l0 j=0

k0 + l0 B k0 +l0 −j j

cj0

0=c0 ∈Fq

  if 0 ≤ k0 + l0 < q − 1;  0, k +l −q+1 k0 +l0 0 0 = − q−1 B , if q − 1 ≤ k0 + l0 < 2q − 2;   −k0 +l0 B q−1 − 1, if k0 + l0 = 2q − 2, q−1   if 0 ≤ k0 + l0 < q − 1;  0, = −1, if k0 + l0 = q − 1;   −1, if 0 ≤ k + l = 2q − 2. 0

0

Therefore Σ1 − δ0 Σ2 =

0,

if k0 + l0 = q − 1;

−1,

if k0 + l0 = q − 1.

Orthonormal systems

29

Now, if k0 + l0 = q − 1, then Gk (m)gl (m) deg m<ν

=−

ν−1

c1 ,...,cν−1 ∈Fq i=1

ki +li cν−1 Hiν−1−i + · · · + ci+1 Hi1 + ci Hi0

ν−1−i 1 0 ki . + · · · + ci+1 Hi + ci Hi − δi cν−1 Hi

Summing up on c1 , . . . , cν−1 , we get the identity (−1)ν , if k0 + l0 = . . . = kν−1 + lν−1 = q − 1; Gk (m)gl (m) = 0, otherwise, deg m<ν which coincides with (1.45). (iv) The proof of the identity (1.46) is similar to that of (1.45).

We use the orthogonality relation (1.45) to obtain an explicit formula for the coeﬃcients of the expansion f (t) =

∞

ai Gi (t),

t ∈ O,

(1.49)

i=0

of an arbitrary function f ∈ C(O, K). Theorem 1.15 The coeﬃcients ai → 0 of the expansion (1.49) are expressed as follows: gqν −i−1 (m)f (m), (1.50) ai = (−1)ν deg m<ν

for any integer ν, such that q ν > i. Proof. By (1.49), the right-hand side of (1.50) equals ∞ i=0

ai (−1)ν

gqν −i−1 (m)Gi (m),

deg m<ν

which is equal to ai by virtue of (1.45). Remark 1.2 Following Remark 1.1(4) of Sect. 1.2.1, we conclude from the formulas for general Carlitz polynomials and the expression (1.50) that the sequence {Gi }∞ 0 forms an A-basis of the A-module Int(A) of all

30

Chapter 1

integer-valued polynomials from A[t]. The coeﬃcients of the expansion of polynomials in the basis {Gi } can be obtained by (1.50) in this situation too. It is well known (see, for example, [77], p. 420) that p−1 ≡ (−1)n (mod p), n

(1.51)

if 0 ≤ n ≤ p − 1. Using (1.47) we ﬁnd from (1.51) that for any ν ∈ N ν q −1 ≡ (−1)n (mod p), (1.52) n if 0 ≤ n ≤ q ν − 1. The congruence (1.52) leads to the following corollary to Proposition 1.14. Corollary 1.16 The polynomials Gqν −1 and gqν −1 satisfy the following identities: Gqν −1 (t + s) = (−1)j Gj (t)Gl (s); j+l=q ν −1

Gqν −1 (t − s) =

Gj (t)Gl (s);

j+l=q ν −1

gqν −1 (t + s) =

(−1)j Gj (t)gl (s);

j+l=q ν −1

gqν −1 (t − s) =

Gj (t)gl (s).

j+l=q ν −1

The polynomial system {gqν −1 } has another interesting property. Proposition 1.17 The system of polynomials gqν −1 (t), t ∈ O, ν = 0, 1, 2, . . ., is orthonormal. Proof. By Corollary 1.16, gqν −1 (t) =

ν q −1

Gj (t)gqν −1−j (0)

(1.53)

j=0

where g0 (0) = 1. On the other hand, by the deﬁnition of gj , g

q ν −1

(t) =

ν−1

fiq−1 (t) − 1 .

i=0

(1.54)

Orthonormal systems

31

It follows from (1.54) that gqν −1 ≤ 1 and gqν −1 (0) = (−1)ν , so that gqν −1 = 1 for all ν. In order to prove orthogonality, it suﬃces to show that for any n ∈ N, λ1 , . . . , λn ∈ K, n λk gqk −1 ≥ |λn | k=0

(see the proof of Proposition 50.4 in [98]). By (1.53), n

λk gqk −1 (t) =

k=0

n q −1

j=0

Gj (t)

λk gqk −j−1 (0).

logq (j+1)≤k≤n

Since the sequence {Gj } is orthonormal, n λk gqk −1 = max λk gqk −j−1 (0) ≥ |λn g0 (0)| = |λn | j k=0

logq (j+1)≤k≤n

(for j = q n − 1 the sum consists of a single term), as desired.

1.4.3. Extended hyperdiﬀerentiations. It is easy to see that the results of Section 1.4.2 remain valid if we consider, instead of the general Carlitz polynomials, the general construction (1.41), provided 0, if k < j, k ϕj (x ) = 1, if k = j. In particular, this applies to the case where ϕj = Dj . The sequence of functions obtained from the hyperdiﬀerentiations Dj via the digit construction (1.41) (“extended hyperdiﬀerentiations”) satisfy all the assertions similar to those of Proposition 1.14, Theorem 1.15, Corollary 1.16 and Proposition 1.17. The role of the polynomials gj is played in this case by the functions n−1 di (t) = dαj qj (t) j=0

where, just as above, i = α0 + α1 q + · · · + αn−1 q n−1 , 0 ≤ αj ≤ q − 1, α if 0 ≤ αj < q − 1; Dj j (t), dαj qj (t) = αj Dj (t) − 1, if αj = q − 1.

32

Chapter 1 1.5 Finite places of a global function ﬁeld

1.5.1. Main notions. While classical analysis usually deals with the ﬁelds R and C, in algebraic number theory the main objects are global ﬁelds and objects deﬁned over them, like quadratic forms, diophantine equations, various number-theoretic functions, etc. In characteristic zero, the global ﬁelds are the ﬁeld Q of rational numbers and its ﬁnite algebraic extensions. In positive characteristic p, the global ﬁelds are ﬁnite separable extensions of the function ﬁeld Fq (x) consisting of all rational functions of x with coeﬃcients from Fq . The local ﬁeld K considered above is one of the possible completions of Fq (x). It is clear that global ﬁelds are objects of a discrete nature. Since this book is devoted to analysis on local ﬁelds, the global ﬁelds remain essentially outside its scope. Nevertheless in this section we introduce some basic notions regarding completions of Fq (x), to be able to mention some global aspects of the objects we consider. Let λ be an isomorphic imbedding of Fq (x) into a local ﬁeld L. A couple (λ, L) is called a completion of Fq (x), if λ(Fq (x)) is dense in L. Two completions (λ, L) and (λ , L ) are called equivalent, if there exists an isomorphism ρ of the ﬁeld L onto L such that λ = ρ ◦ λ. An equivalence class of completions of the ﬁeld Fq (x) is called a place. There are the following explicit constructions of completions of Fq (x). θ1 Let 0 = t ∈ Fq (x); then t = , θ1 , θ2 ∈ Fq [x]. Set θ2 |t|∞ = q deg θ1 −deg θ2 ,

|0|∞ = 0.

It is easy to check that | · |∞ is a non-Archimedean absolute value on Fq (x), it deﬁnes a metric on Fq (x), and the completion K∞ of Fq (x) with respect to that metric is a local ﬁeld. Thus, the absolute value | · |∞ deﬁnes a place of Fq (x) often called the inﬁnite place. Note that |x|∞ = q, |x−1 |∞ = q −1 , so that x−1 is a prime element of K∞ (see [17, 38, 122]) for various notions regarding local ﬁelds). Another construction of places of Fq (x) is as follows. Let π ∈ Fq [x] be a τ monic irreducible polynomial, deg π = δ ≥ 1. If t ∈ Fq (x), write t = π n τ where n ∈ Z, τ, τ ∈ Fq [x], and π does not divide τ, τ . In this case we also use the notations n = ordπ t and t ≡ 0 (mod π n ). Set |0|π = 0, |t|π = |π|nπ ,

|π|π = q −δ .

We will denote by Kπ the completion of Fq (x) with respect to the metric

Orthonormal systems

33

determined by this absolute value. Again, this is a non-Archimedean local ﬁeld. For π(x) ≡ x, we obtain the ﬁeld K considered in the preceding sections. Denote by Oπ the ring of integers, and by Pπ the prime ideal in Kπ , that is Oπ = {t ∈ Kπ : |t|π ≤ 1} ,

Pπ = {t ∈ Kπ : |t|π < 1} .

Theorem 1.18 (i) The local ﬁelds K∞ and Kπ give the complete list of diﬀerent places of the global ﬁeld Fq (x). (ii) For each local ﬁeld Kπ , the cardinality of its residue ﬁeld Fπ = Oπ /Pπ equals q δ , and a full system of representatives of the residue classes consists of all polynomials from Fq [x] of degrees < δ. For the proof see [122] (note that our notation is slightly diﬀerent from that in [122]). The places of Fq (x) corresponding to the completions Kπ are often called ﬁnite. Of course, all the ﬁelds Kπ and K∞ are isomorphic to K; to obtain the isomorphism, it suﬃces to take the prime element (respectively, π and x−1 ) for the new variable. Thus, the global ﬁeld structures do not bring anything new on the local ﬁeld level. However they lead to quite diﬀerent properties of objects deﬁned in terms of a global ﬁeld; in particular, the inﬁnite place is usually very diﬀerent from the ﬁnite ones, while properties of ﬁnite places are often similar to each other. 1.5.2. The Carlitz polynomials. It is clear from (1.11) or (1.14) that the Carlitz polynomials ei and the normalized Carlitz polynomials fi are deﬁned globally, they belong to Fq [x][t] and Fq (x)[t], respectively. In this section we prove that the system {fi } and its extension by the digit principle form orthonormal bases in Banach spaces associated with all the ﬁnite places of Fq (x). First we need the following auxiliary result. Lemma 1.19 Let π be irreducible in Fq [x], deg π = δ. If j < δn, then fj (π n g) ≡ 0 (mod π) for all g ∈ Fq [x].

34

Chapter 1

Proof. We may suppose that g = 0, and we have to show that ordπ (ej (π n g)) > ordπ (Dj ). For an integer k ≥ 0, let k ≡ Rk (mod δ), where 0 ≤ Rk ≤ δ − 1. In particular, we write j = δQ + Rj . By Lemma 2.13 from [76], 1, if δ divides n; (1.55) ordπ ([n]) = 0, otherwise. From (1.55) and the deﬁnition (1.1) of Dk we get by an easy computation that q k − q Rk . (1.56) ordπ (Dk ) = δ q −1 Since deg(π n g) > j, we ﬁnd from (1.56) and Lemma 1.1(iii) that ordπ (ej (π n g)) = n + ordπ (g) +

j−1

(q − 1) ordπ (Dk )

k=0

= n + ordπ (g) +

j−1

(q − 1)

k=0 j

= n + ordπ (g) +

q k − q Rk qδ − 1

q − q Rj −Q qδ − 1

= n + ordπ (g) + ordπ (Dj ) − Q > ordπ (Dj ), since n >

j ≥ Q. δ

Let C0 (Oπ , Kπ ) be the vector space over Kπ of all continuous Fq -linear functions Oπ → Kπ , endowed with the supremum norm. Theorem 1.20 For any irreducible polynomial π ∈ Fq [x], δ = deg π ≥ 1, the polynomials fj (t), viewed as elements of C0 (Oπ , Kπ ), form an orthonormal basis in C0 (Oπ , Kπ ). Proof. Let n be an arbitrary natural number. By Lemma 1.19, for j < δn the reduction fj is a well-deﬁned mapping from Fq [x]/(π n ) to Fq [x]/(π) ∼ = Fπ . Here (π n ) and (π) are the principal ideals in Fq [x] generated by π n and π, respectively. For 0 ≤ j, k ≤ δn − 1, the δn × δn matrix fj (xk ) #is triangular, with " all diagonal entries equal to 1. Since 1, x, . . . , xδn−1 is an Fq -basis of

Orthonormal systems

35

∼ Fq [x]/(π n ), it follows that the δn reduced functions fj form Oπ /(π n ) = an Fπ -basis of HomFq (Fπ [[π]]/(π n ), Fπ ). Using Proposition 1.5 we come to the conclusion that {fj } is an orthonormal basis in C0 (Oπ , Kπ ). Of course, for π(x) = x the assertion of Theorem 1.20 coincides with the ﬁrst assertion of Theorem 1.8. Thus, we have just presented a second proof of the latter. An obvious task now is to construct an orthonormal basis of the Banach space C(Oπ , Kπ ) of all the continuous Kπ -valued functions on Oπ . The digit principle does not carry over in a straightforward way, because C0 (Oπ , Kπ ) consists, by deﬁnition, of all Fq -linear functions, while the residue ﬁeld Fπ = Oπ /Kπ is in general bigger than Fq . In spite of this discrepancy, the following result is valid. Theorem 1.21 The system of general Carlitz polynomials is an orthonormal basis of the space C(Oπ , Kπ ). Proof. As before, we denote δ = deg π ≥ 1. Let fj : Oπ → Fπ be reductions of the normalized Carlitz polynomials fj . As in the proof of Theorem 1.13, denote δn−1 & ker fj . Hn = j=0

# We already know that f0 , . . . , fδn−1 is an Fπ -basis of the set of all Fq -linear maps from Fq [x]/(π n ) to Fq [x]/(π). Therefore these functions separate the points of Fq [x]/(π n ). Now the argument from the proof of Theorem 1.13 shows that the set Maps(Oπ /Hn , Fπ ) is spanned over Fπ by the monomials "

f0

β0

· · · fδn−1

βδn−1

,

0 ≤ βj ≤ q δ − 1. qk

k

Note that any fjq is Fq -linear, so in Maps(Oπ /Hn , Fπ ) we can write fj # " as an Fπ -linear combination of f0 , . . . , fδn−1 . Therefore for all n ≥ 1, Maps(Oπ /Hn , Fπ ) is spanned over Fπ by the monomials f0

γ0

· · · fδn−1

γδn−1

,

0 ≤ γj ≤ q − 1,

so this set is an Fπ -basis. It remains to use Proposition 1.5.

The explicit expressions for the coeﬃcients of expansions in the Carlitz

36

Chapter 1

polynomials remain valid in the above global ﬁeld framework (for ﬁnite places). However we have to stress that all the above results fail for the completion K∞ ; the Carlitz polynomials do not even take integral elements to integral elements at this place. For the use of the Carlitz polynomials in an investigation of entire functions on K∞ see [21]. 1.5.3. Hyperdiﬀerentiations. Let us consider, in the framework of Sections 1.5.1 and 1.5.2, the hyperdiﬀerentiations Dk (t) investigated in Sections 1.3 and 1.4.3. The Leibnitz rule (1.29) extends to the case of more than two factors, as Dj1 (t1 ) · · · Djm (tm ). Dk (t1 · · · tm ) = j1 +·jm =k j1 ,...,jm ≥0

It follows that for any polynomials t, s ∈ Fq [x], Dk (tn s) ≡ 0 (mod tn−k ),

n ≥ k.

(1.57)

If, as above, Kπ is a completion of Fq (x) corresponding to a monic irreducible polynomial, then (1.57) implies the continuity of Dk (t) as a function on Oπ with values in Kπ . It can be proved [29] that the functions Dk are ∞ actually not only Fq -linear but even Fπ -linear. The sequence {Dk (t)}k=0 is an orthonormal basis of the space of all Fπ -linear continuous functions Oπ → Kπ . In order to construct an orthonormal basis of C(Oπ , Kπ ), one can use Theorem 1.13. The resulting basis is diﬀerent from the basis of extended hyperdiﬀerentiations considered in Sect. 1.4.3, because here the construction (similar to (1.41)) must use the q δ -base representation of integers, not the q-base representation as before. Just as for the Carlitz polynomials, the inﬁnite place does not ﬁt into the above picture. Though the functions Dk can be extended from Fq (x) onto K∞ , they are not orthonormal in this case. See [29] for further details regarding the global ﬁeld treatment of hyperdiﬀerentiations.

1.6 The Carlitz module 1.6.1. The Carlitz exponential and logarithm. The Carlitz exponential over K is deﬁned by the power series eC (z) =

j ∞ zq

j=0

Dj

.

(1.58)

Orthonormal systems

37

It follows from (1.3) that the series (1.58) is convergent if z ∈ K c , |z| < 1 q − q−1 . Before writing down an explicit formula for the function inverse to eC , ∞ we prove the following lemma. Let {an }∞ 0 , {bn }0 be sequences in K c . Lemma 1.22 Suppose that

i ai bqj

=

i+j=k

Then

0,

for k > 0,

1,

for k = 0.

j bj aqi

=

i+j=k

0,

for k > 0,

1,

for k = 0.

(1.59)

(1.60)

Proof. For k = 0, both the equalities (1.59) and (1.60) mean that a0 b0 = 1. For k = 1, (1.59) means that a0 b1 + a1 bq0 = 0, whence b1 + a1 bq+1 = 0. 0 Then

= 0, b0 a1 + b1 aq0 = aq0 b1 + a1 bq+1 0 and this is the equality required for k = 1 in (1.60). Suppose that (1.60) is true for all k < m. Then, by (1.59), m−1

l

m

bl aqm−l = −aq0

m−1 l=0

l=0

=

m −aq0

bl

l

aqi bqj

m−l=i+j l

bl aqi bqj

l+j

m=l+i+j j>0

=

m −aq0

 

m=k+j j>0

l+j



k=l+j

l bl aqi

  

k

bqj .

Since j > 0, k < m, the equality (1.60) is valid for the inner sum, and we ﬁnd that m−1 l m bl aqm−l = −aq0 bm , l=0

so that (1.60) holds for k = m. This completes the proof.

38

Chapter 1 The Carlitz logarithm logC (z) is deﬁned as ∞

logC (z) =

nz

(−1)

n=0

qn

Ln

.

(1.61)

Since |Ln | = q −n , the series in (1.61) converges if |z| < 1. Below it will be considered for z ∈ K c , |z| < q −1/(q−1) . Theorem 1.23 The Fq -linear functions eC and logC are inverse to each other. Proof. We note ﬁrst of all that eC and logC are mutually inverse as formal power series. Indeed, we have to prove that l 0, for l > 0; (−1)l−n (1.62) qn = Dn Ll−n 1, for l = 0, n=0 l (−1)n = n L Dq n=0 n l−n

0,

for l > 0;

1,

for l = 0.

(1.63)

The left-hand side of (1.62) equals fl (1) (see Proposition 1.7), so that the equalities (1.62) follow from (1.11) and (1.15). Now the equalities (1.63) are consequences of Lemma 1.22. By (1.61) and the ultra-metric inequality, n

|logC (z)| ≤ sup q n |z|q . n≥0

The function ψz (s) = s|z| decreases for s > −(log |z|)−1 ; if |z| < q −1/(q−1) , q−1 n then ψz decreases for s > log q . In particular, ψz (q ) ≤ ψz (q), n ≥ 1. q Hence, |logC (z)| ≤ max(|z|, q|z| ), and we ﬁnd that s

1

1

|logC (z)| < q − q−1 ,

if |z| < q − q−1 .

Therefore eC ◦ logC is well-deﬁned, and eC (logC (z)) = z, 1

1

if |z| < q − q−1 . 1

Similarly, |eC (z)| < q − q−1 , if |z| < q − q−1 , and logC (eC (z)) = z.

It follows immediately from the deﬁnition of the Carlitz diﬀerence operators ∆, ∆(j) (see (1.17), (1.18)) that ∆eC (t) = eC (t)q ,

j

∆(j) eC (t) = eC (t)q .

(1.64)

Orthonormal systems

39

If we introduce an Fq -linear operator √ d= q ◦∆

(1.65)

called the Carlitz derivative, then the ﬁrst equality (1.64) takes the form deC = eC .

(1.66)

The equation (1.66) is the ﬁrst example of a diﬀerential equation with Carlitz derivatives. See Chapter 3 for general results regarding such equations. 1 While the Carlitz exponential eC (t) is deﬁned for |t| < q − q−1 , it is useful to introduce the function 1

z ∈ K c , |z| < q − q−1 ,

wz (t) = eC (tz),

deﬁned for t ∈ O. Using Theorem 1.8 and the identities (1.64) we ﬁnd an explicit expansion of the function wz in the normalized Carlitz polynomials: wz (t) =

∞

qn

(eC (z))

fn (t),

1

t ∈ O, |z| < q − q−1

(1.67)

n=0

(note that when we deal with expansions in the Carlitz polynomials, we have to consider only t ∈ O ⊂ K). 1.6.2. The Carlitz module. Consider the Carlitz module function Cs (t) =

∞

n

fn (s)tq ,

s ∈ O, t ∈ K c , |t| < 1,

(1.68)

n=0

Fq -linear in each of its arguments s, t. If s ∈ Fq [x], then the series in (1.68) is actually a ﬁnite sum, from n = 0 to n = deg s. In the general case, the convergence of the series follows from the fact that fn = 1 for each n = 0, 1, 2, . . . It follows from (1.67) that Cs (eC (z)) = eC (sz),

1

s ∈ O, |z| < q − q−1 .

(1.69)

The identity (1.69) is the main functional equation for the Carlitz exponential. It can also be rewritten as a relation for the Carlitz logarithm: logC (Cs (z)) = s logC (z). If s1 , s2 ∈ O, then by (1.69), Cs1 (Cs2 (eC (z))) = Cs1 s2 (eC (z)).

40

Chapter 1 1

As we know from Theorem 1.23, the set of values !of eC (z), |z| < q − q−1 , 1 coincides with the disk V = z ∈ K c : |z| < q − q−1 . Therefore Cs1 (Cs2 (ζ)) = Cs1 s2 (ζ)

(1.70)

for all ζ ∈ V . In particular, if s1 , s2 ∈ Fq [x], then Cs1 and Cs2 are Fq linear polynomials in ζ. In this case the identity (1.70) means that Cs deﬁnes a homomorphism of the ring Fq [x] into the composition ring K{ζ} of Fq -linear polynomials with coeﬃcients from K. The Carlitz module is the ﬁrst nontrivial example of such a homomorphism. A general theory of such homomorphisms was initiated by Drinfeld [33], so that they are now called (under some natural assumptions) the Drinfeld modules. Note that analogs of the Carlitz exponential exist in very general situations. The theory of Drinfeld modules is the most essential part of function ﬁeld arithmetic; see [45, 111] for detailed expositions in a spirit quite diﬀerent from the above approach. In the literature all the objects are considered in the global ﬁeld setting, usually over K∞ . Cs (z) is introduced only in the polynomial case, s ∈ Fq [x]; on the other hand, instead of the ring Fq [x], much more general rings can be considered. Over K∞ , the Carlitz exponential is an entire function; a detailed study of its properties including a description of its periods (from an extension of K∞ ) and special values, as well as their number-theoretic applications, is given in [45]. We do not try to repeat that material in this book, due to its more analytic and elementary character.

1.7 Canonical commutation relations 1.7.1. The background. In the quantum mechanics of harmonic oscillator (see e.g. [81, 106]) a creation operator a+ transforms a stationary state into a stationary state of the next (higher) energy level; an annihilation operator a− acts in the opposite way. The operators satisfy the canonical commutation relation (CCR) a− a+ − a+ a− = I. In quantum ﬁeld theory these properties are used to obtain operators which change the number of particles. The most widely used representation of the CCR is the Schr¨odinger representation, in which a± are operators on the Hilbert space L2 (R). If the mass, frequency and the Planck constant are all taken equal to 1, then 1 d a± = √ t∓ dt 2

Orthonormal systems

41

where we identify the variable t and the operator of multiplication by t. Let {hn }∞ 0 be the system of Hermite functions, an orthonormal basis of L2 (R). The operators a± act on this basis as follows: √ √ a+ hn = n + 1hn+1 , a− hn = nhn−1 , n = 0, 1, . . . , (1.71) where h−1 = 0. For “the number operator” a+ a− we have a+ a− hn = nhn ,

n = 0, 1, . . .

(1.72)

The above relations hold also for the Bargmann–Fock representation of the CCR. Here the operators a ˜± (satisfying the same relation a ˜− a ˜+ −˜ a+ a ˜− = I) act on the Hilbert space of entire functions u(z) =

∞

zn cn √ , n! n=0

with the inner product (u1 , u2 ) =

1 π

'

z ∈ C,

∞

|cn |2 < ∞,

n=0

2

u1 (z)u2 (z)e−|z| dz.

C

a− u)(z) = u (z). Instead ofthe Hermite Here (˜ a+ u)(z) = zu(z), (˜ ∞ functions, zn the orthonormal basis with the above properties is √ . n! n=0 An important object related to the CCR is the system of coherent states, generalized eigenfunctions (not necessarily belonging to the Hilbert space) of the annihilation operator. In the Bargmann–Fock representation these are the functions z → eλz , λ ∈ C. Analogs of the above constructions are known also in p-adic analysis. An analog of the Schr¨ odinger representation [60] is as follows. The operators ± a act on the space C(Zp , Qp ) of p-adic-valued continuous functions on the ring Zp of p-adic integers, (a+ u)(t) = tu(t − 1),

(a− u)(t) = u(t + 1) − u(t),

t ∈ Zp .

(1.73)

Note that it is a purely non-Archimedean phenomenon that a compact set Zp is preserved under the unit shift of the argument in (1.73). Instead of (1.71)–(1.72) we have in this case a− Pn = Pn−1 , n ≥ 1; a+ Pn = (n + 1)Pn+1 ,

a− P0 = 0, n ≥ 0,

42

Chapter 1

so that (a+ a− )Pn = nPn and, as before, a− a+ − a+ a− = I. Here {Pn } is the Mahler basis of C(Zp , Qp ) (see [98]), that is Pn (t) =

t(t − 1) · · · (t − n + 1) , n ≥ 1; n!

P0 (t) ≡ 1.

The analogs of coherent states are the functions fλ (t) = (1 + λ)t , where |λ|p < 1. Various p-adic analogs of the Bargmann–Fock representation are given in [2, 57, 60]. It is interesting that, in contrast to the classical case, all the above p-adic operators a± are bounded. The problem of existence of linear bounded representations of CCR over valued ﬁelds of arbitrary rank is investigated in [56]. Below we construct a characteristic p representation of CCR. Just as in the above characteristic zero cases, the basic objects happen to be related to the main special functions of the corresponding branch of analysis – this time, to the Carlitz polynomials and the Carlitz exponential. 1.7.2. The representations. Denote τ u = uq . As before, we write √ d = q ◦ ∆. All the operators below are considered on the space C0 (O, K c ) of Fq -linear continuous functions on O with values in K c . Theorem 1.24 Let a+ = τ − I, a− = d. Then a− a+ − a+ a− = [1]1/q I.

(1.74)

The operator a+ a− possesses an orthonormal eigenbasis consisting of the normalized Carlitz polynomials: (a+ a− )fi = [i]fi ,

i = 0, 1, 2, . . . ;

(1.75)

a+ and a− act upon the basis as follows: a+ fi−1 = [i]fi ,

a− fi = fi−1 , i ≥ 1;

a− f0 = 0.

(1.76)

The equation a− u = λu

(1.77)

has solutions (“coherent states”) for any λ ∈ K c ; if λ = 0, then each solution can be written as q

u(t) = λ− q−1

∞ n=0

n

cq fn (t),

c ∈ K c , |c| < 1,

(1.78)

Orthonormal systems

43

for some value of the (q − 1)-th root, and conversely, a− u = λu for the 1 function (1.78). If in (1.78) |c| < q − q−1 , then q

u(t) = λ− q−1 eC (tz),

z = logC (c).

(1.79)

In particular, if q = 2, then every function (1.78) with c ∈ K takes the form (1.79). Proof. The identities (1.74)–(1.76) are checked by a direct computation, with the use of the properties (1.12) and (1.21) of the Carlitz polynomials. Let u be a solution of (1.77). We can write u(t) =

∞

cn fn (t),

t ∈ O, cn → 0.

n=0

Applying the operator d, we get that λu =

∞

c1/q n fn−1 ,

n=1

and, by the uniqueness of the expansion, that 1/q

cn+1 = λcn ,

n = 0, 1, . . . ,

whence cn = λq

n

n +q n−1 +···+q q c0

= µ−1 (c0 µ)q , n

n = 1, 2, . . . ,

q

where µ = λ q−1 . Since cn → 0, we have |c0 µ| < 1, and we obtain the representation (1.78) with c = c0 µ. The representation (1.79) follows from (1.67). If q = 2, c ∈ K, |c| < 1, then |c| ≤ q −1 < q −1/(q−1) , so that in this case (1.78) is equivalent to (1.79). Note that the operators a± are obviously continuous on C0 (O, K c ) but, in contrast to both the classical and p-adic cases, are not linear, only Fq linear. In order to construct an analog of the Bargmann–Fock representation (a representation by operators on a space of holomorphic functions), consider a Banach space H over the ﬁeld K c consisting of Fq -linear power series n ∞ tq , an ∈ K c , |an | → 0. (1.80) an u(t) = Dn n=0

44

Chapter 1

The norm in H is given by u H = sup |an |. n

It follows from (1.3) that a seies (1.80) deﬁnes a holomorphic function for n tq |t| < q −1/(q−1) . It is obvious that the sequence of functions f˜n (t) = , Dn n = 0, 1, 2, . . ., is an orthonormal basis of H. The desired representation is given by the following operators on the space H: a ˜+ = τ,

a ˜− = d.

Note that the form of the operators a ˜± is only slightly diﬀerent from that ± of a , but they act on a diﬀerent Banach space. By a straightforward computation based on the identities n n ∆ tq = [n]tq , Dn+1 = [n + 1]Dnq , we show that the relations (1.74)–(1.76) hold for the operators a ˜± , with f˜n substituted for fn .

1.8 Comments The factorial-like sequences (1.1), (1.2), and (1.4) were introduced, among other basic constructions of the analysis in positive characteristic, by Carlitz [22, 23]. Studying their properties (Lemma 1.1) Carlitz used a diﬀerent technique (the Moore determinants); see also [45]. Our exposition follows, in a little more detailed form, the proof of Proposition 3.1.6 in [45] (see also Theorem 3.20 in [76]). Proposition 1.2 is taken from [45]. Composition rings of locally holomorphic Fq -linear functions and the corresponding skew ﬁelds of “meromorphic” functions (Propositions 1.3, 1.4) were studied by the author [66]. Note that a detailed investigation ∞ k of bi-inﬁnite series ak tq convergent on the whole of K c was carried k=−∞

out by Poonen [85]. The results regarding orthonormal bases (Propositions 1.5, 1.6) are taken from [3, 101]. Properties (1.12)–(1.15) of the additive Carlitz polynomials were proved by Carlitz in the seminal paper [22] where the polynomials were introduced for the ﬁrst time. Our proofs follow [45]. The identity (1.16) is due to Wagner [119]. In fact, Carlitz considered only expansions of polynomials,

Orthonormal systems

45

while the expansions of continuous functions in the system {fi } (Theorem 1.8) were ﬁrst studied by Wagner [118, 119]. There are several diﬀerent proofs of this result [118, 119, 43, 61, 29, 32]; we followed [119]. Theorem 1.10 was proved in [64]. Hyperdiﬀerentiations were introduced by Hasse [48] and studied in a more general context in [108, 49, 99]; for various generalizations see [112] and references therein. An interpretation of hyperdiﬀerentiations as functions on K was ﬁrst given by Voloch [117] who proved Proposition 1.11 (i), (iii). Other results of Section 1.3 were obtained by Jeong [53, 55] and Snyder [105]. The connections (1.36), (1.39) of hyperdiﬀerentiations with powers of the operator ∆ were used by Jeong [54] to solve some diﬀerence equations containing ∆. The digit construction (1.41) was ﬁrst proposed by Carlitz [23] for the case of the Carlitz polynomials. In [23], Carlitz proved all the results of Section 1.4.2 (see also [43]) except Proposition 1.17 proved in [62]. Our proof of the crucial orthogonality relation (1.45) is taken from the dissertation by Yang [123]. It was noticed by Jeong [53] that all the constructions are applicable also to the hyperdiﬀerentiations. The general digit principle (Theorem 1.13) was proved by Conrad [29] who also found its analog for the case of characteristic 0. Apart from the above explicit constructions of bases for the positive characteristic case, there are also various general results, applicable both for the cases of our ﬁeld K, the ﬁeld Qp of p-adic numbers, and its ﬁnite extensions. See [3, 7, 15, 19, 20, 29, 104, 107, 118]. Further constructions of orthonormal bases in C0 (O, K) are provided by an appropriate version of the umbral calculus; see Chapter 2 below. The global ﬁeld interpretation of the Carlitz polynomials was initiated by Wagner [118]; see also [43, 29]. Our proofs of Theorems 1.20 and 1.21 are taken from [29] where the case of hyperdiﬀerentiations is also considered in detail. Note that the Carlitz expansions at ﬁnite places have been used by Goss [44] to study, in the spirit of the Iwasawa isomorphism, Oπ -valued measures on Oπ . See Section 2.6. The notions of the Carlitz exponential and logarithm, as well as the Carlitz module identity (1.69) (for s ∈ Fq [x]) were introduced by Carlitz [22]; his deﬁnitions had slightly diﬀerent normalizations. We follow [43, 45]. The function ﬁeld representations of the canonical commutation relations

46

Chapter 1

were constructed by the author [61, 62]. In this book we do not touch on non-Archimedean models in physics; see [57, 115].

2 Calculus

This chapter is devoted to function ﬁeld counterparts of smoothness and analyticity, and their interplays with the rate of decay of the coeﬃcients of the Fourier–Carlitz series. This is connected with the Carlitz diﬀerence operators and the Carlitz derivatives (see Chapter 1). The latter notion, via the appropriate notion of an antiderivative, leads to the Volkenborn-type integration theory for Fq -linear functions. There is another extension that is a kind of fractional derivative. Wider classes of operators on Fq -linear functions are introduced in the Fq -linear umbral calculus and applied to construction of new orthonormal polynomial systems. 2.1 Fq -Linear calculus 2.1.1. Taylor coeﬃcients of Fq -linear holomorphic functions. Let u(t) =

∞

n

cn tq ,

cn ∈ K c ,

(2.1)

n=0

and the series (2.1) have a nonzero radius of convergence. The classical formula for the coeﬃcients of a power series cannot be used to reconstruct the coeﬃcients of (2.1) – in higher order terms, both the numerators, the higher derivatives, and the denominators, the factorials, vanish. The next result shows that the Carlitz diﬀerence operators (1.18) and the Carlitz factorials (1.1) are the appropriate replacements. Theorem 2.1 The coeﬃcients of the series (2.1) can be found as follows: cn =

1 ∆(n) u(t) lim , Dn t→0 tqn 47

n = 0, 1, 2, . . .

(2.2)

48

Chapter 2

Proof. It will be convenient to rewrite the series (2.1) as u(t) =

∞

n

an

n=0

tq , Dn

an = cn Dn . We may also assume that an → 0; the general case is reduced to this one by substituting λt for t with an appropriate λ (transforming the disk of convergence into the unit disk). In the notation of Section 1.7.2, now u belongs to the Banach space H. n tq If f˜n (t) = , then by the identities of the Bargmann–Fock-type repreDn sentation of Section 1.7.2,

d λf˜n = λ1/q f˜n−1 , n ≥ 1, λ ∈ K c ; df˜0 = 0. It follows that dn u(t) =

∞

1/q n

ak

k=n

k−n

tq , Dk−n

so that 1/q n

ak

dn u(t) . t→0 t

= lim

(2.3)

On the other hand, from the identities (1.76) and (1.23) for the normalized Carlitz polynomials we ﬁnd that τ n dn = ∆(n) .

(2.4)

Applying τ n to both sides of (2.3) and using (2.4) we come to (2.2).

2.1.2. Smooth functions. Let u ∈ C0 (O, K c ), Dk u(t) = t−q ∆(k) u(t), k

t ∈ O \ {0}.

We will say that u ∈ C0k+1 (O, K c ) if Dk u can be extended to a continuous function on O. C0k+1 (O, K c ) can be considered as a Banach space over K c , with the norm sup |u(t)| + |Dk u(t)| . t∈O

Note that coincides with the set of all diﬀerentiable Fq -linear functions O → K c . In this section we will obtain a characterization of functions from C01 (O, K c )

Calculus

49

C0k+1 (O, K c ) in terms of coeﬃcients of the expansion u =

∞

cn fn . We

n=0

begin with several auxiliary results. In particular, the next lemma, a necessary condition of the diﬀerentiability of a Fq -linear function, is a special case of the main result of this section, which will be proved later. Lemma 2.2 If u ∈ C01 (O, K c ), then |cn |q n −→ 0,

n → ∞.

(2.5)

Proof. First we prove that the sequence {|cn |q n } is bounded. Assuming the opposite, we ﬁnd a strictly increasing sequence {nr } such that lim |cnr |q nr = ∞. Choosing, if necessary, an appropriate subsequence, we r→∞ may assume, for 0 ≤ n < nr , that |cn |q n < |cnr |q nr .

(2.6)

Suppose that u (0) = λ. It follows from the identity (1.54) that t−1 fn (t) =

gqn −1 (t) , Ln

n = 0, 1, 2, . . .

(2.7)

(the polynomials at both sides of (2.7) have the same roots t = xl , l < n, the same degrees and the same leading coeﬃcients). If ϕ(t) = t−1 u(t), then by (2.7), nr cn gqn −1 (xnr ) = λ. r→∞ L n n=0

lim ϕ(xnr ) = lim

r→∞

(2.8)

On the other hand, by (2.7) and the representation (1.14) of the Carlitz polynomials, we ﬁnd that n n (qj −1)n Ln x (−1)n−j gqn −1 (xn ) = j Dn j=0 = (−1)n +

n j=1

where

L n Dj Lqj

n−j

(−1)n−j

Ln j Dj Lqn−j

= q −mn ,

x(q

j

−1)n

50

Chapter 2

mn = n + (q j − 1)n −

qj − 1 − q j (n − j) = jq j − (1 + q + · · · + q j−1 ) > 0, q−1

so that gqn −1 (xn ) ≡ (−1)n (mod x). Therefore n r cn gqn −1 (xnr ) lim r→∞ L n=0 n n −1 r cnr cn nr nr = lim gqn −1 (x ) + ((−1) + Br ) r→∞ L L n n r n=0 cn cn where |gqn −1 (xnr )| ≤ 1, = |cn |q n , |Br | ≤ q −1 , r = |cnr | q nr > Lnr Ln |cn |q n (due to (2.6)), so that n r cn nr gqn −1 (x ) = lim |cnr | q nr = ∞, lim r→∞ r→∞ Ln n=0

which contradicts (2.8). Thus, the sequence {|cn |q n } is bounded. Suppose that it does not tend to 0. Since we may multiply the coeﬃcients cn by a ﬁxed power of x, and we may also change ﬁnitely many coeﬃcients in an arbitrary way, we may assume that |cn |q n ≤ 1,

lim sup |cn |q n = 1. n→∞

We have r cn gqn −1 (xr ) = λ, r→∞ L n n=0

lim

so that, given ε > 0, there exists r0 ∈ N, such that r cn r gqn −1 (x ) − λ < ε, r ≥ r0 . L n n=0 The above argument shows that gqn −1 (xr ) ≡ (−1)n

(mod x),

if 0 ≤ n ≤ r. Therefore, assuming that ε < 1, we get r 0 +n n=0

(−1)n

cn = λ + σ(n) Ln

(2.9)

Calculus

51

where |σ(n)| < 1 for all n ≥ 0. Hence for any n ≥ 1, (−1)r0 +n cr0 +n = |σ(n) − σ(n − 1)| < 1, Lr0 +n

in contradiction to (2.9).

The next result is an extension of Proposition 1.17 regarding the ortho∞ normality of the sequence {gqn −1 }n=0 . Lemma 2.3 If a function v : O \ {0} → K c is continuous and bounded, and v admits a pointwise convergent expansion v(t) =

∞

vn gqn −1 (t),

t ∈ O \ {0},

(2.10)

n=0

vn ∈ K c , then sup |v(t)| = sup |vn |. 0≤n≤∞

t∈O\{0}

Proof. It is clear that |v(t)| ≤ sup |vn |, n

t = 0.

In order to prove the inverse inequality, we use the identity (1.46) (with k = 0, l = q ν − 1), which shows that 0, if n < m, gqn −1 (t) = m (−1) , if n = m. deg t=m t monic

If n > m, then gqn −1 (t) = 0 for deg t = m, due to the identity (2.7) and the corresponding property of the Carlitz polynomials. Now the summation in both sides of (2.10) yields v(t) = (−1)m vm , deg t=m t monic

so that |vm | ≤

sup |v(t)|. t∈O\{0}

52

Chapter 2

Lemma 2.4 Let a function w ∈ C(O, K c ) be such that the function γ(t) = tw(t) is Fq -linear. Then w(t) =

∞

wn ∈ K c , wn → 0,

wn gqn −1 (t),

(2.11)

n=0

and the series (2.11) is uniformly convergent on O. Proof. Consider the expansion γ(t) = γn fn (t),

t ∈ O,

(2.12)

where γn ∈ K c , γn → 0. The fact that the function t−1 γ(t) is continuous at t = 0 means that γ(t) is diﬀerentiable at t = 0. By Lemma 2.2, |γn |q n → 0 for n → ∞. Dividing both sides of (2.12) by t we obtain the expansion (2.11) with wn = L−1 n γn , so that wn → 0. Now we are in a position to prove the characterization result. ∞

Theorem 2.5 The function u =

cn fn ∈ C0 (O, K c ) belongs to

n=0

C0k+1 (O, K c ) if and only if k

q nq |cn | → 0

for n → ∞.

(2.13)

In this case k

sup |Dk u(t)| = sup q (n−k)q |cn |. t∈O

Proof. We use the identity

k

d fn = Since dk =

√

qk

(2.14)

n≥k

fn−k , if n ≥ k, 0, if n < k.

◦ ∆(k) , we ﬁnd that Dk u(t) = t−q

k

∞

k

q cn fn−k (t),

t = 0,

(2.15)

n=k

which implies the identity

q−k

D u(t) k

=

∞ n=k

−k

cqn

gqn−k −1 (t) , Ln−k

t = 0.

(2.16)

Calculus

53

Now, if (2.13) is satisﬁed, then we ﬁnd from Proposition 1.17 that the right-hand side of (2.16) is a continuous function on O, which means that u ∈ C0k+1 (O, K c ). The equality (2.14) follows from Lemma 2.3. Conversely, suppose that Dk u is continuous on O. Let w(t) q−k k . By (2.15), the function γ(t) = tw(t) has the form = D u(t) γ(t) =

∞

−k

cqn fn−k (t),

t = 0.

(2.17)

n=k

Since w is continuous on O, we see that γ(0) = 0. On the other hand, fn (0) = 0 for all n, so that the equality (2.17) holds for all t ∈ O, and the function γ is Fq -linear (however, we cannot claim the uniform convergence of the series in (2.17)). Now Lemma 2.4 implies the representation (2.11) with wn → 0. It follows from (2.11) and (2.17) that ∞

−k wn − cqn+k L−1 gqn −1 (t) = 0 n

n=0

for all t = 0. By Lemma 2.3, this means that −k

wn = cqn+k L−1 n for n ≥ 0, which implies (2.13).

2.1.3. Indeﬁnite sum. Viewing the operator d as a kind of derivative, it is natural to introduce an appropriate antiderivative. Following the terminology used in the analysis over Zp (see [98]) we call it the indeﬁnite sum. Consider in C0 (O, K c ) the equation du = f, Suppose that f =

∞

f ∈ C0 (O, K c ).

ϕk fk , ϕk ∈ K c . Looking for u =

k=0

(2.18) ∞

ck fk and using

k=0

the fact that du =

∞ k=1

ϕql ,

1/q

ck fk−1 =

∞

1/q

cl+1 fl

l=0

we ﬁnd that cl+1 = l = 0, 1, 2, . . . Therefore u is determined by (2.18) uniquely up to the term c0 f0 (t) = c0 t, c0 = u(1). Fixing u(1) = 0 we obtain an Fq -linear bounded operator S on C0 (O, K c ), the operator of indeﬁnite sum: Sf = u. It follows from Theorem 2.5 that

54

Chapter 2

S is also a bounded operator on each space C0k (O, K c ), k = 1, 2, . . . Note that Sfk = fk+1 , k = 0, 1, . . ., so that S is not compact. Another possible procedure to ﬁnd Sf is an interpolation. If u = Sf then u(xt) − xu(t) = f q (t),

u(1) = 0.

Setting successively t = 1, x, x2 , . . ., we get u(x) = f q (1), u(x2 ) = f q (x) + xf q (1), u(x3 ) = f q (x2 ) + xf q (x) + x2 f q (1), .................................................... u(xn ) = f q (xn−1 ) + xf q (xn−2 ) + . . . + xn−1 f q (1). Since u is assumed Fq -linear, this determines u(t) for all t ∈ Fq [x]. Extending by continuity, we ﬁnd u(t) for any t ∈ O: if t=

∞

ζn xn ,

ζn ∈ Fq ,

n=0

then u(t) =

∞

n

ζn u(x ) =

n=0

∞ n=1

ζn

n

j−1 q

x

u(t) =

f (x

j=1

so that ∞ n=0

x f

)=

∞ ∞

ζn xj−1 f q (xn−j ),

j=1 n=j

n q

n−j

∞

m

ζm+n+1 x

.

m=0

2.1.4. The Volkenborn-type integral. The Volkenborn integral of a function on Zp was introduced in [116] (see also [98]), in order to obtain relations between some objects of p-adic analysis resembling classical integration formulas of real analysis. Here we extend this approach to the function ﬁeld situation. Our deﬁnition is based essentially on the Carlitz diﬀerence operator ∆, which again shows its close connection with basic structures of analysis over the ﬁeld K. The integral of a function f ∈ C01 (O, K c ) is deﬁned as ' Sf (xn ) def f (t) dt = lim = (Sf ) (0). n→∞ xn O

Calculus

55

It is clear that the integral is an Fq -linear continuous functional on C01 (O, K c ), ' ' cf (t) dt = cq f (t) dt, c ∈ K c . O

O ∞

Since a power series

n=0

n

an tq with an → 0 converges in C01 (O, K c ), it can

be integrated termwise: ' ∞

an t

qn

dt =

∞

' aqn

n=0

O n=0

n

tq dt. O

The integral possesses the following “invariance” property (related, in contrast to the case of Zp , to the multiplicative structure): ' ' f (xt) dt = x f (t) dt − f q (1). (2.19) O

O

Indeed, let g(t) = f (xt). Then Sg(xn ) = g q (xn−1 ) + xg q (xn−2 ) + · · · + xn−1 g q (1) = f q (xn ) + xf q (xn−1 ) + · · · + xn−1 f q (x) = (Sf )(xn+1 ) − xn f q (1) whence Sg(xn ) (Sf )(xn+1 ) = x · − f q (1), xn xn+1 and (2.19) is obtained by passing to the limit for n → ∞. Using (2.18) we obtain by induction that ' ' f (xn t) dt = xn f (t) dt − xn−1 f q (1) − xn−2 f q (x) − · · · − f q (xn−1 ). O

O

This equality implies the following invariance property. Suppose that a function f vanishes on all elements z ∈ Fq [x] with deg z < n. Then, if g ∈ Fq [x], deg g ≤ n, we have ' ' f (gt) dt = g f (t) dt. O

O

56

Chapter 2

Our next result will contain the calculation of integrals for some important functions on O. For the deﬁnitions of the special functions used below see Chapter 1. Theorem 2.6 (i) For any n = 0, 1, 2, . . . ' n 1 . tq dt = − [n + 1]

(2.20)

O

(ii) For any n = 0, 1, 2, . . . ' fn (t) dt =

(−1)n+1 . Ln+1

(2.21)

O

(iii) If z ∈ K, |z| < 1, then ' Cs (z) ds = logC (z) − z.

(2.22)

O

Proof. (i) We have seen that d

n

tq Dn

n−1

=

tq , Dn−1

n ≥ 1.

It follows from the deﬁnition of the operator S that qn n+1 tq −t t = S Dn Dn+1 whence n

tqn+1 − t Dnq qn+1 S tq = , t −t = [n + 1] Dn+1 n k n+1 x S tq 1 xk(q −1) − 1 −→ − , = xk [n + 1] [n + 1] (ii) We have

' O

fn (t) dt = (Sfn ) (0) = fn+1 (0).

k → ∞.

Calculus

57

According to (1.14), the linear term in the expression for fi is i (−1)i 0 t= t. (−1)i Di Li

(2.23)

Diﬀerentiation yields (2.21). (iii) Applying the operator d (with respect to the variable s) to the function Cs (z) we ﬁnd that ds Cs (z) =

∞

fi−1 (s)z q

i−1

= Cs (z)

i=1

whence Ss Cs (z) = Cs (z) − C1 (z)s =

∞

i

fi (s)z q − zs.

i=0

Fixing z and denoting ϕ(s) = Ss Cs (z) we obtain that ∞

ϕ(xn ) fi (xn ) qi = z − z. xn xn i=0

(2.24)

It is seen from (1.14) and (2.23) that (−1)i fi (xn ) −→ for n → ∞. xn Li On the other hand, it follows from (2.6) and Proposition 1.17 that fi (xn ) i xn ≤ q for all n. If z ∈ O, |z| < 1, then |z| ≤ q −1 , and the series in (2.24) converges uniformly with respect to n. Passing to the limit n → ∞ in (2.24) we come to (2.22). Setting in (2.22) z = eC (t), |t| < 1, we get an identity for the Carlitz exponential: ' eC (st) ds = t − eC (t). O

On the other hand, the formula (2.22) implies a more general formula (conjectured by D. Goss).

58

Chapter 2

Corollary 2.7 If a ∈ O, z ∈ K, |z| < 1, then ' Csa (z) ds = a logC (z) − Ca (z).

(2.25)

O

Proof. Since it is easily shown that (for each ﬁxed z) the mapping O → C01 (O, K c ) of the form a → Csa (z) is continuous, it is suﬃcient to prove (2.25) for a = xn , n = 1, 2, . . . Using (2.22), we ﬁnd that ' n xn−k Cxqk−1 (z). (2.26) Csxn (z) ds = xn (logC (z) − z) − k=1

O

Let t = logC (z). Then z = eC (t) (see Section 1.6.1). It follows from properties of eC that n

xn−k Cxqk−1 (z) =

k=1

n

xn−k (eC (xk t) − xeC (xk−1 t))

k=1

= eC (xn t) − xn eC (t) = Cxn (z) − xn z. Substituting this into (2.26) we come to (2.25). As Csa (z) = Cs (Ca (z)), equation (2.22) also implies that ' Csa (z) ds = logC (Ca (z)) − Ca (z) . O

Comparing this with (2.25) implies a logC (z) = logC (Ca (z)) which is precisely the functional equation of logC (z). 2.1.5. Fractional derivatives. In this section we introduce the operator ∆(α) , α ∈ O, a function ﬁeld analog of the Hadamard fractional d α derivative t dt from real analysis (see [97]). ∞ ∞ Let α ∈ O, α = αn xn , αn ∈ Fq . Denote α (= (−1)n αn xn . The n=0

n=0

transformation α → α ( is an Fq -linear isometry. For an arbitrary continuous Fq -linear function u on O we deﬁne its “fractional derivative” ∆(α) u at a point t ∈ O by the formula

∞

(−1)k Dk (( α)u(xk t) ∆(α) u (t) = k=0

(2.27)

Calculus

59

where {Dk (t)} is the sequence of hyperdiﬀerentiations (see Section 1.3). The series converges for each t, uniformly with respect to α, since |Dk (( α)| ≤ 1 and u(xk t) → 0. Thus ∆(α) u is, for each t, a continuous Fq -linear function in α. Our understanding of ∆(α) as a kind of a fractional derivative is justiﬁed by the following result. Proposition 2.8 ∆(x

n

)

= ∆n , n = 1, 2, . . .

Proof. By the deﬁnition of Dk , it follows from (2.27) that ∆

(xn )

u (t) =

n n k=0

k

(−x)n−k u(xk t).

If n = 1, then ∆(x) u (t) = u(xt) − xu(t) = (∆u)(t). Suppose we have n−1 proved that ∆(x ) = ∆n−1 . Then

n−1 (∆n u) (t) = ∆ ∆(x ) u (t) n−1 n−1 n − 1 n − 1 n−1−k k+1 (−x) (−x)n−1−k u(xk t) = u(x t) − x k k k=0 k=0 n−1 n n−1 n−1 k n−k (−x) (−x)n−k u(xk t) u(x t) + = k−1 k k=1 k=0 n−1 n − 1 n − 1 (−x)n−k u(xk t) + (−x)n u(t) + = u(xn t) + k k−1 k=1

n = ∆(x ) u (t), as desired.

n

m

It follows from Proposition 2.8 that ∆(x ) ◦ ∆(x ) = ∆(x n m = ∆(x ·x ) , which prompts the following composition property. Proposition 2.9 For any α, β ∈ O,

∆(α) ∆(β) u (t) = ∆(αβ) u (t).

n+m

)

60

Chapter 2

Proof. Using the Leibnitz rule (1.29) for hyperdiﬀerentiations we have

∞ ∞

( ∆(α) ◦ ∆(β) u (t) = (−1)k Dk (β) (−1)l Dl (( α)u(xk+l t)

=

=

k=0 ∞ n=0 ∞

l=0

(−1)n u(xn t)

( l (( Dk (β)D α)

k+l=n n ( αβ)u(x t) (−1)n Dn ((

n=0

= ∆(αβ) u (t).

2.2 Umbral calculus 2.2.1. Background. Classical umbral calculus [95, 92, 96, 91] is a set of algebraic tools for obtaining, in a uniﬁed way, a rich variety of results regarding the structure and properties of various polynomial sequences. There exists a lot of generalizations extending umbral methods to other classes of functions. However, there is a restriction common to the whole literature on umbral calculus – the underlying ﬁeld must be of characteristic zero. An attempt to mimic the characteristic zero procedures in the positive characteristic case [37] revealed a number of pathological properties of the resulting structures. More importantly, these structures were not connected with the existing analysis in positive characteristic based on a completely diﬀerent algebraic foundation. Classically, the main notions of umbral calculus are the delta operator, which (like the derivative) decreases degrees of polynomials, and the sequence of binomial type, a polynomial sequence {pn }∞ n=0 , deg pn = n, such that n n pk (s)pn−k (t), n = 0, 1, 2, . . . pn (s + t) = (2.28) k k=0

Many classical polynomial systems satisfy this identity. In order to guess the counterpart of (2.28) in Fq -linear analysis over the ﬁeld K, we look at the system of additive Carlitz polynomials {en }. Let us consider the Carlitz module function (1.68): Cs (z) =

∞ en (s) qn z , Dn n=0

|z| < 1.

(2.29)

Calculus

61

If s ∈ Fq [x], then Cs is a polynomial in z, and (see (1.70)) s, t ∈ Fq [x].

Cts (z) = Ct (Cs (z)),

Let us write the last identity explicitly using (2.29). After rearranging the sums we ﬁnd that Cts (z) =

deg t+deg s

zq

i

i=0

so that ei (st) =

m+n=i m,n≥0 i i n=0

where

n

1 qn q n en (t){em (s)} , Dn Dm

en (t){ei−n (s)}q

n

(2.30)

K

Di i = qn . n K Dn Di−n

(2.31)

In this section we show that the “K-binomial” relation (2.30), a positive characteristic counterpart of the classical binomial formula, can be used for developing umbral calculus in the spirit of [95]. In particular, we introduce and study corresponding (nonlinear) delta operators, obtain a representation for operators invariant with respect to multiplicative shifts, and construct generating functions for polynomial sequences of the K-binomial type. Such sequences are also used for constructing new orthonormal bases of the space C0 (O, K c ) (in particular, a sequence of the Laguerre-type polynomials), in a way similar to the p-adic (characteristic 0) case [113, 114, 90]. 2.2.2. Delta operators and K-binomial sequences. Denote by K c {t} the vector space over K c consisting of Fq -linear polynomials u = k ak tq with coeﬃcients from K c . We will often use the operator of multiplicative shift (ρλ u)(t) = u(λt) on K c {t} and the Frobenius operator τ u = uq . We call a linear operator T on K c {t} invariant if it commutes with ρλ for each λ ∈ K. n

n

Lemma 2.10 If T is an invariant operator, then T (tq ) = cn tq , cn ∈ K c , for each n ≥ 0. Proof. Suppose that qn

T (t ) =

N l=1

cjl tq

jl

62

Chapter 2

where jl are diﬀerent nonnegative integers, cjl ∈ K c . For any λ ∈ K N

jl n n n n n n ρλ T (tq ) = T ρλ (tq ) = T (λt)q = λq T (tq ) = λq cjl tq . l=1

On the other hand, qn

ρλ T (t ) =

N

jl

jl

cjl λq tq .

l=1

Since λ is arbitrary, this implies the required result. If an invariant operator T is such that T (t) = 0, then by Lemma 2.10 the operator τ −1 T on K c {t} is well-deﬁned. Deﬁnition 2.11 An Fq -linear operator δ = τ −1 δ0 , where δ0 is a linear invariant operator on K c {t}, is called a delta operator if δ0 (t) = 0 and n n δ0 (f ) = 0 for deg f > 1, that is δ0 (tq ) = cn tq , cn = 0, for all n ≥ 1. The most important example of a delta operator is the Carlitz derivative d = τ −1 ∆, (∆u)(t) = u(xt) − xu(t). Deﬁnition 2.12 A sequence {Pn }∞ 0 of Fq -linear polynomials is called a basic sequence corresponding to a delta operator δ = τ −1 δ0 , if deg Pn = q n , P0 (1) = 1, Pn (1) = 0 for n ≥ 1, δP0 = 0,

δPn = [n]1/q Pn−1 , n ≥ 1,

(2.32)

q δ0 Pn = [n]Pn−1 , n ≥ 1.

(2.33)

or, equivalently, δ0 P0 = 0,

It follows from well-known identities for the Carlitz polynomials ei (see Proposition 1.7 and Theorem 1.24) that the sequence {ei } is basic with respect to the operator d. For the normalized Carlitz polynomials fi we have the relations df0 = 0,

dfi = fi−1 , i ≥ 1.

The next deﬁnition is a formalization of the property (2.30).

Calculus

63

Deﬁnition 2.13 A sequence of Fq -linear polynomials ui ∈ K c {t} is called a sequence of K-binomial type if deg ui = q i and for all i = 0, 1, 2, . . . ui (st) =

i i n=0

n

qn

un (t) {ui−n (s)}

s, t ∈ K.

,

(2.34)

K

If {ui } is a sequence of K-binomial type, then ui (1) = 0 for i ≥ 1, u0 (1) = 1 (so that u0 (t) = t). Indeed, for i = 0 the formula (2.34) gives u0 (st) = u0 (s)u0 (t). Setting s = 1 we have u0 (t) = u0 (1)u0 (t), and since deg u0 = 1, so that u0 (t) ≡ 0, we get u0 (1) = 1. If i > 0, for all t 0 = ui (t) − ui (t) =

i−1 i n=0

n

qn

{ui−n (1)}

un (t),

K

and the linear independence of the polynomials un means that ul (1) = 0 for l ≥ 1. Theorem 2.14 For any delta operator δ = τ −1 δ0 , there exists a unique basic sequence {Pn }, which is a sequence of K-binomial type. Conversely, given a sequence {Pn } of K-binomial type, deﬁne the action of δ0 on Pn by the relations (2.33), extend it onto K c {t} by linearity and set δ = τ −1 δ0 . Then δ is a delta operator, and {Pn } is the corresponding basic sequence. Proof. Let us construct a basic sequence corresponding to δ. Set P0 (t) = t and suppose that Pn−1 has been constructed. It follows from Lemma 2.10 that δ is surjective, and we can choose Pn satisfying (2.32). For any c ∈ K c , Pn + ct also satisﬁes (2.32), and we may redeﬁne Pn choosing c in such a way that Pn (1) = 0. Hence, a basic sequence {Pn } indeed exists. If there is another basic sequence {Pn } with the same delta operator, then δ(Pn − Pn ) = 0, whence Pn (t) = Pn (t) + at, a ∈ K c , and setting t = 1 we ﬁnd that a = 0. In order to prove the K-binomial property, we introduce some operators having an independent interest. (l) Consider the linear operators δ0 = τ l δ l .

64

Chapter 2

Lemma 2.15 (i) The identity (l)

δ0 Pj =

Dj ql Dj−l

l

q Pj−l

(2.35)

holds for any l ≤ j. (ii) Let f be an Fq -linear polynomial, deg f ≤ q n . Then a generalized Taylor formula

(l) n δ0 f (s) Pl (t) (2.36) f (st) = Dl l=0

holds for any s, t ∈ K. Proof. By (2.32), −1

δ l Pj = δ l−1 [j]q Pj−1 −l

= [j]q δ l−1 Pj−1 −l

−(l−1)

−l

−(l−1)

= [j]q [j − 1]q = [j]q [j − 1]q

δ l−2 Pj−2 = . . . . . . [j − (l − 1)]q

−1

Pj−l ,

so that (l)

δ0 Pj = [j][j − 1]q . . . [j − (l − 1)]q

l−1

l

q Pj−l

which is equivalent to (2.35). Since deg Pj = q j , the polynomials P1 , . . . , Pn form a basis of the vector space of all Fq -linear polynomials of degrees ≤ n (because its dimension equals n). Therefore f (st) =

n

bj (s)Pj (t)

(2.37)

j=0

where bj (s) are, for each ﬁxed s, some elements of K c . (l) Applying the operator δ0 , 0 ≤ l ≤ n, in the variable t to both sides of (2.37) and using (2.35) we ﬁnd that

n

Dj ql (l) δ0 f (st) = bj (s) ql Pj−l (t) Dj−l j=l

Calculus (l)

(note also that δ0 account that

65

commutes with ρs ). Setting t = 1 and taking into Pj−l (1) =

0,

if j > l;

1,

if j = l,

we come to the equality

(l) δ0 f (s)

bl (s) =

Dl

,

0 ≤ l ≤ n,

which implies (2.36).

Note that the formulas (2.35) and (2.36) for the Carlitz polynomials ei were established long ago; see [43]. It is important that, in contrast to the classical umbral calculus, the linear operators involved in (2.36) are not powers of a single linear operator. Proof of Theorem 2.14 (continued). In order to prove that {Pn } is a sequence of K-binomial type, it suﬃces to take f = Pn in (2.36) and to use the identity (2.35). To prove the second part of the theorem, we calculate the action in the variable t of the operator δ0 , deﬁned by (2.33), upon the function Pn (st). Using the relation Dn+1 = [n + 1]Dnq we ﬁnd that n n qj Pn−j (s) (δ0 Pj ) (t) δ0,t Pn (st) = j K j=0 =

=

=

n

Dn qj j=1 Dn−j Dj

n−1

Dn [i + 1]

q i+1 Pn−i−1 (s)Piq (t) q i+1 i=0 Dn−i−1 Di+1 n−1 Dn−1 q qi+1 Pn−i−1 (s)Piq (t) [n] q D D i n−i−1 i=0 n−1 q

= [n] =

j

q q Pn−j (s)[j]Pj−1 (t)

n−1 i i=0

q [n]Pn−1 (st)

i

q Pn−i−1 (s)Pi (t) K

= (δ0 Pn ) (st),

that is, δ0 commutes with multiplicative shifts.

66

Chapter 2

It remains to prove that δ0 (f ) = 0 if deg f > 1. Assuming that δ0 (f ) = 0 n for f = aj Pj we have j=0

0=

n

q aj [j]Pj−1 =

j=0

n−1

q 1/q ai+1 [i

+ 1]1/q Pi

i=0

whence a1 = a2 = . . . = an = 0 due to the linear independence of the sequence {Pi }. 2.2.3. Invariant operators. Let T be a linear invariant operator on K c {t}. Let us ﬁnd its representation via an arbitrary ﬁxed delta operator δ = τ −1 δ0 . By (2.36), for any f ∈ K c {t}, deg f = q n ,

(l) n δ0 f (s) (T Pl ) (t) . (T f )(st) = (ρs T f ) (t) = T (ρs f ) (t) = Tt f (st) = Dl l=0

Setting s = 1 we ﬁnd that T =

∞

(l)

σl δ0

(2.38)

l=0

(T Pl )(1) . The inﬁnite series in (2.38) actually becomes a Dl ﬁnite sum if both sides of (2.38) are applied to any Fq -linear polynomial f ∈ K c {t}. Conversely, any such series deﬁnes a linear invariant operator on K c {t}. Below we will consider in detail the case where δ is the Carlitz derivative (l) d, so that δ0 = ∆, and the operators δ0 = ∆(l) are given recursively:

l−1 ∆(l) u (t) = ∆(l−1) u (xt) − xq u(t); (2.39) where σl =

see (1.18) (the formula (2.38) for this case was proved by a diﬀerent method in [54]). Using (2.39) with l = 0, we can compute for this case the coeﬃcients cn n from Lemma 2.10. We have ∆(l) (tq ) = 0, if n < l, n

n

∆(tq ) = [n]tq ,

n ≥ 1;

∆(2) (tq ) = τ 2 d2 (tq ) = τ ∆τ −1 ∆(tq )

n−1 n = τ ∆ [n]1/q tq = [n][n − 1]q tq , n

n

n

n ≥ 2,

Calculus

67

and by induction n

∆(l) (tq )[n][n − 1]q . . . [n − l + 1]q

l−1

Dn

n

tq =

ql Dn−l

n

tq ,

n ≥ l.

(2.40)

The explicit formula (2.40) makes it possible to ﬁnd out when an operator θ = τ −1 θ0 , with θ0 =

∞

σl ∆(l) ,

(2.41)

l=1

is a delta operator. We have θ0 (t) = 0, n

n

θ0 (tq ) = Dn Sn tq , where Sn =

n

σl

l=1

q Dn−l

l

. Thus θ is a delta operator if and only if Sn = 0 for

all n = 1, 2, . . . Example 1. Let σl = 1 for all l ≥ 1, that is θ0 =

∞

∆(l) .

(2.42)

l=1 q i −1

Since |Di | = q − q−1 , we have l q n −q l q Dn−l = q − q−1 , q n −q

so that |Sn | = q q−1 (= 0) by the ultra-metric property of the absolute value. Comparing (2.42) with a classical formula from [95] we may see the polynomials Pn for this case as analogs of the Laguerre polynomials. (−1)l+1 . Now Example 2. Let σl = Ll Sn =

n

(−1)l+1

l=1

1 l

q Ll Dn−l

.

Let us use the identity h−1

(−1)j

qj j=0 Lj Dh−j

=

(−1)h+1 Lh

(2.43)

68

Chapter 2

proved in [40]. It follows from (2.43) that h (−1)j qj j=1 Lj Dh−j

=

(−1)h+1 1 (−1)h 1 − + =− , Lh Dh Lh Dh

so that Sn = Dn−1 (= 0), n = 1, 2, . . . In this case θ0 (tq ) = tq for all j ≥ 1 n n−1 for n ≥ 1. (of course, θ0 (t) = 0), and P0 (t) = t, Pn (t) = Dn tq − tq j

j

2.2.4. Orthonormal bases. Let {Pn } be the basic sequence corresponding to a delta operator δ = τ −1 δ0 , δ0 =

∞

σl ∆(l)

l=1

(the operator series converges on any polynomial from K c {t}). Pn , n = 0, 1, 2, . . . Then for any n ≥ 1 Let Qn = Dn δQn = Dn−1/q δPn =

[n]1/q 1/q Dn

Pn−1 =

Pn−1 = Qn−1 , Dn−1

and the K-binomial property of {Pn } implies the identity Qi (st) =

i

qn

Qn (t) {Qi−n (s)}

,

s, t ∈ K.

(2.44)

n=0

The identity (2.44) may be seen as another form of the K-binomial property. Though it resembles its classical counterpart, the presence of the Frobenius powers is a feature speciﬁc for the case of a positive characteristic. We will call {Qn } a normalized basic sequence. Theorem 2.16 If |σ1 | = 1, |σl | ≤ 1 for l ≥ 2, then the sequence {Qn }∞ 0 is an orthonormal basis of the space C0 (O, K c ) – for any f ∈ C0 (O, K c ) there is a uniformly convergent expansion f (t) =

∞

ψn Qn (t),

t ∈ O,

(2.45)

n=0

(n) where ψn = δ0 f (1), |ψn | → 0 as n → ∞, f = sup |ψn |. n≥0

(2.46)

Calculus

69

Proof. We have Q0 (t) = P0 (t) = t, so that Q0 = 1. Let us prove that Qn = 1 for all n ≥ 1. Our reasoning will be based on expansions in the normalized Carlitz polynomials fn . Let n = 1. Since deg Qn = q n , we have Q1 = a0 f0 + a1 f1 . We know that Q1 (1) = f1 (1) = 0, hence a0 = 0, so that Q1 = a1 f1 . Next, δQ1 = Q0 = f0 . Writing this explicitly we ﬁnd that f0 = a1 τ −1 1/q

∞

1/q 1/q

1/q 1/q

σl ∆(l) f1 = a1 σ1 df1 = a1 σ1 f0 ,

l=1

whence a1 = σ1−1 , Q1 = σ1−1 f1 , and Q1 = 1. Assume that Qn−1 = 1 and consider the expansion Qn =

n

aj fj

j=1

(the term containing f0 is absent since Qn (1) = 0). Applying δ we get δQn =

n

1/q

aj

j=1

By (1.22), ∆ ej =

1/q −1

σl

τ

∆(l) fj .

l=1

 

(l)

∞

Dj ql Dj−l

l

eqj−l , if l ≤ j,

0,

so that

(l)

∆ fj =

if l > j, l

q , if l ≤ j, fj−l

0,

if l > j.

Therefore δQn =

n

1/q aj

j=1

j

1/q q l−1 fj−l .

σl

(2.47)

l=1

q It follows from the identity fi−1 = fi−1 + [i]fi (a consequence of (1.21)) that l−1 q l−1 fj−l = ϕj,l,k fj−l+k k=0

where ϕj,l,0 = 1, |ϕj,l,k | < 1 for k ≥ 1. Substituting into (2.47) we ﬁnd

70

Chapter 2

that Qn−1 =

n

1/q aj

j=1

=

n

1/q

aj

whence

j−1

n

fi

j=i+1

l−1

ϕj,l,k fj−l+k

k=0

fi

i=0

n−1 i=0

1/q

σl

l=1

j=1

=

j

j

1/q

σl

ϕj,l,i−j+l

l=j−i 1/q

aj

j

1/q

σl

ϕj,l,i−j+l

l=j−i

n j 1/q 1/q max aj σl ϕj,l,i−j+l = 1 0≤i≤n−1 j=i+1 l=j−i

(2.48)

by the inductive assumption and the orthonormal basis property of the normalized Carlitz polynomials. For i = n − 1, we obtain from (2.48) that n 1/q 1/q σl ϕn,l,l−1 ≤ 1. an l=1

We have ϕn,1,0 = 1, |σ1 | = 1, and n 1/q σl ϕn,l,l−1 < 1, l=2

so that

n 1/q σl ϕn,l,l−1 = 1 l=1

whence |an | ≤ 1. Next, for i = n − 2 we ﬁnd from (2.48) that n−1 n 1/q 1/q 1/q 1/q σl ϕn−1,l,l−1 + an σl ϕn,l,l−1 ≤ 1. an−1 l=1

l=2

We have proved that the second summand on the left is in O; then the ﬁrst summand is considered as above, so that |an−1 | ≤ 1. Repeating this reasoning we come to the conclusion that |aj | ≤ 1 for all j. Moreover, |aj | = 1 for at least one value of j; otherwise we would come to a contradiction with (2.48). This means that Qn = 1.

Calculus

71

If f is an arbitrary Fq -linear polynomial, deg f = q N , then by the generalized Taylor formula (2.36) f (t) =

N

ψl Ql (t),

t ∈ O,

l=0

(l) where ψl = δ0 f (1).

Since Ql = 1 for all l, we have f ≤ sup |ψl |. On the other hand, l l (l) δ0 f = τ l τ −1 δ0 f , and if we prove that δ0 f ≤ f , this will imply the (l) inequality δ0 f ≤ f . We have

l−1 (l) ∆(l−1) f (t) ∆ f = max ∆(l−1) f (xt) − xq t∈O

≤ max ∆(l−1) f (t) ≤ . . . ≤ max |(∆f ) (t)| ≤ f , t∈O

so that

t∈O

∞ (l) σl ∆ f ≤ sup |σl | · ∆(l) f ≤ f δ0 f = l l=0

(l)

whence δ0 f ≤ f and sup |ψl | ≤ f . l

Thus, we have proved (2.46) for any polynomial. By a well-known result of non-Archimedean functional analysis (see Theorem 50.7 in [98]), the uniformly convergent expansion (2.45) and the equality (2.46) hold for any f ∈ C0 (O, K c ).

(n)

The relation ψn = δ0 f (1) also remains valid for any f ∈ C0 (O, K c ).

Indeed, denote

by ϕn (f ) a continuous linear functional on C0 (O, K c ) of the (n) form δ0 f (1). Suppose that {FN } is a sequence of Fq -linear polynomials uniformly convergent to f . Then FN = ϕn (FN )Qn , n

so that F − FN =

∞

{ψn − ϕn (FN )} Qn ,

n=0

and by (2.46), F − FN = sup |ψn − ϕn (FN )| . n

72

Chapter 2

For each ﬁxed n we ﬁnd that |ψn − ϕn (FN )| ≤ F − FN , and passing to the limit as N → ∞ we get that ψn = ϕn (f ), as desired. By Theorem 2.16, the Laguerre-type polynomial sequence from Example 1 is an orthonormal basis of C0 (O, K c ). The sequence from Example 2 does not satisfy the conditions of Theorem 2.16. Note that the conditions |σ1 | = 1, |σl | ≤ 1, l = 2, 3, . . ., imply that Sn = 0 for all n, so that the series (2.41) considered in Theorem 2.16 always correspond to delta operators. Let us write a recurrence formula for the coeﬃcients of the polynomials Qn . Here we assume only that Sn = 0 for all n. Let Qn (t) =

n

(n)

j

γj tq .

(2.49)

j=0 (0)

We know that γ0 = 1. n n Using the relation δ0 tq = Dn Sn tq we ﬁnd that for n ≥ 1 Qn−1 = δQn = τ −1

n

j

(n)

γj Dj Sj tq =

n−1

(n)

γi+1

1/q

1/q

1/q

i

Di+1 Si+1 tq .

i=0

j=1

Comparing this with the equality (2.49), with n − 1 substituted for n, we get

1/q (n−1) (n) 1/q 1/q = γi+1 Di+1 Si+1 γi whence

(n)

γi+1 =

(n−1)

q

γi

Di+1 Si+1

,

i = 0, 1, . . . , n − 1; n = 1, 2, . . .

(2.50) (n)

The recurrence formula (2.50) determines all the coeﬃcients γi (if the (n) polynomial Qn−1 is already known) except γ0 . The latter can be found from the condition Qn (1) = 0: (n)

γ0

=−

n

(n)

γj .

j=1

2.2.5. Generating functions. The deﬁnition (1.68) of the Carlitz module can be seen as a generating function for the normalized Carlitz polynomials fi . Here we give a similar construction for the normalized

Calculus

73

basic sequence in the general case. As in Section 2.2.4, we consider a delta operator of the form δ = τ −1 δ0 , δ0 =

∞

σl ∆(l) .

l=1

We assume that Sn = 0 for all n. Let us deﬁne the generalized exponential eδ (t) =

∞

b j tq

j

(2.51)

j=0

by the conditions δeδ = eδ , b0 = 1. Substituting (2.51) we come to the recurrence relation bqj bj+1 = (2.52) Dj+1 Sj+1 which determines eδ as a formal power series. Since b0 = 1, the composition inverse logδ to the formal power series eδ has a similar form: logδ (t) =

∞

n

βn tq ,

βn ∈ K,

(2.53)

n=0

(see Section 19.7 in [83] for a general treatment of formal power series of this kind). A formal substitution gives the relations m β0 = 1, bm βnq = 0, l = 1, 2, . . . , m+n=l

whence l

βl = −

m

q bm βl−m ,

l = 1, 2, . . .

(2.54)

m=1

Theorem 2.17 Suppose that |σ1 | = 1 and |σl | ≤ 1 for #the " all l. Then both −1 , if series (2.51) and (2.53) converge on the disk D = t ∈ O : |t| ≤ q q # " q = 2, or D2 = t ∈ O : |t| ≤ q −2 , if q = 2, and eδ (t logδ z) =

∞ n=0

n

Qn (t)z q ,

t ∈ O, z ∈ Dq .

(2.55)

74

Chapter 2

Proof. Since

D n ql D

n−l

q l −1 = q − q−1 ,

under our assumptions we have |Dn Sn | = q −1 for all n. By (2.52), |bj+1 | = q|bj |q , j = 0, 1, 2, . . ., and we prove easily by induction that |bj | = q

q j −1 q−1

,

j = 0, 1, 2, . . .

(2.56)

For the sequence (2.54) we obtain the estimate |βj | ≤ q

q j −1 q−1

,

j = 0, 1, 2, . . .

(2.57)

Indeed, this is obvious for j = 0. If (2.57) is proved for j ≤ l − 1, then |βl | ≤ max |bm | · |βl−m |q

m

1≤m≤l

≤ max q

q m −1 m q l−m −1 q−1 q−1 +q

1≤m≤l

=q

q l −1 q−1

.

It follows from (2.56) and (2.57) that both the series (2.51) and (2.53) are convergent for t ∈ Dq (in fact they are convergent on a wider disk from K c , but here we consider them only on K). Note also that 1

qn 1 = |t| | logδ (t)| ≤ max q − q−1 q q−1 |t| n≥0

if t ∈ Dq . If λ ∈ Dq , then the function t → eδ (λt) is continuous on O, and by Theorem 2.16 eδ (λt) =

∞

ψn (λ)Qn (t)

n=0

(n) (n) where ψn (λ) = δ0 eδ (λ·) (1) = δ0 eδ (λ) due to the invariance of the (n)

(n)

operator δ0 . Since δ0 Therefore

(n)

n

= τ n δ n and δeδ = eδ , we ﬁnd that δ0 eδ = eqδ .

eδ (λt) =

∞

Qn (t) {eδ (λ)}

qn

(2.58)

n=0

for any t ∈ O, λ ∈ Dq . Setting in (2.58) λ = logδ (z) we come to (2.55).

Calculus

75

2.3 Locally analytic functions 2.3.1. Interpolation series. The Mahler basis of C(Zp , Qp ) mentioned in Section 1.7.1 is a very special case of the general theory of interpolation series developed by Amice [3]. Here we give a brief description of its main results (for the case of a local ﬁeld). Let L be a non-Archimedean local ﬁeld, and qL be its residue ﬁeld cardinality. We denote by OL its ring of integers, and by πL its prime element (see [59] for a summary of basic notions regarding local ﬁelds, as well as references for further reading). Let M ⊂ OL be an inﬁnite compact subset; more generally, it would be possible to consider any compact subset of L of diameter ≤ 1. The distance function d(t1 , t2 ) on L induces a distance −k on M ; for any α ∈ M , denote by Vk (α) the ball in M of radius qL , k ≥ 0, centered in α. The compact set M is called regular, if every ball Vk−1 (α) is a disjoint union of νk balls Vk (αi ), i = 1, . . . , νk . Note that for M = OL , νk = qL for any k. Let us ﬁx a sequence {tn } ⊂ M . Consider the polynomials Pn (t) = (t − t0 )(t − t1 ) · · · (t − tn−1 ). Suppose that the mapping n → tn is injective. Then we may introduce also the normalized polynomials Qn (t) =

Pn (t) (t − t0 )(t − t1 ) · · · (t − tn−1 ) = . (tn − t0 )(tn − t1 ) · · · (tn − tn−1 ) Pn (tn )

Given a function f ∈ C(M, L), we can introduce its interpolation polynomials Fn ∈ L[t], deg Fn ≤ n, by the classical formulas   n k f (t ) j  Pk (tk ), Fn (t) = bk Qk (t), bk =  (t ) P j=0 k+1 j k=0

equivalent to the Lagrange interpolation formula

n f (tk ) Pn+1 (t). Fn (t) = (t − tk )Pn+1 (tk ) k=0

In order to characterize those interpolation points {tn }, for which the above polynomial systems possess basis properties, Amice introduced the notion of a very well distributed sequence {tn }. Denote by Mk , k ≥ 0, the quotient of M with respect to the equivalence relation: α ∼ β (α, β ∈ M ), if and

76

Chapter 2

only if Vk (α) = Vk (β). Let prk : M → Mk be the canonical projection. For an element µ ∈ Mi and n ≥ 1 set Ti (µ) = {k ∈ Z+ : prk (tk ) = µ} , νi (µ, n) = card(Ti (µ) ∩ {0, 1, . . . , n − 1}). A sequence {tn } is called well distributed of order h ∈ N, if, for any i ≤ h, n , νi (µ, n) ≥ int card Mi where int(z) is the integral part of the real number z. Finally, {tn } is called very well distributed if it is well distributed of order h, for any h ≥ 1. The signiﬁcance of this notion becomes clear from the following fundamental result. Theorem 2.18 (Amice) The polynomial sequence {Qk } is an orthonormal basis of the L-Banach space C(M, L), if and only if the sequence {tn } is very well distributed.

The above notions are useful also for the investigation of spaces of locally analytic functions. A function f ∈ C(M, L) is called locally analytic, if for any µ ∈ M there exists a disk Dµ of nonzero radius containing µ, such that the restriction of f to Dµ ∩ M can be extended to a function on Dµ represented by a convergent power series. More precisely, f is called locally analytic of order h (h ∈ N) if the radius of the disk is greater than or equal −h to qL . The set of all such functions is denoted Ah (M ). Suppose that M is a union of disks Vi of radius q −h . Then the set Ah (M ) of locally analytic functions on M of order h can be made a Banach space with the norm |f | = sup |fi | where fi is a restriction of f to Vi , and |fi | is i

the supremum of the absolute values of the Taylor coeﬃcients of fi . Below we assume that diam M = 1. Theorem 2.19 (i) Let h ∈ N. Suppose that elements sn,h ∈ L are chosen 1 Pn have norm 1 in Ah (M ). in such a way that the polynomials Rn,h = sn,h The polynomials Rn,h form an orthonormal basis of Ah (M ), if and only if

Calculus

77

the sequence {tn } is well distributed of order h. In this case |sn,h | =

−λn qL ,

λn =

h

int

k=1

n card Mk

.

(ii) The polynomial system {Rn,h } is an orthonormal basis of Ah (M ) for any h ≥ 0, if and only if the sequence {tn } is very well distributed.

For the proofs of Theorems 2.18 and 2.19 see [3]. It is easy to reformulate Theorem 2.19 to obtain conditions for local analyticity of a function in terms of the coeﬃcients of its expansion in the polynomial system {Pn }. There are two major examples for Amice’s theory. In the ﬁrst example L = Qp , M = Zp , tn = n, n = 0, 1, 2, . . . We get the Mahler basis in C(Zp , Qp ) and the local analyticity conditions in terms of the Mahler expansions. See [3] for the details. In the second example, related to the material of this book, L = K, M = O, card Mk = q k . As shown by Wagner [118], a very well distributed sequence {tn } for this case can be constructed as follows. Let S = {α0 , α1 , . . . , αq−1 } ⊂ O be a complete system of representatives of classes from O/P . For any nonnegative integer n, consider the digit expansion n = n0 + n1 q + · · · + ns q s ,

0 ≤ ni ≤ q − 1,

and set tn = αn0 + αn1 x + · · · + αns xs . With this sequence, the polynomials Qn form an orthonormal basis of the 1 Banach space C(M, K). The polynomials sn,h Pn with sn,h ∈ K, form an orthonormal basis of Ah (O), if and only if |sn,h | = q

−λn

,

λn =

h k=1

int

n qk

.

This result can be extended to the case of locally analytic functions with values in a Banach space E over K, in particular to the case E = K c . 2.3.2. The Carlitz expansions. Let us consider expansions of continuous locally analytic functions in the general Carlitz polynomials Gi (see Section 1.4.2).

78

Chapter 2 Let h be a nonnegative integer. Denote ∞ n ; int µn,h = qi

n ∈ Z+ .

i=h+1

Theorem 2.20 The polynomials xµn,h Gn (t), n ≥ 0, form an orthonormal basis of the Banach space Ah (O). Proof. Let us ﬁrst consider relations between {Qn } and {Gn } as orthonormal bases of C(O, K). Both Qn and Gn are polynomials of the same degree. Therefore for each n ≥ 0, there exist elements gnj ∈ K, j = 0, 1, . . . , n, such that n Gn (t) = gnj Qj (t), max |gnj | = 1. j=0

0≤j≤n

Writing in matrix form, we get      G0 (t) Q0 (t) 0 0 ... 0 g00  G1 (t)   g10 g11   0 ... 0     Q1 (t)     . . .  =  . . . . . . . . . . . . . . . . . . . . . . . . .  . . .  . Gn (t) gn0 gn1 gn2 . . . gnn Qn (t) By Theorem 2.19(i), {xµn,h Qn (t)} is an orthonormal basis of Ah (O). For any n ≥ 0, we write Rn (t) = xµn,h Qn (t), Hn (t) = xµn,h Gn (t), and ρij = gij xµi,h −µj,h . We have      H0 (t) R0 (t) 0 0 ... 0 ρ00  H1 (t)   ρ10 ρ11   0 ... 0     R1 (t)     . . .  = . . . . . . . . . . . . . . . . . . . . . . . . .  . . .  . Hn (t) ρn0 ρn1 ρn2 . . . ρnn Rn (t) where the diagonal elements ρjj have absolute value 1, while all other elements of the matrix (ρij ) belong to O, because ∞ i j int − int ≥ 0, µi,h − µj,h = l q ql i=h+1

if i ≥ j. This proves that Hj belongs, for each j, to the unit ball of Ah (O), and the reductions of these polynomials form a basis of the appropriate reduced Fq -vector space. The desired result is a consequence of Proposition 1.5.

Calculus

79

Note that Theorem 2.20 remains valid for the space of functions with values in K c . It can be extended also to the case of a completion of Fq (x) at an arbitrary ﬁnite place; see [124].

Corollary 2.21 Let f (t) =

∞

an Gn (t), an ∈ K c , be a continuous function

n=0

on O, and let |an | = q −αn ,

γ = lim inf n→∞

αn . n

Then: (i) f is locally analytic of order h, if and only if αn −

∞

int n/q i −→ ∞, as n → ∞.

i=h+1

(ii) f is locally analytic, if and only if γ > 0. If γ > 0 and l = max(0, int(− log(q − 1) + log γ)/ log q) + 1), then f is locally analytic of order h ≥ l. Proof. The ﬁrst part follows directly from Theorem 2.20. In order to prove the second part, we give an estimate of the numbers µn,h . Obviously, µn,h ≤

∞ n n = . i q (q − 1)q h

(2.59)

i=h+1

On the other hand, let us write the q-digit expansion n = nw q w + · · · + n1 q + n0 ,

nw = 0.

Let s(n, h) = nw + · · · + nh+2 + nh+1 and t(n, h) = nh q h + · · · + n1 q + n0 . Then n int = nw q w−h−1 + · · · + nh+2 q + nh+1 , q h+1 n int = nw q w−h−2 + · · · + nh+2 , q h+2 ...... n int = nw . qw

80

Chapter 2

Adding up these equalities and summing the progressions in the right-hand side, we get q w−h − 1 q w−h−1 − 1 q−1 + nw−1 + · · · + nh+1 q−1 q−1 q−1 −h h n − t(n, h) − q s(n, h) q (n − t(n, h)) − s(n, h) = . = q−1 (q − 1)q h

µn,h = nw

It is easy to see that s(n, h) ≤ (q − 1)(w − h) ≤ (q − 1)(logq n − h), t(n, h) ≤ (q − 1)q h , whence n + h − logq n − 1. (2.60) µn,h ≥ (q − 1)q h It follows from (2.59) and (2.60) that 1 µn,h = . (q − 1)q h n

lim

n→∞

(2.61)

Now, since a locally analytic function is locally analytic of some order h, the equality (2.61) implies the equivalence of local analyticity and the relation γ > 0. If γ > 0, then the inequality γ−

1 >0 (q − 1)q h

for a nonnegative integer h implies f ∈ Ah (O). Solving the inequality with respect to h we come to the last assertion of our corollary. 2.3.3. Fq -linear functions. In applications, the result of Corollary 2.21 is often used for continuous Fq -linear functions given by expansions in the normalized Fq -linear Carlitz polynomials, u(t) =

∞

an fn (t),

t ∈ O,

(2.62)

n=0

an ∈ K c , an → 0. For this case, the above result is formulated as follows. Corollary 2.22 The function u is locally analytic if and only if " # γ = lim inf −q −n logq |an | > 0, n→∞

and if (2.63) holds, then u is analytic on any ball of radius q −l , l = max(0, int(− log(q − 1) + log γ)/ log q) + 1).

(2.63)

Calculus

81

An important special case is that of l = 0, that is of functions (2.62), analytic on O. The result of Corollary 2.22 can be expressed in the following simpler form. Corollary 2.23 A function (2.62) is analytic on O, if and only if qq

n

/(q−1)

|an | −→ 0,

as n → ∞.

(2.64)

Let us give an elementary Proof of this fact, avoiding the use of Amice’s theory. Suppose that u is analytic on O, that is u(t) =

∞

n

K c bn → 0.

b n tq ,

(2.65)

n=0

By Theorem 1.8, n

tq =

n

βnj ej (t)

j=0

where βnj

(j) qn ∆ t t=1 . = Dj

We have seen in Section 1.7.2 (and used it in Section 2.1.1) that n−j

tqn tq √ (j) qj ◦ ∆ = , Dn Dn−j whence

∆(j) tq

n

= t=1

Dn j

q Dn−j

,

so that n

tq =

n Dn j

j=0

q Dn−j

fj (t).

Substituting this into (2.65) and changing the order of summation we ﬁnd that ∞ Dk an = q n bk . Dk−n k=n

82

Chapter 2

Since |Dk | = q −(q

k

−1)/(q−1)

, we obtain that

|an | ≤ q −(q

n

−1)/(q−1)

sup |bk |, k≥n

which implies (2.64). Conversely, (2.64) means that an Dn → 0, as n → ∞. Using the explicit formula (1.14) for the Carlitz polynomials we can write n ∞ j an Dn u(t) = (−1)n−j tq . qj D n j=0 Dj Ln−j n=0

It is easy to get that

D n Dj Lqj

n−j

where

snj =

(2.66)

= q snj

(n − j)q j − q j − · · · − q n−1 ,

if n ≥ j + 1,

0,

if n = j.

" # Since snj ≤ q j (n − j − 1) − q n−j−1 , if n ≥ j + 1, and the function z → q z − z increases for z ≥ 1, we have snj < 0. Now we may rewrite (2.66) in the form (2.65) with ck −→ 0, n → ∞. |bn | ≤ max k≥n Dk

2.4 General smooth functions 2.4.1. Smoothness in the sense of Schikhof. The notion of a smooth Fq -linear function introduced in Section 2.1.2 is closely connected with the structures speciﬁc for Fq -linear functions on subsets of a function ﬁeld K – it involves the Carlitz higher diﬀerence operators. Unfortunately, no analog of the latter is known for general continuous functions on K or O. However, there exists a notion of smoothness proposed by Schikhof [98] for arbitrary non-Archimedean ﬁelds. As we will see, for our ﬁeld K the smoothness property in the sense of Schikhof can be expressed equivalently in terms of coeﬃcients of expansions in the general Carlitz polynomials. For a positive integer n, set ∇n O = {(t1 , . . . , tn ) ∈ On : ti = tj if i = j} .

Calculus

83

The n-th order diﬀerence quotient Φn f : ∇n+1 O → K of a function f : O → K is inductively deﬁned by Φ0 f = f , (Φn f )(t1 , t2 , . . . , tn+1 ) = (t1 − t2 )−1 {(Φn−1 f )(t1 , t3 , . . . , tn+1 ) − (Φn−1 f )(t2 , t3 , . . . , tn+1 )} for n ≥ 1 and (t1 , t2 , . . . , tn+1 ) ∈ ∇n+1 O. A function f is called a C n function (f ∈ C n (O, K)), if and only if Φn f can be extended to a continuous function Φn f : On+1 → K. For f ∈ C n (O, K) we deﬁne its n-th hyperderivative D(n) f setting

D(n) f (t) = Φn f (t, t, . . . t). C n -functions admit a characterization in terms of their Taylor expansions [98]. If f ∈ C n (O, K), then for all t, y ∈ O, f (t) = f (y) +

= f (y) +

n−1

(t − y)j D(j) f (y) + (t − y)n Φn f (t, y, . . . y)

j=1 n

(t − y)j D(j) f (y) + (t − y)n Φn f (t, y, . . . y) − D(n) f (y) .

j=1

Conversely, let f : O → K. If there exist continuous functions λ1 , . . . , λn−1 : O → K and Λn : O × O → K, such that f (t) = f (y) +

n−1

(t − y)j λj (y) + (t − y)n Λn (t, y),

t, y ∈ O,

j=1

then f ∈ C n (O, K). The above diﬀerence quotients can be written in another way. Let " ˜ n+1 O = (t, y1 , . . . , yn ) ∈ On+1 : yj + · · · + yj+k = 0 ∇ for 1 ≤ j ≤ n and 0 ≤ k ≤ n − j} . ˜ n+1 O → K by Deﬁne Ψn f : ∇ (Ψn f ) (t, y1 , . . . , yn ) = (Φn f ) (y1 +· · ·+yn +t, y1 +· · ·+yn−1 +t, . . . , y1 +t, t). Then f ∈ C n (O, K) if and only if Ψn f can be extended to a continuous function Ψn f : On+1 → K.

84

Chapter 2 The functions Ψn f can also be deﬁned inductively by Ψ0 f = f , (Ψn f ) (t, y1 , . . . , yn ) = yn−1 {(Ψn−1 f ) (t, y1 , . . . , yn−1 + yn ) − (Ψn−1 f ) (t, y1 , . . . , yn−1 )} ,

n ≥ 1. 2.4.2. Some auxiliary constructions. For any positive integer n, we write its q-digit expansion n = n0 +n1 q+· · ·+nw q w with nw = 0 and denote by ν(n) the largest integer ν, such that q ν divides n. Let l(n) = nw q w . Let Γj be the factorial-like sequence deﬁned by (1.4). Note that

αν(n) q ν(n) − 1 = αν(n) − 1 q ν(n) + (q − 1) q ν(n)−1 + · · · + 1 . Using this identity together with Lemma 1.1(i) we ﬁnd that Γn−1 1 = . Γn Lν(n)

(2.67)

Deﬁne a new sequence of polynomials {Hn (t)} setting H0 (t) = 1, Hn (t) =

Γn+1 Gn+1 (t) , Γn t

n ≥ 1.

Lemma 2.24 The sequence {Hn (t)}n≥0 is an orthonormal basis of C(O, K). Proof. {Hn (t)} is a polynomial of degree n, and its leading coeﬃcient is Γ−1 n , the same as the one for Gn (t). It follows from (2.7) and (2.67) that Hn (t) = gqν(n+1) −1 (t)Gn+1−qν(n+1) (t). Taking into account Proposition 1.17 we see that Hn ≤ 1, and the required property is (see the proof of Theorem 2.20) a consequence of Proposition 1.5. In fact, in the last proof we used the representation Hn (t) =

∞ j=0

θnj Gj (t)

(2.68)

Calculus

85

where θnj ∈ O for each n, j ≥ 0, θnj = 0 for j > n, and |θnn | = 1 for each n. Inverting the triangular matrix of coeﬃcients we ﬁnd that Gi (t) =

∞

γij Hj (t)

(2.69)

j=0

where γij ∈ O, |γii | = 1 for all i, j, γij = 0 for j > i, ∞

θik γkj =

∞

γik θkj = δij .

(2.70)

k=0

k=0

Lemma 2.25 If a function f : O \ {0} → K can be expressed as a ∞ convergent series f (t) = an Hn (t), for all t ∈ O \ {0}, then the sequence n=0

{an } is uniquely determined by the values f (t), t ∈ O \ {0}. Proof. Let f (t) = 0 for all t ∈ O \ {0}. We will show that an = 0 for any n ≥ 0. Let m ∈ Z+ , m < q w − 1, w ∈ N. Denote by S(w) the set of all nonzero polynomials from Fq [x] with degrees < w. From the deﬁnition of the polynomials Hn , changing the summation index we get w q −1

an−1

n=1

Γn Gn (α) = 0 for all α ∈ S(w). Γn−1 α

Therefore

gqw −1−(m+1) (α)

w q −1

n=1

α∈S(w)

an−1

Γn Gn (α) = 0, Γn−1 α

so that w q −1

n=1

an−1

Γn Γn−1

gqw −1−(m+1) (α)Gn (α) = 0,

deg α<w

since Gn (0) = 0 for any n ≥ 1. Using Proposition 1.14 we ﬁnd that an = 0 for all n. 2.4.3. The Carlitz expansions. The main result of this section is as follows.

86

Chapter 2

Theorem 2.26 Let f (t) =

∞

aj Gj (t)

j=0

be a continuous function from O to K. Then f ∈ C m (O, K), if and only if lim |aj |j m = 0.

j→∞

(2.71)

Proof. Using the identities (1.43) and (2.67) we get, for y1 = 0, that Ψ1 f (t, y1 ) = Φ1 f (y1 + t, t) = ∞

1 (f (y1 + t) − f (t)) y1

1 (Gn0 (y1 + t) − Gn0 (t)) y 1 n0 =0 n0 ∞ n0 1 = Gn0 −j1 (t) an0 Hj1 −1 (y1 ) L j1 n0 =0 j1 =1 ν(j1 ) ∞ ∞ n0 1 = Hj1 −1 (y1 )Gn0 −j1 (t) an0 Lν(j1 ) j1 j1 =1 n0 =j1 ∞ ∞ n0 + j1 + 1 an0 +j1 +1 = Hj (y1 )Gn0 (t). j1 + 1 Lν(j1 +1) 1 n =0 j =0

=

0

an0

(2.72)

1

The above transformations are justiﬁed by the fact that in the last series the sequence of terms tends to zero, as n0 + j1 → ∞, for any y1 = 0, t ∈ O. Suﬃciency. Suppose that (2.71) holds. Since 1 ≤ n, L ν(n)

(2.73)

the coeﬃcients at Hj1 (y1 )Gn0 (t) in (2.72) tend to 0, as n0 + j1 → ∞. Let us rewrite (2.72) using (2.68). We ﬁnd that n0 + j1 + 1 an +j +1 0 1 θj ,n Gn (y1 )Gn0 (t). Ψ1 f (t, y1 ) = j1 + 1 Lν(j1 +1) 1 1 1 n0 ,n1 ,j1 ≥0

Using the recursion formula for the operations Ψm we obtain the expression bn0 ,n1 ,...,nm−1 ,jm Hjm (ym ) Ψm f (t, y1 , . . . , ym ) = nm−1 ,...,n1 ,n0 ,jm ≥0

× Gnm−1 (ym−1 ) . . . Gn1 (y1 )Gn0 (t),

(2.74)

Calculus

87

where bn0 ,n1 ,...,nm−1 ,jm nm−1 + jm + 1 n0 + j1 + 1 n1 + j2 + 1 ··· = jm + 1 j1 + 1 j2 + 1 j1 ,...,jm−1 ≥0

×

an0 +j1 +1 θj ,n +j +1 · · · θjm−1 ,nm−1 +jm +1 . Lν(j1 +1) Lν(j2 +1) · · · Lν(jm +1) 1 1 2 (2.75)

Since θij = 0 for i < j, it follows from (2.73) that bn0 ,n1 ,...,nm−1 ,jm ≤ |an0 +j1 +1 | (n0 + j1 + 1)m . In addition, we must have j1 ≥ n1 + j2 + 1, j2 ≥ n2 + j3 + 1, ··············· jm−1 ≥ nm−1 + jm + 1, if bn0 ,n1 ,...,nm−1 ,jm = 0. Therefore bn0 ,n1 ,...,nm−1 ,jm −→ 0, as n0 + n1 + . . . + nm−1 + jm → ∞. # " The sequence Hjm (ym )Gnm−1 (ym−1 ) . . . Gn1 (y1 )Gn0 (t) , jm , nm−1 , . . . , n0 ≥ 0, is an orthonormal basis of C(Om+1 , K) (see [3]). Therefore Ψm f , thus Φm f can be extended to a continuous function from Om+1 to K, so that f ∈ C m (O, K). Necessity. Let f ∈ C m (O, K). Then f ∈ C k (O, K) for 0 ≤ k ≤ m. Therefore, in particular, Ψ1 f (t, y1 ) can be extended to a continuous function Ψ1 f from O × O to K. Since, as before, {Hj1 (y1 )Gn0 (y0 )}j1 ,n0 ≥0 is an orthonormal basis of C(O × O, K), we have Ψ1 f (t, y1 ) = bn0 ,j1 Hj1 (y1 )Gn0 (t), n0 ,j1 ≥0

bn0 ,j1 → 0, as n0 + j1 → ∞. Comparing this with (2.72) we ﬁnd that n0 + j1 + 1 an +j +1 0 1 − bn0 ,j1 Hj1 (y1 )Gn0 (t) = 0 j1 + 1 Lν(j1 +1) n0 ≥0 j1 ≥0

88

Chapter 2

for any 0 = y1 ∈ O, t ∈ O. By Lemma 2.25, n0 + j1 + 1 an0 +j1 +1 = bn0 ,j1 , j1 + 1 Lν(j1 +1)

n0 , j1 ≥ 0.

This equality means that the coeﬃcients of Hj1 (y1 )Gn0 (t) in (2.72) tend to 0, as j1 + n0 → ∞. By (2.68), the continuation Ψ1 f of Ψ1 f (for simplicity, now we preserve the notation Ψ1 f ) can be written as Ψ1 f (t, y1 ) =

n0 + j1 + 1 an0 +j1 +1 θj ,n Gn (y1 )Gn0 (t). j1 + 1 Lν(j1 +1) 1 1 1

n0 ,n1 ,j1 ≥0

Calculating inductively the further diﬀerence quotients, we obtain the expansion (2.74) with the coeﬃcients bn0 ,n1 ,...,nm−1 ,jm of the form (2.75) and bn0 ,n1 ,...,nm−1 ,jm −→ 0,

as n0 + n1 + . . . + nm−1 + jm → ∞.

It follows from (2.75) that j1 ,...,jm−1

×

n0 + j1 + 1 j1 + 1

n1 + j2 + 1 nm−1 + jm + 1 ··· j2 + 1 jm + 1

an0 +j1 +1 θj ,n +j +1 · · · θjm−1 ,nm−1 +jm +1 Lν(j1 +1) Lν(j2 +1) · · · Lν(jm−1 +1) 1 1 2 = Lν(jm +1) bn0 ,n1 ,...,nm−1 ,jm ,

(2.76)

for all jm , n0 , n1 , . . . , nm−1 ≥ 0. Let im−1 ≥ 1, nm−1 + jm + 1 = km−1

(km−1 ≥ 1),

jm + 1 = l(km−1 ). By Lucas’ theorem about binomial coeﬃcients modulo a prime number (see [46]), in this case

km−1 nm−1 + jm + 1 = = 1. jm + 1 l(km−1 )

Calculus

89

Recalling (2.70) we multiply both sides of (2.76) by γkm−1 ,im−1 (km−1 = 1, 2, . . .), and add up to get n0 + j1 + 1 nm−3 + jm−2 + 1nm−2 + jm−1 + 1 ··· j1 + 1 jm−2 + 1 im−1 + 1 j ,...,j 1

m−2

×

an0 +j1 +1 θj ,n +j +1 · · · θjm−2 ,nm−2 +im−1 +1 Lν(j1 +1) · · · Lν(im−1 +1) 1 1 2 = Lν(jm +1) bn0 ,n1 ,...,nm−1 ,jm γkm −1,im −1 km−1 ≥1 jm =l(km−1 )−1 nm−1 =km−1 −jm −1

for all im−1 ≥ 1, and all n0 , n1 , . . . nm−2 ≥ 0. We change the index im−1 to jm−1 , and keep doing the previous step on jm−2 , . . . , j2 , j1 . After that we obtain the identity n0 + i1 + 1 an0 +i1 +1 i1 + 1 Lν(i1 +1) = Lν(j2 +1) · · · Lν(jm +1) k1 ,k2 ,...,km−1

× bn0 ,n1 ,...,nm−1 ,jm γkm −1,im −1 · · · γk2 ,j2 γk1 ,i1

(2.77)

where the summation is made over those km−1 ≥ 1, km−2 ≥ 2, . . . , k1 ≥ m − 1, for which km−1 ≥ jm−1 , . . ., k2 ≥ j2 , k1 ≥ i1 , and jm = l(km−1 ) − 1, nm−1 = km−1 − l(km−1 ), . . ., j2 = l(k1 ) − 1, n1 = k1 − l(k1 ). Let us write the q-digit expansion k = cw q w + cw−1 q w−1 + · · · + c0 , with cw = 0 and w suﬃciently large, and i1 + 1 = l(k) = cw q w , n0 = k − l(k). Then ν(i1 + 1) = w, and we have inductively: k1 ≥ i1 ≥ q w − 1, k2 ≥ j2 ≥ q

w−1

− 1,

j2 + 1 = l(k1 ) ≥ q w−1 , j3 + 1 = l(k2 ) ≥ q

w−2

ν(j2 + 1) ≥ w − 1, ,

ν(j3 + 1) ≥ w − 2,

......... km−1 ≥ jm−1 ≥ q w−m+2 − 1,

jm + 1 = l(km−1 ) ≥ q w−m+1 ,

ν(jm + 1) ≥ w − m + 1. From the equality (2.77) we ﬁnd that ak = Lw Lw−1 · · · Lw−m+1

ck1 ,...,km−1 bn0 ,n1 ,...,nm−1 ,jm

k1 ,...,km−1

with ck1 ,...,km−1 ≤ 1. We have kq −1 < q w ≤ k, so that k −m ≤ |Lw Lw−1 · · · Lw−m+1 | ≤ k −m C(m),

90

Chapter 2

where C(m) depends only on m. Therefore m ck1 ,...,km−1 bn0 ,n1 ,...,nm−1 ,jm |ak |k = Ω k1 ,...,km−1 where Ω is a bounded sequence. As jm → ∞ when k → ∞, we ﬁnd that lim |ak |k m = 0,

k→∞

as desired.

Theorem 2.25 remains valid for functions with values in K c . Comparing k with Theorem 2.5 we see that C0k+1 (O, K c ) ⊂ C q (O, K c ). 2.5 Entire functions 2.5.1. Deﬁnitions. In this section we consider functions on the ﬁeld K∞ (see Section 1.5) and the completion Ω of its algebraic closure K∞ . Just as in the construction of the ﬁeld K c used throughout this book, the absolute value | · |∞ extends to an ultra-metric absolute value on Ω, and Ω is algebraically closed [98]. Below we often use the following notation. Let 0 = t ∈ Fq (x); then t = θ1 , θ1 , θ2 ∈ Fq [x], and we write |t|∞ = q −ν(t) where ν(t) = deg θ2 − deg θ1 θ2 is called the valuation on Fq (x). In particular, ν(t) = − deg t for t ∈ Fq [x]. As we identify K∞ with the ﬁeld Fq ((x−1 )) of formal power series in x−1 , we can extend ν onto K∞ writing ∞

ν as xs = − sup{r ∈ Z, ar = 0}. s=−∞

Note that ν(t) ≥ 0 if and only if |t|∞ ≤ 1; if |t|∞ = 1, then ν(t) = 0. The relation |t|∞ = q −ν(t) is maintained for the extensions of | · |∞ and ν onto Ω. On the other hand, we may call the number deg t = −ν(t) the degree of an element t ∈ Ω. We agree that deg 0 = −∞. Let us consider a series ∞ f (t) = bn tn , bn ∈ Ω. (2.78) n=0

For any real number r, denote Mr (f ) = sup{rn − ν(bn ); n ∈ N}

(2.79)

Calculus

91

where the supremum may take inﬁnite values; for example, Mr (0) = −∞. If |z|∞ ≤ q r , then obviously |f (t)|∞ ≤ q rn−ν(bn ) , so that the function Mr (f ) deﬁned in terms of the coeﬃcients bn describes the growth of the function f . For a precise description of this connection see Theorem 42.2 in [98]. As in classical analysis, a function (2.78) is called entire if the series in (2.78) converges everywhere on Ω. For a function f to be entire, it is necessary and suﬃcient that ν(bn ) + nν(t) → ∞,

as n → ∞,

(2.80)

for any t ∈ Ω. Note that for an entire function f , ν(bn ) is positive for large values of n. Denote by Ent(Ω) the set of all entire functions with coeﬃcients in Ω. It follows from (2.80) that Mr (f ) is ﬁnite for any f ∈ Ent(Ω). For f ∈ Ent(Ω), denote by A(q, f ) the set of real numbers a possessing the following property: there exist real numbers b = b(a) > 0 and r(a), such that for any r ≥ r(a), one has a

Mr (f ) ≤ bq r .

(2.81)

If the set A(q, f ) is empty, we say that the q-order of f is +∞; otherwise the lower bound ω(q, f ) of the set A(q, f ), possibly equal to −∞, is called the q-order of f . Below we use for brevity the notation ω(f ). Suppose that f ∈ Ent(Ω), ω(f ) = ±∞. Denote by B(f ) the set of real numbers b ≥ 0 with the following property: there exists a real number r(b) such that for any r ≥ r(b), Mr (f ) ≤ bq r

ω(f )

.

(2.82)

If B(f ) = ∅, we say that the type of f is +∞; otherwise the type is the lower bound ρ(f ) of the set B(f ). 2.5.2. The q-order and type of an entire function. We begin with an estimate for the coeﬃcients bn of an entire function (2.78). Proposition 2.27 Suppose that the series (2.78) deﬁnes an entire function on Ω with ﬁnite q-order. Then: (i) for any real number θ > ω(f ), there exists an integer n1 ∈ N such that n logq n , (2.83) ν(bn ) ≥ θ for any integer n ≥ n1 ;

92

Chapter 2

(ii) if ω(f ) = 1, and if f is of ﬁnite type, then for any α > ρ(f ), there exists an integer n2 ∈ N such that n (1 + log(α log q)), (2.84) ν(bn ) ≥ n logq n − log q for any integer n ≥ n2 . Proof. Let ω(f ) < ω < θ. There exists r0 ∈ R such that sup{nr − ν(bn ); n ≥ 0} ≤ q rω log n . Then ω log q n logq n ω 1− , ν(bn ) ≥ ω logq n

for r ≥ r0 . Let n ≥ q r0 ω , r =

and taking

θω n ≥ max q r0 ω , q θ−ω

we come to (2.83). Suppose that ω(f ) = 1 and ρ(f ) < ∞. Let α > ρ(f ). There exists a real number r1 such that sup{nr − ν(bn ); n ≥ 0} ≤ αq r for r ≥ r1 . For n ≥ αq r1 log q, we get (2.84) taking r = logq

n . α log q

The next result is an explicit description of the order and type of an entire function given by its coeﬃcients bn . Proposition 2.28 Let f be an entire function on Ω given by the series (2.78). Then n logq n . (2.85) ω(f ) = lim sup ν(bn ) n→∞ If the q-order of f is ﬁnite, then

ρ(f ) = lim sup r→∞

Mr (f ) q rω(f )

.

(2.86)

Calculus

93

Proof. Denote by λ(f ) the right-hand side of (2.85). If λ(f ) = +∞, then it follows from the deﬁnition that ω(f ) = +∞. Suppose that λ(f ) < +∞. Let λ > λ(f ). Then there exists an integer n1 such that n logq n <λ ν(bn )

for n ≥ n1 .

Then rn − ν(bn ) < rn − λ1 n logq n. Investigating (in an elementary way) the function ϕ(n) = rn − λ1 n logq n, we ﬁnd that sup ϕ(n) ≤ bq λr , b > 0. n

This means that λ ∈ A(q, f ), so that ω(f ) ≤ λ(f ). The inverse inequality follows from (2.83). The proof of (2.86) is obvious. As an example of an entire function on Ω, consider the Carlitz exponential ∞ qn t eC (t) = . D n n=0 It follows from the deﬁnition (1.1) of the sequence {Dn } that deg(Dn ) = nq n .

(2.87)

By (2.85) and (2.87), we ﬁnd that ω(eC ) = 1. Next, Mr (eC ) = sup(r − n)q n . n≥1

Finding the maximum of the function [1, ∞) z → (r − z)q z we see that 1 1 q r attained at z = r − . By (2.86), this implies the it equals e log q log q 1 equality ρ(eC ) = . e log q By the ultra-metric Liouville theorem, a bounded entire function over Ω is a constant ([98], Theorem 42.6). In this section we prove a K∞ -analog of a “more delicate” result by P´ olya [84] which states that an entire function ϕ of (exponential) order less than 1 or order 1 and type less than log 2, such that ϕ(N) ⊂ Z, is a polynomial. First we formulate some auxiliary results, in fact, very interesting ones but lying beyond the scope of this book. 2.5.3. Factorials and interpolation. A very general notion of a factorial, covering both the classical and Carlitz factorials, was proposed by Bhargava [13, 14].

94

Chapter 2

Let K be a local ﬁeld with a prime element π and the ring of integers R (in fact, Bhargava’s construction deals with a more general framework of Dedekind domains), S ⊂ R be an arbitrary subset. A π-ordering of S is a sequence Λ = {a0 , a1 , a2 , . . .} ⊂ S where a0 is arbitrary and an is chosen recursively to minimize the valuation of (an − a0 ) · · · (an − an−1 ). The element n!Λ = (an − a0 ) · · · (an − an−1 ) called the generalized factorial generates the same ideal for any choice of Λ ([13], Theorem 1). The n-th generalized binomial polynomial is then deﬁned as (t − a0 ) · · · (t − an−1 ) t t = ; = 1. (2.88) 0 Λ n Λ n!Λ t maps S into R for all n ≥ 0. Moreover [13, 14], n Λ the generalized binomial polynomials (2.88) form an R-basis for the ring Int(S, R) of polynomials over K mapping S into R. If S = R = Zp , then Λ = {0, 1, 2, . . .} is a p-ordering, and n!Λ = n! (in the classical sense). The above basis assertion in this case is close to Mahler’s theorem about the binomial coeﬃcient basis of C(Zp , Zp ) (see [98]). There are some rings for which the above construction works in the “global”, purely algebraic, setting. Such is, in particular, the ring Fq [x]. Deﬁne a one-to-one correspondence between N and Fq [x] as follows. For s every n ∈ N, let n = ni q i be its q-adic expansion. Then, put By construction,

i=0

un =

s

uni xi

i=0

where u0 , u1 , . . . , uq−1 are the elements of Fq . This sequence can be interpreted as a kind of ordering [13, 14] deﬁning Bhargava’s factorial n!Fq [x] =

n−1

(un − uk ).

(2.89)

k=0

It can be shown that, up to nonzero factors from Fq , the elements (2.89)

Calculus

95

coincide with Carlitz’s general factorials Γi =

s

Dini

i=0

(see Chapter 1). The sequence of polynomials n−1 t = Γ−1 (t − uk ) n n Fq [x]

(2.90)

k=0

is a basis of the Fq -module Int(Fq [x]) of polynomials over Fq (x) taking values from Fq [x] on elements Fq [x]. from Note that the polynomial qtm , m ≥ 0, coincides with the normalized Fq [x]

Carlitz polynomial fm (t). A detailed study of the polynomials (2.90) is given in [1]. Here we present some of the results without proofs. Deﬁne the elements an,k and bn,k of Fq [x] (n, k ∈ N) by n t bn,k , tn = n Fq [x] k=0

Γn

n t = (−1)n−k an,k tk . n Fq [x] k=0

We have bn,k = an,k = 0 for k > n and k < 0. We see also that bn,0 = an,0 = 0 for n > 0. Lemma 2.29 The coeﬃcients bn,k and an,k satisfy the recursive relations an+1,k = un an,k + an,k−1 , bn,k = Le(k) bn−1,k−1 + uk bn−1,k , where e(k) denotes the highest power of q dividing k. The above recursive relations are used to obtain estimates of the coeﬃcients an,k , bn,k . Proposition 2.30 If 1 ≤ k ≤ n, then 2q − 1 an,k k − k logq k; ≤ − logq n + deg Γn q−1 deg bn,k ≤ n logq k.

(2.91) (2.92)

96

Chapter 2

It follows from (2.91) that values of the polynomial nt Fq [x] satisfy the estimate 2q−1 q q−1 +δ t deg , n ≥ 1, (2.93) ≤ − logq n + n Fq [x] e log q if x ∈ Ω is of degree δ. 2.5.4. Interpolation series for entire functions. Consider a polynomial g ∈ Ω[t] of degree k, k

g(t) =

cn tn .

n=0

We have

t g(t) = cn bn,j , j Fq [x] n=0 j=0 k

so that g(t) =

n

∆j (g)

j≥0

where ∆j (g) =

t j Fq [x]

cn bn,j .

(2.94)

n≥j

Let f (t) =

∞

cn tn be an entire function on Ω.

If j ≥ 0 and

n=0

lim deg(cn bn,j ) = −∞, we put, extending (2.94),

n→∞

∆j (f ) =

∞

cn bn,j .

n=j

Theorem 2.31 Let f ∈ Ent(Ω) and ω(f ) < 1, or ω(f ) = 1 and ρ(f ) < 1 . Then for all j ∈ N, ∆j (f ) exists, and for all t ∈ Ω, e log q ∞ t f (t) = ∆j (f ) . (2.95) j Fq [x] j=0

Calculus

97

Proof. The assumption can be rephrased as τ (f ) < τ (f ) = lim sup r→∞

Let τ > 0 be such that τ (f ) < τ <

1 where e log q

Mr (f ) . qr

1 . By Proposition 2.27, e log q

deg cn ≤ nθ − n logq n,

n ≥ N1 ,

where θ = logq (eτ log q) < 0. Suppose that n ≥ j. It follows from (2.92) that deg(cn bn,j ) ≤ nθ − n logq n + n logq j → −∞, as n → ∞. This proves the existence of ∆j (f ). An elementary estimate yields the bound deg ∆j (f ) ≤ θj,

j ≥ N1 .

(2.96)

Let t ∈ Ω be of degree δ. By (2.93) and (2.96),

2q−1 t q q−1 +δ + θj → −∞, ≤ − logq j + deg ∆j (f ) j Fq [x] e log q as j → ∞. This implies the convergence of the series in the right-hand side of (2.95). Denote ∞ t , f (t) = ∆j (f ) j Fq [x] j=0

fN (t) =

N −1

cn tn ,

f N (t) =

n=0

N −1 j=0

t . j Fq [x]

∆j (f )

Suppose that t ∈ Ω is of degree δ. Let A > 0. Since fN (t) → f (t), as N → ∞, there exists a natural number N2 such that deg(f (t) − fN (t)) ≤ −A,

N ≥ N2 .

(2.97)

N ≥ N3 .

(2.98)

There exists also N3 ∈ N, such that deg(f (t) − f N (t)) ≤ −A,

98

Chapter 2

Considering N ≥ N1 we have fN (t) − f N (t) =

N −1

t j Fq [x]

[∆j (fN ) − ∆j (f )]

j=0

where ∆0 (fN ) = ∆0 (f ) and ∆j (fN ) − ∆j (f ) =

cn bn,j .

n≥N

For all n ≥ N , deg(cn bn,j ) ≤ n(θ + logq j) − n logq n ≤ N θ. Therefore deg(∆j (fN ) − ∆j (f )) ≤ N θ,

2q−1 t q q−1 +δ , ≤ N θ − logq j + deg (∆j (fN ) − ∆j (f )) j Fq [x] e log q

2q−1

deg(fN (t) − f N (t)) ≤ N θ +

q q−1 +δ . e log q

This results in the existence of N4 ∈ N such that deg(fN (t) − f N (t)) ≤ −A,

N ≥ N4 .

(2.99)

It follows from (2.97), (2.98), and (2.99) that deg(f (t) − f (t)) ≤ −A. Since A is arbitrary, f (t) = f (t), which means the required equality (2.95). The analog of P´ olya’s theorem is as follows. Theorem 2.32 Let f ∈ Ent(Ω) satisfy the conditions of Theorem 2.3.1 and, in addition, f (Fq [x]) ⊂ Fq [x]. Then f is a polynomial. Proof. By Theorem 2.31, for any t ∈ Ω we have the representation (2.95). Substituting t = un , n = 0, 1, 2, . . ., we conclude that the sequence {∆j (f )} can be seen as a solution of the linear system f (u0 ) = ∆0 (f ); u1 u1 f (u1 ) = ∆0 (f ) + ∆1 (f ) , 0 Fq [x] 1 Fq [x] .............................. un un un f (un ) = ∆0 (f ) + ∆1 (f ) + · · · + ∆n (f ) 0 Fq [x] 1 Fq [x] n Fq [x] ..............................

Calculus

99 ∗ n Fq [x] ∈ Fq (by un factorials),and j ∈ Fq [x]

For all n ∈ N, f (un ) ∈ Fq [x] (by our assumption), the relation between Bhargava’s and Carlitz’s

un

Fq [x] for all j < n (since the polynomials (2.90) belong to Int(Fq [x])). By induction, we deduce that ∆j (f ) ∈ Fq [x] for all j ∈ N. Meanwhile we know that deg ∆j (f ) ≤ θj, for j large enough. Since θ < 0, this means that ∆j (f ) = 0 for j large enough, so that f is a polynomial. It can be shown [21] that the function ∞ ∞ t = fn (t) f (t) = q n Fq [x] n=0 n=0 1 is an entire function on Ω, f (Fq [x]) ⊂ Fq [x], ω(f ) = 1, ρ(f ) = e log q , and, 1 of course, f is not a polynomial. Thus, the bound e log q in Theorem 2.32 cannot be improved.

2.6 Measures and divided power series 2.6.1. Non-Archimedean measures. Let Kπ be a completion of the ﬁeld Fq (x) at a ﬁnite place, corresponding to an irreducible polynomial π (see Section 1.5). As usual, we denote by Oπ the ring of integers in Kπ . Let H be a compact totally disconnected Abelian topological group. It is well known (see e.g. [58]) that H has a fundamental system of subgroup neighborhoods {Hj }∞ j=1 of the identity e, such that Hj+1 ⊂ Hj ,

∞ &

Hj = {e},

H/Hj is ﬁnite,

j=1

and H can be represented as a projective limit H = lim H/Hj . ←

A π-adic measure on H is a ﬁnitely additive Oπ -valued function µ on the algebra of compact open subsets of H. Let prj : H → H/Hj be the canonical mapping. For any h ∈ H/Hn , the inverse image pr−1 n (h) is open in H, and the collection of all such sets forms a base of the projective limit topology in H. If ϕ is a function from H/Hj to Oπ , and f : H → Oπ is a function of the form f (h) = ϕ(prj h), h ∈ H, then the function f is called locally constant (obviously, f is constant on a neighborhood of each point).

100

Chapter 2

For any locally constant function f , we can deﬁne its integral as ' f (h) dµ(h) = ϕ(t)µ(pr−1 j (t)), t∈H/Hj

H

and this deﬁnition does not depend on the number j admissible in the deﬁnition of local constancy. More generally, a continuous Oπ -valued function on H can be approximated uniformly by locally constant functions. Then +an obvious approximation procedure leads to a deﬁnition of the integral f (h) dµ(h) for any H

continuous function. This integral establishes the duality of the space of measures to the space of continuous functions. It is important that the measure µ(S) is bounded in absolute value by a constant independent of the compact open set S. In our case this constant equals 1, since we have assumed that µ is Oπ -valued. We omit the details of the above integration theory, since they are given in any textbook on non-Archimedean analysis; see, for example, [98]. Note however that the Volkenborn-type integration of Section 2.1.4 is not covered by the above scheme (it can actually be interpreted as integration with respect to an unbounded measure). Let M [H] be the Oπ -module of Oπ -valued measures on H. Then M [H] can be furnished with the convolution product µ ∗ ν deﬁned by the relation ' ' ' f (h) d(µ ∗ ν)(h) = f (h + s) dµ(h)dν(s). (2.100) H

H H

At this point we specialize to the case H = Oπ (see [43, 44] for the case H = Zp ). As we know (Theorem 1.21), any continuous function f : Oπ → Oπ can be written uniquely as f (t) =

∞

aj Gj (t),

Oπ aj → 0.

j=0

Here Gj are the general Carlitz polynomials. Thus an Oπ -valued measure µ is uniquely determined by the sequence ' Gj (t) dµ(t). (2.101) bj = Oπ

This implies the following result.

Calculus

101

Theorem 2.33 An Oπ -valued measure µ on Oπ is uniquely determined by its moments ' ti dµ(t), i = 0, 1, 2, . . . Oπ

2.6.2. Divided power series. Given an Oπ -valued measure µ on Oπ , we associate to it the formal divided power series Pµ (z) =

∞ j=0

bj

zj j!

(2.102)

zj is a formal symbol which must not be j! understood literally (since j! = 0 in Oπ for j ≥ p). However the multiplication of such symbols is deﬁned in a natural way: j+k z j+k zj zk ∗ = . j j! k! (j + k)! where bj is given by (2.101), and

Deﬁning the addition in a termwise manner and assuming the distributive property we make the set of all formal divided power series a commutative ring. Thus, if we have two formal divided power series, P (z) =

∞

(1) z

bj

j=0

then

j

j!

,

Q(z) =

∞ j=0

(2) z

bj

j

j!

,

 k + l (1) (2) z j  b k bl  . (P ∗ Q)(z) = k j! j=0 ∞



(2.103)

k+l=j

Comparing (2.103) with an expression for the formal divided power series corresponding to the convolution (2.100) (here we use the identity (1.43)) we see that the mapping ∞ zj µ → bj , j! j=0 where the coeﬃcients bj are given by (2.101), is the ring isomorphism from the ring Mπ of Oπ -valued measures on Oπ onto the ring F Dπ of formal divided power series with coeﬃcients from Oπ . For related subjects in the characteristic zero case see [4, 98].

102

Chapter 2

As an example, consider the Dirac measure µ = δα supported in α ∈ Oπ . Proposition 2.34 The formal divided power series corresponding to the measure δα is ∞ zj Pδα (z) = (2.104) Gj (α) . j! j=0 Proof. By the deﬁnition of δα , ' Gj (t) dδα (t) = Gj (α). Oπ

Now the equality (2.103) is a consequence of (2.101) and (2.102).

The next nice property follows directly from the above deﬁnitions and the identity (1.43). Proposition 2.35 For any α, β ∈ Oπ , Pδα+β (z) = Pδα (z)Pδβ (z). 2.6.3. The ring of hyperderivatives. Another interpretation of the ∞ Dk above objects is given in terms of formal diﬀerential operators ak k! k=0 d . Just as formal divided power series, these obwhere ak ∈ Oπ , D = dt jects, called hyper derivatives, do not make literal sense. Nevertheless, these formal objects form a ring Dif f π , and, for example, their action on polynomials is well deﬁned. On the other hand, there is an isomorphism F Dπ → Dif f π given by Dk zk → P (D) = ak . k! k! Note that the ring Dif f π is not generated by D. P (z) =

ak

2.7 Comments Theorems 2.1 and 2.5 were proved by the author in [62]. Note that the case k = 0 (the diﬀerentiability) of Theorem 2.5 was established much earlier by Wagner [119], and our proof for the general case is based on Wagner’s

Calculus

103

result (see Lemma 2.2). The results about the Volkenborn-type integration (Sections 2.1.3 and 2.1.4) are also taken from [62]. The construction of the fractional derivatives (Section 2.1.5) was given in [67]. Proposition 2.8 was proved earlier by Jeong [55]; our elementary proof follows [67]. The version of umbral calculus expounded in Section 2.2 was developed in [65]. The characterization of locally analytic functions (Section 2.3) was given by Yang [124]; the case of functions analytic on O was considered also in the paper [62], from which the elementary proof of Corollary 2.23 is taken. A characterization of general C n -functions (in the sense of Schikhof [98]) in terms of their Carlitz expansions (Theorem 2.26) is a special case of the results by Yang [125] who considered not only the ﬁeld K, but all ﬁnite places of Fq (x). Wagner [120, 121] studied the conditions for diﬀerentiability of a continuous function at a given point from O (in terms of the Carlitz expansion); he also proved Lemma 2.24. General results about entire functions over K∞ are taken from [21]; see also [43, 126]. The analog of P´ olya’s theorem was proved by Adam [1]; for earlier results in this direction see [21, 31]. Our exposition of Oπ -valued measures and divided power series follows Goss [43, 44].

3 Diﬀerential equations

In this chapter we consider various classes of Fq -linear diﬀerential equations with the Carlitz derivatives. We prove existence and uniqueness theorems for regular systems of such equations (the counterparts of regular systems of linear diﬀerential equations over C), as well as for strongly nonlinear equations containing self-compositions y ◦ y ◦ · · · ◦ y of an unknown function y. We consider the behavior of solutions of singular equations; in particular, we introduce and investigate a kind of regular singularity. As a ﬁrst step in developing a theory of partial diﬀerential equations with the Carlitz derivatives, we consider some analogs of classical evolution equations.

3.1 Existence and uniqueness theorems 3.1.1. The framework. As we saw in Section 1.6, the Carlitz exponential eC satisﬁes the simplest equation deC = ec containing the Carlitz derivative d deﬁned in (1.65). In Chapter 4 we will consider many other special functions satisfying diﬀerential equations of this kind. Here our task is to develop a general theory of such equations. In fact, we consider equations (or systems) with holomorphic or polynomial coeﬃcients. The meaning of a polynomial (holomorphic) coeﬃcient in the function ﬁeld case is not a usual multiplication by a polynomial (holomorphic function), but the action of a polynomial (holomorphic function) in the operator τ , τ u = uq . As in classical theory, we have to make a distinction between the regular and singular cases. In the analytic theory of linear diﬀerential equations over C a regular equation has a constant leading coeﬃcient (which can be assumed equal to 1). A leading coeﬃcient of a singular equation is a holomorphic function 105

106

Chapter 3

having zeros at some points. One can divide the equation by its leading coeﬃcient, but then poles would appear at other coeﬃcients, and the solution can have singularities (not only poles but in general also essential singularities) at those points. Similarly, in our case we understand a regular equation as one with the coeﬃcient 1 at the highest order derivative. As usual, a regular higherorder equation can be transformed into a regular ﬁrst-order system. For the regular case we obtain a local existence and uniqueness theorem, which is similar to analogous results for equations over C or Qp (for the latter see [78]). The only diﬀerence is a formulation of the initial condition, which is speciﬁc for the function ﬁeld case. The leading coeﬃcient Am (τ ) of a singular Fq -linear equation of an order m is a nonconstant holomorhic function of the operator τ . Now one cannot divide the equation Am (τ )dm u(t) + Am−1 (τ )dm−1 u(t) + . . . + A0 (τ )u(t) = f (t) for an Fq -linear function u(t) (note that automatically u(0) = 0) by Am (τ ). ∞ If Am (τ ) = ami τ i , ami ∈ K c , then i=0

Am (τ )dm u =

∞

qi

ami (dm u) ,

i=0

and even when Am is a polynomial, in order to resolve our equation with respect to dm u one has to solve an algebraic equation. Thus for the singular case the situation looks even more complicated than in the classical theory. However we show that the behavior of the solutions cannot be too intricate. Namely, in striking contrast to the classical theory, any formal series solution converges in some (suﬃciently small) neighbourhood of the singular point t = 0. Note that in the p-adic case a similar phenomenon takes place for equations satisfying certain strong conditions upon zeros of indicial polynomials [26, 102, 8, 86]. In our case such behavior is proved for any equation, which resembles the (much simpler) case [86] of diﬀerential equations over a ﬁeld of characteristics zero, whose residue ﬁeld also has characteristic zero. Note however that in general a singular equation need not possess a formal power series solution. 3.1.2. Equations without singularities. Let us consider an equation dy(t) = P (τ )y(t) + f (t)

(3.1)

Diﬀerential equations m where for each z ∈ K c , t ∈ K,

107

P (τ )z =

∞

k

πk z q ,

f (t) =

∞ j=0

k=0

j

ϕj

tq , Dj

(3.2)

m πk are m × m matrices with elements from K c , ϕj ∈ K c , and it is assumed that the series (3.2) have positive radii of convergence. The action of the operator k τ uponka vector or a matrix is deﬁned componentwise, so k q that z q = z1q , . . . , zm for z = (z1 , . . . , zm ). Similarly, if π = (πij ) is a k k q matrix, we write π q = πij . The norm |π| of a matrix π with elements from K c is deﬁned as the maximum of the absolute values of the elements. We will seek an Fq -linear solution of the (3.1) in some neighborhood of the origin, of the form y(t) =

∞

i

yi

i=0

tq , Di

m yi ∈ K c ,

(3.3)

where y0 is a given element, so that the “initial” condition for our situation is lim t−1 y(t) = y0 .

t→0

(3.4)

Note that the function (3.3), provided the series has a positive radius of convergence, tends to zero for t → 0, so that the right-hand side of (3.1) makes sense for small |t|. m the equation (3.1) has a unique local Theorem 3.1 For any y0 ∈ K c solution of the form (3.3), which satisﬁes (3.4), with the series having a positive radius of convergence. Proof. Making (if necessary) the substitutions t = c1 t , y = c2 y , with suﬃciently small |c1 |, |c2 |, we may assume that the coeﬃcients in (3.2) are such that ϕj → 0 for j → ∞, |πkq | · q − Using the identities d

i

tq Di

q k+1 −q q−1

≤ 1,

k = 0, 1, . . .

(3.5)

i−1

tq = , i ≥ 1; Di−1

d(const) = 0;

(3.6)

108

τ

Chapter 3

i−1 i tq tq = [i] , i ≥ 1, Di−1 Di

(3.7)

we substitute (3.3) into (3.1), which results in the recurrence formula for the coeﬃcients yi : q k+1 k πk ynq [n + 1]q . . . [n + k]q + ϕql , l = 0, 1, 2, . . . , (3.8) yl+1 = n+k=l

where the expressions in square brackets are omitted if k = 0. It is seen from (3.8) that a solution of (3.1), (3.4) (if it exists) is unique. Since |[n]| = q −1 for all n > 0, we ﬁnd that q k+1 −q k k [n + 1]q . . . [n + k]q = q −(q +···+q) = q − q−1 , and it follows from (3.5),(3.8) that |yl+1 | ≤ max |ϕl |q , |y0 |q

l+1

! l , |y1 |q , . . . , |yl |q .

Since ϕn → 0, there exists a number l0 such that |ϕl | ≤ 1 for l ≥ l0 . Now either |yl | ≤ 1 for all l ≥ l0 (and then the series (3.3) is convergent in a neighborhood of the origin), or |yl1 | > 1 for some l1 ≥ l0 . In the latter case ! l+1 l |yl+1 | ≤ max |y0 |q , |y1 |q , . . . , |yl |q , l ≥ l1 . Let us choose A > 0 in such a way that l

|yl | ≤ Aq ,

l = 1, 2, . . . , l1 . l

Then it follows easily by induction that |yl | ≤ Aq for all l, which implies the convergence of (3.3) near the origin. 3.1.3. Singular equations. We will consider scalar equations of arbitrary order m Aj (τ )dj u = f (3.9) j=0

where ∞

n

tq f (t) = ϕn , Dn n=0 Aj (τ ) are power series having (as well as that for f ) positive radii of convergence.

Diﬀerential equations

109

It will be convenient to start from the model equation m

aj ∈ K c , am = 0.

aj τ j dj u = f,

(3.10)

j=0

Suppose that u(t) is a formal solution of (3.10), of the form u(t) =

∞

n

un

n=0

tq . Dn

(3.11)

Then a0

∞ n=0

un

n n n ∞ m ∞ tq tq tq = ϕn . + aj un [n − j + 1] . . . [n] Dn Dn Dn j=1 n=j n=0

Changing the order of summation we ﬁnd that for n ≥ m   m un a0 + aj [n − j + 1] . . . [n] = ϕn .

(3.12)

j=1

Let us consider the expression m

Φn = a0 +

aj [n − j + 1] . . . [n],

n ≥ m.

j=1

Using repeatedly the identity [i]q + [1] = [i + 1] we ﬁnd that m Φqn

=

m aqo

+

m

m aqj

j=1 m

= aqo +

m

j−1

[n − k]q

k=0

aqj

m

j=1

j−1

m

[n]q

m−k

−

k=0

k

[1]q

m−l

,

l=1

m

that is Φqn = Φ(m) ([n]) where m

Φ(m) (t) = aqo +

m j=1

m aqj

j−1

k=0

tq

m−k

−

k

[1]q

m−l

l=1

is a polynomial on K c of a certain degree N not depending on n. Let θ1 , . . . , θN be its roots. Then m

Φ(m) ([n]) = aqm

N ν=1

([n] − θν ).

110

Chapter 3

As n → ∞, [n] → −x in K c . We may assume that θν = [n] for all ν, if n is large enough. If θν = −x for all ν, then for large n, say n ≥ n0 ≥ m, (m) Φ ([n]) ≥ µ > 0. If k ≤ N roots θν coincide with −x, then n (m) Φ ([n]) ≥ µq −kq ,

n ≥ n0 .

Combining the inequalities and taking the root we get |Φn | ≥ µ1 q −µ2 q , n

n ≥ n0 .

(3.13)

where µ1 , µ2 > 0. Now it follows from (3.12) and (3.13) that the series (3.11) has (together with the series for f ) a positive radius of convergence. Turning to the general equation (3.9) we note ﬁrst of all that one can ∞ αk τ k (even without assuming its conapply an operator series A(τ ) = k=0

vergence) to a formal series (3.11), setting k

τ u(t) =

∞

n+k

k uqn [n

q k−1

+ 1]

n=0

tq . . . [n + k] , Dn+k

k ≥ 1,

and ∞ ql k k−1 t αk uqn [n + 1]q . . . [n + k] A(τ )u(t) = Dl l=0

n+k=l

k−1

where the factor [n + 1]q . . . [n + k] is omitted for k = 0. Therefore the notion of a formal solution (3.11) makes sense for equation (3.9). We will need the following elementary estimate. Lemma 3.2 Let k ≥ 2 be a natural number, with a given partition k = i1 + · · · + ir , where i1 , . . . , ir are positive integers, r ≥ 1. Then q i1 +···+ir + q i2 +···+ir + · · · + q ir ≤ q k+1 . Proof. The assertion is obvious for k = 2. Suppose it has been proved for some k and consider a partition k + 1 = i1 + · · · + ir .

Diﬀerential equations

111

If i1 > 1 then k = (i1 − 1) + i2 + · · · + ir , so that q (i1 −1)+i2 +···+ir + q i2 +···+ir + · · · + q ir ≤ q k+1 whence q i1 +i2 +···+ir + q i2 +···+ir + · · · + q ir ≤ q k+2 . If i1 = 1 then k = i2 + · · · + ir , q i2 +···+ir + q i3 +···+ir + · · · + q ir ≤ q k+1 and q i1 +···+ir + q i2 +···+ir + · · · + q ir ≤ 2q k+1 ≤ q k+2 .

Now we are ready to formulate the main result of this section. Theorem 3.3 Let u(t) be a formal solution (3.11) of equation (3.9), where the series for Aj (τ )z, z ∈ K c , and f (t), have positive radii of convergence. Then the series (3.11) has a positive radius of convergence. Proof. Applying (if necessary) the operator τ a suﬃcient number of times to both sides of (3.9) we may assume that Aj (τ ) =

∞

aji τ i+j ,

aji ∈ K c , j = 0, 1, . . . , m,

i=0

where aj0 = 0 at least for one value of j. Let us assume, for example, that am0 = 0 (otherwise the reasoning below would need an obvious adjustment). Denote by L the operator at the left-hand side of (3.9), and by L0 its “principal part”, L0 u =

m

aj0 τ j dj u

j=0

(the model operator considered above; we will maintain the notation introduced there). Note that L0 is a linear operator. As we have seen, qn n t tq = Φn , n ≥ n0 , L0 Dn Dn

112

Chapter 3

where Φn satisﬁes the inequality (3.13). This means that L0 is an automorphism of the vector space X of formal series ∞

u=

n

un

n=n0

tq , Dn

un ∈ K c ,

as well as of its subspace Y consisting of series with positive radii of convergence. Let us write the formal solution u of equation (3.9) as u = v + w, where v=

n 0 −1

n

un

n=0

tq , Dn

w=

∞

n

un

n=n0

tq . Dn

Then (3.9) takes the form Lw = g,

(3.14)

qn

with g = gn tDn ∈ Y . In order to prove our theorem, it is suﬃcient to verify that w ∈ Y . For any y ∈ X we can write Ly = (L0 − L1 )y = L0 (I − L−1 0 L1 )y where L1 y = −

m ∞

aji τ i+j dj y.

(3.15)

j=0 i=1

In particular, it is seen from (3.14) that −1 (I − L−1 0 L1 )w = L0 g,

L−1 0 g ∈ Y.

Writing formally −1 = (I − L−1 0 L1 )

∞ −1 k L0 L1 k=0

and noticing that L−1 0 L1 : X → τ X, we ﬁnd that w=

∞ −1 k L0 L1 h,

(3.16)

k=0

where h = L−1 0 g =

∞ n=n0

qn

hn tDn , hn = Φ−1 n gn , and the series in (3.16)

converges in the natural non-Archimedean topology of the space X.

Diﬀerential equations

113

A direct calculation shows that for any λ ∈ K c

L−1 0 L1

qn ∞ q n+i t (n) t q i −1 =− λ λ Φn+i Ψi Dn Dn+i i=1

where (n)

Ψi

= [n + 1]q

i−1

[n + 2]q

i−2

. . . [n + i]

m

i

i

[n − j + 1]q . . . [n]q aji ,

j=0

and the coeﬃcient at aj0 in the last sum is assumed to equal 1. Proceeding by induction we get qn −1 r t L0 L1 λ Dn ∞

qi2 +···+ir

qi3 +···+ir (n+i ) (n) r = (−1) Ψi2 1 Ψi1 ...

i1 ,...,ir =1 (n+i1 +···+ir−1 )

× Ψir

λq

i1 +···+ir

i2 +···+ir

Φ−q n+i1

i3 +···+ir

−q Φn+i 1 +i2

. . . Φ−1 n+i1 +···+ir

n+i1 +···+ir

tq × , Dn+i1 +···+ir

r = 1, 2, . . .

Substituting this into (3.16) and changing the order of summation we ﬁnd an explicit formula for w(t): w(t) =

l ∞ tq Dl

l=n0

l−n

(−1)r hqn

(n)

qi2 +···+ir

Ψi1

(n+i1 )

qi3 +···+ir

Ψi2

...

n+i1 +···+ir =l n≥n0 , ii ,...,ir ≥1

i2 +···+ir i3 +···+ir (n+i +···+ir−1 ) Φ−q Φ−q . . . Φ−1 × Ψir 1 n+i1 +···+ir . (3.17) n+i1 n+i1 +i2 Observe that i−1 i−2 (n) Ψi ≤ (q −1 )q +q +···+1 sup |aji |, j

n

|gn | ≤ M1q ,

i

|aji | ≤ M2q ,

M1 , M2 ≥ 1 (due to positivity of the corresponding radii of convergence). We have l−n l i1 +···+ir q M1q , hn ≤ |Φn |−q

114

Chapter 3

and by Lemma 3.2, |Φn |−q

i1 +···+ir

i1 +···+ir

≤ µq1

|Φn+i1 |−q

i2 +···+ir

· · · |Φn+i1 +···+ir |−1

+q i2 +···+ir +···+q ir +1 µ2 (q n+i1 +···+ir +q n+i2 +···+ir +···+q n+ir +q n )

q

≤

l−n+1 +1 µ2 q n (q l−n+1 +1) µq1 q

≤ q µ3 q

l+1

,

µ3 > 0.

Lemma 3.2 also yields i2 +···+ir i3 +···+ir (n+i1 ) q (n+i +···+ir−1 ) (n) q . . . Ψir 1 Ψi2 Ψi1 ≤ M2q

i1 +···+ir +q i2 +···+ir +···+q ir

≤ M2q

l+1

.

Writing (3.17) as w(t) =

∞ l=n0

we ﬁnd that lim sup |wl |q

−l

l→∞

l

wl

tq Dl

−l l l+1 q l+1 ≤ lim sup q µ3 q M1q M2q < ∞, l→∞

which implies the positivity of the radius of convergence.

3.2 Strongly nonlinear equations 3.2.1. Recurrence relations. Studying strongly nonlinear equations and implicit functions we encounter recurrence relations of the same form ci+1 = µi

∞ j+l=i l=0

k=1

Bjkl

qj+λ n1

n1 +···+nk−1

cn1 cqn2 · · · cnq k

+ ai ,

n1 +···+nk =l

i = 1, 2, . . . , (3.18) (here and below n1 , . . . , nk ≥ 1 in the internal sum), with coeﬃcients from j K c , such that |µi | ≤ M , M > 0, |Bjkl | ≤ B kq , B ≥ 1, |ai | ≤ M for all i, j, k, l; the number λ is either equal to 1, or λ = 0, and in that case |B01l | ≤ 1. Proposition 3.4 For an arbitrary element c1 ∈ K c , the sequence detern mined by the relation (3.18) satisﬁes the estimate |cn | ≤ C q , n = 1, 2, . . ., with some constant C ≥ 1.

Diﬀerential equations

115

Proof. Set cn = σdn , |σ| < 1, n = 1, 2, . . ., and substitute this into (3.18). We have di+1 = µi

∞ j+l=i l=0

Bjkl

k=1

σ (1+q

n1

q j+λ

+···+q n1 +···+nk−1 )

n1 +···+nk =l

×

n1 dn1 dqn2

n1 +···+nk−1 · · · dnq k

qj+λ

−1

+ σ −1 ai .

Here (1+qn1 +···+qn1 +···+nk−1 )qj+λ −1 ≤ |σ|kqj+λ −1 , σ and (under our assumptions) choosing σ such that |σ| is small enough we reduce (3.18) to the relation

di+1 = µi

∞ j+l=i l=0

k=1

bjkl

n1

n1 +···+nk−1

dn1 dqn2 · · · dqnk

qj+λ

n1 +···+nk =l

+ σ −1 ai ,

i = 1, 2, . . . , (3.19)

where |bjkl | ≤ 1. It follows from (3.19) that |di+1 | ≤ M max sup j+l=i l=0

max

k≥1 n1 +···+nk =l

n1

|dn1 | · |dqn2 | · · ·

max

q n1 +···+nk−1

× |dnk |

qj+λ ,M

−1

−1 σ ai .

−1 −1 Let B = max 1, M, |d1 |, M sup |σ ai | . Let us show that i

|dn | ≤ B q

n−1

+q n−2 +···+1

,

n = 1, 2, . . . .

(3.20)

This is obvious for n = 1. Suppose that we have proved (3.20) for n ≤ i.

116

Chapter 3

Then

|di+1 | ≤ M max sup

Bq

max

j+l=i k≥1 n1 +···+nk =l

× Bq

n1 +n2 −1

× · · · Bq

+q n1 +n2 −2 +···+q n1

n1 +···+nk−1 +nk −1

≤ M max B q

n1 −1

j+l

+···+q j+1

j+l=i

+q n1 −2 +···+1

···

+q n1 +···+nk−1 +nk −2 +···+1

≤ B · Bq

i

+q i−1 +···+q

+ q n1 +···+nk−1

= Bq

i

+q i−1 +···+1

qj+1

,

and we have proved (3.20). Therefore |cn | ≤ |σ|B for some C, as desired.

q n −1 q−1

≤ Cq

n

3.2.2. Implicit functions of algebraic type. In this section we look for Fq -linear locally holomorphic solutions of equations of the form P0 (t) + P1 (t) ◦ z + P2 (t) ◦ (z ◦ z) + · · · + PN (t) ◦ (z ◦ z ◦ · · · ◦ z) = 0 (3.21) -. / , N

where P0 , P1 , . . . PN ∈ RK c (see Section 1.1.2 for the deﬁnitions). Suppose j ajk tq is such that a00 = 0, a01 = 0; these that the coeﬃcient Pk (t) = j≥0

assumptions are similar to those guaranteeing the existence and uniqueness of a solution in classical complex analysis. Then (see Section 1.1.2) P1 is invertible in RK c , and we can rewrite (3.21) in the form z + Q2 (t) ◦ (z ◦ z) + · · · + QN (t) ◦ (z ◦ z ◦ · · · ◦ z) = Q0 (t) -. / ,

(3.22)

N

where Q0 , Q2 , . . . , QN ∈ RK c , that is Qk (t) =

∞

j

bjk tq ,

j

|bjk | ≤ Bkq ,

j=0

for some constants Bk > 0, and b00 = 0. Proposition 3.5 Equation (3.21) has a unique solution z ∈ RK c satisfying the “initial condition” z(t) −→ 0, t → 0. t

Diﬀerential equations

117

Proof. Let us look for a solution of the transformed equation (3.22), of the form ∞ i z(t) = ci tq , ci ∈ K c ; (3.23) i=1

our initial condition is automatically satisﬁed for a function (3.23). Substituting (3.23) into (3.22) we come to the system of equalities q j  N n1 +···+nk−1  n1  bjk  cn1 cqn2 · · · cqnk ci = −  + bi0 , i ≥ 1. k=2

n1 +···+nk =l nj ≥1

j+l=i j≥0,l≥1

(3.24) In each of them the right-hand side depends only on c1 , . . . , ci−1 , so that the relations (3.24) determine the coeﬃcients of a solution (3.23) uniquely. By Proposition 3.4, z ∈ RK c . More generally, let P1 (t) =

j

ν ≥ 0,

aj1 tq ,

aν1 = 0.

j≥ν

Then equation (3.21) has a unique solution in RK c , of the form z(t) =

∞

i

ci tq ,

ci ∈ K c .

i=ν+1

The proof is similar. 3.2.3. Existence and uniqueness of solutions. Let us consider the equation dz(t) =

∞ ∞ j=0 k=1

ajk τ j (z ◦ z ◦ · · · ◦ z)(t) + , -. / k j

∞

aj0 tq

j

(3.25)

j=0

j

where ajk ∈ K c , |ajk | ≤ Akq (k ≥ 1), |aj0 | ≤ Aq , A ≥ 1. We look for a solution in the class of Fq -linear locally holomorphic functions of the form z(t) =

∞

k

ck tq ,

ck ∈ K c ,

k=1

thus assuming the initial condition t−1 z(t) → 0, as t → 0.

(3.26)

118

Chapter 3

Theorem 3.6 A solution (3.26) of equation (3.25) exists with a non-zero radius of convergence, and is unique. Proof. We may assume that |aj0 | ≤ 1,

aj0 → 0,

as j → ∞.

(3.27)

Indeed, if that is not satisﬁed, we can perform a time change t = γt1 j obtaining an equation of the same form, but with the coeﬃcients aj0 γ q instead of aj0 , and it remains to choose γ with |γ| small enough. Note that, in contrast to the case of the usual derivatives, the operator d commutes with the above time change. (3.27) we substitute (3.26) into (3.25) using the fact that Assuming

k k k−1 1/q d ck tq = ck [k]1/q tq , k ≥ 1, where [k] = xq − x. Comparing the coeﬃcients we come to the recursion ci+1 = [i + 1]−1

∞ j+l=i j≥0,l≥1

k=1

qj+1

aqjk

n1 +···+nk−1

n1

cn1 cqn2 · · · cqnk

n1 +···+nk =l

+ ai0 ,

i ≥ 1,

where c1 = [1]−1 aq00 . This already shows the uniqueness of a solution. The i fact that |ci | ≤ C q for some C follows from Proposition 3.4. Using Proposition 3.5 we can easily reduce to the form (3.25) some classes of equations given in the form not resolved with respect to dz. As in the classical case of equations over C (see [51]), some of equations (3.25) can have also nonholomorphic solutions, in particular those which are meromorphic in the sense of Section 1.1.2. As an example, we consider Riccati-type equations dy(t) = λ(y ◦ y)(t) + (P (τ )y)(t) + R(t)

(3.28)

2

where λ ∈ K c , 0 < |λ| ≤ q −1/q , (P (τ )y)(t) =

∞

k

pk y q (t),

R(t) =

k=1 2

∞ k=0

2

pk , rk ∈ K c , |pk | ≤ q −1/q , |rk | ≤ q −1/q for all k.

k

r k tq ,

Diﬀerential equations

119

Theorem 3.7 Under the above assumptions, equation (3.28) possesses solutions of the form ∞

y(t) = ct1/q +

n

c, an ∈ K c , c = 0,

an tq ,

(3.29)

n=0

where the series converges on the open unit disk |t| < 1. Proof. For the function (3.29) we have dy(t) = c1/q [−1]1/q tq

−2

+

∞

1/q q a1/q t n [n]

n−1

[−1] = x1/q − x,

,

n=1

(y ◦ y)(t) = c ct1/q +

∞

1/q an tq

n

+

n=0 1+ q1

=c

tq

−2

1/q

+ ca0

∞

n=0

+ ca0 tq

−1

ct1/q +

an

+

∞

qn am tq

m=0

∞

1/q

can+1 + cq

n+1

m

n an+1 tq

n=0

+

∞ l=0

tq

l

n

an aqm .

m+n=l m,n≥0

Finally, (P (τ )y)(t) =

∞

pk+1 cq

k+1

k

tq +

k=0

∞

tq

l

l=0

i

pi aqj .

i+j=l i≥1,j≥0

Comparing the coeﬃcients we ﬁnd that c = λ−1 [−1]1/q , 1/q

al+1 ([l+1]1/q −λc)−λcq

l+1

al+1 = λ

1/q

a0

m+n=l m,n≥0

+ a0 = 0, n

q an am +

(3.30) i

pi aqj +rl ,

l ≥ 0.

i+j=l i≥1,j≥0

(3.31) By (3.30), we have |c| ≥ 1, and either a0 = 0, or |a0 | = 1. Next, (3.31) is a recurrence relation (with an algebraic equation to be solved at each step) giving values of al for all l ≥ 1. Let us prove that |aj | ≤ 1 for all j. Suppose we have proved that for j ≤ l. It follows from (3.31) that l+2 q (3.32) al+1 [l + 1] − λq cq al+1 − λq cq al+1 ≤ q −1/q .

120

Chapter 3

Suppose that |al+1 | > 1. We have λq cq = [−1], so that |λq cq | = q −1/q , and since |[l + 1]| = q −1 and |c| ≥ 1, we ﬁnd that l+2 |al+1 [l + 1]| < |λq cq al+1 | < λq cq aql+1 . l+1 Therefore the left-hand side of (3.32) equals |λq cq | · cq · aql+1 > q −1/q , and we have come to a contradiction. 3.3 Regular singularity 3.3.1. Background. Let us consider an analog of the class of equations with regular singularity, the most thoroughly studied class of singular equations (see [27] or [47] for the classical theory of diﬀerential equations over C; the case of non-Archimedean ﬁelds of characteristic zero was studied in [35, 89]). A typical class of systems with regular singularity at the origin ζ = 0 over C consists of systems of the form

∞ k ζy (ζ) = B + y(ζ) Ak ζ k=1

where B, Aj are constant matrices, and the series converges on a neighborhood of the origin. Such a system possesses a fundamental matrix solution of the form W (ζ)ζ C where W (ζ) is holomorphic on a neighborhood of zero, C is a constant matrix, and ζ C = exp(C log ζ) is deﬁned by the obvious power series. Under some additional assumptions regarding the eigenvalues of the matrix B, one can take C = B. For similar results over Cp see Section III.8 in [35]. In order to investigate such a class of equations in the framework of Fq -linear analysis over K, one has to go beyond the class of functions represented by power series. An analog of the power function need not be holomorphic, and cannot be deﬁned as above. Fortunately, we have another option here – instead of power series expansions we can use the expansions in Carlitz polynomials on the compact ring of integers O ⊂ K. It is important to stress that our approach would fail if we consider equations over K c instead of K (our solutions may take their values from K c , but they are deﬁned over subsets of K). In this sense our techniques are diﬀerent from those developed for both the characteristic zero cases. 3.3.2. A model scalar equation. We begin with the simplest scalar

Diﬀerential equations

121

equation λ ∈ K c,

τ du = λu,

(3.33)

whose solution may be seen as a function ﬁeld counterpart of the power function t → tλ . We look for a continuous Fq -linear solution u(t) =

∞

ci fi (t),

t ∈ O,

(3.34)

i=0

where K c ci → 0, as i → ∞, {fi } is the sequence of orthonormal Carlitz polynomials. It follows from the identities (1.12) and (1.21) that ∆fi = [i]fi + fi−1 ,

i ≥ 1;

(3.35)

as we know, ∆f0 = 0. Therefore ∆u(t) =

∞

(cj+1 + [j]cj )fj (t).

j=0

Substituting into (3.33) and using uniqueness of the Carlitz expansion we ﬁnd a recurrence relation cj+1 + [j]cj = λcj ,

j = 0, 1, 2, . . . ,

whence, given c0 , the solution is determined uniquely by cn = c0

n−1

(λ − [j]).

j=0

Suppose that |λ| ≥ 1. Since |[j]| = q −1 for j ≥ 1, we see that |cn | = |c0 | · |λ|n 0 if c0 = 0. This contradiction shows that equation (3.33) has no continuous solutions if |λ| ≥ 1. Therefore we shall assume that |λ| < 1. Let u(t, λ) be the solution of (3.33) with c0 = 1; note that the ﬁxation of c0 is equivalent to the initial condition u(1, λ) = c0 . The function u(t, λ) is a function ﬁeld counterpart of the power function tλ . Theorem 3.8 The function t → u(t, λ), |λ| < 1, is continuous on O. It is analytic on O if and only if λ = [j] for some j ≥ 0; in this case j u(t, λ) = u(t, [j]) = tq . If λ = [j] for any integer j ≥ 0, then u(t, λ) is locally analytic on O if and only if λ = −x, and in that case u(t, −x) = 0 for |t| ≤ q −1 . The relation m

m

u(tq , λ) = u(t, λq + [m]),

t ∈ O,

(3.36)

122

Chapter 3

holds for all λ, |λ| < 1, and for all m = 0, 1, 2, . . . j

Proof. If u(t) = tq , j ≥ 0, then

j

j j j ∆u(t) = (xt)q − xtq = xq − x tq = [j]u(t), j

so that u(t, [j]) = tq . Suppose that λ = [j], j = 0, 1, 2, . . . Then |cn | ≤ {max(|λ|, q −1 )}n → 0 as n → ∞, so that u(t, λ) is continuous. More precisely, if λ = −x, then |λ + x| = q −ν for some ν > 0, j |λ − [j]| = (λ + x) − xq = q −ν , j ≥ j0 , if j0 is large enough. This means that for some positive constant C |cn | = Cq −nν ,

n ≥ j0 .

(3.37)

On the other hand, if λ = −x, then |λ − [j]| = q −q , j

|cn | = q −

q n −1 q−1

.

(3.38)

By Corollary 2.22, the function u is locally analytic if and only if # " γ = lim inf −q n logq |cn | > 0. n→∞

(3.39)

If λ = −x, then by (3.37) γ = 0, so that u(t, λ) is not locally analytic. If λ = −x, we see from (3.39) and (3.38) that γ = (q − 1)−1 , l = 1, and u(t, −x) is analytic on any ball of radius q −1 . We have u(t, −x) =

∞

(−1)n x

q n −1 q−1

fn (t),

n=0

and u(t, −x) is not identically zero on O due to the uniqueness of the Fourier–Carlitz expansion. At the same time, since u(t, −x) is analytic on the ball {|t| ≤ q −1 }, we can write ∞ m u(t, −x) = am tq , |t| ≤ q −1 . m=0

Substituting this into equation (3.33) with λ = −x, we ﬁnd that am = 0 for all m, that is u(t, −x) = 0 for |t| ≤ q −1 .

Diﬀerential equations

123

In order to prove (3.36), note ﬁrst that (3.36) holds for λ = [j], j = 0, 1, 2, . . . Indeed, m q j m m+j = tq u(tq , [j]) = tq and

j

q m m m + xq − x = [m + j]. [j]q + [m] = xq − x

Let us ﬁx t ∈ O. We have u(t, λ) =

 ∞ n−1 n=0



j=0

(λ − [j])

  

fn (t),

and the series converges uniformly with respect to λ ∈ P r where " # P r = λ ∈ K c : |λ ≤ r , for any positive r < 1. Thus u(t, λ) is an analytic element on P r (see m m Chapter 10 of [36]). Similarly, u(tq , λ) and u(t, λq + [m]) are analytic elements on P r (for the latter see Theorem 11.2 from [36]). Suppose that q −1 ≤ r < 1. Since both sides of (3.36) coincide on an inﬁnite sequence of points λ = [j], j = 0, 1, 2, . . ., they coincide on P r (see Corollary 23.8 in [36]). This implies their coincidence for |λ| < 1. The paradoxical fact that a locally analytic function u(t, −x) vanishes on an open subset is a good illustration of the violation of the principle of analytic continuation in the non-Archimedean case. It is also interesting that the nonlinear equation du = λu is much simpler than the linear equation τ du = λu. If in (3.33) λ is an m × m matrix with elements from K c (we shall write λ ∈ Mm (K c )), and we look for a solution u ∈ Mm (K c ), then we can ﬁnd a continuous solution (3.34) with matrix coeﬃcients    i−1  (λ − [j]Im ) c0 , i ≥ 1 (3.40) ci =   j=0

def

(Im is a unit matrix) if |λ| = max |λij | < 1. Note that c0 = u(1), so that if c0 is an invertible matrix, then u is invertible on a certain neighborhood of 1.

124

Chapter 3

3.3.3. First-order systems. Let us consider a system τ du − P (τ )u = 0

(3.41)

with the coeﬃcient P (τ ) given in (3.2). We assume that |πk | ≤ γ, γ > 0, for all k, |π0 | < 1. Denote by g(t) a solution of the equation τ dg = π0 g. Let λ1 , . . . , λm ∈ K c be the eigenvalues of the matrix π0 . Theorem 3.9 If k

λi − λqj = [k],

i, j = 1, . . . , m; k = 1, 2, . . . ,

(3.42)

then the system (3.41) has a matrix solution u(t) = W (g(t)),

W (s) =

∞

k

wk sq ,

w0 = Im ,

(3.43)

k=0

where the series for W has a positive radius of convergence. Proof. Substituting (3.43) into (3.41), using the fact that ∆ = τ d is a k derivation of the composition ring of Fq -linear series, and that ∆(tq ) = k [k]tq , we come to the identity ∞

∞ ∞ ∞ k k k j k wk τ (π0 )τ (g(t)) − πj τ wk τ (g(t)) = 0. [k]wk τ (g(t)) + k=0

j=0

k=0

k=0

(3.44) If the series for W has indeed a positive radius of convergence (which will be proved later), then all the expressions in (3.44) make sense for small |t|, since g(t) → 0 as |t| → 0. Since w0 = Im , the ﬁrst summand in the second sum in (3.44) and the summand with j = k = 0 in the third sum are cancelled. Changing the order of summation we ﬁnd that (3.44) is equivalent to the system of equations k πj τ j (wk−j ), wk [k]Im + τ k (π0 ) − π0 wk =

k = 1, 2, . . . ,

(3.45)

j=1

with respect to the matrices wk . The system (3.45) is solved step by step – if the right-hand side of (3.45) with some k is already known, then wk is determined uniquely, provided the spectra of the matrices [k]Im + τ k (π0 ) and π0 have an empty intersection ([39], Section VIII.1). This condition is equivalent to (3.42), and it remains to prove that the series for W has a nonzero radius of convergence.

Diﬀerential equations

125

Let us transform π0 to its Jordan normal form. We have U −1 π0 U = A where U is an invertible matrix over K c , and A is block-diagonal: A=

l 0

λ(α) Idα + H (α)

α=1

where λ(α) are eigenvalues from the collection {λ1 , . . . , λm }, H (α) is a Jordan cell of order dα having zeros on the principal diagonal and ones on the q k (α) one below it. Denote µk = λ(α) + [k]. If Vk = τ k (U ), then l

0 (α) µk Idα + H (α) . Vk−1 [k]Im + τ k (π0 ) Vk =

(3.46)

α=1

If Bk is the matrix (3.46), and Ck is the matrix in the right-hand side of (3.45), then (3.45) takes the form wk Vk Bk Vk−1 − U AU −1 wk = Ck or, if we use the notation w 1k = U −1 wk Vk , 1k , 1k = C w 1k Bk − Aw where

 1k = U −1 Ck Vk = U −1  C

k

k = 1, 2, . . . ,

(3.47) 

−1  1k−j Vk−j Vk πj τ j U w

j=1

= U −1

k

πj τ j (U )τ j (w 1k−j ) ,

j=1

w 10 = Im . We may assume that |U | ≤ 1, |U −1 | ≤ ρ, ρ > 0. In accordance with the quasidiagonal form of the matrices A and Bk we can decompose the matrix w 1k into dα × dβ blocks

(αβ) , α, β = 1, . . . , l. 1k w 1k = w

1 (αβ) . Then the system (3.47) is decoupled 1k = C Similarly we write C k into a system of equations for each block:

(β) (αβ) (αβ) (αβ) 1 (αβ) . µk − λ(α) w 1k − H (α) w 1k +w 1k H (β) = C (3.48) k Equation (3.48) can be considered as a system of scalar equations with (αβ) respect to elements of the matrix w 1k . Let us enumerate these elements

126

(αβ) w 1k

Chapter 3 lexicographically (in i, j) with the inverse enumeration order of ij (αβ)

(αβ)

1k is obtained from w 1k by the the second index j. The product H (α) w shift of all the rows one step upwards, the last row being ﬁlled by zeros. (αβ) Similarly, the product w 1k H (β) is the result of shifting all the columns (αβ) of w 1k one step to the right and ﬁlling the ﬁrst column by zeros ([39], Section I.3). This means that the system (3.48) (with ﬁxed α, β) with the above enumeration of the unknowns is upper triangular. Indeed, the latter is equivalent to the fact that each equation contains, together with some unknown, only the unknowns with larger numbers, and this property is the result of the above shifts. (αβ) of the system (3.48) equals Therefore the determinant Dk

dα dβ (β) . By our assumption |π0 | < 1, and if λ(α) is an eigenµk − λ(α) value of π0 with an eigenvector f = 0, then |λ(α) | · |f | = |π0 f | < |f |, so that |λ(α) | < 1. This means that all the coeﬃcients on the left in (3.48) have absolute values ≤ 1. (β) It follows from (3.42) that λi = −x, i = 1, . . . , n. As k → ∞, µk = k (β) q k (β) λ + xq − x → −x. Thus µk − λ(α) ≥ σ1 > 0 for all k, whence (αβ) Dk ≥ σ2 > 0 where σ2 does not depend on k. Now we obtain an estimate for the solution of the system (3.47), 1 |w 1k | ≤ ρ1 C k ,

(3.49)

with ρ1 > 0 independent of k. Looking at (3.49) we ﬁnd that qj

1k−j | |w 1k | ≤ ρ2 max |w 1≤j≤k

where ρ2 does not depend on k. We may assume that ρ2 ≥ 1. Now we ﬁnd that k−1

|w 1k | ≤ ρ2q

+q k−2 +···+1

,

k = 1, 2, . . .

(3.50)

Indeed, (3.50) is obvious for k = 1. Suppose that we have proved the inequalities j−1

|w 1j | ≤ ρq2

+q j−2 +···+1

,

1 ≤ j ≤ k − 1.

Diﬀerential equations

127

Then

q k−1 q |w 1k | ≤ ρ2 max 1, |w 11 | , . . . , |w 1k−1 |

k−1 k−1 k−2 (q+1)q k−2 (q k−2 +···+1)q = ρ2q +q +···+1 , , . . . , ρ2 ≤ ρ2 max 1, ρq2 , ρ2 and (3.50) is proved. 1k τ k (U −1 ), we have Therefore, since wk = U w k

k−1

|wk | ≤ ρq · ρ2q

+···+1

q k+1 −1 q−1

≤ ρ3

,

ρ3 > 0,

which means that the series in (3.43) has a positive radius of convergence. Remarks 3.1 (1) If ϕ ∈ Fm q , then, as usual, v = uϕ is a vector solution of the system τ dv − P (τ )v = 0, since the system is Fq -linear. However, the system is nonlinear over K c , so that we cannot obtain a vector solution in such a way for an arbitrary ϕ ∈ K c . (2) Analogs of the condition (3.42) occur also in the analytic theory of differential equations over C (see Corollary 11.2 in [47]) and Qp (Section III.8 in [35]). For systems over C, it is requested that diﬀerences of the eigenvalues of the leading coeﬃcient π0 must not be nonzero integers. Over Qp , in addition to that, the eigenvalues must not be nonzero integers themselves. In both the characteristic zero cases it is possible to get rid of such conditions by using special changes of variables called shearing transformations. For example, let m = 1, and the equation over Qp has the form

∞ k ζu (ζ) = nu(ζ) + u(ζ), n ∈ N. πk ζ k=1

Then the change of variables u(ζ) = ζv(ζ) gives the transformed equation

∞ ζv (ζ) = (n − 1)v(ζ) + πk ζ k v(ζ). k=1

Repeating the transformation, we remove the term violating the condition. A modiﬁcation of this approach works for systems of equations. In our case the situation is diﬀerent. Let us consider again the scalar case m = 1. Here the condition (3.42) is equivalent to the condition π0 = −x (the general solution of the equation π0 −π0q = [k] has the form π0 = −x+ξ k where ξ − ξ q = 0, that is either ξ = 0, or |ξ| = 1; the latter contradicts

128

Chapter 3

our assumption |π0 | < 1). If, on the contrary, π0 = −x, then, as we saw in Theorem 3.8, g(t) = 0 for |t| ≤ q −1 , and the construction (3.43) does not make sense. On the other hand, a formal analog of the shearing transformation for this case is the substitution u = τ (v). However it is easy to see that v satisﬁes an equation with the same principal part, as the equation for u. 3.3.4. Euler-type equations. Classically, the Euler equation has the form ζ m u(m) (ζ) + βm−1 ζ m−1 u(m−1) (ζ) + · · · + β0 u(ζ) = 0 where β0 , β1 , . . . βm−1 ∈ C. It can be reduced to a ﬁrst-order linear system with a constant matrix. The solutions are linear combinations of functions of the form ζ λ (log ζ)k . Of course, such functions have no direct Fq -linear counterparts, and our study of the Euler-type equations will again be based on expansions in the Carlitz polynomials. Let us consider a linear equation τ m dm u + bm−1 τ m−1 dm−1 u + · · · + b0 u = 0

(3.51)

where b0 , b1 , . . . , bm−1 ∈ K c . In order to reduce (3.51) to a ﬁrst-order system, it is convenient to set ϕk = τ k−1 dk−1 u,

k = 1, . . . , m.

It is easy to prove by induction that dτ k−1 − τ k−1 d = [k − 1]1/q τ k−2 . Therefore τ dϕk = τ k dk u + [k − 1]τ k−1 dk−1 u = ϕk+1 + [k − 1]ϕk , k = 1, . . . , m − 1. Next, by (3.51), τ dϕm = τ m dm u + [m − 1]τ m−1 dm−1 u = ([m − 1] − bm−1 )ϕm − bm−2 ϕm−1 − · · · − b0 ϕ1 . Thus equation (3.51) can be written as a system τ dϕ = Bϕ,

ϕ = (ϕ1 , . . . , ϕm ),

(3.52)

Diﬀerential equations where

129



 0 1 0 0 ... 0  0  [1] 1 0 ... 0    0  0 [2] 1 ... 0     B= 0 0 0 [3] . . . 0 .    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .    0  0 0 0 ... 1 −b0 −b1 −b2 −b3 . . . [m − 1] − bm−1

This time we cannot directly use the above results, since |B| ≥ 1. However in some cases it is possible to proceed in a slightly diﬀerent way. Suppose that all the eigenvalues of the matrix B lie in the open disk {|λ| < 1}. Transforming B to its Jordan normal form we ﬁnd that B = X −1 (B0 + N )X where X is an invertible matrix, B0 is a diagonal matrix, |B0 | = µ < 1, N is nilpotent, that is N κ = 0 for some natural number κ, and N commutes with B0 . If Ψ is a matrix solution of the system τ dΨ = (B0 + N )Ψ, then Φ = X −1 ΨX is a matrix solution of (3.52). On the other hand, we can obtain Ψ just as in the case N = 0 considered in Section 3.3.2, writing Ψ(t) =

∞

ci fi (t),

(3.53)

i=0

ci =

  i−1 

j=0

 

(B0 + N − [j]Im )



c0 .

(3.54)

Indeed, the product in (3.54) is the sum of the expressions (−[j])ν1 B0ν2 N ν3 where ν1 + ν2 + ν3 = i, ν3 < κ. Therefore in (3.53) " #i−κ |ci | ≤ |c0 | · |N |κ−1 max(µ, q −1 ) −→ 0,

i → ∞.

Let us consider in detail the case where m = 2. Our equation is τ 2 d2 u + b1 τ du + b0 u = 0.

(3.55)

130

Chapter 3

Now we have the system (3.52) with B=

1 . [1] − b1

0 −b0

The characteristic polynomial of B is D2 (λ) = λ2 + λ(b1 − [1]) + b0 , with the discriminant δ = (b1 − [1])2 − 4b0 . We assume that the eigenvalues are such that |λ1 |, |λ2 | < 1. This condition is satisﬁed, for example, if p = 2, |b0 | < 1, |b1 | < 1. The greatest common divisor of the ﬁrst order minors of B − λI2 is 1. This means that B is diagonalizable if and only if λ1 = λ2 , that is if δ = 0 (see, e.g. [41]). In this case B=X

−1

λ1 0

0 X λ2

(3.56)

for some invertible matrix X, and our system has a matrix solution Φ, such that   ∞ n−1 n−1  XΦ(t)X −1 = tI2 + diag (λ1 − [j]), (λ2 − [j]) fn (t)   n=1

j=0

j=0

def

= diag{ψ1 (t), ψ2 (t)}.

It is easy to see that ψ1 (t) and ψ2 (t) are solutions of equation (3.55). ξ12 , then Indeed, if X −1 = ξξ11 21 ξ22 Φ(t)X

−1

1 = (ξ11 ψ1 (t), ξ21 ψ2 (t)), 0

and (for ψ1 ) it is suﬃcient to show that ξ11 = 0. However X −1 X = I2 , and χ12 if ξ11 = 0, then writing X = ( χχ11 21 χ22 ) we ﬁnd that ξ12 χ22 = 0, ξ12 χ21 = 1. At the same time, by (3.56), ξ12 λ2 χ21 = 0, and ξ12 λ2 χ22 = 1, and we come to a contradiction. Similar reasoning works for ψ2 (t). It follows from the uniqueness of the Fourier–Carlitz expansion that ψ1 and ψ2 are linearly independent. def

If λ1 = λ2 = λ, then B is similar to the Jordan cell λ 1 N= . 0 λ

Diﬀerential equations It is proved by induction that  n−1 (λ − [j])  n−1  j=0 (N − [j]I2 ) =    j=0 0

131

 (λ − [i]) j=0 0≤i≤n−1  . i=j  n−1  (λ − [j])

n−1

j=0

In this case we have the following two linearly independent solutions of (3.55):   ∞ n−1  (λ − [j]) fn (t), ψ1 (t) = t +   n=1

 

ψ2 (t) = t +

∞  n−1



j=0

n=1   j=0 0≤i≤n−1 i=j

(λ − [i])

      

fn (t).

Thus, for the case of the eigenvalues from the disk {|λ| < 1}, we have given an explicit construction of solutions for the Euler type equations.

3.4 Evolution equations 3.4.1. Classes of functions and operators. After the above basic properties of the Carlitz diﬀerential equations are established, a natural next step is to try to consider partial diﬀerential equations. However, here we encounter serious diﬃculties, some phenomena absent in the classical situation. For example, if we consider the natural action of the Carlitz m n derivatives ds and dt on an Fq -linear monomial f (s, t) = sq tq , we notice immediately that dm s f is not a polynomial, nor even a holomorphic function in t, if m > n (since the action of d is not linear and taking the involves m m−1 q-th root). Moreover, it follows from the relation d sq = [m]1/q sq and the commutation property dλ = λ1/q d, where a scalar λ ∈ K c is identiﬁed with the operator of multiplication by λ, that ds and dt do not commute even on monomials f with m < n. It appears that nevertheless there is a class of partial diﬀerential operators (acting on an appropriate class of functions of several variables), which possess reasonable properties. Operators from this class contain the derivative d in only one distinguished variable, and the linear operator ∆ in every other variable. Such operators d and ∆ (in diﬀerent variables) do

132

Chapter 3

not commute either but satisfy a simple commutation relation. Algebraic structures related to this class of operators will be considered in detail in Chapter 5. Denote by Fn+1 (n ≥ 1) the set of all germs of functions of the form u=

∞ i1 =0

...

∞

min(i1 ,...,in )

in =0

m=0

m

i1

in

cm,i1 ,...,in sq1 . . . sqn

zq Dm

(3.57)

where cm,i1 ,...,in ∈ K c are such that all the series are convergent on some neighborhoods of the origin. Let Fˆn+1 be the set of polynomials from Fn+1 . Below d will denote the Carlitz derivative in the variable z, while ∆j will mean the diﬀerence operator ∆ in the variable sj . In the action of each operator d, ∆j on a function from Fn+1 (acting in a single variable) other variables are treated as scalars. Obviously, linear operators ∆j commute with multiplications by scalars: ∆j λ = λ∆j , λ ∈ K c , while dλ = λ1/q d. 3.4.2. The Cauchy problem. Let us consider equations of the form {P (∆1 , . . . , ∆n ) + Q(∆1 , . . . , ∆n )d} u = 0

(3.58)

where P, Q are nonzero polynomials with coeﬃcients from K c . We look for a solution u ∈ Fn+1 of the form (3.57) satisfying an “initial condition” lim z −1 u(z, s1 , . . . , sn ) = u0 (s1 , . . . , sn )

z→0

(3.59)

where u0 (s1 , . . . , sn ) is an Fq -linear holomorphic function on a neighborhood of the origin. The condition (3.59) (similar to the initial conditions for “ordinary” diﬀerential equations with Carlitz derivatives) means actually that the coeﬃcients c0,i1 ,...,in of the solution (3.57) are prescribed for any i1 , . . . , in . Below we use the notation [∞] = −x. Then [n] → [∞], as n → ∞. Theorem 3.10 Suppose that Q([i1 ], . . . , [in ]) = 0

for all i1 , . . . , in = 0, 1, . . . , ∞.

(3.60)

Then the Cauchy problem (3.58)–(3.59) has a unique solution u ∈ Fn+1 . Proof. It is easy to see that

q ij

∆j s

=

ij

[ij ]sq , if ij = 0;

0,

if ij = 0.

Diﬀerential equations

133

q implies the relation The identity Dm = [m]Dm−1

d

m

zq Dm

m−1

zq Dm−1

=

0,

, if m = 0; if m = 0.

Therefore for the function (3.57) we get ∞

du =

i1 =1 ∞

=

∞

min(i1 ,...,in )

in =1

m=1

...

1/q

i1 −1

cm,i1 ,...,in sq1

∞ min(j 1 ,...,jn )

...

j1 =0

in −1

. . . sqn

ν

j1

1/q

jn

cν+1,j1 +1,...,jn +1 sq1 . . . sqn

ν=0

jn =0

m−1

zq Dm−1 zq . Dν

Next, P (∆1 , . . . , ∆n )u =

∞

...

i1 =0

∞

min(i1 ,...,in )

in =0

m=0

m

i1

in

cm,i1 ,...,in P ([i1 ], . . . , [in ])sq1 . . . sqn

zq . Dm

Writing a similar formula for Q(∆1 , . . . , ∆n )u and substituting all this into (3.58) we ﬁnd that ∞ i1 =0

...

∞ in =0

min(i1 ,...,in )

cm,i1 ,...,in P ([i1 ], . . . , [in ])

m=0

+

1/q cm+1,i1 +1,...,in +1 Q([i1 ], . . . , [in ])

m

i1 sq1

in . . . sqn

zq =0 Dm

for arbitrary values of the variables. Hence, we come to the recursion cm+1,i1 +1,...,in +1 =

−cqm,i1 ,...,in

P ([i1 ], . . . , [in ]) Q([i1 ], . . . , [in ])

q ,

m ≤ min(i1 , . . . , in ).

(3.61)

Since all the elements c0,i1 ,...,in (i1 , . . . , in = 0, 1, 2, . . .) are given, from (3.61) we ﬁnd all the coeﬃcients of (3.57). The set {[i], i = 0, 1, . . . , ∞} is compact in K. Therefore the condition (3.60) implies the inequality |Q([i1 ], . . . , [in ])| ≥ µ > 0

(3.62)

134

Chapter 3

for all i1 , . . . , in = 0, 1, 2, . . . , ∞. Note also that |[i]| = q −1 for any i, and |c0,i1 ,...,in | ≤ Crq

i1 +···+q in

(3.63)

for some positive constants C and r, since the series for the initial condition u0 (s1 , . . . , sn ) =

∞

...

i1 =0

∞

i1

in

c0,i1 ,...,in sq1 . . . sqn

in =0

converges near the origin. By (3.62) and (3.63), |c1,i1 +1,...,in +1 | ≤ C1q rq

i1 +1 +···+q in +1

(where C1 > 0 does not depend on i1 , . . . , in ), thus |c2,i1 +2,...,in +2 | ≤ C1q

2

+q q i1 +2 +···+q in +2

r

,

and, by induction |cl,i1 +l,...,in +l | ≤ C1q

l

+q l−1 +···+q q i1 +l +···+q in +l

r

for any l ≥ 0. This means that for any j1 , . . . , jn ≥ l, l

|cl,j1 ,...,jn | ≤ C2q rq

j1

+···+q jn

,

C2 > 0,

which, together with the equality |Dm | = q −

q m −1 q−1

,

implies the convergence of the series in (3.57) near the origin.

Remark 3.2 It is easy to generalize Theorem 3.10 to the case of systems of equations, where P and Q are matrices whose elements are polynomials of ∆1 , . . . , ∆n . In this case, instead of (3.60) we have to require the invertibility of Q([i1 ], . . . , [in ]) for all i1 , . . . , in = 0, 1, . . . , ∞. In an obvious way, this generalization covers also the case of a scalar equation of a higher order in d. Equations (3.58) can be seen as function ﬁeld analogs of classical evolution equations of mathematical physics. Speciﬁc examples of equation (3.58) related to hypergeometric functions will be considered in Chapter 4. In Chapter 5, we will consider modules of holonomic type corresponding to the general equations (3.58).

Diﬀerential equations

135

3.5 Comments The general theory of the Carlitz diﬀerential equations (Section 3.1) was initiated by the author [63]. An important motivation was to cover, by general theorems, the analogs of the classical hypergeometric equations introduced by Thakur [109, 110]. See Chapter 4 for an exposition of this subject. The material given in the subsequent sections is taken from [66] (strongly nonlinear equations), [64] (regular singularity), and [71] (evolution equations).

4 Special functions

This chapter is devoted to some Fq -linear special functions deﬁned or interpreted in terms of the Carlitz diﬀerential equations. In particular, we consider hypergeometric functions, analogs of the Bessel functions, polylogarithms and K-binomial coeﬃcients. We discuss overconvergence properties of some special functions.

4.1 Hypergeometric functions 4.1.1. Hypergeometric equations. Let us consider an evolution equation (in the sense of Section 3.4.2) {P (∆1 , . . . , ∆n ) + Q(∆1 , . . . , ∆n )d} u = 0

(4.1)

with n ≥ max(r, s), r, s ∈ N, P (t1 , . . . , tn ) = Q(t1 , . . . , tn ) =

r

(ti − ai ),

i=1 s

(tj − bj ),

j=1

where ai , bj ∈ K c , and the elements bj do not coincide with any of the elements [ν], ν = 0, 1, . . . , ∞. The condition (3.60) is satisﬁed, and the Cauchy problem for equation (4.1) is well-posed. Let us specify the initial condition (in terms of prescribing the values of c0,i1 ,...,in in the expansion (3.57) of a solution) as follows: c0,0,...,0 = 1,

c0,i1 ,...,in = 0 137

138

Chapter 4

for all other values of i1 , . . . , in . Then cm,i1 ,...,in = 0 for all sets of indices (m, i1 , . . . , in ) except those with m = i1 = . . . = in . Denote σm = cm,m,...,m . By (3.61), we ﬁnd that q  r   r  (−1) a i   i=1 σ1 = , s    bj    (−1)s j=1

q2  r  q r  ([1] − a )   r   ai  i      (−1)  σ2 =

i=1 i=1 , s s      bj  ([1] − bj )      (−1)s  j=1

j=1

and, by induction, after rearranging the factors we get r

σm =

m

([0] − ai )q ([1] − ai )q

i=1 s

m−1

· · · ([m − 1] − ai )q ,

m = 1, 2, . . . .

([0] − bj )qm ([1] − bj )qm−1 · · · ([m − 1] − bj )q

j=1

(4.2) For a ∈ K c , denote a0 = 1, m

am = ([0] − a)q ([1] − a)q

m−1

· · · ([m − 1] − a)q ,

m≥1

(4.3)

(of course, [0] = 0, but we maintain the symbol [0] to have an orderly notation). The Pochhammer-type symbol ·m satisﬁes the recurrence am+1 = ([m] − a)q aqm .

(4.4)

It follows from (4.2) that the solution u(z; t1 , . . . , tn ) of the above Cauchy problem is given by the formula u(z; t1 , . . . , tn ) = r Fs (a1 , . . . , ar ; b1 , . . . , bs ; t1 · · · tn z)

(4.5)

where r Fs is the hypergeometric function r Fs (a1 , . . . , ar ; b1 , . . . , bs ; z)

=

m ∞ a1 m · · · ar m z q . b1 m · · · bs m Dm m=0

(4.6)

The series (4.6) has a positive radius of convergence since b1 , . . . , bs do not coincide with any of the elements [ν], ν = 0, 1, . . . , ∞ (such parameters will be called admissible).

Special functions

139

Let hm be the coeﬃcients of the power series (4.5), that is a1 m · · · ar m , m = 0, 1, 2, . . . . b1 m · · · bs m Dm

m −1 q = [m + 1] = xq − xq = ([m] − [−1])q , we ﬁnd that hm =

Since

Dm+1 q Dm

hm+1 = hqm

([m] − a1 ) · · · ([m] − ar ) ([m] − b1 ) · · · ([m] − bs )([m] − [−1])

q .

(4.7)

hm+1 is the q-th power of a rathqm ional function of [m], which is a clear analog of the basic property of the classical hypergeometric function. Note that any rational function of [m] may appear in (4.7), except those for which (4.7) does not make sense. It follows from (4.4) that ∆i u = ∆u where ∆ = τ d is the Carlitz diﬀerence operator in the variable z. Therefore r Fs satisﬁes the equation     r s   (∆ − ai ) −  (∆ − bj ) d r Fs = 0. (4.8)   The identity (4.7) means that the ratio

i=1

j=1

In particular, for the Gauss-like hypergeometric function 2 F1 we have {(∆ − a)(∆ − b) − (∆ − c)d}2 F1 = 0. Substituting ∆ = τ d and using the commutation relations dτ − τ d = [1]1/q I,

∆τ − τ ∆ = [1]τ,

we can rewrite this equation in the form

! τ (1 − τ )d2 − c − ([1]1/q + a + b)τ d − ab 2 F1 = 0,

(4.9)

resembling the classical hypergeometric equation. If b = [ν], ν = 0, 1, . . . , ∞, then, in particular, |b + x| ≥ µ > 0, whence ν |b − [ν]| = (b + x) − xq = |b + x| ≥ µ for large values of ν. This means that |b − [ν]| ≥ µ1 > 0,

ν = 0, 1, 2, . . . ,

so that ν

|bν | ≥ µ1q

+q ν−1 +···+q

q ν+1 −1 −1 q−1

= µ1

−(q−1)−1 −1

= µ1

q

µ1q−1

q ν .

140

Chapter 4

Therefore, if |z| is small enough, then the series (4.6) converges uniformly with respect to the parameters ai ∈ K c and bj ∈ K c \ {[ν], ν = 0, 1, . . . , ∞} on any compact set. Thus, the function r Fs is a locally analytic function of its parameters. All the above constructions make sense if, instead of the ﬁeld K, we consider a completion of the global ﬁeld Fq (x) with respect to any of its places. Only minor changes are needed in the case of a ﬁnite place corresponding to an irreducible polynomial π ∈ Fq [x] (the ﬁeld K is obtained if π(x) = x). In the case of the “inﬁnite” valuation some additional assumptions are needed to guarantee convergence of the series deﬁning the hypergeometric series (4.6). 4.1.2. Thakur’s hypergeometric functions. The ﬁrst deﬁnitions of hypergeometric functions over K (or, globally, over completions of Fq (x)) were proposed by Thakur [109, 110]. For αi , βi ∈ Z, such that the series below makes sense, Thakur’s hypergeometric function is deﬁned as r Fs (α1 , . . . , αr ; β1 , . . . , βs ; z)

where

=

∞ (α1 )n · · · (αr )n qn z (β1 )n · · · (βs )n Dn n=0

 −(α−1) q  if α ≥ 1;  Dn+α−1 , n (α)n = (−1)n−α L−q , if α ≤ 0, n ≤ −α; −α−n   0, if α ≤ 0, n > −α.

(4.10)

(4.11)

If a = [−α], α ∈ Z, then

m m+α

q−(α−1) −α q −(α−1) ([m] − a)q = xq − xq = xq −x = [m + α]q , and it is easy to check that the Pochhammer–Thakur symbols (4.11) satisfy the same recurrence (4.4) as the above symbols (4.3). The normalization a0 = 1 is diﬀerent from (4.11) and resembles the classical one. Therefore, if ai = [−αi ], bj = [−βj ] (i = 1, . . . , r; j = 1, . . . , s),

αi ∈ Z, βj ∈ N,

then the functions (4.10) and (4.6) coincide up to a change of variable z ⇒ ρz where ρ depends on all the parameters but does not depend on z. The appropriate counterparts of equations (4.8) and (4.9) were among the ﬁrst nontrivial examples of Carlitz diﬀerential equations.

Special functions

141

As for the rationality property (4.7), Thakur’s hypergeometric function corresponds to the case of rational functions with zeros and poles of the form [ν], ν ∈ Z. Note that, in addition to the function (4.10), Thakur considered other possible solutions of the hypergeometric equation. In some cases such solutions can be nonholomorphic, even strongly singular [64]. Apart from (4.10), in [109, 110] Thakur introduced the second analog of the hypergeometric function, with parameters ai , bj from completions of Fq (x), and the expressions en (a) instead of the symbols (a)n . Both kinds of hypergeometric function satisfy various identities resembling the classical ones; however no connections of Thakur’s second analog with the Carlitz diﬀerential equations have been found so far. 4.1.3. Contiguous relations. Among many identities for classical hypergeometric functions, a special role belongs to relations between contiguous functions, that is the hypergeometric functions 2 F1 whose parameters diﬀer by ±1 [9]. For Thakur’s hypergeometric functions, analogs of the contiguous relations were found in [109, 110, 111]. Here we present analogs of the contiguous relations for our more general situation. Of 15 possible relations, we give, just as Thakur did, only two, which is suﬃcient to demonstrate speciﬁc features of the function ﬁeld case; other relations can be obtained in a similar way. Note that the specializations of our identities for the case of parameters like [−α], α ∈ Z, are slightly diﬀerent from those in [109, 110, 111], due to a diﬀerent normalization of our Pochhammer-type symbols. Our ﬁrst task is to ﬁnd an appropriate counterpart of the shift by 1 for parameters from K c . Denote T1 (a) = (a − [1])1/q ,

a ∈ K c.

(4.12)

If a = [−α], α ∈ Z, then T1 ([−α]) = ([−α] − [1])1/q = xq

−α−1

− x = [−α − 1],

so that the transformation T1 indeed extends the unit shift of integers. The inverse transformation is given by T−1 (a) = aq + [1],

a ∈ K c.

(4.13)

Theorem 4.1 The following identities for the Pochhammer-type symbol

142

Chapter 4

and the Gauss-type hypergeometric function are valid for every m ∈ N and any admissible parameters from K c : T1 (a)m = a−q (a − [m])am , m

am+1 = −aq

m+1

T1 (a)qm ; q qm

T−1 (a)m = −([1] + a ) T−1 (a)m = − 2 F1 (T1 (a), b; c; az) 2 F1 (a, b; c; z)

a = 0;

aqm−1 ;

(4.14) (4.15) (4.16)

q qm

([1] + a ) am ; ([m − 1] − a)q

− 2 F1 (a, T1 (b); c; bz) = (a − b)2 F1 (a, b; c; z);

(4.17) (4.18)

− 2 F1 (a, b; c; z)q + (cq − bq )2 F1 (a, b; T1 (c); c−1 z)q − (aq + [1])2 F1 (T−1 (a), b; c; (aq + [1])−1 z) = 0.

(4.19)

Proof. Substituting (4.12) into (4.3) and using the fact that [ν] + [1]1/q = ν xq − x1/q = [ν + 1]1/q , we get T1 (a)m = ([1] − a)q

m−1

([2] − a)q

m−2

· · · ([m] − a),

(4.20)

which implies (4.14). If we raise both sides of (4.20) to the power q and compare the resulting identity with (4.4), we come to (4.15). The proofs of (4.16) and (4.17) are similar, based on the identity [ν] − [1] = [ν − 1]q . Using (4.14) we ﬁnd that 2 F1 (T1 (a), b; c; az)

=

∞

(a − [m])

am bm qm z , cm Dm

(b − [m])

am bm qm z , cm Dm

m=0

2 F1 (a, T1 (b); c; bz) =

∞ m=0

which implies (4.18). Similarly, if we write down all the terms involved in (4.19) and use the identities (4.4), (4.14), (4.15), and (4.17), after rather lengthy but quite elementary calculations we verify the required identity (4.19). 4.2 Analogs of the Bessel functions and Jacobi polynomials 4.2.1. The Bessel-type functions. Let Jn (t) =

∞

k

(−1)

k=0

tq

n+k n

Dn+k Dkq

,

n = 0, 1, 2, . . .

(4.21)

Special functions

143

2

The series (4.21) converges for |t| < q − q−1 . Computing the action of the Carlitz diﬀerence operator ∆ on Jn we ﬁnd that ∆2 Jn (t) − [n]∆Jn (t) + τ Jn (t) = 0.

(4.22)

The equation (4.22) can be reduced to a ﬁrst order system (3.41) with regular singularity. Writing also q −n

J−n (t) = (−1)n {Jn (t)} we get, for all n ∈ Z, the identity ∆(r) Jn (t) = τ r Jn−r (t),

r = 0, 1, 2, . . . ,

and the recurrence relation Jn+1 (t) − [n]Jn (t) + τ Jn−1 (t) = 0. See [24] for some identities involving the second linearly independent solution of the equation (4.22), as well as an analog of the modiﬁed Bessel function. The latter is deﬁned by the series In (t) =

∞

tq

n+k n

k=0

Dn+k Dkq

,

n = 0, 1, 2, . . . .

The functions In and Jn are connected as follows. Let γ ∈ K be a solution r of the equation γ q−1 = −1. Then γ q = (−1)r γ for any natural number r, so that In (γt) = (−1)n γJn (t). On the other hand, In can be expressed via Thakur’s hypergeometric function; see [109]. Just as the Carlitz exponential eC , the function Jn becomes an entire function over K∞ . See [24]. Let us consider some expansions related to the Bessel–Carlitz functions Jn . 2

Proposition 4.2 If |t| < q − q−1 , then ∞

1 tq = n Jl+n (t). Dn Lql n

l=0

(4.23)

144 Proof. We have

Chapter 4 2q n+k −q n −1 2q n+k 1 q−1 ≤ q q−1 , =q n q Dn+k Dk 2

so that for |t| < q − q−1 −ε , ε > 0, |Jm (t)| ≤ sup q −εq

m+k

= q −εq . m

(4.24)

k≥0

1 n Since qn = q lq , we ﬁnd from the inequality (4.24) with m = l + n that Ll the series (4.23) converges on the above region of t. The right-hand side of (4.23) equals k qn n ∞ ∞ ∞ q l+n+ν q k+n (−1)k−r tq 1 t t ν (−1) = , n l+n = qr Dk+n r=0 Lr Dk−r Dn Lql ν=0 Dl+n+ν Dνq l=0 k=0 as desired, due to the identity (1.63).

We will obtain further expansions after we introduce some polynomials having an independent interest. (n,k)

4.2.2. A class of polynomials. The polynomials Pm (t), m, n, k ∈ Z+ , are deﬁned as follows: qn+k m Dr+n+k qr+n+k m−r m (n,k) (−1) t , (4.25) Pm (t) = qn r Dr+k r=0 m Dm where = appears also in the explicit formula (1.14) for the qr r Dr Dm−r additive Carlitz polynomials ei . Comparing (4.25) and (1.14) we see that for n = 0, qk

(0,k) Pm (t) = (em (t))

.

Moreover, q n+k

(n,k) Pm (t) = ∆(n) (em (t))

. (n,k)

There are some other interesting special cases of the polynomials Pm For example, if we allow the value k = −n, n ≤ m, we ﬁnd that (n,−n) (t) = Pm

Dm qn em−n (t). qn Dm−n

.

Special functions

145

Some nice identities are obtained for the modiﬁed polynomials !q−n (n,k) (n,k) (t) = Pm (t) . P1m It can be veriﬁed directly that (n+1,k−1) (n,k) ∆P1m (t) = P1m (t)

!q .

The above connection with the Carlitz polynomials em leads to various (n,k) (n,k) identities involving them together with Pm or P1m (see [24]). (n,k) Let us prove an expansion formula involving Jk (t) and Pm (t). 2

Proposition 4.3 If s, t ∈ K, |s| ≤ 1, |t| < q − q−1 , then n

Jkq (st) =

∞ (−1)m n+k

m=0

q Dm

(n,k) Jm+k+n (t)Pm (s).

(4.26)

Proof. By the deﬁnition of Jk , n

Jkq (st) =

∞

(st)q

(−1)r

Dr

r=0

r+k+n

q k+n

n

q Dr+k

.

Using (4.23) we get n

Jkq (st) = =

∞

(−1)r

r=0 ∞ m=0

Dr+k+n k+n qn Drq Dr+k

Jm+k+n (t)

∞ r=0

sq

r+k+n

(−1)r

∞

1

q r+k+n l=0 Ll

Jr+l+k+n (t)

r+k+n Dr+k+n sq . q k+n q n q r+k+n Dr Dr+k Lm−r

(n,k)

Now the identity (4.26) follows from the deﬁnition of Pm

(s).

See [24] for further identities containing the above functions and polynomials.

4.3 Polylogarithms 4.3.1. A logarithm-like function. We begin with an analog of the function − log(1 − t) deﬁned via the equation (1 − τ )du(t) = t,

t ∈ K,

(4.27)

146

Chapter 4

a counterpart of the classical equation (1 − t)u (t) = 1 (note that the function i(t) = t is a unit element in the composition ring). Let us look for an Fq -linear holomorphic solution u(t) =

∞

an tq

n

(4.28)

n=0

of equation (4.27). We have du(t) =

∞

1/q q a1/q t n [n]

n−1

.

n=1

Substituting into (4.27) we ﬁnd that ∞

j 1/q aj+1 [j + 1]1/q − aj [j] tq = t,

j=0

and equation (4.27) is satisﬁed if and only if a0 is arbitrary, a1 = [1]−1 , aj+1 = aqj

[j]q , [j + 1]

j ≥ 1,

By induction, we ﬁnd that aj = [j]−1 . Let l1 (t) be the solution (4.28) of equation (4.27) with a0 = 0. Then l1 (t) =

∞ qn t . [n] n=1

(4.29)

The series in (4.29) converges if t ∈ K, |t| ≤ q −1 . The function l1 (t) is clearly diﬀerent from the Carlitz logarithm logC (see Chapter 1). From the composition ring viewpoint, logC (t) is an analog of e−t , though in other respects it is a valuable analog of the logarithm. By the way, another possible analog of the logarithm is a continuous (on O) solution of the equation ∆u(t) = t, an analog of tu (t) = 1, such that u(1) = 0. By the identity (1.36), u = D1 , the ﬁrst hyperdiﬀerentiation; see Chapter 1. Now we consider continuous nonholomorphic extensions of l1 . We will use the following simple lemma. Lemma 4.4 Consider the equation z q − z = ξ,

ξ ∈ K c.

(4.30)

Special functions

147

If |ξ| = 1, then for all the solutions z1 , . . . , zq of equation (4.30), |z1 | = . . . = |zq | = 1. If |ξ| < 1, then there exists a unique solution z1 of equation (4.30) with |z1 | = |ξ|. This solution can be written as z1 = −

∞

j

ξq .

(4.31)

j=0

For all other solutions we have |zj | = 1, j = 2, . . . , q. Proof. Let |ξ| = 1. If some solution zj of equation (4.30) is such that |zj | < 1, then the ultra-metric inequality would imply |ξ| < 1. If |zj | > 1, then |zj |q > |zj |, so that |ξ| = |zj |q > 1, and we again come to a contradiction. Now suppose that |ξ| < 1. Then the series in (4.31) converges and deﬁnes a solution of (4.30), such that |z1 | = |ξ|. All other solutions are obtained by adding elements of Fq to z1 . Therefore |z2 | = . . . = |zq | = 1. Theorem 4.5 Equation (4.27) has exactly q continuous solutions on O coinciding with (4.29) as |t| ≤ q −1 . These solutions have the expansions in the normalized Carlitz polynomials u=

∞

ci fi

(4.32)

i=0

where the coeﬃcient c1 is an arbitrary solution of the equation cq1 − c1 + 1 = 0,

(4.33)

higher coeﬃcients are found from the relation cn =

∞

q j+1

(cn−1 [n − 1])

,

n ≥ 2,

(4.34)

j=0

and the coeﬃcient c0 is determined by the relation c0 =

∞ i=1

(−1)i+1

ci , Li

(4.35)

Proof. Writing equation (4.27) in the form du(t) − ∆u(t) = t and using the relations (1.21) and (3.35) we ﬁnd that ∞

1/q ci+1 − ci+1 − ci [i] fi (t) = f0 (t),

i=0

t ∈ O.

148

Chapter 4

This is equivalent to equation (4.33) for c1 and the relations cqn − cn + [n − 1]q cqn−1 = 0,

n ≥ 2.

(4.36)

By Lemma 4.4, there are q solutions of equation (4.33), and for any of them |c1 | = 1. For n = 2, we ﬁnd from Lemma 4.4 and the equality |[1]| = q −1 that the corresponding equation (4.36) has the solution (4.34) with |c2 | = q −q and q − 1 other solutions with absolute value 1. Choosing at each subsequent step the solution (4.34) we obtain the sequence cn , such that |cn | ≤ q −q

n−1

,

n ≥ 2,

(4.37)

so that |cn | → 0, and the series (4.32) indeed determines a continuous Fq -linear function on O. Since fi (t) =

i

(−1)i−j

j=0

1

j

j Dj Lqi−j

tq ,

we see that (−1)i fi (t) = . t→0 t Li lim

Therefore, if we choose c0 according to (4.35), then our solution u is such that lim t−1 u(t) = 0.

(4.38)

t→0

Note that |Ln | = q −n , so that the series in (4.35) is convergent. By Corollary 2.22, it follows from (4.37) that u is locally analytic; specifically, it is analytic on any ball of radius q −1 . Thus it can be represented for |t| ≤ q −1 by convergent power series (4.28), in which a0 = 0 by (4.38). Therefore u(t) = l1 (t) for |t| ≤ q −1 , as desired. Any other continuous solution of equation (4.27) on O is obtained inevitably by the same procedure, but with |c1 | = . . . = |cN | = 1, |cn | < 1, if n ≥ N + 1, with some N > 1, and with some c0 ∈ K c . In this case by Lemma 4.4, |cN +1 | = q −q and, by induction, |cN +l | = q −(q

l

+q (l−1) +···+q )

.

(4.39)

Indeed, this was shown above for l = 1. If (4.39) is true for some l, then |cN +l+1 | = |cN +l |q q −q = q −(q

l+1

+q l +···+q 2 +q )

,

Special functions

149

and we have proved (4.39) for any l ≥ 1. It will be convenient to write (4.39) in the form |cN +l | = q −αl ,

αl = q l + q (l−1) + · · · + q.

(4.40)

Suppose that our solution coincides with the series (4.29) for |t| ≤ q −1 . By Fq -linearity, this means the analyticity of the solution on any ball of radius q −1 . Using a specialization of Corollary 2.21 for the Fq -linear case we ﬁnd that ∞ − logq |cn | − q n−i −→ ∞ as n → ∞, i=2

that is − logq |cn | −

q n−1 −→ ∞ q−1

as n → ∞.

In particular, we obtain that αl −

q N +l−1 −→ ∞ q−1

as l → ∞.

(4.41)

However by (4.40), αl =

q l+1 − q , q−1

which contradicts (4.41), since N ≥ 2.

In fact, continuous solutions which satisfy (4.35) and have coeﬃcients cn of the form (4.34), but starting from some larger values of n, are also extensions of the function (4.29), but from smaller balls. Below we denote by l1 (t) a ﬁxed solution of equation (4.27) on O coinciding with (4.29) for |t| ≤ q −1 , as described in Theorem 4.5. 4.3.2. Polylogarithms. The polylogarithms ln (t) are deﬁned recursively by the equations ∆ln = ln−1 ,

n ≥ 2,

(4.42)

which agree with the classical ones tln (t) = ln−1 (t). If we look for analytic Fq -linear solutions of (4.42), such that t−1 ln (t) → 0 as t → 0, we obtain easily by induction that ln (t) =

j ∞ tq , [j]n j=1

|t| ≤ q −1 .

(4.43)

150

Chapter 4

In order to ﬁnd continuous extensions of ln onto O, we consider the Carlitz expansions ln =

∞

(n)

ci fi ,

n = 2, 3, . . .

(4.44)

i=0

Consider ﬁrst the dilogarithm l2 . We have ∆l2 =

∞

(2)

(2)

ci+1 + [i]ci

fi ,

i=0

so that (2)

(2)

ci+1 + [i]ci

= ci ,

i = 0, 1, 2, . . . ,

(4.45)

where ci are the coeﬃcients described in Theorem 4.5. The recursion (4.45) (2) leaves c0 arbitrary and determines all other coeﬃcients in a unique way: n c(2) n = (−1) Ln−1

∞

(−1)j

j=n

cj , Lj

n ≥ 1.

(4.46)

Indeed, the series in (4.46) is convergent, since cn satisﬁes the estimate (4.37), while |Ln | = q −n . For n = 1 the equality (4.46) means, due to (2) (4.35), that c1 = c0 , which coincides with (4.45) for i = 0. If (4.46) is proved for some n, then (2)

n+1 Ln cn+1 = cn − [n]c(2) n = cn + (−1)

∞ j=n

= (−1)n+1 Ln

∞ j=n+1

as desired. We have

(−1)j

(−1)j

cj Lj

cj , Lj

cj = q j−qj−1 , Lj

so that for n > 1, ∞ c (−1)j j ≤ sup cj ≤ q sup q j−1 q −qj−1 . Lj j≥n Lj j≥n j=n

Special functions

151

The function z → zq −z is monotone decreasing for z ≥ 1. Therefore ∞ c j (−1)j ≤ q n · q −qn−1 , n > 1, L j j=n so that by (4.46)

n−1 (2) cn ≤ q · q −q ,

n > 1.

Using Corollary 2.22 as above, we ﬁnd that l2 is analytic on all balls of the (2) radius q −1 . If we choose c0 in such a way that (2)

c0 =

∞

(2)

(−1)i+1

i=1

ci , Li

the solution (4.44) of equation (4.42) with n = 2 is a continuous extension of the dilogarithm l2 given by the series (4.43) with n = 2. Repeating the above reasoning for each n, we come to the following result. Theorem 4.6 For each n ≥ 2, there exists a unique continuous Fq -linear solution of equation (4.42) coinciding for |t| ≤ q −1 with the polylogarithm (4.43). The solution is given by the Carlitz expansion (4.44) with i−1 (n) ci ≤ Cn q −q , i > 1, Cn > 0, (n)

c0

=

∞

(n)

(−1)i+1

i=1

ci . Li

The above construction and properties of polylogarithms can be implemented also in the global ﬁeld setting, for any ﬁnite place of Fq (x). See [67] for the details. 4.3.3. Zeta function. We deﬁne ζ(t), t ∈ K, setting ζ(0) = 0, ζ(x−n ) = ln (1),

n = 1, 2, . . . ,

and then using the fractional derivative of Section 2.1.5 and setting

ζ(t) = ∆(θ0 +θ1 x+··· ) ln (1), n = 1, 2, . . . , if t = x−n (θ0 +θ1 x+· · · ), θj ∈ Fq . The correctness of this deﬁnition follows

152

Chapter 4

from Proposition 2.9. It is clear that ζ is a continuous Fq -linear function on K with values in K c . In particular, we have ζ(xm ) = ∆m+1 l1 (1), m = 0, 1, 2, . . . . The above deﬁnition is of course inspired by the classical polylogarithm relation ∞ ∞ zn d zn . = z dz n=1 ns ns−1 n=1 Let us write down some relations for “special values” ζ(xn ), n ∈ N. Let us consider the expansion of ln (t) in the sequence of hyperdiﬀerentiations. By Theorem 1.12, ∞ i ∆ ln (1)Di (t). ln (t) = i=0

Therefore ln (t) =

∞

ζ(x−n+i )Di (t),

n ∈ N, t ∈ O.

(4.47)

i=0

In particular, combining (4.47) and (4.43) we get j ∞ ∞ tq = ζ(x−n+i )Di (t), n [j] i=0 j=1

|t| ≤ q −1 .

Let us consider the double sequence An,r ∈ K, An,1 = (−1)n−1 Ln−1 , 1 , r ≥ 2. An,r = (−1)n+r Ln−1 [i ][i ] . . . [ir−1 ] 0
r−1

This sequence appears in the expansion of a hyperdiﬀerentiation Dr in the normalized Carlitz polynomials (see (1.31)), as well as in the expression [54] of the Carlitz diﬀerence operators ∆(n) via the iterations ∆n : ∆(n) =

n

An,r ∆r ,

n ≥ 1.

r=1 (n)

For coeﬃcients of the expansion (4.44) we have ci (see Theorem 1.8), and by (4.48), (n)

ci

=

i r=1

Ai,r (∆r ln ) (1) =

i r=1

(4.48) = ∆(i) ln (1), i ≥ 1

Ai,r ζ(xr−n ).

(4.49)

Special functions (n)

Since c0

153

= ζ(x−n ), we have (see Theorems 4.5 and 4.6) ζ(x−n ) =

∞

(−1)i+1 L−1 i

i

Ai,r ζ(xr−n ).

(4.50)

r=1

i=1

The identity (4.50) may be seen as a distant relative of Riemann’s functional equation for the classical zeta function. Since Dr (t) is not diﬀerentiable (Section 1.3.1), the interpretation of the sequence {Ai,r } given in (1.31) shows, by Lemma 2.2, that L−1 i Ai,r 0 as i → ∞. Thus it is impossible to change the order of summation in (4.50). Finally, consider the coeﬃcients of the expansion (4.32) for l1 . As in (4.49), we have an expression ci =

i

Ai,r ζ(xr−1 ).

r=1

By Theorem 4.5, for i ≥ 2 we have ci =

∞

j

zi = cqi−1 [i − 1]q ∈ K c .

(zi )q ,

(4.51)

j=0

The series in (4.51) may be seen as an analog of

j −z . This analogy

j

becomes clearer if, for a ﬁxed z ∈ K c , |z| < 1, we consider the set S of all ∞ jn convergent power series z q corresponding to sequences {jn } of natural n=1 i

j

ij

numbers. Let us introduce the multiplication ⊗ in S setting z q ⊗z q = z q and extending the operation distributively (for a similar construction in the framework of q-analysis in characteristic 0 see [80]). Denoting by ⊗ the p

product in S of elements indexed by prime numbers we obtain in a standard way the identity ∞ pn ⊗ (zi )q ci = p

n=0

(the inﬁnite product is understood as a limit of the partial products in the topology of Ωx ), an analog of the Euler product formula. The above polylogarithms and zeta function are diﬀerent from the existing objects playing an essential part at the function ﬁeld arithmetic (see [45, 94, 111]) and having no relations to the Carlitz diﬀerential equations. At present the author is unaware of any arithmetic applications, though

154

Chapter 4

hopes for them to emerge. This hope is based on the fact that the above ζ is purely an object of the characteristic p arithmetic, while Goss’s zeta function is deﬁned initially on Z and interpolated onto Zp .

4.4 K-binomial coeﬃcients 4.4.1. Absolute values. Let us consider the K-binomial coeﬃcients Dk k = 0 ≤ m ≤ k, (4.52) qm , m K Dm Dk−m introduced in the positive characteristic version of umbral calculus (Section 2.2). A straightforward calculation shows that k = 1, 0 ≤ m ≤ k. m K k ∈ Fq (x), it is natural to consider also other places of Fq (x) Since m K corresponding to monic irreducible polynomials π ∈ Fq [x] (see Section 1.5). Let δ = deg π, and denote by |·|π the absolute value on Fq (x) corresponding to π. Below we prove that the expression (4.52) belongs to the ring of integers not only for the ﬁeld K, but for any ﬁnite place of the global function ﬁeld Fq (x). Together with Section 2.2, this property supports the case for considering the expressions (4.52) as “proper” analogs of the classical binomial coeﬃcients. For other analogs see [111]. Proposition 4.7 For any monic irreducible polynomial π ∈ Fq [x], the K-binomial coeﬃcients (4.52) satisfy the inequality k ≤ 1, 0 ≤ m ≤ k. m K π Proof. First we compute |Dm |π . It follows from Lemma 2.13 of [76] that q −δ , if δ divides i, |[i]|π = 1, otherwise.

Special functions

155

Writing m = jδ + i, with i, j ∈ Z+ , 0 ≤ i < δ, we ﬁnd that i

(j−1)δ+i

δ+i

|Dm |π = |[jδ]|qπ |[(j − 1)δ]|qπ . . . |[δ]|πq i −δ q(j−1)δ q −δ qδ −δ · ... · q = q · q =

q

(j−1)δ δ −δ 1+q +···+q

q i =q

jδ −1 q −1

−δq i q δ

.

(we already made similar computations in the proof of Lemma 1.19). Similarly we can write k − m = κδ + λ, with κ, λ ∈ Z+ , 0 ≤ λ < δ, and get that |Dk−m | = q

κδ −1 q δ −1

−δq λ q

.

If i + λ < δ, then we obtain a similar representation for k simply by adding those for m and k − m, so that k logq m K π !

δ q i+λ q (j+κ)δ − 1 − q i q jδ − 1 − q λ q κδ − 1 q jδ+i =− δ q −1 δ =− δ q i 1 + q λ+jδ − q λ − q jδ q −1 δ =− δ q i q λ − 1 q jδ − 1 ≤ 0. q −1 If i + λ ≥ δ, then k = (j + κ + 1)δ + ν where 0 ≤ ν = i + λ − δ < δ. In this case k logq m K π !

δ q ν q (j+κ+1)δ − 1 − q i q jδ − 1 − q λ q κδ − 1 q jδ+i =− δ q −1 δ i =− δ q + q λ+jδ+i − q i+jδ − q ν < 0, q −1 since ν < i + λ.

Below we will use only the valuation with π(x) = x, that is, as above, consider the ﬁeld K.

156

Chapter 4

4.4.2. Identities. Let us derive, for the K-binomial coeﬃcients (4.52), analogs of the classical Pascal and Vandermonde identities.

Proposition 4.8 The identity q q k k−1 k−1 = + Dq−1 (4.53) m−1 K m K m m K k−1 k = = 0. holds, if 0 ≤ m ≤ k and it is assumed that k −1 K K Proof. Let em (t) = Dm fm (t) be the “nonnormalized” Carlitz polynomials. They satisfy the main K-binomial identity (see (2.30)) k k qm em (s) {ek−m (t)} , m K m=0

ek (st) =

(4.54)

which holds, for example, for any s, t ∈ Fq [x]. It is known (see Proposition 1.7) that q−1 ek−1 . ek = eqk−1 − Dk−1

(4.55)

Let us rewrite the left-hand side of (4.54) in accordance with (4.55), and apply to each term the identity (4.54) with k − 1 substituted for k. We have k−1 k − 1q q i+1 q ei (s)eqk−i−1 (t). ek−1 (st) = i K i=0 q−1 By (4.55), eqi = ei+1 + Diq−1 ei , eqk−i−1 = ek−i + Dk−i−1 ek−i−1 , whence

eqk−1 (st)

=

q k k−1 j=1

+

k−1 i=0

Note that

j−1

qj ej (s)ek−j (t)

K

k−1 i

q

+

k−1 i=0

q i (q−1)

k−1 i

q

i

Diq−1 ei (s)eqk−i (t) K

i

Diq−1 Dk−i−1 ei (s)eqk−i−1 (t). K

q k−1 q i (q−1) q−1 k − 1 Diq−1 Dk−i−1 = Dk−1 . i i K K

(4.56)

Special functions

157

Indeed, the left-hand side of (4.56) equals q Dk−1

−q q Diq−1 Dk−i−1 = i+1

q i+1 Diq Dk−i−1

Dk−1

i

i

q Di Dk−i−1

q−1 Dk−1

and coincides with the right-hand side. Therefore the last sum in the expression for eqk−1 (st) equals q−1 Dk−1

k−1 i=0

k−1 i

i

q−1 ek−1 (st). ei (s)eqk−i−1 (t) = Dk−1 K

Using (4.55) again we ﬁnd that ek (st) =

q k k−1 i=0

i−1

i ei (s)eqk−i (t)

+

K

q k k−1 i=0

i

i

Diq−1 ei (s)eqk−i (t),

K

and the comparison with (4.54) yields q q k m k k−1 k−1 q−1 em (s)eqk−m (t) = 0 − − Dm m K m−1 K m K m=0 for any s, t. Since the Carlitz polynomials are linearly independent, we obtain that q q m k−1 k−1 k q−1 eqk−m (t) = 0 − − Dm m−1 K m K m K for any t, and it remains to note that ek−m (t) = 0 if t ∈ Fq [x], deg t ≥ k, by the deﬁnition of the Carlitz polynomials. More generally, we have the following Vandermonde-type identity. Let k, m be integers, 0 ≤ m ≤ k. (m)

Proposition 4.9 Deﬁne cli

(m)

∈ K by the recurrence relation (m)

(m)

q−1 cl+1,i = cl,i−1 + cli Dm−i (m)

(4.57) (m)

and the initial conditions cli = 0 for i < 0 and i > l, c00 = 1. Then, for any l ≤ m, ql l k (m) k − l = cli . (4.58) m−i K m K i=0

158

Chapter 4

Proof. The identity (4.58) is trivial for l = 0. Suppose it has been proved for some l. Let us transform the right-hand side of (4.58) using the identity (4.53). Then we have ql+1 ql+1 l l k (m) k − l − 1 (m) k − l − 1 q−1 = cli + cli Dm−i m − i − 1 m − i m K K K i=0 i=0 l+1 l+1 l l+1 q (m) k − l − 1 q (m) k − l − 1 q−1 Dm−i . + cli = cl,j−1 m − i m − j K K i=0 j=1 (m)

(m)

Since we assume that cl,−1 = cl,l+1 = 0, the summation in both the above sums can be performed from 0 to l + 1. Using (4.57) we obtain the required identity (4.58) with l + 1 substituted for l. 4.4.3. A Carlitz diﬀerential equation. Now we consider a function f ∈ F2 (see (3.57)) associated with the K-binomial coeﬃcients, that is ∞ k m k k f (s, t) = sq tq . m K m=0

(4.59)

k=0

Proposition 4.10 The function (4.59) satisﬁes the equation ds f (s, t) = ∆t f (s, t) + [1]1/q f (s, t). Proof. Let us compute ds f . We have ∞ k 1/q m−1 k−1 k [m]1/q sq tq ds f (s, t) = m K k=1 m=1 ν ∞ ν + 11/q µ ν = [µ + 1]1/q sq tq . µ + 1 K ν=0 µ=0 Using Proposition 4.8 we ﬁnd that ds f = Σ1 + Σ2 where ν ∞ µ ν ν [µ + 1]1/q sq tq , Σ1 = µ K ν=0 µ=0 Σ2 =

ν ∞ ν 1−q −1 q µ q ν [µ + 1]1/q Dµ+1 s t . µ+1 K ν=0 µ=0

(4.60)

Special functions

159

Note that µ+1

1/q µ

1/q −x = xq − x + (xq − x) = [µ] + [1]1/q , [µ + 1]1/q = xq so that Σ1 =

∞ ν ν ν=0 µ=0

µ

µ

ν

[µ]sq tq + [1]1/q f (s, t).

(4.61)

K

Next, we have 1/q [µ + 1] Dν ν 1−q −1 [µ + 1]1/q Dµ+1 = D µ+1 q µ+1 Dµ+1 µ+1 K Dµ+1 Dν−µ−1 Dν = , q µ+1 Dµ Dν−µ−1 and also q Dν−µ−1 =

1 Dν−µ q [ν − µ]Dν−µ−1 , = [ν − µ] [ν − µ]

whence µ

q Dν−µ . [ν − µ]qµ

µ+1

q = Dν−µ−1

Therefore Σ2 =

ν ∞ ν ν=0 µ=0

µ

As above, [ν − µ]q = xq Σ2 =

ν−µ

µ

µ

µ

ν

[ν − µ]q sq tq .

K

q µ −x = [ν] − [µ], so that

ν ∞ ν ν=0 µ=0

µ

µ

ν

([ν] − [µ])sq tq .

K

Together with (4.61), this implies (4.60).

4.5 Overconvergence properties 4.5.1. The background. The idea of overconvergence is among the basic ones in contemporary p-adic analysis. In contrast to analysis over R and C, the power series for principal special functions over Qp or Cp converge only on ﬁnite disks or annuli. For example, the exponential series

160 exp(t) =

Chapter 4 ∞ n=0

tn n! ,

t ∈ Cp , converges if and only if |t|p < p−1/(p−1) (see [35]

or [90]). At the same time, for many special functions there exist some expressions combining their values in various points (usually connected by the Frobenius power t → tp ), for which the corresponding power series converge on wider regions. The simplest example is the Dwork exponential θ(t) = exp(π(t − tp ))

(4.62)

where π is a root of the equation z p−1 + p = 0. The power series for θ(t), in p−1

the variable t, converges for |t|p < p p2 (> 1), though the formula (4.62) is not valid outside the unit disk. The special value θ(1) is a primitive p-th root of unity. Other examples involve the exponential function of q-analysis [5], some hypergeometric functions [34], polylogarithms [28], and many others. The overconvergent functions usually satisfy equations possessing special algebraic properties called the Frobenius structures (see [5, 89]). In this section we consider the overconvergence phenomena in the positive characteristic situation, that is for Fq -linear functions on subsets of the ﬁeld K. In particular, using the Carlitz exponential eC we construct an analog of the Dwork exponential and prove its overconvergence, consider the overconvergence problems for some other special functions. These problems are much simpler than those in the characteristic zero case. The reason is that those functions satisfy diﬀerential equations with the Carlitz derivatives; the diﬀerence structure of the latter leads immediately to overconvergence properties of some linear combinations of solutions. 4.5.2. The Dwork–Carlitz exponential. Let σ be an arbitrary solution of the equation z q−1 = −x. Let us consider the function E(t) = eC (σ(t − tq )), 1 − q−1

(4.63)

(recall that we denote by | · | also the deﬁned initially for |t| < q extension of the absolute value from K onto K c ). Note that, in spite of the formal resemblance, the formulas for the Dwork exponential (4.62) and “the Dwork–Carlitz exponential” (4.63) have a quite diﬀerent meaning – the function θ(t) is a multiplicative combination of values of the classical exponential, while E(t) is an additive combination of values of the Carlitz exponential. This diﬀerence from classical overconvergence theory appears also in some other examples given below.

Special functions

161

From the deﬁnition of eC , after a simple transformation we ﬁnd that n

n−1 ∞ n σq σq tq . − (4.64) E(t) = σt + D D n n−1 n=1 In order to investigate the convergence of the series (4.64), we have to study the structure of elements Dn . Proposition 4.11 For any n ≥ 1, q n −1 n−1 n−1 Dn − (−1)n x1+q+···+q ≤ q − q−1 −(q−1)q .

(4.65)

Proof. We will prove that n−1 Dn − (−1)n x1+q+···+q ≤ q −ln

(4.66)

where the sequence {ln } is determined by the requrrence ln = qln−1 + 1,

l1 = q.

(4.67)

Indeed, if n = 1, then D1 = xq − x, so that |D1 + x| = q −q . Suppose that we have proved (4.66) for some value of n. We have 2 n q Dn − (−1)n xq+q ···+q ≤ q −qln , whence

2 n [n + 1]Dnq − (−1)n [n + 1]xq+q ···+q ≤ q −(qln +1) .

Since Dn+1 = [n + 1]Dnq , we ﬁnd that

n 2 n+1 Dn+1 − (−1)n+1 x1+q+···+q − (−1)n xq+q ···+q ≤ q −ln+1 .

(4.68)

It is easy to check that ln =

qn − 1 + (q − 1)q n−1 q−1

(4.69)

satisﬁes (4.67); the expression (4.69) can also be deduced from a general formula for a solution of a diﬀerence equation; see [42]. On the other hand, q n+2 −q 2 n+1 (4.70) (−1)n xq+q ···+q = q − q−1 ,

162

Chapter 4

and, by a simple computation, q n+2 − q − ln+1 = q n − 1 > 0, q−1

n ≥ 1.

(4.71)

It follows from (4.68), (4.70), (4.71), and the ultra-metric property of the absolute value, that n Dn+1 − (−1)n+1 x1+q+···+q ≤ q −ln+1 ,

which proves the inequalities (4.66) and (4.65) for any n. Now we can prove the overconvergence of E(t).

Proposition 4.12 The series in (4.64) converges for |t| < ρ, where ρ = q

q−1 q2

n

σq n→∞ Dn

> 1. In particular, E(1) = lim

Proof. Let us write n

n−1

n−1

σq σq σq − = Dn Dn−1 Dn−1 We have σ q

n

−q n−1

= −xq

n−1

σq

is deﬁned, and E(1) = σ.

n

−q n−1

Dn

Dn−1

−1 .

, so that

n−1

σq σq σq xq Dn−1 + Dn − =− Dn Dn−1 Dn−1 Dn n−1

n−1 n−2 σq xq Dn−1 − (−1)n−1 x1+···+q =− Dn−1 Dn

! n−1 + Dn − (−1)n x1+···+q . n

n−1

n−1

If n ≥ 2, then by Proposition 4.11, n−1

n−1 − q n−1 + q q−1−1 +(q−1)q n−2 q n−1 1+···+q n−2 Dn−1 − (−1) x , x ≤q n −1 q n−1 n−1 Dn − (−1)n x1+···+q ≤ q − q−1 −(q−1)q . Comparing the right-hand sides we check that the ﬁrst of them is bigger; therefore n n−1 σq n−1 n−1 n−1 n −1 σq − q n−1 + q q−1−1 +(q−1)q n−2 − qq−1 − q q−1−1 − qq−1 − ·q ·q ·q , ≤q Dn Dn−1

Special functions so that

n n−1 σq q n−2 (q−1)2 +1 σq q−1 − , ≤ q− Dn Dn−1

For n = 1, we get

163

n ≥ 2.

(4.72)

q q σ = |σ| −x − 1 = |σ| x , − σ [1] [1] D1

whence

q σ 1 −(q−1) ≤ q − q−1 − σ . D1

(4.73)

It follows from (4.72) that the series in (4.64) converges for |t| < ρ. For t = 1, we obtain that n

n−1 n ∞ σq σq σq = lim − . (4.74) E(1) = σ + n→∞ Dn Dn Dn−1 n=1 Note that

qn σ 1 − q−1 Dn = q

for all values of n. Now n+1

n+1

n

σq σq σq = lim [n + 1] = lim [n] = −xE(1) q n→∞ n→∞ n→∞ Dn Dn+1 Dn

E(1)q = lim because

qn qn x σ 1 − q−1 −q n −→ 0, Dn = q

as n → ∞. By (4.74), E(1) = 0, so that E(1)q−1 = −x, thus E(1) satisﬁes the same equation as σ. All the solutions of this equation are obtained by multiplying σ by nonzero elements ξ ∈ Fq . Therefore E(1) = σξ, ξ ∈ Fq , ξ = 0. If ξ = 1, then 1

|E(1) − σ| = |(1 − ξ)σ| = |σ| = q − q−1 .

(4.75)

On the other hand, by (4.74),

n n−1 σq σq |E(1) − σ| ≤ sup − , Dn−1 n≥1 Dn

and we see that (4.75) contradicts (4.72) and (4.73).

164

Chapter 4

It is interesting that the special value σ = E(1), just as the special value of the Dwork exponential in the characteristic 0 case, generates a cyclotomic extension of the function ﬁeld (related in this case to the Carlitz module); see [94]. 4.5.3. Examples of overconvergence. The Carlitz exponential eC satisﬁes the simplest equation deC = eC , so that eC (t)q + xeC (t) = eC (xt).

(4.76)

The right-hand side of (4.76) obviously converges on a wider disk than eC itself (note that, in contrast to the p-adic case, E(t) is not known to satisfy a homogeneous equation with the Carlitz derivative). Similarly, the Bessel–Carlitz function Jn (t) satisﬁes the identity ∆Jn = q Jn−1 , so that q (t) + xJn (t) = Jn (xt), Jn−1

(4.77)

and we have an overconvergence for the right-hand side of (4.77). In this sense equations with the Carlitz derivatives may be seen themselves as analogs of the Frobenius structures of p-adic analysis. The next two examples (of an essentially similar nature) are just a little more complicated. Consider the polylogarithms ln (t) =

j ∞ tq , [j]n j=1

n = 1, 2, . . .

(4.78)

The series in (4.78) converges for |t| < 1. In Section 4.3, we constructed their continuous extensions to the disk {t ∈ K : |t| ≤ 1}. Here we give the following overconvergence result resembling Coleman’s theorem [28] about classical polylogarithms. Proposition 4.13 The power series for the function Ln (t) = ln (t) − ln (tq ) converges for |t| < q 1/q . Proof. By a simple transformation, we get ∞ j 1 tq 1 tq . − Ln (t) = n + n n [1] [j] [j − 1] j=2

(4.79)

Special functions

165

We have,

([j − 1] − [j]) [j]n−1 + [j − 1][j]n−2 + · · · + [j − 1]n−1 1 1 . − = [j]n [j − 1]n [j]n [j − 1]n j−1 j j−1 For any j ≥ 2, |[j]| = q −1 , |[j − 1] − [j]| = xq − xq = q −q , so that 1 1 −q j−1 +n−1 , [j]n − [j − 1]n ≤ q and the convergence radius of the series (4.79) equals q 1/q .

Let us consider the hypergeometric function F (a, b; c; t) =

∞ an bn qn t , cn Dn n=0

(4.80)

/ {[0], [1], . . . , [∞]}, [∞] = −x. For the notation see where a, b, c ∈ K c , c ∈ Section 4.1. If |a| = |b| = |c| = 1,! then the disk of convergence of the series (4.80) is 1

t ∈ K c : |t| < q − q−1 . In this case |T1 (a)| = |T1 (b)| = |T1 (c)| = 1.

Proposition 4.14 The identity ab τ F T1 (a), T1 (b); T1 (c); t − xF (a, b; c; t) = −F (a, b; c; xt) c

(4.81)

holds for any values of the variable and parameters, such that all the terms of (4.81) make sense. In particular, if |a| = |b| = |c| = 1, then the righthand side of (4.81) is overconvergent, that is the series for the right-hand 1 side converges for |t| < q 1− q−1 (> 1). Proof. Changing the index of summation we ﬁnd that τ F (T1 (a), T1 (b); T1 (c); z) =

∞ T1 (a)qn−1 T1 (b)qn−1 qn t q T1 (c)qn−1 Dn−1 n=0

for any z from the convergence disk. Using the identity (4.15) and the fact q that Dn = [n]Dn−1 we get −qn ∞ n an bn [n] ab τ F (T1 (a), T1 (b); T1 (c); z) = − zq cn Dn c n=0 (note that [0] = 0), which implies (4.81).

166

Chapter 4 4.6 Comments

The hypergeometric functions (4.6) were introduced by the author [71], as an application of the equations of evolution type (Section 3.4.2). A decisive motivation was to extend Thakur’s deﬁnition [109, 110] of hypergeometric functions in such a way that the parameters would belong to K c , just as the argument and the values. The Bessel–Carlitz functions and the associated polynomials were deﬁned and studied by Carlitz [24]. See also [103]. The deﬁnitions and results regarding polylogarithms and the zeta function are taken from [67]. In this book we do not touch on the much richer theory of Goss’s zeta function; see [45, 111] and references therein. In [45, 111] a theory of the function ﬁeld gamma function is also developed. The results regarding K-binomial coeﬃcients are taken from [68]. Our exposition of overconvergence phenomena follows [72].

5 The Carlitz rings

In this chapter we consider the Carlitz rings, the rings of diﬀerential operators with Carlitz derivatives. After a description of basic algebraic properties of these rings, we study modules over them, introduce a class of quasiholonomic modules, similar in many respects to the class of holonomic modules of the characteristic zero theory. We show that some special functions of analysis over a local ﬁeld of positive characteristic, as well as partial diﬀerential operators of evolution type, generate quasiholonomic modules.

5.1 Algebraic preliminaries Here we collect some well-known results about noncommutative rings and modules over them, which will be needed in the sequel. The results are given without proofs, which can be found, together with further details, in [30, 79, 10, 12]. 5.1.1. Filtered rings. A ﬁltered ring is a ring R with a family {Rn , n = 0, 1, 2, . . .} of its additive subgroups, such that (i) for each i, j, Ri Rj ⊂ Ri+j ; (ii) for i < j, Ri ⊂ Rj ; ∞ 2 Rn = R. (iii) n=0

The family {Rn } is called a ﬁltration of R. A graded ring is a ring T together with a family {Tn , n = 0, 1, 2, . . .} of its additive subgroups, such that (i) Ti Tj ⊂ Ti+j ; ∞ $ Tn , as an Abelian group. (ii) T = n=0

The family {Tn } is called a grading of T ; a nonzero element of Tn is said 167

168

Chapter 5

to be homogeneous of degree n. If a and b are homogeneous elements and ab = 0, then deg(ab) = deg a + deg b. If T is a graded ring, then T has a natural ﬁltration {Rn } with Rn = T0 ⊕ · · · ⊕ Tn . Conversely, given a ﬁltered ring R, one can construct an associated graded ring T = gr R as follows. Set Tn = Rn /Rn−1 , R−1 = {0}, T = ⊕Tn . To deﬁne multiplication in T , it suﬃces to consider homogeneous elements. If a ∈ Rn \Rn−1 , c ∈ Rm \Rm−1 , we consider a = a+Rn−1 ∈ Tn , c = c + Rm−1 ∈ Tm , and set ac = ac + Rm+n−1 . It is easy to check that this multiplication is well-deﬁned and makes T a ring.

Proposition 5.1 If R is a ﬁltered ring, and gr R is right (left) Noetherian, then R is right (left) Noetherian. Let R be a ﬁnitely generated extension of a subring R0 with generators x1 , . . . , xn . An element of R of the form r0 xi1 r1 xi2 r2 · · · xil rl with ri ∈ R0 is called a word of length l. The standard ﬁltration on R is obtained if Rj (j ≥ 1) is an additive subgroup of R generated by all words of length ≤ j. A ﬁnitely generated extension R of a subring R0 with generators x1 , . . . , xn ∈ / R0 is called an almost normalizing extension, if for any i, j (i) R0 xi + R0 = xi R0 + R0 ; n (ii) xi xj − xj xi ∈ xl R0 + R0 . l=1

Proposition 5.2 Let R be an almost normalizing extension of a right (left) Noetherian ring R0 , with the standard ﬁltration. Then both R and gr R are right (left) Noetherian rings.

5.1.2. Filtered modules. Let M be a left module over a ﬁltered ring R with a ﬁltration {Rn }. A ﬁltration in M is a system of subgroups Mj , j = 0, 1, 2, . . ., such that (i) for each i, j, Rj Mi ⊂ Mi+j (as usual, we denote by Rj Mi the set of all ﬁnite sums rl ml , rl ∈ Rj , mj ∈ Mi ); l

(ii) for each i < j, Mi ⊂ Mj ; ∞ 2 (iii) Mn = M . n=0

A module with a ﬁltration is called a ﬁltered module.

The Carlitz rings

169

The ﬁltration of a module is not unique. In particular, given any submodule M0 of M which generates M , one can deﬁne a standard ﬁltration setting Mn = Rn M0 . ∞ $ Tn be a graded ring. A grading of a left T -module M is an Let T = n=0

Abelian group decomposition M =

∞ $

Mn , such that Tj Mi ⊂ Mi+j for

n=0

any i, j. A graded module is a module together with a ﬁxed grading. The nonzero elements of the subgroup Mn are called homogeneous of ∞ mn with mn ∈ Mn , then mn is the n-th homogedegree n, and if m = n=0

neous component of m. A graded submodule N of M is a submodule with ∞ $ a grading N = Nn , such that Nn ⊂ Mn . A graded homomorphism θ n=0

is a T -homomorphism M → M between two graded T -modules, such that θ(Mn ) ⊂ Mn for each n. Similarly, if M, M are ﬁltered R-modules, with the ﬁltrations {Mn }, {Mn }, then a ﬁltered homomorphism θ is a R-homomorphism θ : M → M such that θ(Mn ) ⊂ Mn for each n. For a ﬁltered module M over a ﬁltered ring R, we can deﬁne a graded module gr M over the graded ring gr R, as follows. Let (gr M )n = Mn /Mn−1 ∞ $ (M−1 = {0}), gr M = (gr M )n . The multiplication is deﬁned, via n=0

the distributive law, by the multiplication of representatives of the residue classes. Classes of ﬁltered modules over a given ﬁltered ring and graded modules over a given graded ring, with appropriate classes of homomorphisms, form categories. Then the correspondence M → gr M is a functor between these categories. If M, N are ﬁltered R-modules, θ : M → N is a ﬁltered homomorphism, then θ(Mj ) ⊂ θ(M ) ∩ Nj . If θ(Mj ) = θ(M ) ∩ Nj for each j, then θ is called strict. θ

ϕ

Proposition 5.3 Let R be a ﬁltered ring, and let L −→ M −→ N be an exact sequence of ﬁltered modules and ﬁltered homomorphisms. Then gr θ gr ϕ gr L −→ gr M −→ gr N is an exact sequence if and only if θ and ϕ are strict.

Corollary 5.4 Let R be a ﬁltered ring, and ϕ :

M → N a ﬁltered

170

Chapter 5

homomorphism. Then gr ϕ is injective (surjective) if and only if ϕ is injective (resp. surjective) and ϕ is strict. A ﬁltration {Mn } in a module M over a ﬁltered ring is called good if the graded module gr M is ﬁnitely generated. Proposition 5.5 (i) If a ﬁltered module M has a good ﬁltration, then M is ﬁnitely generated. (ii) A standard ﬁltration in a ﬁnitely generated module over a ﬁltered ring is good. Let M be a ﬁltered R-module, and N be a submodule in M . Let us construct and study ﬁltrations in N and M/N . A ﬁltration in N can be deﬁned as Nk = N ∩ Mk . In M/N , we consider the subgroups Fk = Mk /(Mk ∩N ), the images of Mk in M/N . The sequence {Fk , k = 0, 1, 2, . . .} is a ﬁltration in M/N , and Fk /Fk−1 ∼ = Mk /(Mk−1 + Mk ∩ N ). This deﬁnes the canonical projection πk : Mk /Mk−1 −→ Fk /Fk−1 . On the other hand, the imbedding N ⊂ M generates a monomorphism ϕk : (N ∩ Mk )/(N ∩ Mk−1 ) −→ Mk /Mk−1 . Passing to the maps of the associated graded modules and using Proposition 5.3 we obtain the exact sequence ϕ

π

0 −→ gr N −→ gr M −→ gr M/N −→ 0

(5.1)

where ϕ = ⊕ϕk , π = ⊕πk . Suppose that R is a Noetherian ring and M is a ﬁltered R-module with a good ﬁltration. Then gr M is a ﬁnitely generated module over the ring gr R, which is Noetherian by Proposition 5.1. Therefore gr M is a Noetherian module (see [75]). Due to the exact sequence (5.1), gr N is isomorphic to a submodule of gr M , while gr M/N is isomorphic to a quotient module of gr M . Thus, the modules gr N and gr M/N are Noetherian, so that the above ﬁltrations in N and M/N (called the induced ﬁltrations) are good. 5.1.3. Generalized Weyl algebras (GWA). We will consider only a

The Carlitz rings

171

GWA of degree 1 [10, 11, 12]. Let D be a ring with an automorphism σ and a central element a. A GWA A = D(σ, a) is the ring generated by D and two indeterminates X, Y , with the relations Xλ = σ(λ)X, Y λ = σ −1 (λ)Y (for all λ ∈ D), Y X = a, XY = σ(a). We have A =

∞ $ n=−∞

An where An = Dvn ,  n  if n > 0;  X , vn = 1, if n = 0;   Y −n , if n < 0.

It follows from the above relations that vn vm = (n, m)vn+m = vn+m n, m for some (n, m) = σ −n−m (n, m) ∈ D. If n > 0 and m > 0, then for n≥m (n, −m) = σ n (a) · · · σ n−m+1 (a),

(−n, m) = σ −n+1 (a) · · · σ −n+m (a).

For n ≤ m, (n, −m) = σ n (a) · · · σ(a),

(−n, m) = σ −n+1 (a) · · · a.

In other cases (n, m) = 1. If D is a Noetherian domain, then so is A too. The ﬁrst and basic example of a GWA was the ﬁrst Weyl algebra of ordinary diﬀerential operators (in a variable ξ) with polynomial coeﬃcients over C. In this case D = C[H], σ : H → H − 1, a = H, X is the operator of multiplication by ξ, Y = d/dξ, H = Y X. Below, the ring of diﬀerential operators with Carlitz derivatives (in one variable) will be interpreted as a GWA. Suppose that the ring D is commutative. The cyclic group G, generated by the automorphism σ, acts on the set Specm(D) of maximal ideals of D. An orbit O is cyclic of length n (respectively, linear) if it contains a ﬁnite (respectively, inﬁnite) number n = card O of elements. The set of all cyclic (linear) orbits is denoted by Cyc (resp. Lin). An A-module M (below we consider left modules) is called a weight module if V is semisimple as a D-module, so that 0 M= Mp p∈Specm(D)

172

Chapter 5

where Mp (the component of M of weght p) is the sum of simple Dsubmodules isomorphic to D/p. The support Supp(M ) of the weight module M is the set of maximal ideals p, such that Mp = {0}. Each weight A-module M is decomposed into the direct sum of Asubmodules 0 M= {MO : O is an orbit} , $ {Mp : p ∈ O}. Hence, the support of a simple weight where MO = ( module belongs exactly to one orbit. Now the set A(weight) of isomorphism ( classes of simple weight A-modules consists of the sets A(weight, linear) and ( A(weight, cyclic) of isomorphism classes of simple weight A-modules with support from a linear resp. cyclic orbit. An orbit O is called degenerate if it contains an ideal p (a marked ideal) such that a ∈ p. Denote by Linn (Cycn) the set of all nondegenerate linear (cyclic) orbits. Each linear orbit O(p), p ∈ O may be identiﬁed with Z via the mapping i σ (p) → i. Therefore the notation used for Z (like an interval, semiaxis, i j etc) may be "used for O(p). # For example, σ (p) ≤ σ (p) if and only if i ≤ j; i (−∞, p] = σ (p), i ≤ 0 . Marked ideals p1 < . . . < ps of a degenerate linear orbit O divide it into s + 1 parts, Γ1 = (−∞, p1 ], Γ2 = (p1 , p2 ], . . . , Γs+1 = (ps , ∞). For maximal ideals belonging to linear orbits (we write p ∈ Specm. lin(D)), we introduce the equivalence relation: p ∼ q if and only if p, q ∈ Supp M ( for some isomorphism class [M ] ∈ A(weight, linear). It can be proved that p ∼ q if and only if p and q both belong either to a nondegenerate linear orbit or to some Γi . Theorem 5.6 The mapping ( Specm. lin(D)/ ∼ −→ A(weight, linear),

Γ → [L(Γ)],

is bijective, with the inverse [L] → Supp L. Here (i) if Γ ∈ Linn is a nondegenerate orbit, then L(Γ) = A/Ap, p ∈ Γ; (ii) if Γ = (−∞, p], then L(Γ) = A/A(p, X); (iii) if Γ = (σ −n (p), p], n ∈ N, then L(Γ) = A/A(p, X, Y n ); (iv) if Γ = (p, ∞), then L(Γ) = A/A(σ(p), Y ). ( See [12] for a description of A(weight, cyclic).

The Carlitz rings

173

Let D be a Dedekind domain (for example, a principal ideal domain [6]), M be an A-module. Let tor(M ) = {m ∈ M : am = 0 for some 0 = a ∈ D} be the D-torsion submodule of M . Every simple A-module is either Dtorsion (M = tor(M )) or D-torsion-free (tor(M ) = {0}). Moreover, M is a weight module if and only if it is D-torsion. Theorem 5.7 All simple A-modules are D-torsion if and only if there are iniﬁnitely many cyclic orbits.

5.2 The Carlitz rings 5.2.1. Deﬁnitions and general properties. As in Chapter 3 (see (3.57)) we denote by Fn+1 the set of all germs of functions of the form f (s, t1 , . . . , tn ) =

∞ k1 =0

...

∞

min(k1 ,...,kn )

kn =0

m=0

m

k1

am,k1 ,...,kn sq tq1

kn

. . . tqn

(5.2)

where am,k1 ,...,kn ∈ K c are such that all the series are convergent on some neighborhoods of the origin. We do not exclude the case n = 0 where F1 m am sq convergent on a will mean the set of all Fq -linear power series m

neighborhood of the origin. F(n+1 will denote the set of all polynomials from Fn+1 , that is the series (5.2) in which only a ﬁnite number of coeﬃcients is diﬀerent from zero. The ring An+1 is generated by the operators τ, d = ds , ∆t1 , . . . , ∆tn on Fn+1 , and the operators of multiplication by scalars from K c . To simplify the notation, we will write ∆j instead of ∆tj and identify a scalar λ ∈ K c with the operator of multiplication by λ. The operators ∆j are K c -linear, so that ∆j λ = λ∆j ,

λ ∈ K c,

(5.3)

while the operators τ, ds satisfy the commutation relations dτ − τ d = [1]1/q , τ λ = λq τ, dλ = λ1/q d (λ ∈ K c ).

(5.4)

In the action of each operator ds , ∆j (acting in a single variable), other variables are treated as scalars. The operator τ acts simultaneously on all the variables and coeﬃcients, so that q m+1 q k1 +1 kn +1 τf = am,k1 ,...,kn sq t1 . . . tqn .

174

Chapter 5

An+1 is a Fq -algebra, but is not a K-algebra. We have k [k]tqj , if k ≥ 1; qk ∆j tj = 0, if k = 0;

(5.5)

the second equality can be included in the ﬁrst one, if we set [0] = 0. Similarly ds sq

m

= [m]1/q sq

m−1

,

m ≥ 0.

(5.6)

Since |[m]| = q −1 for any m ≥ 1, the action of operators from An+1 does not spoil convergence of the series (5.2). The identity [k + 1] − [k]q = [1], together with (5.5) and (5.6), implies the commutation relations ∆j τ − τ ∆j = [1]τ,

ds ∆j − ∆j ds = [1]1/q ds ,

j = 1, . . . , n,

(5.7)

veriﬁed by applying both sides of each equality to an arbitrary monomial. As we saw in Section 3.4.1, operators from An+1 possess properties resembling those of a class of partial diﬀerential operators of classical analysis. In many respects, the rings An+1 may be seen as the function ﬁeld counterparts of the Weyl algebras of polynomial diﬀerential operators [16, 30]. Using the commutation relations (5.3), (5.4), and (5.7), we can write any element a ∈ An+1 as a ﬁnite sum (5.8) a= cl,µ,i1 ,...,in τ l dµs ∆i11 . . . ∆inn .

Proposition 5.8 The representation (5.8) of an element a ∈ An+1 is unique. Proof. Suppose that

cl,µ,i1 ,...,in τ l dµs ∆i11 . . . ∆inn = 0.

(5.9)

l,µ,i1 ,...,in k1

kn

Applying the left-hand side of (5.9) to the function stq1 . . . tqn with k1 , . . . , kn > 0 we ﬁnd that   k1 +l l l l kn +l  cl,0,i1 ,...,in [k1 ]i1 q . . . [kn ]in q  sq tq1 . . . tqn =0 l

i1 ,...,in

The Carlitz rings whence

l

175 l

cl,0,i1 ,...,in [k1 ]i1 q . . . [kn ]in q = 0

i1 ,...,in

for each l. Writing this in the form ρ(in )y in = 0

(5.10)

in

where ρ(in ) =

l

l

cl,0,i1 ,...,in [k1 ]i1 q . . . [kn−1 ]in−1 q ,

l

y = [kn ]q ,

i1 ,...,in−1

and taking into account that (5.10) holds for arbitrary kn ≥ 1, that is for an inﬁnite set of values of y, we ﬁnd that ρ(in ) = 0. Repeating this reasoning we get the equality cl,0,i1 ,...,in = 0 for all l, 0, i1 , . . . , in . Suppose that cl,µ,i1 ,...,in = 0 for µ ≤ µ0 and arbitrary l, i1 , . . . , in . µ0 +1

k1

kn

tq1 . . . tqn Then we apply the left-hand side of (5.9) to the function sq and proceed as before coming to the equality cl,µ0 +1,i1 ,...,in = 0 for all l, i1 , . . . , in . The basic algebraic property of the ring An+1 is given by the following result. Theorem 5.9 An+1 is a Noetherian domain. Proof. It is clear that An+1 is an almost normalizing extension of the ﬁeld K c , so that, by Proposition 5.2, it is a Noetherian ring. Let us prove that An+1 has no zero-divisors. We begin with the case n = 0. For brevity, we write d instead of ds . Let a, b ∈ A1 , ab = 0, a=

m1 n1

λij τ i dj ,

b=

i=0 j=0

n2 m2

µkl τ k dl ,

k=0 l=0

and a = 0, b = 0, that is m1

λin1 τ = 0, i

i=0

m2

µkn2 τ k = 0.

(5.11)

k=0

It is easily proved (by induction) that dτ i − τ i d = [i]1/q τ i−1 ,

−j

dj τ − τ dj = [j]q dj−1 ,

(5.12)

176

Chapter 5

for any natural numbers i, j. It follows from (5.12) that dj τ k = τ k dj + O(dj−1 ) where O(dj−1 ) means a polynomial in the variable d, of degree ≤ j − 1, with coeﬃcients from the composition ring K c {τ } of polynomials in the operator τ . Therefore the coeﬃcient of dn1 +n2 in the expression for the operator ab equals 1 i+k λin1 µi−n = P1 (P2 (τ )) kn2 τ i,k

where P1 (τ ) =

m1

λin1 τ i ,

P2 (τ ) =

i=0

m2

−n1

µqkn2 τ k ,

k=0

which contradicts (5.11), since the ring K c {τ } has no zero-divisors [45]. Suppose that it has been proved that An+1 has no zero-divisors for some n. Let us prove that An+2 has no zero-divisors. Let a, b ∈ An+2 , a = 0, b = 0, and ab = 0. Let us write the representations (5.8) for the elements a and b in the form of (noncommutative) polynomials a=

k i=0

ai ∆in+1 ,

b=

l

bj ∆jn+1 ,

j=0

with coeﬃcients from An+1 , such that ak = 0, bl = 0. It follows from the commutation relations (5.3) and (5.7) that the degrees of polynomials in ∆n+1 obtained after multiplication of monomials ai ∆in+1 and bj ∆jn+1 cannot exceed i + j. In particular, ∆kn+1 bl = bl ∆kn+1 + O ∆k−1 n+1 (here the symbol O means a polynomial with coeﬃcients from An+1 ), so that k+l−1 ak ∆kn+1 bl ∆ln+1 = ak bl ∆k+l . n+1 + O ∆n+1 Now the assumption ab = 0 means, since An+1 has no zero-divisors, that ak = 0 or bl = 0, and we come to a contradiction. 5.2.2. Filtration in An+1 . Let us introduce a ﬁltration in An+1 denoting by Γν , ν ∈ Z+ , the K c -vector space of operators (5.8) with

The Carlitz rings

177

max{l + µ + i1 + · · · + in } ≤ ν where the maximum is taken over all the terms contained in the representation (5.8). It is clear that An+1 is a ﬁltered ring. Setting T0 = K c , Tν = Γν /Γν−1 , ν ≥ 1, we introduce the associated graded ring ∞ 0 gr(An+1 ) = Tν . ν=0

¯ 1, . . . , ∆ ¯ n ∈ T1 of It is generated by scalars λ ∈ T0 and the images τ¯, d¯s , ∆ the elements τ, ds , ∆1 , . . . , ∆n ∈ Γ1 respectively, which satisfy, by virtue of (5.3), (5.4), and (5.7), the relations d¯s τ¯ − τ¯d¯s = 0, τ¯λ = λq τ¯, d¯s λ = λ1/q d¯s , ¯ j λ = λ∆ ¯ j (j = 1, . . . , n). ¯ j d¯s = 0, ∆ ¯ j τ¯ − τ¯∆ ¯ j = 0, ∆ ¯j − ∆ d¯s ∆ By Proposition 5.2, gr(An+1 ) is Noetherian. Let us compute the dimension of the K c -vector space Γν . Note that dim Γν = dim

ν 0

Tj ,

j=1

coincides with the dimension of so that dim Γν appropriate space appearing in the natural ﬁltration in gr(An+1 ). Lemma 5.10 For any ν ∈ N,

the

ν+n+2 . n+2

dim Γν =

Proof. The number dim Γν coincides with the number of nonnegative integral solutions (l, µ, i1 , . . . , in ) of the inequality l + µ + i1 + · · · + in ≤ ν, so that ν N (j, n + 2) dim Γν = j=0

where N (j, k) is the number of diﬀerent representations of j as sums of k nonnegative integers. It is known (Proposition 6.1 in [74]) that N (j, k) = j+k−1 . Then (see Section 1.3 from [88]) k−1 ν ν ν+n+2 j+n+1 ν+n+1−i , = = dim Γν = n+2 n+1 n+1 j=0 i=0

178

Chapter 5

as desired.

5.3 The ring A1

5.3.1. A1 as a GWA. Let us consider a GWA A = D(σ, H) where D = K c [H], σ is the Fq -linear automorphism of K c [H] deﬁned by the assignments: σ : H → H − λ1 , λ → λq (λ ∈ K c ), λ1 = [1]1/q = x − x1/q ; see Section 5.1.3 for the notation related to a GWA. It is easy to check that the mapping A1 → D(σ, H), τ → X, d → Y, dτ → H, λ → λ (λ ∈ K c ), is an Fq -algebra isomorphism. Thus A1 can be identiﬁed with the above GWA. Let G = σ be the subgroup of the group of ring automorphisms Aut(D) of D generated by the element σ. The group G acts in the obvious way on the set of maximal ideals of the algebra D, Specm(D) = {D(H − λ)| λ ∈ K c} ∼ = K c , D(H − λ) ↔ λ. The orbit O of an element p = D(H − λ) ∈ Specm(D) for some λ ∈ K c is determined by the elements n−1

σ n (H − λ) = H − λ1 − λq1 − · · · − λq1 1

σ −n (H − λ) = H + λ1 + λ1q + · · · + λ1q

n

− λq ,

1 n−1

1

− λ qn ,

(5.13) (5.14)

n ≥ 1. Note that, for n ≥ 2, n

n

λ1 + λq1 + · · · + λq1 = −x1/q + xq , 1

1 n

1

λ1 + λ1q + · · · + λ1q = −x qn+1 + x.

Lemma 5.11 Let λ ∈ K c , n ∈ N. Then σ n (D(H − λ)) = D(H − λ) if and only if λ ∈ −x1/q + Fqn . Proof. If n = 1, then σ(D(H − λ)) = D(H − λ) if and only if g1 (λ) = 0 where g1 (λ) = λq − λ + λ1 = λq − λ − x1/q + x. The polynomial g1 of degree q has q distinct roots, since its derivative equals −1. Since g1 (λ + γ) = g1 (λ) + γ q − γ, we ﬁnd that for a root λ of g1 , λ + γ is a root for any γ ∈ Fq .

The Carlitz rings

179

Noticing that −x1/q is a root we obtain that −x1/q + Fq is the set of all roots. Similarly, if n ≥ 2 then, by (5.13), σ n (D(H − λ)) = D(H − λ) if and only if gn (λ) = 0, n−1

n

gn (λ) = λq − λ + λ1 + λq1 + · · · + λq1

n

= λq − λ − x1/q + xq

As above, we ﬁnd that −x1/q + Fqn are the roots of gn .

n−1

.

Below we will need the M¨ obius function µ : N → {0, ±1} given by the rule: µ(1) = 1, µ(p1 · · · pr ) = (−1)r if p1 , . . . , pr are distinct primes, and µ(n) = 0 otherwise. The M¨obius inversion formula (Theorem 3.24 in [76]) is as follows. Let f be a function on N with values in an Abelian group, and g(n) = f (d). Then d|n

f (n) =

µ(d)g

d|n

n d

.

(5.15)

The Euler function ϕ on N is deﬁned as ϕ(1) = 1 and ϕ(n) being the number of natural numbers m that are coprime to n and 1 ≤ m < n. The next result gives a classiﬁcation of ﬁnite orbits in Specm(D). Theorem 5.12 Let O = Oλ (λ ∈ K c ) be the orbit of a maximal ideal D(H − λ) of the polynomial algebra D = K c [H] under the action of the cyclic group G. Then: (i) The orbit Oλ consists of a single element if and only if λ ∈ −x1/q +Fq . Thus there are exactly q distinct maximal σ-invariant ideals of the algebra D. (ii) The orbit Oλ contains exactly n ≥ 2 elements, if and only if   3 λ ∈ −x1/q + Fqn \ Fq m  . m|n,m=n

Thus there are exactly ln = n−1

µ(d)q d = n−1 ϕ(q n − 1) n

d|n

distinct orbits, each of them contains exactly n ≥ 2 elements, and such that 1 > 0. ln ≥ n−1 q n 1 − q−1

180

Chapter 5

Proof. The statement (i) follows immediately from Lemma 5.11. By the same lemma, the orbit Oλ contains exactly n ≥ 2 elements if and only if   3 λ ∈ −x1/q + Fqn \ Fq m  . m|n,m=n

Let ln be the number of distinct orbits containing exactly n elements. Then f (n) = nln is the number of all maximal ideals of D belonging to the above orbits. It follows from Lemma 5.11 that the function g(n) = f (d) is equal d|n 1/q to card −x + Fqn = card (Fqn ) = q n . Therefore, by (5.15), we have n µ(d)q d , f (n) = d|n

n whence ln = n−1 µ(d)q d . Next, d|n

nln ≥ card −x1/q + Fqn − card −x1/q + Fqn−1

− · · · − card −x1/q + Fq = q n − q n−1 − · · · − q 1 q n−1 − 1 qn n n n =q 1− > 0. >q − =q −q q−1 q−1 q−1

2 Fqm = n−1 ϕ(q n − 1), because Clearly, ln = n−1 card Fqn \ m|n,m=n

ϕ(q n − 1) equals the number of primitive elements of Fqn [76].

5.3.2. Simple A1 -modules. The above description of cyclic orbits, together with the general results of Section 5.1.3, leads to a classiﬁcation of simple A1 -modules. First of all, note that, by Theorems 5.7 and 5.12, all simple A1 -modules are weight. ( 1 (linear) and A ( 1 (cyclic) the sets of isomorphism classes of Denote by A simple A1 -modules with support from a linear and a cyclic orbit respectively. The ideal (H) of the polynomial algebra D = K c [H] generated by the element H is a maximal ideal of D. By (5.13), we see that the orbit n O(H) of the maximal ideal (H) is an inﬁnite orbit (since λq1 = λ1 for all n ≥ 1). This linear orbit is the only degenerate one. Therefore, in terms of the equivalence introduced in Section 5.1.3, there are only two equivalence classes in O(H): Γ− = (−∞, (H)] and Γ+ = ((H), ∞). Correspondingly,

The Carlitz rings

181

the set of equivalence classes in Specm. lin(D) consists of Γ− , Γ+ , and the nondegenerate linear orbits. Now Theorem 5.6 specializes to the following description.

Theorem 5.13 The mapping ( 1 (linear), Specm. lin(D)/ ∼ −→ A

Γ → [L(Γ)],

is a bijection, with inverse [L] → Supp L (in particular, Supp L(Γ) = Γ) where (i) if Γ ∈ Linn, then L(Γ) = A1 /A1 p, for any p ∈ Γ; (ii) if Γ = Γ− = (−∞, (H)], then L(Γ− ) = A1 /(A1 H + A1 X); (iii) if Γ = Γ+ = ((H), ∞), then L(Γ+ ) = A1 /(A1 σ(H) + A1 Y ). All the modules listed in Theorem 5.13 are inﬁnite-dimensional as vector spaces over K c . They can be written in an explicit form as follows.

$ $ i i (i) L = A1 /A1 p = ⊕ K c1 ⊕ K cY K cX i≥1

i≥1 0

0

where u = u + A1 p, 1 = X = Y , p = (H − λ) ∈ Γ ∈ Linn, and the action of the generators of A1 (as a GWA) is given by the rules: i

i+1

i

i+1

, i ≥ 0; XX = X , Y Y = Y

i−1 i q q i−1 XY = λ − λ1 − λ1 − · · · − λ1 , i ≥ 1; Y 1 1 i i−1 i−1 Y X = λ + λ1 + λ1q + · · · + λ1q X , i ≥ 2; Y X = λ1. (ii) L− =

$

i

i

K c Y , Y = Y i + A1 H + A1 X,

i≥0

with the action 0

HY = 0, 0

XY = 0,

i i−1 i HY = − λ1 + λq1 + · · · + λq1 Y , i ≥ 1;

i−1 i i−1 XY = − λ1 + λq1 + · · · + λq1 Y , i ≥ 1; i

YY =Y

i+1

,

i ≥ 0.

182

Chapter 5

(iii) L+ =

$

i

i

K c X , X = X i + A1 σ(H) + A1 Y ,

i≥0

with the action 0

0

1

1

i

HX = λ1 X , HX = 2λ1 X , HX = i ≥ 2; 0

0

1

i

1 q i−1

1 q

2λ1 + λ1 + · · · + λ1

1 q i−2

1 q

2λ1 + λ1 + · · · + λ1

Y X = 0, Y X = λ1 X , Y X =

X

i

X ,

i−1

,

i ≥ 2; i

XX = X

i+1

,

i ≥ 0.

For a description of all the simple cyclic A1 -modules see [11]. It is very important that, in contrast to the case of the Weyl algebra of polynomial diﬀerential operators over a ﬁeld of characteristic zero, there exist ﬁnitedimensional A1 -modules. A simple construction is given in the proof of the following theorem. Theorem 5.14 For any k = 1, 2, . . ., there exists a non-trivial A1 -module M whose dimension, as of a vector space over K c , equals k. Proof. Let M = (K c )k . Denote by e1 , . . . , ek the standard basis in M , that is ej = (0, . . . , 0, 1, 0, . . . , 0), with 1 at the j-th place. Let (λij ) be a k × k matrix over K c , such that λij ∈ Fq if i = j, while the diagonal elements satisfy the equation λq − λ + [1]1/q = 0. We deﬁne the action of τ and ds on M as follows: τ (cej ) = cq ej ; d(ej ) =

n

c ∈ K c , j = 1, . . . , k,

λij ei ; d(cej ) = c1/q ej ,

i=1

with subsequent additive continuation onto M . k If x = cj ej , cj ∈ K c , then we have j=1

τ d(x) =

k j=1

cj

n i=1

λqij ei ,

dτ (x) =

k j=1

cj

n i=1

λij ei ,

The Carlitz rings

183

so that dτ (x) − τ d(x) = [1]1/q x, and we indeed have an A1 -module.

5.3.3. The center and ideals of A1 . Let K c (H) = S −1 D, S = D\{0}, be the ﬁeld of fractions of D. The localization B = S −1 A1 at S is the skew Laurent polynomial ring B = K c (H)[X, X −1 , σ], that is (see [79]) the ring of polynomials over K c (H) in X and X −1 subject to the relation aX = Xσ(a), a ∈ K c (H). The ring A1 can be identiﬁed with a subring of B via the ring monomorphism A1 → B, X → X, Y → HX −1 , δ → δ (δ ∈ D). The ring B is a Euclidean ring (the left and right division with remainder algorithms hold), hence a principal left and right ideal domain [79]. Returning to the cyclic orbits on Specm(D), denote by Cycn the set of all the ﬁnite orbits containing exactly n elements. The quantity of such orbits was counted in Theorem 5.12. For each orbit O ∈ Cycn and any λ, such that D(H − λ) ∈ O, let αO =

n−1

σ i (H − λ).

i=0

It follows from (5.13) that αO is a central element of A1 (since σ n (H −λ) = H − gn (λ) − λ = H − λ; see the proof of Lemma 5.11). The next result describes the center of the ring A1 and its localization B. Proposition 5.15 (i) For each n ∈ N, the skew Laurent extension K c [t, t−1 ; σ n ] is a simple Fq -algebra with center Fqn . (ii) The ring B = S −1 A1 = K c (H)[X, X −1 ; σ] is a simple ring with center n(O) αO : n(O) ∈ Z, Z(B) = Fq O∈Cyc

and n(O) = 0 for all but ﬁnitely many O

which containes countably many elements.

184

Chapter 5

(iii) The center Z(A1 ) =

Fq

O∈Cyc

n(O)

αO

: n(O) ≥ 0, and all but ﬁnitely many n(O) = 0

contains countably many elements. Proof. (i) Let Z be the center of K c [t, t−1 ; σ n ]. First we prove that Z ⊂ K c . Indeed, otherwise we would have a nonzero central element z = λtm + z where 0 = m ∈ Z, 0 = λ ∈ K c , and z contains elements of strictly higher or strictly lower degrees in t. For any µ ∈ K c \ Fqn|m| , we have µz − zµ = (µ − σ nm (µ)) λtm + z = 0 (z does not contain terms with tm ), and we come to a contradiction. Next, an element λ ∈ K c commutes with t if and only if λ ∈ Fqn . Therefore Z = Fqn . The fact that K c [t, t−1 ; σ n ] is a simple ring follows from a general result about skew Laurent polynomial rings (Theorem 1.8.5 in [79]). (ii) Similarly, B is a simple ring. The above reasoning leads to the n(O) inclusion Z(B) ⊂ K c (H). Clearly, Fq αO ⊂ Z(B). We have to O∈Cyc

prove that any nonzero element z ∈ Z(B) can be written in this form. The rational function z is equal to γ fg where f, g ∈ Fq [H] are coprime monic polynomials, and γ ∈ K c . If z = γ, then γ = XγX −1 = σ(γ), so that γ ∈ Fq , as desired. Let z = γ. Since z = X n zX −n = σ n (z) for all n ∈ Z, and since f and g are coprime, n(O) we see that f and g are equal to ﬁnite products of the form αO with n(O) ≥ 0. Then also γ ∈ Z(B), whence γ ∈ Fq . (iii) We have Z(A1 ) ⊂ Z(B), and the required result follows from the preceding one. The main properties of ideals of A1 are summarized in the following theorem. Theorem 5.16 (i) Every nonzero prime ideal of A1 is a maximal ideal. (ii) Each nonzero ideal of A1 is a unique ﬁnite product of maximal ideals. All ideals commute.

The Carlitz rings

185

(iii) The mapping Cyc → Specm(A1 ), def O → mO = A1 p = A1 αO , p∈O

is a bijection with inverse m → Supp(A1 /m). For the proof and further details regarding the ideal structure and other algebraic properties of A1 see [11].

5.4 Quasi-holonomic modules 5.4.1. Dimensions. Let M be a ﬁltered left module with a good ﬁltration {Mν } over the Carlitz ring An+1 (the ﬁltration in An+1 was introduced in Section 5.2.2). Below we always assume that the submodules Mν are ﬁnitedimensional as vector spaces over K c . Let us consider the associated graded module gr M =

∞ 0

(gr M )ν ,

(gr M )ν = Mν /Mν−1 (M−1 = {0}).

ν=0

As we know, the good ﬁltration assumption means that gr M is ﬁnitely generated. The next results show that some of the basic properties of ﬁltered and graded modules over An+1 resemble those for modules over the Weyl algebras [16, 30], though a contrasting phenomenon found in Theorem 5.14 will lead to some diﬀerent properties too. It will be convenient to consider, together with the rings An+1 , n ≥ 0, the ring A0 generated by τ and the scalars λ ∈ K c , and the modules over A0 . In the proof of the next theorem we will use an auxiliary result (see [30]). A polynomial p ∈ Q[t] is called numerical if it takes values from Z at all suﬃciently big natural numbers. Lemma 5.17 (i) Let f : Z → Z be a function such that f (l)−f (l−1) = γ(l) for l ≥ l0 > 0, where γ(l) is a numerical polynomial. Then there exists a number l1 > 0 such that f coincides, for l ≥ l1 , with some numerical polynomial. (ii) Let γ(t) be a numerical polynomial. Then there exist integers

186

Chapter 5

c0 , . . . , ck such that

t ck−i γ(t) = i i=0 k

where

t t(t − 1) · · · (t − i + 1) = . In particular, γ(ν) ∈ Z for all ν ∈ Z. i i!

Theorem 5.18 There exist a polynomial χ ∈ Q[t], and a positive number σ such that s dim(Mi /Mi−1 ) = χ(s) for s ≥ σ. i=0

Here dim means the dimension over K c . Proof. First we consider the case n = −1, that is ﬁltered and graded A0 -modules. Let F0 : Z → Z, F0 (s) =

s

dim Mi

i=−∞

where Mi = Mi /Mi−1 , as i ≥ 0, Mi = {0}, as i < 0. The graded ring gr(A0 ) is generated (compare Section 5.2.2) by scalars λ ∈ K c and the element τ . Let ϕi−1 : Mi−1 → Mi , ϕi (m) = τ (m), m ∈ Mi−1 . Denote Qi−1 = ker ϕi , Li = Mi / Im ϕi−1 and consider the exact sequence ϕi−1

0 −→ Qi−1 −→ Mi−1 −→ Mi −→ Li −→ 0

(5.16)

of vector spaces over K c . The mappings ϕi are semilinear with respect to τ considered as an automorphism of the algebraically closed ﬁeld K c (see [18]). As explained in Appendix 1 to Chapter 2 of [18], basic notions of linear algebra remain valid for semilinear mappings – a semilinear mapping of a vector space into itself can be interpreted as a linear mapping between two diﬀerent vector spaces, and, for instance, dimensions of the kernel and cokernel are not changed in this interpretation. Now we ﬁnd from (5.16) that dim Mi−1 = dim Qi−1 + dim (Im ϕi−1 ) , dim Mi = dim (Im ϕi−1 ) + dim Li ,

The Carlitz rings

187

whence (5.17) dim Qi−1 − dim Mi−1 + dim Mi − dim Li = 0. $ $ Note that Q = Qi and L = Li are graded ﬁnitely generated i≥0

i≥0

modules over gr(A0 ) which are annihilated by the operator τ . Therefore Q and L are ﬁnite-dimensional vector spaces over K c , thus Qi = Li = {0} s s for i big enough, so that dim Qi = const, dim Li = const, if s is big i=0

i=0

enough. Summing up the identities (5.17) we ﬁnd σ0 > 0 such that F0 (s) − F0 (s − 1) = const ∈ Z,

s ≥ σ0 ,

and F0 (s) coincides, for s big enough, with a numerical polynomial of degree 1. Let n = 0, that is we will consider A1 -modules. We repeat the above reasoning with the mappings ψi : Mi−1 → Mi , ψi (m) = d(m), semilinear s with respect to the automorphism τ −1 . If F1 (s) = dim Mi (the notai=−∞

tion F1 corresponds to A1 -modules), we ﬁnd, as above, on the basis of the result for A0 , that F1 (s) − F1 (s − 1) = γ(s),

s ≥ σ1 > 0,

where γ(s) is a numerical polynomial, and it remains to use Lemma 5.17. The same reasoning (for the linear mappings corresponding to ∆1 , . . . , ∆n ), with induction on n, brings the required result for the general situation. The number d(M ) = deg χ is called the dimension of M , while the leading coeﬃcient of χ multiplied by d(M )! is called the multiplicity of M . It follows from Lemma 5.17 (ii) that the multiplicity (just as the dimension) is always a natural number. These deﬁnitions follow the notions related to modules over the Weyl algebras and their generalizations [16, 30, 73]; just as for modules over algebras (see Lemma 6.1 in [73]), it can be shown that d(M ) agrees with the general notion of the Gelfand–Kirillov dimension. Following the existing tradition, we call χ the Hilbert polynomial. Note that χ(i) = dim Mi for big enough values of i. Below we return to the earlier notation considering the rings An+1 for n ≥ 0. Let us look more closely at the notion of a good ﬁltration. Recall that the ﬁltration {Γν } in An+1 was introduced in Section 5.2.2.

188

Chapter 5

Proposition 5.19 Let M be a ﬁltered module with a ﬁltration {Mν } over An+1 . The ﬁltration {Mν } is good if and only if there exists a number k0 such that Mi+k = Γi Mk , for all i ∈ Z+ , k ≥ k0 .

(5.18)

Proof. Suppose that the condition (5.18) holds. Since dim Mk0 < ∞, Mk0 possesses a ﬁnite K c -basis, and images of its elements generate gr(M ). Thus the ﬁltration is good. Conversely, let gr(M ) be ﬁnitely generated, with generators represented by the elements u1 , . . . , us , where uj ∈ Mkj \ Mkj −1 , j = 1, 2, . . . , s. Set k0 = max{k1 , . . . , ks }. We will prove (5.18) by induction on i. The case i = 0 is trivial. Suppose that Mi−1+k = Γi−1 Mk ,

k ≥ k0 ,

(5.19)

for some i. Let v ∈ Mi+k . Denote by µl and σl the canonical projections µl : Ml → Ml /Ml−1 ,

σl : Γl → Γl /Γl−1 .

By the ﬁnite generation assumption, µi+k (v) ∈

s

σi+k−kj Γi+k−kj µkj (uj ).

j=1

Since Γi+k−kj = Γi · Γk−kj (that is every element from the left-hand set is a sum of products of elements from Γi and Γk−kj ), we see that v∈

s

Γi · Γk−kj uj + Mi+k−1 .

j=1

Now we have Γk−kj uj ∈ Mk , and it remains to use (5.19) to obtain (5.18). The next results deal with the comparison of ﬁltrations. Proposition ! 5.20 Suppose ! that M is a left An+1 -module with two ﬁltra(1) (2) tion, Mi and Mi . ! (1) (i) If Mi is a good ﬁltration, then there exists a number k1 such that (1)

Mi

(2)

⊂ Mi+k1 , for all i.

(5.20)

The Carlitz rings

189

(ii) If both ﬁltrations are good, then there exists a number k2 such that (1)

(2)

Mi−k2 ⊂ Mi (1)

(1)

is good, then there exists a number k0 such that Γi Mj

for all j ≥ k0 , i ≥ 0. Since

ﬁnd that

(1) Mk0

(5.21)

!

Proof. If Mi (1) Mi+j ,

(2)

⊂ Mi+k2 , for i big enough.

⊂

(2) Mk1

(1) Mk0

=

is ﬁnite-dimensional over K c , we

for some k1 . Then (1)

(1)

(2)

(2)

Mi+k0 = Γi · Mk0 ⊂ Γi · Mk1 ⊂ Mi+k1 , so that (1)

Mi

(1)

(2)

⊂ Mi+k0 ⊂ Mi+k1 ,

i ≥ 0,

and we have proved (5.20). The second assertion is a consequence of the ﬁrst one.

Corollary 5.21 The dimension d(M ) and the multiplicity m(M ) are the same for any good ﬁltration of a An+1 -module M . Proof. Suppose we have two good ﬁltrations, as in Proposition 5.20 (ii). Let χ(1) and χ(2) be the corresponding Hilbert polynomials. It follows from (5.21) that χ(2) (i − k2 ) ≤ χ(1) (i) ≤ χ(2) (i + k2 ) for large values of i. Since the behavior of a polynomial at inﬁnity is determined by its degree, we get that the degrees and leading coeﬃcients of the polynomials χ(1) and χ(2) coincide. Let us consider the behavior of the dimension d(M ) in the context of some operations on An+1 -modules. Proposition 5.22 (i) Let M be a ﬁltered An+1 -module with a good ﬁltration, N ⊂ M a submodule, M/N a quotient module (with the induced ﬁltrations). Then d(M ) = max{d(N ), d(M/N )},

(5.22)

and if d(N ) = d(M/N ), then m(M ) = m(N ) + m(M/N ).

(5.23)

190

Chapter 5

(ii) Let L1 , . . . , Lk be ﬁltered An+1 -modules with good ﬁltrations, L = k $ Lj . Then j=1

d(L) = max{d(L1 ), . . . , d(Lk )},

(5.24)

and if d(L) = d(Li ) for all i, then m(L) =

k

m(Li ).

(5.25)

i=1

Proof. (i) Let {Nk } and {Fk } be the induced ﬁltrations in N and M/N respectively (see Section 5.1.2). By (5.1), for each k we have the exact sequence 0 → Nk /Nk−1 → Mk /Mk−1 → Fk /Fk−1 → 0 which implies the equality dim Nk /Nk−1 + dim Fk /Fk−1 = dim Mk /Mk−1 . Summing up by k, we obtain the equality for the corresponding Hilbert polynomials χN (s) + χM/N (s) = χM (s) valid for large s, thus for all s. This implies (5.22) and (5.23). (ii) The proof of (5.24) and (5.25) follows by induction from part (i) applied to the exact sequence 0 → Lk → L → L1 ⊕ · · · ⊕ Lk−1 → 0.

5.4.2. Dimension: calculations and estimates. First of all, consider An+1 as a left module over itself. It follows from Lemma 5.10 that d(An+1 ) = n + 2,

m(An+1 ) = 1.

(5.26)

Next, let us consider direct sums Arn+1 = An+1 ⊕ · · · ⊕ An+1 . , -. / r

By (5.24), we ﬁnd that d Arn+1 = n + 2 for any r = 1, 2, . . .. This leads to the following general result.

The Carlitz rings

191

Proposition 5.23 For any ﬁnitely generated left An+1 -module M , d(M ) ≤ n + 2.

(5.27)

Proof. Let M be generated by r elements. Then there exists an epimorphism ϕ : Arn+1 → M , thus M ∼ = Arn+1 / ker ϕ, and, by virtue of Proposition 5.22 (i), d Arn+1 = max{d(M ), d(ker ϕ)}, whence d(M ) ≤ d Arn+1 = n + 2. By (5.26), in general the bound (5.27) cannot be improved. Nevertheless, there are cases with smaller values of d(M ). An important special case is given by the next proposition. Proposition 5.24 If I is a nonzero left ideal in An+1 , then d (An+1 /I) ≤ n + 1.

(5.28)

Proof. First we consider the case of a principal left ideal I = An+1 γ, γ ∈ An+1 . Then we have an exact sequence θ

0 → An+1 → An+1 → An+1 /An+1 γ → 0 where θ(a) = aγ for all a ∈ An+1 . Here we used the fact (Theorem 5.9) that An+1 has no zero-divisors. Now, if d (An+1 /I) = n + 2 = d (An+1 ), then, by (5.23), we would have m (An+1 ) = m(I) + m (An+1 /I) which contradicts the equality m (An+1 ) = 1, since the multiplicity is a natural number. Let us consider the set F(n+1 of polynomials (5.2) as a An+1 -module. A ﬁltration (0) (1) Fn+1 ⊂ Fn+1 ⊂ . . . ⊂ F(n+1 (j)

can be introduced by setting Fn+1 to be the collection of all the polynomials (5.2), in which the maximal indices k1 , . . . , kn corresponding to nonzero coeﬃcients am,k1 ,...,kn do not exceed j. This ﬁltration is obviously good.

192

Chapter 5

Proposition 5.25 For the module F(n+1 ,

d F(n+1 = n + 1, m F(n+1 = n!

(5.29)

(j)

Proof. Let us compute dim Fn+1 . For a ﬁxed µ, the quantity of n-tuples (k1 , . . . , kn ) of nonnegative integers, for which min(k1 , . . . , kn ) = µ, is added up from those n-tuples where i numbers are equal to µ while n − i numbers are strictlylarger and can take j − µ values. Therefore the above n n (j − µ)n−i . Next, µ + 1 possible values of m in quantity equals i=1 i (5.2) correspond to each n-tuple. Thus, (j)

dim Fn+1 =

j

(µ + 1)

µ=0

=

j

n n i=1

i

(j − µ)n−i

(µ + 1) {(j − µ + 1)n − (j − µ)n } .

µ=0

Denote rµ = (j −µ+1)n −(j −µ)n , Ri = r0 +r1 +· · ·+ri = (j +1)n −(j −i)n . Performing the Abel transformation we get (j)

dim Fn+1 = (j + 1)Rj −

j−1

Ri

i=0

= (j + 1)n+1 − j(j + 1)n +

j−1

(j − i)n

i=0

= (j + 1)n +

j

k n = (j + 1)n + Sn (j + 1)

k=1

where Sn (N ) = 1 + 2 + · · · + (N − 1)n . It is known ([52], Chapter 15) that n 1 n+1 Bk N n+1−k Sn (N ) = k n+1 n

n

k=0

where Bk are the Bernoulli numbers. Therefore we ﬁnd that (j)

dim Fn+1 =

(j + 1)n+1 + Pn (j) n+1

where Pn is a polynomial of degree n. This implies (5.29).

The Carlitz rings

193

It is instructive to compare the above calculations with those for modules over the Weyl algebras [16, 25, 30]. For such modules, possible values of d(M ) are also bounded from above by the dimension of the algebra considered as a module over itself. Moreover, by the celebrated Bernstein inequality, d(M ) is also bounded from below by the dimension of the module consisting of polynomials. Modules with the minimal possible dimension are called holonomic, and it appears that just such modules emerge in many important applications. In our situation, an analog of the Bernstein inequality is impossible, since there are examples of ﬁnite-dimensional A1 -modules M (see Theorem 5.14), for which obviously d(M ) = 0. However, the above analogy makes it natural to call an An+1 -module M quasiholonomic, if d(M ) = n + 1. As we will see, lower dimensions correspond in applications to degenerate cases, while quasiholonomic modules play roles similar to those of holonomic modules over the Weyl algebras. Returning to possible values of d(M ) for A1 -modules, we mention a special case opposite to the one considered in Theorem 5.14. Proposition 5.26 Let M be a ﬁnitely generated A1 -module with a good ﬁltration. Suppose that there exists a “vacuum vector” v ∈ M , such that ds v = 0 and τ m (v) = 0 for all m = 0, 1, 2, . . .. Then d(M ) ≥ 1. Proof. It follows from (5.12) that ds τ m v = [m]1/q τ m−1 v,

m = 1, 2, . . . ,

that is τ m−1 v is an eigenvector of a linear operator ds τ on M (considered as a K c -vector space) corresponding to the eigenvalue [m]1/q . Therefore the vectors τ m−1 v are linearly independent. It follows from the existence of the Hilbert polynomial χ implementing the dimension d(M ) that d(M ) ≥ 1. 5.4.3. Evolution operators. Let R ∈ An+1 be an operator of the form R = P (∆1 , . . . , ∆n ) + Q(∆1 , . . . , ∆n )d where P, Q are non-zero polynomials. As we know (Theorem 3.10), under some nondegeneracy conditions the Cauchy problem for the equation Ru = 0 is well-posed. Let I = An+1 R.

194

Chapter 5

Theorem 5.27 The module M = An+1 /I is quasiholonomic. Proof. Due to (5.28), we have to show only that d(M ) ≥ n + 1. First we prove two lemmas. Lemma 5.28 An operator A= al,µ,i1 ,...,in τ l dµ ∆i11 . . . ∆inn ,

al,µ,i1 ,...,in ∈ K c ,

(5.30)

is linear if and only if al,µ,i1 ,...,in = 0 for l = µ. Proof. Let σ ∈ K c . Suppose that Aσ = σA, that is l−µ σal,µ,i1 ,...,in τ l dµ ∆i11 . . . ∆inn = al,µ,i1 ,...,in σ q τ l dµ ∆i11 . . . ∆inn . By the uniqueness of the representation (5.30) (Proposition 5.8), we ﬁnd l−µ that σ q = σ, whenever al,µ,i1 ,...,in = 0. Since σ is arbitrary, that is possible if and only if l = µ. Lemma 5.29 The ideal I does not contain nonzero linear operators. Proof. Using the commutation relations deﬁning An+1 we can rewrite the representation (5.30) of an arbitrary operator A ∈ An+1 in the form A= al,µ,i1 ,...,in ∆i11 . . . ∆inn τ l dµ , al,µ,i1 ,...,in ∈ K c . (5.31) Just as in (5.30), the coeﬃcients al,µ,i1 ,...,in are determined in a unique way, and Lemma 5.28 remains valid for the representation (5.31). Suppose that an operator A ∈ An+1 is such that AR is linear. Let us write (5.31) in the form A=

N N

αlµ τ l dµ

l=0 µ=0

where αlµ are the appropriate elements of the commutative K c -algebra D (without zero-divisors) generated by the linear operators ∆1 , . . . , ∆n . We have also R = γ + δd, γ, δ ∈ D, γ = 0, δ = 0. In this new notation,

N N αlµ τ l dµ (γ + δd). AR = l=0 µ=0

As an element γ ∈ D is permuted with powers of τ and d, in additional terms (appearing in accordance with the commutation rules) the powers of

The Carlitz rings

195

τ and d respectively are the same, while the degrees of elements from D (as polynomials of ∆1 , . . . , ∆n ) decrease. Therefore AR =

N N

αlµ γ τ l dµ +

l=0 µ=0

N N

αlµ δ τ l dµ+1

l=0 µ=0

where γ , δ ∈ D, γ = 0, δ = 0, whence AR =

N N

(αlν γ + αl,ν−1 δ ) τ l dν +

l=0 ν=1

N αl0 γ τ l + αlN δ τ l dN +1 . l=0

By Lemma 5.28, αl0 = 0, αlN = 0,

l = 1, . . . , N ;

(5.32)

l = 0, 1, . . . , N.

(5.33)

Considering terms with l < N, ν = N , we ﬁnd that αlN γ + αl,N −1 δ = 0, and (5.33) yields αl,N −1 = 0, 0 ≤ l ≤ N − 1. Repeating the reasoning we obtain that αlν = 0 for l ≤ ν. On the other hand, for l ≥ 2, ν = 1 we get αl1 γ + αl0 δ = 0, and, by (5.32), αl1 = 0. Repeating we come to the conclusion that αlν = 0 for l > ν, so that A = 0. Proof of Theorem 5.27 (continued). Let us consider the induced ﬁltration in M . The subspace Mν is generated by images in M of the elements (5.30) with max(l + µ + i1 + · · · + in ) ≤ ν; those two elements whose diﬀerence belongs to I are identiﬁed. Let us consider elements with l = µ. Elements of the form τ l dl ∆i11 · · · ∆inn with diﬀerent collections of parameters (l, i1 , . . . , in ) are linearly independent in An+1 . If some linear combination of their images equals zero in M , then the corresponding linear combination of the elements themselves must belong to I, which (by Lemma 5.29) is possible only if it is equal to zero. Therefore the images of the above elements are linearly independent, so that " # : 2l + i1 + · · · + in ≤ ν dim Mν ≥ card (l, i1 , . . . , in ) ∈ Zn+1 + " # ≥ card (l, i1 , . . . , in ) ∈ Zn+1 : l + i1 + · · · + in ≤ int(ν/2) . +

196

Chapter 5

Evaluating the number of nonnegative integral solutions of the above inequality as in the proof of Lemma 5.10 we ﬁnd that [ν/2] + n + 1 ≥ c1 [ν/2]n+1 ≥ c2 ν n+1 dim Mν ≥ n+1 for large values of ν (c1 , c2 are positive constants independent of ν). Thus, d(M ) ≥ n + 1, as desired. 5.4.4. Quasiholonomic functions. Let 0 = f ∈ Fn+1 , If = {ϕ ∈ An+1 : ϕ(f ) = 0} . If is a left ideal in An+1 . The left An+1 -module Mf = An+1 /If is isomorphic to the submodule An+1 f ⊂ Fn+1 – an element ϕ(f ) ∈ An+1 f corresponds to the class of ϕ ∈ An+1 in Mf . A natural good ﬁltration in Mf is induced from that in An+1 – the subspace Mj is generated by elements τ l dµs ∆i11 . . . ∆inn f with l + µ + i1 + · · · + in ≤ j. As we know, if If = {0}, then d(Mf ) ≤ n + 1. We call a function f quasiholonomic if the module Mf is quasiholonomic, that is d(Mf ) = n + 1. The condition If = {0} means that f is a solution of a “diﬀerential equation” ϕ(f ) = 0, ϕ ∈ An+1 . For n = 0, we have the following easy result. Theorem 5.30 If a nonzero function f ∈ F1 satisﬁes an equation ϕ(f ) = 0, 0 = ϕ ∈ A1 , then f is quasiholonomic. Proof. " l #∞It is suﬃcient to show that dim Mf = ∞. In fact, the sequence τ f l=0 is linearly independent because otherwise we would have a ﬁnite collection of elements c0 , c1 , . . . , cN ∈ K c , some of which are diﬀerent from zero, such that N

c0 f (s) + c1 f q (s) + · · · + cN f q (s) = 0

(5.34)

for all s from a neighborhood of the origin in K c . It follows from (5.34) that f takes only a ﬁnite number of values. By the uniqueness theorem for nonArchimedean holomorphic functions, f (s) ≡ const on some neighborhood of the origin. Due to the Fq -linearity, f (s) ≡ 0, and we have come to a contradiction. In particular, any Fq -linear polynomial of s is quasiholonomic, since it is annihilated by dm s , with a suﬃciently large m. If n > 0, the situation is more complicated. We call the module Mf

The Carlitz rings

197

(and the corresponding function f ) degenerate if d(Mf ) < n + 1 (by the Bernstein inequality, there are no degeneracy phenomena for modules over the complex Weyl algebra). We give an example of degeneracy for the case n = 1. Let f (s, t1 ) = g(st1 ) ∈ F2 where the function g belongs to F1 and satisﬁes an equation ϕ(g) = 0, ϕ ∈ A1 . Then f is degenerate. Indeed, by the general rule, Mj is spanned by elements τ l dµs ∆i1i f with l + µ + i1 ≤ j. In the present situation, ∆1 f = g(xst1 ) − xg(st1 ) = τ ds g, that an element τ l dµs ∆i1i f is a linear combination of elements so l+λ τ dµ+ν g (s, t) with λ ≤ i1 , ν ≤ i1 . Therefore Mj is contained in s the linear hull of elements τ k dm s g, k + m ≤ 2j. By Theorem 5.30, the K c -dimension of the latter does not exceed a linear function of 2j, so that d(Mf ) ≤ 1. On the other" hand, #∞ since, as in the proof of Theorem l 5.30, the system of functions τ f l=0 is linearly independent, we ﬁnd that d(Mf ) = 1. In order to exclude the degenerate case, we introduce the notion of a nonsparse function. A function f ∈ Fn+1 of the form (5.2) is called nonsparse if there exists a sequence ml → ∞ such that, for any l, there exist sequences (i) (i) (i) (i) k1 , k2 , . . . , kn ≥ ml (depending on l), such that kν → ∞ as i → ∞ (ν = 1, . . . , n), and am ,k(i) ,...,k(i) = 0. l

1

n

Lemma 5.31 If a function f is nonsparse, then the system of functions (τ ds )λ ∆j11 . . . ∆jnn f (λ, j1 , . . . , jn = 0, 1, 2, . . .) is linearly independent over K c. Proof. Suppose that J1 Λ

...

λ=0 j1 =0

Jn

cλ,j1 ,...,jn (τ ds )λ ∆j11 . . . ∆jnn f = 0

(5.35)

jn =0

for some cλ,j1 ,...,jn ∈ K c , Λ, J1 , . . . , Jn ∈ N. Substituting (5.2) into (5.35) and collecting coeﬃcients of the power series we ﬁnd that J1 Λ λ=0 j1 =0

for all l, i.

...

Jn jn =0

(i)

cλ,j1 ,...,jn [ml ]λ [k1 ]j1 . . . [kn(i) ]jn = 0

(5.36)

198

Chapter 5

We see from (5.36) that the polynomial   Jn−1 J1 Jn  Λ  (i) (i) ... cλ,j1 ,...,jn [ml ]λ [k1 ]j1 . . . [kn−1 ]jn−1 z jn   jn =0

λ=0 j1 =0

jn−1 =0

has an inﬁnite sequence of diﬀerent roots, so that J1 Λ λ=0 j1 =0

Jn−1

...

(i)

(i)

cλ,j1 ,...,jn [ml ]λ [k1 ]j1 . . . [kn−1 ]jn−1 = 0

jn−1 =0

for all l, i, and for each jn = 0, 1, . . . , Jn . Repeating this reasoning we ﬁnd that all the coeﬃcients cλ,j1 ,...,jn are equal to zero. Now the above arguments regarding d(Mf ) yield the following result. Theorem 5.32 If a function f is nonsparse, then d(Mf ) ≥ n + 1. If, in addition, f satisﬁes an equation ϕ(f ) = 0, 0 = ϕ ∈ An+1 , then f is quasiholonomic. As in the classical situation, one can construct quasiholonomic functions by addition. Proposition 5.33 If the functions f, g ∈ Fn+1 are quasiholonomic, and f + g is nonsparse, then f + g is quasiholonomic. Proof. Consider the An+1 -module M2 = (An+1 f ) ⊕ (An+1 g). Since f and g are both quasiholonomic, we have d(M2 ) = n + 1. Next, let N2 be a submodule of M2 consisting of such pairs (ϕ(f ), ϕ(g)) that ϕ(f )+ϕ(g) = 0. Then d(M2 ) = max{d(N2 ), d(M2 /N2 )}, so that d(M2 /N2 ) ≤ n + 1. On the other hand, we have an injective mapping An+1 (f +g) → M2 /N2 , which maps ϕ(f + g) to the image of (ϕ(f ), ϕ(g)) in M2 /N2 . Therefore d(An+1 (f + g)) ≤ d(M2 /N2 ) ≤ n + 1. It remains to use Theorem 5.32. It was noticed by Zeilberger (see [25] and references therein) that many classical sequences of functions generate holonomic modules, if, as a preparation, a transform is made with respect to the discrete variables, reducing the continuous–discrete case to the purely continuous one (simultaneously

The Carlitz rings

199

in all the variables). In our situation, if, for example, {Pk (s)} is a sequence of Fq -linear polynomials with deg Pk ≤ q k , we set f (s, t) =

∞

k

Pk (s)tq .

(5.37)

k=0

Then f ∈ F2 , and we may consider the question of whether f is quasiholonomic. It is remarkable that several functions obtained this way from the basic special functions of Fq -linear analysis, are indeed quasiholonomic. (a) The Carlitz polynomials. Considering the construction (5.37) for Pk = fk , the normalized Carlitz polynomials, we obtain the Carlitz module (Section 1.6.2), that is Cs (t) =

∞

k

fk (s)tq .

(5.38)

k=0

By Proposition 1.7, fk (s) =

k (−1)k−i qi i=0 Di Lk−i

It is easy to check that i − q −1 +(k−i)q i qi , Di Lk−i = q q−1

i

sq .

0 ≤ i ≤ k.

For large values of k, an elementary investigation of the function z → (k − z)q z , z ≤ k, shows that max (k − i)q i ≤ αq k ,

0≤i≤k

α > 0,

so that |fk (s)| ≤ q αq

k

for all s ∈ K c with |s| ≤ q −1 . Therefore the series (5.38) converges for small |t|, so that the Carlitz module function belongs to F2 . Since ds fi = fi−1 for i ≥ 1, and ds f0 = 0, we see that ds Cs (t) = Cs (t). Clearly, the function Cs (t) is nonsparse. Therefore the Carlitz module function is quasiholonomic, jointly in both its variables. (b) Thakur’s hypergeometric polynomials. We consider the polynomial

200

Chapter 5

case of Thakur’s hypergeometric function (Section 4.1.2): (−a1 )m . . . (−al )m m (1) zq l Fλ (−a1 , . . . , −al ; −b1 , . . . , −bλ ; z) = (−b1 )m . . . (−bλ )m Dm m (5.39) where a1 , . . . , al , b1 , . . . , bλ ∈ Z+ , m (−1)a−m L−q a−m , if m ≤ a, (−a)m = , a ∈ Z+ . (5.40) 0, if m > a, It is seen from (5.40) that the terms in (5.39), which make sense and do not vanish, are those with m ≤ min(a1 , . . . , al , b1 , . . . , bλ ). Let f (s, t1 , . . . , tl , u1 , . . . , uλ ) ∞ ∞ ∞ ∞ q k1 = ... ... l Fλ (−k1 , . . . , −kl ; −ν1 , . . . , −νλ ; s)t1 . . . k1 =0

kl =0 ν1 =0

νλ =0 kl

ν1

× tql uq1

νλ

. . . uqλ .

(5.41)

We prove as above that all the series in (5.41) converge near the origin. Thus, f ∈ Fl+λ+1 . It is not diﬃcult to check ([111], Section 6.5) that ds (l Fλ (−k1 , . . . , −kl ; −ν1 , . . . , −νλ ; s)) = l Fλ (−k1 + 1, . . . , −kl + 1; −ν1 + 1, . . . , −νλ + 1; s)

(5.42)

if all the parameters k1 , . . . , kl , ν1 , . . . , νλ are diﬀerent from zero. If at least one of them is equal to zero, then the left-hand side of (5.42) equals zero. This property implies the identity ds f = f , the same as that for the Carlitz module function. Since f is nonsparse, it is quasiholonomic. Next we will see that the K-binomial coeﬃcients (4.52) correspond to a quasiholonomic function satisfying a more complicated equation containing also the operator ∆t . (c) K-binomial coeﬃcients. By Proposition 4.10, the function k ∞ m k k f (s, t) = sq tq , m K m=0 k=0

belonging to F2 , satisﬁes the equation ds f (s, t) = ∆t f (s, t) + [1]1/q f (s, t). Obviously, f is nonsparse. Therefore f is quasiholonomic.

The Carlitz rings

201

5.5 Comments The general framework of ﬁltered and graded rings and modules (Sections 5.1.1, 5.1.2) is well known. See, for example, [16, 30, 79]. Generalized Weyl algebras were introduced by Bavula [10] and studied in his subsequent works, especially in the paper [12] by Bavula and van Oystaeyen. Investigation of the algebraic structure of the rings of diﬀerential operators with the Carlitz derivatives was initiated by the author [63] and continued on a larger scale by Bavula [11] who applied successfully his GWA techniques. Our Section 5.3 is an introductory exposition of part of the results from [11]. Note that Proposition 5.15 (iii) (by Bavula [11]) corrects an erroneous statement from [63]. Properties of diﬀerential operators (in the usual sense) over ﬁelds of positive characteruistic are quite diﬀerent; see, for example [87]. Theorem 5.14 is taken from the author’s paper [68], which contained also the notion of a quasiholonomic module and the whole related material (in the surveys [69, 70] quasiholonomic modules are called holonomic). The general technique of ﬁltered modules over the Carlitz rings is quite similar to the case of modules over the Weyl algebras, though some of the results are diﬀerent, in particular due to the existence of nontrivial ﬁnitedimensional modules.

Bibliography

[1] D. Adam, Car-P´ olya and Gel’fond theorems for Fq [T ], Acta Arith. 115 (2004), 287–303. [2] S. Albeverio and A. Yu. Khrennikov, Representations of the Weyl group in spaces of square integrable functions with respect to p-adic valued Gaussian distributions, J. Phys. A: Math. and Gen. 29 (1996), 5515–5527. [3] Y. Amice, Interpolation p-adique, Bull. Soc. Math. France 92 (1964), 117–180. [4] Y. Amice, Duals. In: Proc. Conf. p-Adic Analysis, Nijmegen, 1978, pp. 1–15. [5] Y. Andr´e and L. Di Visio, q-Diﬀerence equations and p-adic local monodromy, Ast´erisque 296 (2004), 55–111. [6] M. F. Atiyah and I. G. Macdonald, Introduction to Commutative Algebra, Addison-Wesley, Reading, 1969. [7] A. Baker, p-Adic continuous functions on rings of integers and a theorem of K. Mahler, J. London Math. Soc. 33 (1986), 414–420. [8] F. Baldassarri, Diﬀerential modules and singular points of p-adic diﬀerential equations, Adv. Math. 44 (1982), 155–179. [9] H. Bateman and A. Erd´elyi, Higher Transcendental Functions, Vol. 2, McGrawHill, New York, 1953. [10] V. Bavula, Generalized Weyl algebras and their representations, St. Petersburg Math. J. 4, No. 1 (1993), 71–92. [11] V. Bavula, The Carlitz algebras, J. Pure Appl. Algebra 212 (2008), 1175– 1186. [12] V. Bavula and F. van Oystaeyen, The simple modules of certain generalized crossed products, J. Algebra 194 (1997), 521–566. [13] M. Bhargava, P -orderings and polynomial functions on arbitrary subsets of Dedekind rings, J. Reine Angew. Math. 490 (1997), 101–127. [14] M. Bhargava, The factorial function and generalizations, Amer. Math. Monthly 107 (2000), 783–799. [15] M. Bhargava and K. S. Kedlaya, Continuous functions on compact subsets of local ﬁelds, Acta Arith. 91 (1999), 191–198. [16] J.-E. Bj¨ ork, Rings of Diﬀerential Operators, North-Holland, Amsterdam, 1979. [17] N. Bourbaki, Alg`ebre Commutative, Hermann, Paris, 1961–65. [18] N. Bourbaki, Algebra I, Springer, Berlin, 1989. [19] A. Buium, Continuous π-adic functions and π-derivations, J. Number Theory

203

204

Bibliography

84 (2000), 34–39. [20] P.-J. Cahen and J.-L. Chabert, Integer-Valued Polynomials, American Mathematical Society, Providence, 1997. [21] M. Car, Polya’s theorem for Fq [T ], J. Number Theory 66 (1997), 148–171. [22] L. Carlitz, On certain functions connected with polynomials in a Galois ﬁeld, Duke Math. J. 1 (1935), 137–168. [23] L. Carlitz, A set of polynomials, Duke Math. J. 6 (1940), 486–504. [24] L. Carlitz, Some special functions over GF (q, x), Duke Math. J. 27 (1960), 139–158. [25] P. Cartier, D´emonstration “automatique” d’identit´es et fonctions hyperg´eometriques (d’apr`es D. Zeilberger), Ast´erisque 206 (1992), 41–91. [26] D. N. Clark, A note on the p-adic convergence of solutions of linear diﬀerential equations, Proc. Amer. Math. Soc. 17 (1966), 262–269. [27] E. A. Coddington and N. Levinson, Theory of Ordinary Diﬀerential Equations, McGraw-Hill, New York, 1955. [28] R. F. Coleman, Dilogarithms, regulators and p-adic L-functions, Invent. Math. 69 (1982), 171–208. [29] K. Conrad, The digit principle, J. Number Theory 84 (2000), 230–257. [30] S. C. Coutinho, A Primer of Algebraic D-modules, Cambridge University Press, 1995. [31] L. Delamette, Th´eor`eme de P´ olya en caract´eristique ﬁnie, Acta Arith. 106 (2003), 159–170. [32] B. Diarra, The Hopf algebra structure of the space of continuous functions on power series over Fq and Carlitz polynomials, Contemp. Math. 319 (2003), 75–97. [33] V. G. Drinfeld, Elliptic modules, Math. USSR Sbornik 23 (1976), 561–592. [34] B. Dwork, Lectures on p-Adic Diﬀerential Equations, Springer, New York, 1982. [35] B. Dwork, G. Gerotto, and F. J. Sullivan, An Introduction to G-Functions, Princeton University Press, 1994. [36] A. Escassut, Analytic Elements in p-Adic Analysis, World Scientiﬁc, Singapore, 1995. [37] L. Ferrari, An umbral calculus over inﬁnite coeﬃcient ﬁelds of positive characteristic, Comp. Math. Appl. 41 (2001), 1099–1108. [38] I. B. Fesenko and S. V. Vostokov, Local Fields and Their Extensions, American Mathematical Society, Providence, 2002. [39] F. R. Gantmacher, Matrizentheorie, Springer, Berlin, 1986. [40] E.-U. Gekeler, Some new identities for Bernoulli–Carlitz numbers, J. Number Theory 33 (1989), 209–219. [41] I. M. Gelfand, Lectures on Linear Algebra, Interscience, New York, 1961. [42] A. O. Gelfond, Calculus of Finite Diﬀerences, Hindustan Publishing Corporation, Dehli, 1971. [43] D. Goss, Fourier series, measures, and divided power series in the theory of function ﬁelds, K-Theory 1 (1989), 533–555. [44] D. Goss, A formal Mellin transform in the arithmetic of function ﬁelds, Trans. Amer. Math. Soc. 327 (1991), 567–582. [45] D. Goss, Basic Structures of Function Field Arithmetic, Springer, Berlin, 1996. [46] A. Granville, Arithmetic properties of binomial coeﬃcients. I. In: Organic

Bibliography

[47] [48] [49] [50] [51] [52] [53] [54] [55] [56]

[57] [58] [59] [60] [61] [62] [63] [64] [65] [66] [67] [68] [69]

205

Mathematics. CMS Conf. Proc. Vol. 20. American Mathematical Society, Providence, 1997, pp. 253–276. P. Hartman, Ordinary Diﬀerential Equations, Wiley, New York, 1964. H. Hasse, Theorie der h¨ oheren Diﬀerentiale in einem algebraischen Funktionk¨ orper mit vollkommenem Konstantenk¨ orper bei beliebiger Charakterictik, J. Reine Angew. Math. 175 (1936), 50–54. H. Hasse and F. K. Schmidt, Noch eine Begr¨ undung der Theorie der h¨ oheren Diﬀerentialquotienten in einem algebraischen Funktionk¨ orper einer Unbestimmten, J. Reine Angew. Math. 177 (1937), 215–237. I. N. Herstein, Noncommutative Rings, Carus Math. Monograph No. 15, Math. Assoc. of America, J. Wiley and Sons, 1968. E. Hille, Lectures on Ordinary Diﬀerential Equations, Addison-Wesley, Reading, 1969. K. Ireland and M. Rosen, A Classical Introduction to Modern Number Theory, Springer, New York, 1982. S. Jeong, A comparison of the Carlitz and digit derivatives bases in function ﬁeld arithmetic, J. Number Theory 84 (2000), 258–275. S. Jeong, Continuous linear endomorphisms and diﬀerence equations over the completion of Fq [T ], J. Number Theory 84 (2000), 276–291. S. Jeong, Hyperdiﬀerential operators and continuous functions on function ﬁelds, J. Number Theory 89 (2001), 165–178. H. Keller, H. Ochsenius and W. H. Schikhof, On the commutation relation AB − BA = I for operators on non-classical Hilbert spaces. In p-Adic Functional Analysis (A. K. Katsaras et al., eds.), Lect. Notes Pure Appl. Math., Vol. 222, Marcel Dekker, New York, 2001, pp. 177–190. A. Khrennikov, Non-Archimedean Analysis: Quantum Paradoxes, Dynamical Systems and Biological Models, Kluwer, Dordrecht, 1997. H. Koch, Galois Theory of p-Extensions, Springer, Berlin, 2002. A. N. Kochubei, Pseudo-Diﬀerential Equations and Stochastics over NonArchimedean Fields, Marcel Dekker, New York, 2001. A. N. Kochubei, p-Adic commutation relations, J. Phys. A: Math. and Gen. 29 (1996), 6375–6378. A. N. Kochubei, Harmonic oscillator in characteristic p, Lett. Math. Phys. 45 (1998), 11–20. A. N. Kochubei, Fq -linear calculus over function ﬁelds, J. Number Theory 76 (1999), 281–300. A. N. Kochubei, Diﬀerential equations for Fq -linear functions, J. Number Theory 83 (2000), 137–154. A. N. Kochubei, Diﬀerential equations for Fq -linear functions II: Regular singularity, Finite Fields Appl. 9 (2003), 250–266. A. N. Kochubei, Umbral calculus in positive characteristic, Adv. Appl. Math. 34 (2005), 175–191. A. N. Kochubei, Strongly nonlinear diﬀerential equations with Carlitz derivatives over a function ﬁeld, Ukrainian Math. J. 57 (2005), 794–805. A. N. Kochubei, Polylogarithms and a zeta function for ﬁnite places of a function ﬁeld, Contemporary Math. 384 (2005), 157–167. A. N. Kochubei, Quasi-holonomic modules in positive characteristic, J. Algebra 302 (2006), 826–844. A. N. Kochubei, Umbral calculus and holonomic modules in positive

206

[70]

[71] [72] [73] [74] [75] [76] [77] [78] [79] [80] [81] [82] [83] [84] [85] [86] [87] [88] [89] [90] [91] [92] [93] [94] [95]

Bibliography

characteristic, In: p-Adic Mathematical Physics, AIP Conference Proceedings, Vol. 826 (2006), 254–266. A. N. Kochubei, Hypergeometric functions and Carlitz diﬀerential equations over function ﬁelds, In: Arithmetic and Geometry around Hypergeometric Functions, Progress in Mathematics, Vol. 260, Birkh¨ auser, Basel, 2007, pp. 163–187. A. N. Kochubei, Evolution equations and functions of hypergeometric type over ﬁelds of positive characteristic, Bull. Belg. Math. Soc. 14 (2007), 947959. A. N. Kochubei, Dwork-Carlitz exponential and overconvergence for additive functions in positive characteristic, Intern. J. Number Theory 4 (2008), 453–460. G. R. Krause and T. H. Lenagan, Growth of Algebras and Gelfand–Kirillov Dimension, AMS, Providence, 2000. S. K. Lando, Lectures on Generating Functions, AMS, Providence, 2003. S. Lang, Algebra, Addison-Wesley, Reading, 1965. R. Lidl and H. Niederreiter, Finite Fields, Cambridge University Press, 1983. E. Lucas. Th´ eorie des Nombres, Jacques Gabay, Paris, 1995. E. Lutz, Sur l’equation y 2 = x3 − Ax − B dans les corps p-adiques, J. Reine Angew. Math. 177 (1937), 238–247. J. C. McConnell and J. C. Robson, Noncommutative Noetherian Rings, AMS, Providence, 2001. M. B. Nathanson, Additive number theory and the ring of quantum integers In: General Theory of Information Transfer and Combinatorics, Lect. Notes Computer Sci. Vol. 4123 (2006), 505–511. A. Perelomov, Generalized Coherent States and Their Applications, Springer, Berlin, 1986. C. Perez-Garcia, Locally convex spaces over non-Archimedean valued ﬁelds, Contemporary Math. 319 (2003), 251–279. R. S. Pierce, Associative Algebras, Springer, New York, 1982. ¨ G. P´ olya, Uber ganzwertige ganze Funktionen, Rend. Circ. Mat. Palermo 40 (1915), 1–16. B. Poonen, Fractional power series and pairings on Drinfeld modules, J. Amer. Math. Soc. 9 (1996), 783–812. M. van der Put, Meromorphic diﬀerential equations over valued ﬁelds, Indag. Math. 42 (1980), 327–332. M. van der Put, Diﬀerential equations in characteristic p, Compositio Math. 97 (1995), 227–251. J. Riordan, Combinatorial Identities, Wiley, New York, 1968. ´ P. Robba and G. Christol, Equations Diﬀ´ erentielles p-Adiques, Hermann, Paris, 1994. A. M. Robert, A Course in p-Adic Analysis, Springer, New York, 2000. S. Roman, The Umbral Calculus, Academic Press, Orlando, 1984. S. M. Roman and G.-C. Rota, The umbral calculus, Adv. Math. 27 (1978), 95–188. A. van Rooij, Non-Archimedean Functional Analysis, Marcel Dekker, New York, 1978. M. Rosen, Number Theory in Function Fields, Springer, New York, 2002. G.-C. Rota, D. Kahaner and A. Odlyzko, On the foundations of combinatorial

Bibliography

207

theory. VIII. Finite operator calculus, J. Math. Anal. Appl. 42 (1973), 684– 760. [96] G.-C. Rota and B. D. Taylor, The classical umbral calculus, SIAM J. Math. Anal. 25 (1994), 694–711. [97] S. G. Samko, A. A. Kilbas, and O. I. Marichev, Fractional Integrals and Derivatives: Theory and Applications, Gordon and Breach, New York, 1993. [98] W. Schikhof, Ultrametric Calculus, Cambridge University Press, 1984. [99] F. K. Schmidt, Die Wronskische Determinante in beliebigen diﬀerenzierbaren Funktionk¨ orpern, Math. Z. 45 (1939), 62–74. [100] P. Schneider, Nonarchimedean Functional Analysis, Springer, Berlin, 2002. [101] J.-P. Serre, Endomorphismes compl`etement continus des espaces de Banach p-adiques, Publ. Math. IHES 12 (1962), 69–85. [102] M. Setoyanagi, Note on Clark’s theorem for p-adic convergence, Proc. Amer. Math. Soc. 125 (1997), 717–721. [103] D. Sinnou and D. Laurent, Ind´ependence algebrique sur les T -modules, Compositio Math. 122 (2000), 1–22. [104] S. De Smedt, Some new bases for p-adic continuous functions, Indag. Math. New Ser. 4 (1993), 91–98. [105] B. Snyder. Hyperdiﬀerential Operators on Function Fields and Their Applications. PhD thesis, Ohio State University, Columbus, 1999. [106] A. Sudbery, Quantum Mechanics and the Particles of Nature, Cambridge University Press, 1986. [107] K. Tateyama, Continuous functions on discrete valuation rings, J. Number Theory 75 (1999), 23–33. [108] O. Teichm¨ uller, Diﬀerentialrechnung bei Charakteristik p, J. Reine Angew. Math. 175 (1936), 89–99. [109] D. S. Thakur, Hypergeometric functions for function ﬁelds, Finite Fields Appl. 1 (1995), 219–231. [110] D. S. Thakur, Hypergeometric functions for function ﬁelds II, J. Ramanujan Math. Soc. 15 (2000), 43–52. [111] D. S. Thakur, Function Field Arithmetic, World Scientiﬁc, Singapore, 2004. [112] Y. Uchino and T. Satoh, Function ﬁeld modular forms and higher derivations, Math. Ann. 311 (1998), 439–466. [113] L. Van Hamme, Continuous operators, which commute with translations, on the space of continuous functions on Zp . In: p-Adic Functional Analysis (J. M. Bayod et al., eds.), Lect. Notes Pure Appl. Math. Vol. 137, Marcel Dekker, New York, 1992, pp. 75–88. [114] A. Verdoodt, Umbral calculus in non-Archimedean analysis. In: p-Adic Functional Analysis (A. K. Katsaras et al., eds.), Lect. Notes Pure Appl. Math. Vol. 222, Marcel Dekker, New York, 2001, pp. 309–322. [115] V. S. Vladimirov, I. V. Volovich and E. I. Zelenov, p-Adic Analysis and Mathematical Physics, World Scientiﬁc, Singapore, 1994. [116] A. Volkenborn, Ein p-adisches Integral und seine Anwendungen I, II, Manuscripta Math. 7 (1972), 341–373; 12 (1974), 17–46. [117] J. F. Voloch, Diﬀerential operators and interpolation series in power series ﬁelds, J. Number Theory 71 (1998), 106–108. [118] C. G. Wagner, Interpolation series for continuous functions on π-adic completions of GF (q, x), Acta Arithm. 17 (1971), 389–406. [119] C. G. Wagner, Linear operators in local ﬁelds of prime characteristic, J.

208

Bibliography

Reine Angew. Math. 251 (1971), 153–160. [120] C. G. Wagner, Interpolation series in local ﬁelds of prime characteristic, Duke Math. J. 39 (1972), 203–210. [121] C. G. Wagner, Diﬀerentiability in local ﬁelds of prime characteristic, Duke Math. J. 41 (1974), 285–290. [122] A. Weil, Basic Number Theory, Springer, Berlin, 1967. [123] Z. Yang, Non-Archimedean Analysis over Function Fields with Positive Characteristic, PhD thesis, Ohio State University, Columbus, 1999. [124] Z. Yang, Locally analytic functions over completions of Fr (U ), J. Number Theory 73 (1998), 451–458. [125] Z. Yang, C n -functions over completions of Fr [T ] at ﬁnite places of Fr (T ), J. Number Theory 108 (2004), 346–374. [126] J. Yu, Transcendence theory over function ﬁelds, Duke Math. J. 52 (1985), 517–527.

Index

C n -function, 83 K-binomial coeﬃcients, 154 Fq -linear function, 4 π-ordering, 94

Finite place, 33 Formal divided power series, 101 Fractional derivative, 58

Additive Carlitz polynomials, 13 Admissible parameters, 138 Almost normalizing extension, 168 Bargmann–Fock representation, 41, 42, 43 Basic sequence, 62 Bessel–Carlitz functions, 143 Carlitz derivative, 39 Carlitz factorials, 2 Carlitz module function, 39 Coherent states, 41, 42 Completion, 32 Contiguous relations, 141 Cyclic orbit, 171

General Carlitz polynomials, 24 Generalized binomial polynomial, 94 Generalized factorial, 94 Generalized Taylor formula, 64 Generalized Weyl algebra, 171 Global ﬁelds, 32 Good ﬁltration, 170 Graded module, 169 Graded ring, 167 Grading, 167 Grading of a module, 169 Hilbert polynomial, 187 Hyperdiﬀerentiations, 17 hypergeometric equation, 139 hypergeometric function, 138

Degenerate module, 197 Degenerate orbit, 172 Delta operator, 62 Diﬀerence quotient, 83 Digit principle, 23 Dimension of a ﬁltered module, 187 Dwork exponential, 160 Dwork–Carlitz exponential, 160

Induced ﬁltrations, 170 Inﬁnite place, 32 Interpolation polynomials, 75 Invariant operator, 61

Entire function, 91 Euler function, 179 Extended hyperdiﬀerentiations, 31

M¨ obius function, 179 Multiplicity of a ﬁltered module, 187

Filtered module, 168 Filtered ring, 167 Filtration in a module, 168 Filtration in a ring, 167

Linear orbit, 171 Locally analytic function, 76 Locally constant function, 99

Non-Archimedean Banach space, 9 Non-Archimedean property, 2 Nonsparse function, 197 Normalized basic sequence, 68 Normalized Carlitz polynomials, 14

209

210 Numerical polynomial, 185

Index

Order of an entire function, 91 Orthonormal basis, 9

Schr¨ odinger representation, 40, 41 Sequence of K-binomial type, 63 Standard ﬁltration, 169 Strongly singular function, 16

Place, 32 Pochhammer-type symbol, 138 Prime ideal, 9

Thakur’s hypergeometric function, 140 Topological tensor product, 10 Type of an entire function, 91

Quasiholonomic function, 196 Quasiholonomic module, 193

Ultra-metric inequality, 2

Reduction, 9 Regular compact set, 75 Residual space, 9 Residue ﬁeld, 9

Very well distributed sequence, 76 Volkenborn-type integral, 55 Weight module, 171 Well distributed sequence, 76

Analysis in positive characteristic

Read more

Positive Ray Analysis of Magnesium

Read more

Characteristic Functions

Read more

Characteristic functions

Read more

Positive Teaching, Positive Learning

Read more

Characteristic classes

Read more

Positive psychology in practice

Read more

Positive psychology in practice

Read more

Positive polynomials in control

Read more

Positive Psychology in Practice

Read more

Geometry of characteristic classes

Read more

Lectures on characteristic classes

Read more

q-clan geometries in characteristic 2

Read more

Curvature and Characteristic Classes

Read more

Algebroid Curves in Positive Characteristics

Read more

Geometry of characteristic classes

Read more

Characteristic Classes. (AM-76)

Read more

Curvature and Characteristic Classes

Read more

Positive Operators

Read more

Positive Reinforcement

Read more

Positive Possibilities

Read more

Proof Positive

Read more

Positive Psychology

Read more

Q-Clan Geometries in Characteristic 2

Read more

Foliated Bundles and Characteristic Classes

Read more

Positive Operators

Read more

Proof Positive

Read more

Proof Positive

Read more

Positive Philosophy

Read more

Proof Positive

Read more

Recommend Documents

Analysis in positive characteristic

This page intentionally left blank CAMBRIDGE TRACTS IN MATHEMATICS General Editors ´ S , W . FU LTO N , A . K AT O K ...

Positive Ray Analysis of Magnesium

PROCEEDINGS OF THZ NATIONAL ACADEMVIY OF SCIENCES Volume 7 FEBRUARY 15, 1921 Number 2 POSITIVE RAY ANALYSIS OF MAGNE...

Characteristic Functions

...

Characteristic functions

G R IF F IN B OOKS O F CO G N A T E I N T EREST Applications of characteristic functions • • E. LUKACS and R. G. l....

Positive Teaching, Positive Learning

Positive Teaching, Positive Learning Positive Teaching, Positive Learning is an ideas book for the classroom with prac...

Characteristic classes

Positive psychology in practice

Positive Psychology in Practice Edited by P. Alex Linley and Stephen Joseph John Wiley & Sons, Inc. Positive Psychol...

Positive psychology in practice

Positive Psychology in Practice Edited by P. Alex Linley and Stephen Joseph John Wiley & Sons, Inc. Positive Psychol...

Positive polynomials in control

Lecture Notes in Control and Information Sciences Editors: M. Thoma · M. Morari 312 Lecture Notes in Control and Info...

Positive Psychology in Practice