Table Of Content

Differential Galois Theory of Linear Difference Equations 8 0 0 2 Charlotte Hardouin∗, Michael F. Singer† n a J December 24, 2007 9 ] A C Abstract . h t a We present a Galois theory of difference equations designed to measure the differ- m ential dependencies among solutions of linear difference equations. With this we are [ abletoreproveHölder’stheoremthattheGammafunctionsatisfiesnopolynomialdif- 1 ferential equation andareableto give general resultsthatimply, forexample, thatno v differential relationship holds among solutions of certain classes of q-hypergeometric 3 9 functions. 4 1 . 1 0 8 0 : v i X r a ∗IWR, Im Neuenheimer Feld 368, 69120 Heidelberg, Germany, [email protected] heidelberg.de. †North Carolina State University, Department of Mathematics, Box 8205, Raleigh, North Carolina 27695-8205, USA, [email protected]. This material is based upon work supported by the National Science Foundation under Grant No. CCF-0634123. 1 Contents 1 Introduction 3 2 Galois Theory 5 3 Differential Relations Among Solutions of Difference Equations 8 3.1 First order equations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 3.2 Higher order equations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 4 Inverse Problem 21 5 Parameterized Difference Equations 25 6 Appendix 28 6.1 Rational Solution of Difference Equations . . . . . . . . . . . . . . . . . . . 28 6.2 Σ∆Π-Galois Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 6.2.1 Σ∆Π-Picard-Vessiot Extensions . . . . . . . . . . . . . . . . . . . . 34 6.2.2 Σ∆Π-Galois Groups and the Fundamental Theorem . . . . . . . . . 39 6.2.3 Torsors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 References 47 2 1 Introduction In 1887, Otto Hölder [23] proved that the Gamma function Γ(x) satisfies no differential polynomial equation, that is, there is no nonzero polynomial P(x,y,y′,...) such that P(x,Γ(x),Γ′(x),...) = 0. This result has been reproved and generalized over the years by many researchers (for example, [4, 13, 19, 30, 31, 40]; see also [41]). Most recently, Hardouin ([16] and [17]) (and subsequently van der Put, see the appendix to [17]) proved and generalized this result using the Galois theory of difference equations (as developed in particular in [37]). This Galois theory associates a linear algebraic group to a linear difference equation and, using properties of linear algebraic groups, Hardouin was able to derive her results. In this paper we develop a general Galois theory of difference and differential equations where the Galois groups are linear differential groups, that is groups of matrices whose entries lie in a differential field and satisfy a set of polynomial differential equations. This general theory encompasses the usual Galois theory of linear differential equations [38], the Galois theory of linear difference equations [37] and the Galois theory of parameterized differential equations [12]1. We will develop this theory in its full generality in Section 6.2. We will use this theory in a restricted setting, that is, when we wish to analyze the differential behavior of a solution of a linear difference equation and describe this restricted theory and the tools we need for subsequent sections in Section 2. In Section 3, we describe possible differential relations among solutions of linear difference equations. The key idea is that the form of possible differential relations among solutions of linear difference equations is determined by the form of the differential equations defining the associated Galois group. We begin by considering first order equations. We prove in our setting a general result which implies the following result (cf. Corollary 3.2 below). This result (and its q-analogue) already appears in Hardouin’s work ([17], Prop. 2.7). Let C(x) be the field of rational functions over the complex numbers and F the fieldof 1-periodicfunctions meromorphicon the complexplane. Leta (x),...,a (x) ∈ 1 n C(x) and let z (x),...,z (x) be functions, meromorphic on C (resp. C∗) such 1 n that z (x+1)−z (x) = a (x), i = 1,...,n. i i i The functions z (x),...,z (x) are differentially dependent over F(x) if and only 1 n if thereexists a nonzerohomogeneouslineardifferentialpolynomialL(Y ,...,Y ) 1 n with coefficients in C such that L(a (x),...,a (x)) = g(x+1)−g(x). 1 n 1 In [3], André develops a Galois theory that encompases the difference and differential Galois theories andconsidersthedifferentialGaloistheoryasalimitingcaseofthedifferenceGaloistheory. Ourtheorydoes not have that feature while André’s theory does not address the questions studied in this paper. Another approach to describing analytic properties of solutions of difference equations involving pseudogroups has been announced by Umemura in [48]. 3 We also give a similar result for difference equations of the form z(x+1) = a(x)z(x) and q-difference versions (as does Hardouin in [16]). These results can be considered as an analogue of the Kolchin-Ostrowski Theorem [27] which characterizes the possible algebraic relations among solutions of first order differential equations. This latter result follows (using the Picard-Vessiot Theroy) from a description of the algebraic subgroups of products of one dimensional linear algebraic groups. In our setting these results follow from a description of the differential algebraic subgroups of products of one dimensional linear algebraic groups in the same general way once the machinery of our Galois theory is established. The theorem of Hölder follows from these results. Continuing with first order equations, we use facts about solvable differential subgroups of GL to reprove and generalize (cf. 2 Propositions 3.5, 3.9, and 3.10 below) the following result of Ishizaki [24]. If a(x),b(x) ∈ C(x) and z(x) ∈/ C(x) satisfies z(qx) = a(x)z(x)+b(x), |q| =6 1 and is meromorphic on C, then z(x) is not differentially algebraic over G(x), where G is the field of q-periodic functions meromorphic on C∗. We then turn to higher order equations. Using facts about differential subgroups of simple algebraic groups, we can characterize differential relationships among solutions of difference equations whose difference Galois group is a simple, noncommutative, algebraic group (cf. Proposition 3.11 and Corollary 3.12). Using this and Roques’s [39] computation of the difference Galois groups of the q-hpergeometric equations, we can show, for example (cf. Example 3.14), Let y (x),y (x) be linearly independent solutions of the hypergeometric equation 1 2 2ax−2 x−1 y(q2x)− y(qx)+ y(x) = 0 a2x−1 a2x−q2 wherea ∈/ qZ anda2 ∈ qZ and|q| =6 1. Theny (x),y (x),y (qx) are differentially 1 2 1 independent over G(x), where G is the field of q-periodic functions meromorphic on C∗. In Section 4, we consider a special case of the problem of determining which linear differen- tialalgebraicgroupsoccurasGaloisgroups. InSection5,weconsidercertainparameterized families Y(qx,t) = A(x,t)Y(x,t) of difference equations and show that the associated con- nection matrix ([15]) is independent of t if and only if the Galois group we associate with this equation (a priori alinear differential algebraicgroup) is conjugateto a linear algebraic group. The paper ends with an appendix containing two subsections, In Section 6.1 we gather together some facts about rational solutions of difference equations that are used throughout the preceding sections and, as mentioned above, in Section 6.2 we present our general Galois theory of linear difference and differential equations. We wish to thank Carsten Schneider for references to the literature on finding rational solutions of difference equations as well as Daniel Bertrand for useful comments and advice. 4 2 Galois Theory In this section we will give the basic definitions and results needed in Sections 3 - 5. These results follow from a more general approach to the Galois theory of differential and difference equations that we present in Section 6.2, where complete proofs are also given. The parenthetical references indicate the relevant general statements and results from the appendix. Definition 2.1 (Definition 6.6) A σ∂-ring is a commutative ring R with unit together with an automorphism σ and a derivation ∂ satisfying σ(∂(r)) = ∂(σ(r)) ∀r ∈ R. A σ∂-field is defined similarly2. Examples 2.2 1. C(x),σ(x) = x+1,∂(x) = d , dx 2. C(x),σ(x) = qx, (q ∈ C\{0}),∂ = x d , dx 3. C(x,t),σ(x) = x+1,σ(t) = t,∂ = ∂ . ∂t For k a σ∂-field we shall consider difference equations of the form σ(Y) = AY, A ∈ GL (k) (1) n Definition 2.3 (Definition 6.10) A σ∂-Picard-Vessiot ring (σ∂-PV-ring) over k for equation (1) is a σ∂-ring R containing k satisfying: 1. R is a simple σ∂-ring, i.e., R has no ideals, other than (0) and R, that are invariant under σ and ∂ 2. There exists a matrix Z ∈ GL (R) such that σ(Z) = AZ. n 3. R is generated as a ∂-ring over k by the entries of Z and 1/det(Z), i.e., R = k{Z,1/det(Z)} ∂ Note that when ∂ is identically zero, this corresponds to the usual definition of a Picard- Vessiot extension for a difference equation. To prove existence and uniqueness of Picard- Vessiot extensions, one needs to assume that the field of σ-invariant elements of k is algebraically closed. In the case of σ∂-PV extensions, kσ = {c ∈ k | σ(c) = c} is a differential field with derivation ∂ and we need to assume that this field is differentially closed (see Section 9.1 of [12] for the definition and references.) With this assumption, we have Proposition 2.4 (Propositions 6.14 and 6.16) Let k be a σ∂-field with kσ a differentially closed field. There exists a σ∂-PV ring for (1) and it is unique up to σ∂-k-isomorphism. Furthermore, Rσ = kσ. 2 All fields considered in this paper are of characteristic 0. 5 Definition 2.5 The σ∂-Galois group Aut (R/k) of the σ∂-PV ring R (or of (1)) is σ∂ Aut (R/k) = {φ | φ is a σ∂-k-automorphism of R} . σ∂ Asintheusualtheoryoflineardifferenceequations, onceonehasselected afundamental solution matrix of (1) in R, this group may be identified with elements of GL (kσ). The n next result states that both Aut (R/k) and R have additional structure. σ∂ Theorem 2.6 (Propositions 6.18 and 6.24) Let k be a σ∂-field and assume that kσ is a differentially closed field. Let R = k{Z, 1 } , σ(Z) = AZ be a σ∂-PV extension of k. detZ ∂ 1. We may identify Aut (R/k) with the set of kσ-points of a linear ∂-differential alge- σ∂ braic group G defined over kσ. 2. R is a reduced ring and is the coordinate ring of a G-torsor V defined over k. The action of G on V induces an action of G(kσ) on R that corresponds to the action of Aut (R/k) on R under the above identification. σ∂ In the above result we use the terms “coordinate ring” and “G-torsor” in the differential sense (see Sections 4 and9.4 of [12]for definitions of these terms as well as other definitions, facts and references concerning linear differential algebraic groups.) In the Appendix, we will furthermore show (Lemma 6.8) that R = R ⊕ ... ⊕ R is the finite direct 0 t−1 sum of integral differential k-algebras R where σ : R → R isomorphically. i imodt i+1modt In particular, the quotient fields of the R all have the same ∂-differential transcendence i degree over k (see [28], Ch. II.10). Abusing language, we define this to be the ∂-differential dimension, ∂−dim. (R) of R over k. The above theorem implies that the ∂−dim. (R/k) k k is the same as the differential dimension of (the identity component) of G, ∂−dim. (G) C (see Proposition 6.26). Finally, we have the usual Galois correspondence (cf., Theorem 1.29 of [37]). We note that if R is a σ∂-PV-extension of k and K is the total ring of quotients, then any σ∂- k-automorphism of K must leave R invariant. Therefore one can identify the group of σ∂-k-automorphisms Aut (K/k) of K with the σ∂-Galois group Aut (R/k). σ∂ σ∂ Theorem 2.7 (Theorem 6.20) Let k and R be as in Theorem 2.6 and let K be the total ring of quotients of R. Let F denote the set of σ∂-rings F with k ⊂ F ⊂ K such that every non-zerodivisor of F is a unit in F and let G denote the set of ∂-differential subgroups of Aut (R/k). There is a bijective correspondence α : F → G given by α(F) = G(K/F) = σ∂ {φ ∈ Aut (K/k) | φ(u) = u ∀u ∈ F}. The map β : G → F given by β(H) = {u ∈ σ∂ K | φ(u) = u ∀φ ∈ H} is the inverse of α. In particular, an element of K is left fixed by all φ in Aut (K/k) if and only if it is in σ∂ k and for a differential subgroup H of Aut (K/k) we have H = Aut (K/k) if and only σ∂ σ∂ if KH = k. 6 In the next section we will need the following consequences of the above result. Let R = k{Z, 1 } be a σ∂-PV-ring where σ(Z) = AZ. We can consider the σ-ring S = detZ ∂ k[Z, 1 ] ⊂ R. Note that this ring is not necessarily closed under the action of ∂. In the detZ next result we show that S is the usual PV ring (as in [37]) for the difference equation σ(Y) = AY. To avoid confusion we will use the prefix “σ-” to denote objects (e.g., σ- PV-extension, σ-Galois group) from the Galois theory of difference equations described in [37]. Proposition 2.8 (Proposition 6.21) Let k and R be as in Theorem 2.6 and let S = k[Z, 1 ] ⊂ R. detZ 1. S is the σ-PV-extension of k corresponding to σ(Y) = AY, and 2. Aut (R/k) is Zariski dense in the σ-Galois group Aut (S/k). σ∂ σ The following result characterizes those difference equations whose σ∂-Galois groups are “constant”. Proposition 2.9 Let k and R be as in Theorem 2.6 and let C = kσ∂ = {c ∈ k | σ(c) = c and ∂(c) = 0}. The σ∂-Galois group Aut (R/k) ⊂ GL (kσ) is conjugate over kσ to a σ∂ n subgroup of GL (C) if and only if there exists a B ∈ gl (k) such that n n σ(B) = ABA−1 +∂(A)A−1 . In this case, there is a solution Y = U ∈ GL (R) of the system n σ(Y) = AY ∂(Y) = BY Proof. Assume that such a B exists. A calculation shows that σ(∂(Z)−BZ) = A(∂(Z)− BZ). Therefore ∂(Z)−BZ = DZ for some D ∈ gl (kσ). We therefore have ∂(Z)−(B + n D)Z = 0. For any φ ∈ Aut (R/k) we will denote by [φ] ∈ GL (kσ) the matrix such σ∂ Z n that φ(Z) = Z[φ] . We claim that [φ] ∈ GL (C). To see this note that ∂(φ(Z)) = Z Z n (B +D)Z[φ] +Z∂([φ] ) and φ(∂(Z)) = (B +D)Z[φ] . This implies that Z∂([φ] ) = 0 Z Z Z Z so ∂([φ] ) = 0. Z Now assume that there exists a D ∈ GL (kσ) such that D−1Aut (R/k)D ⊂ GL (C). For n σ∂ n φ ∈ Aut (R/k), let [φ] ∈ GL (kσ) once again be the matrix such that φ(Z) = Z[φ] . σ∂ Z n Z Let U = ZD. For any φ ∈ Aut (R/k) we have that φ(U) = Z[φ] D = ZD(D−1[φ] D). σ∂ Z Z Therefore φ(U) = U[φ] for some [φ] ∈ GL (C). This implies that B = ∂(U)U−1 is left U U n fixed by Aut (R/k) and so B ∈ gl (k). A calculation shows that σ(B) = σ(∂(U)U−1) = σ∂ n ABA−1 +∂(A)A−1. 7 3 Differential Relations Among Solutions of Differ- ence Equations InthissectionweshallshowhowtheGaloistheoryofSection2canbeusedtogivenecessary and sufficient conditions for solutions of linear difference equations to satisfy differential polynomial equations. 3.1 First order equations. The classical Kolchin-Ostrowski Theorem [27] implies that if k ⊂ K are differential fields with the same constants C and z ,...,z ∈ K with z′ = a ∈ k, then z ,...,z are 1 n i i 1 n algebraically dependent over k if and only if there exists a homogeneous linear polynomial L(Y ,...,Y ) with coefficients in C such that L(z ,...,z ) = f ∈ k or, equivalently, 1 n 1 n L(a ,...,a ) = f′, f ∈ k. Kolchin proved this using differential Galois theory and the 1 n fact that algebraic subgroups of (Cn,+) are precisely the vector subspaces. For difference equations, theanalogyofthe indefinite integrals above areindefinite sums, that is, elements y satisfying σ(y)−y = a and one has a similar result characterizing algebraic dependence. The following characterizes differential dependence among indefinite sums. Recall that if k ⊂ K are differential rings, we say that z ,...,z ∈ K are differentially dependent over k 1 n if there is a nonzero differential polynomial P ∈ k{Y ,...,Y } such that P(z ,...,z ) = 0 1 n 1 n (cf., [28], Ch. II). Proposition 3.1 Let k be a σ∂-field with kσ differentially closed and let S ⊃ k be a σ∂-ring such that Sσ = kσ. Let a ,...,a ∈ k and z ,...,z ∈ S satisfy 1 n 1 n σ(z )−z = a i = 1,...,n . i i i Then z ,...,z are differentially dependent over k if and only if there exists a nonzero 1 n homogeneous linear differential polynomial L(Y ,...,Y ) with coefficients in kσ and an 1 n element f ∈ k such that L(a ,...,a ) = σ(f)−f. 1 n Proof. Assuming there exists such an L, one sees that L(z ,...z )−f is left fixed by σ 1 n and so lies in kσ. This yields a relation of differential dependence over k among the z . i Now assume that the z are differentially dependent over k. Differentiating the relations i σ(z ) − z = a , one sees that the ∂-differential ring k{z ,...,z } generated by the z is i i i 1 n ∂ i a σ∂-ring so we may assume that S = k{z ,...,z } . Let M be a maximal σ∂-ideal in S 1 n ∂ and let R = S/M. R is a simple σ∂-ring. Furthermore, R = k{Z, 1 } where σ(Z) = AZ detZ ∂ 8 and 1 a 0 0 ··· 0 0 1 z 0 0 ··· 0 0 1 1  0 1 0 0 ··· 0 0   0 1 0 0 ··· 0 0  0 0 1 a ··· 0 0 0 0 1 z ··· 0 0  2   2  A =  0 0 0 1 ··· 0 0  and Z =  0 0 0 1 ··· 0 0       .. .. .. .. .. .. ..   .. .. .. .. .. .. ..   . . . . . . .   . . . . . . .       0 0 ··· ··· ··· 1 a   0 0 ··· ··· ··· 1 z  n n      0 0 ··· ··· ··· 0 1   0 0 ··· ··· ··· 0 1      with z being the image of z in R. The σ∂-Galois group Aut (R/k) is a differential i i σ∂ subgroup of (kσ,+)n. By assumption, the differential dimension of R over k is less than n, the differential dimension of (kσ,+)n and so Aut (S/k) is a proper differential subgroup of σ∂ (kσ,+)n. Cassidy has classified these groups ([9, 12]): the differential subgroups of (kσ,+)n areprecisely thezero sets ofsystems ofhomogeneouslineardifferential polynomialsover kσ. Therefore there exists a nonzero homogeneous linear differential polynomial L(Y ,...,Y ) 1 n with coefficients in kσ such that Aut (R/k) ⊂ {(c ,...,c ) ∈ (kσ,+)n | L(c ,...,c ) = 0}. σ∂ 1 n 1 n We claim that L(z ,...,z ) = f ∈ k. To prove this it is enough to show that this 1 n element is left fixed by Aut (R/k). Let φ ∈ Aut (R/k). We have that φ(L(z ,...,z ) = σ∂ σ∂ 1 n L(z + c ,...,z + c ) = L(z ....,z ) so the claim is proved. Finally we have that 1 1 n n 1 n L(a ,...,a ) = L(σ(z )−z ,...,σ(z )−z ) = σ(f)−f. 1 n 1 1 n n Theaboveresulthastheratherartificialassumptionthatkσ isdifferentiallyclosed. Nonethe- less, this result can be used to prove results about meromorphic functions. In the following, we denote by F the field of 1-periodic meromorphic functions, that is meromorphic functions f(x) on C such that f(x+1) = f(x) and let G be the field of q-periodic meromorphic functions, that is functions f(x), meromorphic on C∗ = C\{0} such that f(qx) = f(x), where q ∈ C,|q| =6 1. In the following we shall speak of homogeneous linear differential (j) polynomials L(Y ,...,Y ), that is, linear forms in the variables Y . When we substitute 1 n i elements a of a differential field (k,∂) for the variables Y , we will replace Y(j) with ∂j(a ). i i i i In particular, when we are considering the shift σ(x) = x + 1, we will use the derivation ∂ = d and when we consider q-difference equations we will use the derivation ∂ = x d . dx dx Corollary 3.2 Let a (x),...,a (x) ∈ C(x) and let z (x),...,z (x) be functions, mero- 1 n 1 n morphic on C (resp. C∗) such that z (x+1)−z (x) = a (x), (resp. z (qx)−z (x) = a (x)) for i = 1,...,n. i i i i i i The functions z (x),...,z (x) are differentially dependent over F(x) (resp. G(x)) if and 1 n only if there exists a nonzero homogeneous linear differential polynomial L(Y ,...,Y ) with 1 n coefficients in C such that L(a (x),...,a (x)) = g(x+1)−g(x) (resp. L(a (x),...,a (x)) = 1 n 1 n g(qx)−g(x)) for some g(x) ∈ C(x). 9 Proof. We will deal with the case of the shift and apply Proposition 3.1 with σ(x) = x + 1 and ∂ = d . The q-difference case is similar except that we have σ(x) = qx and dx ∂ = x d . Clearly if L(a (x),...,a (x)) = g(x + 1) − g(x) for some g(x) ∈ C(x), then dx 1 n L(z (x),...,z (x))−g(x) ∈ F so the z are differentially dependent over F(x). 1 n i Now assume that the z are differentially dependent over F(x). Note that F(x) is a σ∂ field i with σ(x) = x+1 and ∂ = d . Furthermore, the function f(x) = x is not algebraic over F dx since any polynomial equation over F satisfied by x would also be satisfied by x+n for all n ∈ Z. Let T = F(x){z (x),...,z (x)} . This is a σ∂-domain and Tσ = F. Let C be the 1 n ∂ differential closure of F and define a σ∂-structure on R = T ⊗ C via σ(t⊗c) = σ(t)⊗c F and ∂(t⊗c) = ∂(t)⊗c+t⊗∂(c). We note that Rσ = C. Letting k = C(x) and R as above, we apply Proposition 3.1. We can conclude that there exists a nonzero homogeneous linear differential polynomial L˜(Y ,...,Y ) with co- 1 n efficients in C such that L˜(a (x),...,a (x)) = g˜(x+ 1)− g˜(x) for some g˜(x) ∈ C(x). Re- 1 n place the coefficients of the variables in L˜(Y ,...,Y ) by indeterminates to get a differ- 1 n ential polynomial L¯ with indeterminate coefficients. Replace the coefficients of powers of x in g˜ with indeterminates to get a rational function g¯(x) with indeterminate coefficients. We know that there is a specialization in C of the indeterminates so the equation L¯(a (x),...,a (x)) = g¯(x+1)−g¯(x) is satisfied. By clearing denominators and equating 1 n like powers of x, we see that this latter equation is equivalent to a system of polynomial equations in the indeterminates. Since C is algebraically closed and this system has a solution in some extension, it has a solution in C. Specializing to this solution, yields an L and g that satisfies the conclusion of the corollary. There is also a multiplicative version of the above corollaries. Again, we note that in the shift case ∂ = d and for q-difference equations, ∂ = x d . dx dx Corollary 3.3 Let b (x),...,b (x) ∈ C(x) and let u (x),...,u (x) be nonzero functions, 1 n 1 n meromorphic on C (resp. C∗) such that u (x+1) = b (x)u (x), (resp. u (qx) = b (x)u (x)) for i = 1,...,n. i i i i i i The functions u (x),...,u (x) are differentially dependent over F(x) (resp.G(x)) if and 1 n only if there exists a nonzero homogeneous linear differential polynomial L(Y ,...,Y ) with 1 n coefficientsin Csuch thatL(∂(b1(x)),..., ∂(bn(x))) = g(x+1)−g(x)(resp. L(∂(b1(x)),..., ∂(bn(x))) (b1(x) bn(x) (b1(x) bn(x) = g(qx)−g(x)) for some g(x) ∈ C(x).. Proof. Again, we only prove the corollary in the case of the shift. Let z (x) = u′i(x). Since i ui(x) the domain F(x){u ,...,u } is differentially algebraic over F(x){z ,...,z } , standard 1 n ∂ 1 n ∂ facts concerning differential transcendence degree imply that the z are differentially de- i pendent over F(x) if and only if the u are differentially dependent over F(x). The z i i satisfy b′(x) z (x+1)−z (x) = i . i i b (x) i 10