Integral Transforms and Special Functions
ISSN: 1065-2469 (Print) 1476-8291 (Online) Journal homepage: https://tandfonline.com/loi/gitr20
A note on the asymptotic expansion of the Lerch's transcendent
Xing Shi Cai & José L. López
To cite this article: Xing Shi Cai & José L. López (2019) A note on the asymptotic expansion of the Lerch's transcendent, Integral Transforms and Special Functions, 30:10, 844-855, DOI: 10.1080/10652469.2019.1627530 To link to this article: https://doi.org/10.1080/10652469.2019.1627530
© 2019 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group
Published online: 10 Jun 2019.
Submit your article to this journal
Article views: 208
View related articles
View Crossmark data
Full Terms & Conditions of access and use can be found at https://tandfonline.com/action/journalInformation?journalCode=gitr20 INTEGRAL TRANSFORMS AND SPECIAL FUNCTIONS 2019, VOL. 30, NO. 10, 844–855 https://doi.org/10.1080/10652469.2019.1627530
Research Article A note on the asymptotic expansion of the Lerch’s transcendent
Xing Shi Caia and José L. Lópezb aDepartment of Mathematics, Uppsala University, Uppsala, Sweden; bDepartamento de Estadística, Matemáticas e Informática and INAMAT, Universidad Pública de Navarra, Pamplona, Spain
ABSTRACT ARTICLE HISTORY In Ferreira and López [Asymptotic expansions of the Hurwitz–Lerch Received 21 March 2019 zeta function. J Math Anal Appl. 2004;298(1):210–224], the authors Accepted 1 June 2019 derived an asymptotic expansion of the Lerch’s transcendent KEYWORDS (z s a) |a| a > s > z ∈ C \ ∞) , , for large , valid for 0, 0 and [1, .In Hurwitz–Lerch zeta function; this paper, we study the special case z ≥ 1 not covered in Ferreira and asymptotic expansion; López [Asymptotic expansions of the Hurwitz–Lerch zeta function. J special functions Math Anal Appl. 2004;298(1):210–224], deriving a complete asymp- totic expansion of the Lerch’s transcendent (z, s, a) for z > 1 and AMS CLASSIFICATION s > 0as a goes to infinity. We also show that when a is a posi- 11M35 tive integer, this expansion is convergent for z ≥ 1. As a corollary, m zn/ns we get a full asymptotic expansion for the sum n=1 for fixed z > 1asm →∞. Some numerical results show the accuracy of the approximation.
1. Introduction The Lerch’s transcendent (Hurwitz–Lerch zeta function) [1, §25.14(i)] is defined by means of the power series ∞ zn (z, s, a) = , a = 0, −1, −2, ..., (a + n)s n=0 on the domain |z| < 1foranys ∈ C or |z|≤1for s > 1. For other values of the variables z,s,a,thefunction (z, s, a) is defined by analytic continuation. In particular2 [ ], ∞ − − 1 xs 1 e ax (z, s, a) = dx, a > 0, z ∈ C \ [1, ∞) and s > 0. (1.1) ( ) − −x s 0 1 z e This function was investigated by Erdélyi [3, § 1.11, Equation (1)]. Although using a dif- π ferent notation z = e2 ix, it was previously introduced by Lerch [4]andLipschitz[5] in connection with Dirichlet’s famous theorem on primes in arithmetic progression. If x ∈ Z, the Hurwitz–Lerch zeta function reduces to the meromorphic Hurwitz zeta func- tion ζ(s, a) [6, § 2.3, Equation (2)], with one single pole at s = 1. Moreover, ζ(s,1) is nothing but the Riemann zeta function ζ(s).
CONTACT Xing Shi Cai [email protected], [email protected] © 2019 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivatives License (http://creativecommons.org/licenses/by-nc-nd/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited, and is not altered, transformed, or built upon in any way. INTEGRAL TRANSFORMS AND SPECIAL FUNCTIONS 845
Properties of the Lerch’s transcendent have been studied by many authors. Among other π results, we remark the following ones. Apostol obtains functional relations for (e2 ix, s, a) π and gives an algorithm to compute (e2 ix, −n, a) for n ∈ N intermsofacertainkindof π generalized Bernoulli polynomials [7]. The function (e2 ix, s, a) is used in [8]togen- eralize a certain asymptotic formula considered by Ramanujan. Asymptotic equalities for π some weighted mean squares of (e2 ix, s, a) are given in [9]. Integral representations, as well as functional relations and expansions for (z, s, a) may be found in [6, § 2.5]. See Erdélyi et al. [10] for further properties. Here we want to remark the following two impor- tant properties of the Lerch’s transcendent valid for x, z, s ∈ C, m ∈ N [1, § 25.14.3 and § 25.14.4]: ∞ − 1 zn 1 (z, s,1) = Lis (z) := , |z| < 1, z ns n=1 where Lis(z) is the polylogarithm function [1, § 25.12], and
m−1 zn (z, s, x) = zm (z, s, x + m) + , −x ∈/ N ∪{0} (x + n)s n=0
In particular, when x = 1,thesecondpropertymaybewrittenintheform
m n z m+1 η(z, s, m) := = Lis(z) − z (z, s, m + 1) . (1.2) ns n=1
The finite sum η(z, s, m) for z > 1isofinterestinthestudyofrandom records in full binary trees by Janson [11]. In a full binary tree, each node has two child nodes and each level m+1 of the tree is full. Thus Tm, a full binary tree of height m,hasn = 2 − 1nodes.Inthe random records model, each node in Tm isgivenalabelchosenuniformlyatrandomfrom the set {1, ..., n} without replacement. A node u is called a record when its label is the smallest among all the nodes on the path from u to the root node. Let h(u) be the distance from u to the root. Let X(Tm) be the (random) number of records in Tm.Thenitiseasyto seethattheexpectationofX(Tm) is simply m+1 + 1 2i η (2, 1, m + 1) 2m 1 2m n n = = = + O = + O , h(u) + 1 i + 1 2 m m2 m m2 u∈Tn i=0 (1.3) where the last step follows from elementary asymptotic computations [11, Remark 1.3]. A generalization of random records, called random k-cuts, requires a similar computation which boils down to finding an asymptotic expansion of η(2, b/k, m) for some k ∈ N and 1 ≤ b ≤ k as m →∞,see[12, § 5.3.1]. Or more generally, asymptotic expansions of the function Li (z) F(z, s, a) := (z, s, a) − s , (1.4) za that generalizes the function η(z, s, m − 1) defined in (1.2), from integer to complex val- ues of the variable m: η(z, s, m − 1) =−zmF(z, s, m). Complete asymptotic expansions, 846 X. S. CAI AND J. L. LÓPEZ including error bounds, of (z, s, a) for large a have been investigated in [2]. In particular, for a > 0, s > 0andz ∈ C \ [1, ∞), we have that, for arbitrary N ∈ N [2, Theorem 1],
N−1 ( ) s n −N−s (z, s, a) = cn(z) + O a , (1.5) an+s n=0 | |→∞ ( ) = ( + )...( + − ) as a .Inthisformula, s n : s s 1 s n 1 is the Pochhammer symbol, −1 c0(z) = (1 − z) and, for n = 1, 2, 3, ..., n (−1) Li−n(z) cn(z) := . (1.6) n! From the identities (1.2) and (1.5) we have that, for all N ∈ N, z ∈/ [1, ∞) and s > 0where expansion (1.5) is valid,
− m N1 ( ) z s n −N η(z, s, m − 1) = Lis(z) − cn(z) + O m , (1.7) ms mn n=0 as m →∞. Unfortunately, expansion (1.5) has not been proved for z ∈ [1, ∞),andthen,in principle, the above expansion of η(z, s, m − 1) does not hold in the domain of the variable z where the approximation of η(z, s, a) has a greater interest. Had the expansion (1.5) been proved for z ≥ 1, the asymptotic computations in (1.3) would have become unnecessary and an arbitrarily precise approximation could be achieved automatically from (1.7). However, to our surprise, it seems that the expansion (1.7) is still valid when z > 1. An argument that supports this claim is the following. On the one hand, assuming for the moment that (1.7) holds for z = 2, then m + − η(2, −1, m) = n2n = (m − 1)2m 1 + 2 + O m N , N ∈ N. n=1 On the other hand, using summation by parts [13,pp.56],itiseasytoseethat m + η(2, −1, m) = n2n = (m − 1)2m 1 + 2. n=1 Thus expansion (1.7) seems to be correct for z = 2. Numerical experiments further suggest that this is also true for other values of z > 1. Then, the purpose of this paper is to show that expansion (1.7) holds for z > 1. More generally, to derive an expansion of F(z, s, a) for large a with s > 0andz ≥ 1.
2. An expansion of (z, s, a) for large a and z ≥ 1 The main result of the paper is given in Theorem 2.3. In order to formulate Theorem 2.3, we need to consider the function − − 1 − (z e x)1 a f (z, x, a) := , (2.1) 1 − z e−x and its Taylor coefficients Cn(z, a) at x = 0. We also need the two following Lemmas. INTEGRAL TRANSFORMS AND SPECIAL FUNCTIONS 847
Lemma 2.1: For a, z ∈ C and n = 0, 1, 2, ...,
n ( ) 1−a cn−k z k Cn(z, a) := cn(z) − z pn(z, a), pn(z, a) := (a − 1) , (2.2) k! k=0
−1 where, for z = 1, the coefficients cn(z) have been introduced in (1.6): c0(z) = (1 − z) and, for n = 1, 2, 3, ..., n (−1) Li−n(z) cn(z) := . (2.3) n!
For n = 1, 2, 3, ..., the coefficients cn(z) maybecomputedrecursivelyintheformncn(z) = − ( ) ( ) zcn−1 z .Observethatpn z, a arepolynomialsofdegreeninthevariablea. For z = 1, Definition (2.2) must be understood in the limit sense. More precisely, C0(1, a) = 1 − aand, for n = 1, 2, 3, ...,
n Bn+1 − Bn+1(a − 1) − (n + 1)(a − 1) Cn(1, a) = , (2.4) (n + 1)! where Bn are the Bernoulli numbers and Bn(a) the Bernoulli polynomials [14, §24.2].
Proof: Formulas (2.2)–(2.3) may be derived by combining the Taylor expansions at x = 0 ( − ) − − of e a 1 x and (1 − z e x) 1. Formula (2.4) follows from the generating function of the Bernoulli polynomials Bn(a) [13, Equation (7.81)],
( − ) ∞ − e a 1 x xn 1 = Bn(a) , 1 − e−x n! n=0 the definition of the Bernoulli numbers Bn = Bn(0),andtheTaylorexpansion − −x(1−a) = ( ) =− ( ) of 1 e at x 0. The derivation of the recursion ncn z zcn−1 z is straightforward [2].
Apart from the explicit form of the coefficients Cn(z, a) given above, we may compute them by means of the recurrence relation given in the following lemma.
1−a Lemma 2.2: For z = 1 we have that C0(z, a) = (1 − z )/(1 − z) and, for n = 1, 2, 3, ...,
n−1 − z (−1)n k (a − 1)n Cn(z, a) = C (z, a) − . 1 − z (n − k)! k n!za k=0
For z = 1, we have C0(1, a) = 1 − a, and, for n = 1, 2, 3, ...,
+ n−1 − (a − 1)n 1 (−1)n k Cn(1, a) =− − C (1, a). (n + 1)! (n + 1 − k)! k k=0 848 X. S. CAI AND J. L. LÓPEZ
Proof: Replace ∞ ∞ k − (−x) e x = and f (z, x, a) = C (z, a)xk k! k k=0 k=0 − − − into the identity [1 − z e x]f (z, x, a) = 1 − (z e x)1 a and equate the coefficients of equal powers of x.
Theorem 2.3: For fixed N ∈ N, a > 1 and s > 0,
N−1 Lis(z) (s)n F(z, s, a) := (z, s, a) − = Cn(z, a) + RN(z, s, a), (2.5) za an+s n=0 with Cn(z, a) given in the previous lemmas and, for z > 1, 1−N−s − a RN(z, s, a) = O ( a) + az , a →∞. (2.6)
This means that expansion (2.5) has an asymptotic character for large awhenz> 1.More- over, expansion (2.5) is convergent for a = m = 2, 3, 4, ...and z ≥ 1;i.e.RN(z, s, m) → 0 as N →∞and − ∞ m1 k 1 z (s)n F(z, s, m) :=− = Cn(z, m) , z ≥ 1. (2.7) zm ks mn+s k=1 n=0
Remark 2.1: Note that, both (z, s, a) and Lis(z) are multivalued functions of z.Inour analysis and numeric experiments, we always choose the principle branch for the vari- able z. Thus (2.5) can be transformed into an asymptotic expansion of (z, s, a) without ambiguity as follows:
N−1 Lis(z) (s)n (z, s, a) = + Cn(z, a) + RN(z, s, a). za an+s n=0
Proof: Using the integral representation (2) of (z, s, a) givenin[2] and the integral representation [1, §25.12.11] of the polylogartithm, ∞ − z xs 1 Li (z) = dx, s ( ) x − s 0 e z we find the following integral representation of the function F(z, s, a) defined in (1.4): ∞ Li (z) 1 − − F(z, s, a) := (z, s, a) − s = xs 1 e axf (z, x, a) dx, (2.8) a ( ) z s 0 valid for a > 0, s > 0andz ∈ C,withf (z, x, a) given in (2.1). In principle, the left-hand side of (2.8) is an analytic function of z in {z : |z| < 1; | arg z| <π}.Then,theright-hand side of this equation defines the analytic continuation of F(z, s, a) in the variable z to the cut complex plane C \ (−∞,0]. INTEGRAL TRANSFORMS AND SPECIAL FUNCTIONS 849
The function f (z, x, a) has the following Taylor expansion at x = 0:
n−1 k f (z, x, a) = Ck(z, a)x + rn(z, x, a), (2.9) k=0 where the coefficients Cn(z, a) are given in (2.2)–(2.4). Replacing the function f (z, x, a) in the integral (2.8) by its Taylor expansion (2.9), and interchanging sum and integral, we obtain (2.5) with ∞ ( ) = 1 s−1 −ax ( ) Rn z, s, a : ( ) x e rn z, x, a dx. (2.10) s 0
Using the Cauchy integral formula for the remainder rn(z, x, a),wefind − − xn f (z, w, a) 1 − (z e w)1 a r (z, x, a) = dw, f (z, w, a) := , (2.11) n π ( − ) n − −w 2 i C w x w 1 z e where we choose C tobeaclosedloopthatencirclesthepointsw = 0andw = x,itis traversed in the positive direction, and is inside the region U :={w ∈ C, w > −W, | w| < W}, for some arbitrary but fixed W ∈ (0, 2π).(U is an infinity rectangular region around the positive real line [0, ∞) of width 2W,see[2] for further details.) Figure 1 gives an example of U and C. The function f (z, w, a) is continuous in the variable w for w ∈ U.Thesingularitiesofthis function are w = log z + 2iπn, n ∈ Z \{0} and are located outside U.Defineb := a and β := a − b and write − − − −β − −β − − −β 1 − (z e w)1 a 1 − (z e w) + (z e w) − (z e w)1 b f (z, w, a) = = 1 − z e−w 1 − z e−w −w −β −w 1−b 1 − (z e ) − −β 1 − (z e ) = + (z e w) . (2.12) 1 − z e−w 1 − z e−w We have that − − b−2 − − − − − b−1 k 1 − (z e w)1 b (z e w) k − (z e w) k 1 ew = =− . (2.13) 1 − z e−w 1 − z e−w z k=0 k=1
Figure 1. The integration loop C and the region U. 850 X. S. CAI AND J. L. LÓPEZ
Thefollowingboundsarevalidforw ∈ U and a certain constant M0 > 0 independent of b and w:
− ( −w)−β 1 z e −w −β β w , (z e ) ≤ M0 e ,for0≤ β<1, 1 − z e−w b−1 k ew e w ≤ (b − 1) ,for w ≤ log z, z z (2.14) k=1 b−1 k b−1 ew e w ≤ (b − 1) ,for w ≥ log z. z z k=1
Therefore, from (2.12), (2.13) and (2.14) we have that, for any w ∈ U and a > 1,