Completely Monotonic Gamma Ratio and

Infinitely Divisible H-Function of Fox

D. B. Karp & E. G. Prilepkina

Computational Methods and

Function Theory

ISSN 1617-9447
Volume 16
Number 1

Comput. Methods Funct. Theory (2016)

16:135-153
DOI 10.1007/s40315-015-0128-9

1 23
Your article is protected by copyright and
all rights are held exclusively by Springer-
Verlag Berlin Heidelberg. This e-offprint is
for personal use only and shall not be self-
archived in electronic repositories. If you wish
to self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at link.springer.com”.

1 23
 Author's personal copy
Comput. Methods Funct. Theory (2016) 16:135–153
DOI 10.1007/s40315-015-0128-9

Completely Monotonic Gamma Ratio and Infinitely

Divisible H-Function of Fox

D. B. Karp1,3 · E. G. Prilepkina1,2

Received: 28 January 2015 / Revised: 17 April 2015 / Accepted: 10 May 2015 /

Published online: 22 July 2015
© Springer-Verlag Berlin Heidelberg 2015

Abstract We investigate conditions for logarithmic complete monotonicity of a quo-

tient of two products of gamma functions, where the argument of each gamma function
has a different scaling factor. We give necessary and sufficient conditions in terms of
non-negativity of some elementary functions and some more practical sufficient condi-
tions in terms of parameters. Further, we study the representing measure in Bernstein’s
theorem for both equal and non-equal scaling factors. This leads to conditions on para-
meters under which Meijer’s G-function or Fox’s H -function represents an infinitely
divisible probability distribution on the positive half-line. Moreover, we present new
integral equations for both G-function and H -function. The results of the paper gen-
eralize those due to Ismail (with Bustoz, Muldoon and Grinshpan) and Alzer who
considered previously the case of unit scaling factors.

Keywords Gamma function · Completely monotonic functions ·

Meijer’s G-function · Fox’s H -function · Infinite divisibility

Mathematics Subject Classfication 33B15 · 33C60

Communicated by Stephan Ruscheweyh.

B D. B. Karp
dmkrp@yandex.ru
B E. G. Prilepkina
pril-elena@yandex.ru

1 Far Eastern Federal University, 8 Sukhanova street, Vladivostok 690950, Russia

2 Institute of Applied Mathematics, FEBRAS, 7 Radio Street, Vladivostok 690041, Russia
3 Universidad del Atlántico, Km 7 Antigua via, Puerto Colombia, Colombia

123
 Author's personal copy
136 D. B. Karp, E. G. Prilepkina

1 Introduction

Recall that a non-negative function f defined on (0, ∞) is called completely

monotonic (c.m.) if it has derivatives of all orders and (−1)n f (n) (x) ≥ 0 for n ≥ 1
and x > 0 [18, Def. 1.3]. This inequality is known to be strict unless f is a constant.
By the celebrated Bernstein theorem, a function is completely monotonic if and only
if it is the Laplace transform of a non-negative measure [18, Thm. 1.4]. The above
definition implies the following equivalences

f is c.m. on (0, ∞) ⇔ f ≥ 0 and − f  is c.m. on (0, ∞)

⇔ − f  is c.m. on (0, ∞) and lim f (x) ≥ 0. (1)
x→∞

A positive function f is said to be logarithmically completely monotonic (l.c.m.)

if −(log f ) is completely monotonic [18, Def. 5.8]. According to (1)

f is l.c.m. on (0, ∞) ⇔ (− log f (x)) ≥ 0 and (log f ) is c.m. on (0, ∞)

⇔ (log f ) is c.m. on (0, ∞) and lim (− log f (x)) ≥ 0.
x→∞
(2)

The class of l.c.m. functions is a proper subset of the class of c.m. functions. Their
importance stems from the fact that they represent Laplace transforms of infinitely
divisible probability distributions, see [18, Thm. 5.9] and [17, Sec. 51].
The study of complete monotonicity of the ratio


p
(x + ai )
U (x) = , (3)
(x + bi )
i=1

where  stands for Euler’s gamma function and p ≥ 1 has been initiated by Bustoz
and Ismail who demonstrated in their 1986 paper [4] that for p = 2, a1 = 0 and
a1 + a2 = b1 + b2 this function is l.c.m. on (0, ∞). Eight years later,  Ismail and
p
Muldoon showed in [7] that U (x) is l.c.m. on (0, ∞) for general p if i=1 ai =
p
i=1 bi and b1 = b2 = · · · = b p > 0, ai ≥ 0, or b1 = b2 = · · · = b p−1 = 0,
b p > 0, ai ≥ 0. Their result is, in fact, formulated for the ratio of q-gamma functions,
but the proof works for U (x) just as well. The subject was further pursued by Alzer
[1] who showed in 1997 that U (x) is l.c.m. on (0, ∞) if

0 ≤ a1 ≤ a2 ≤ · · · ≤ a p , 0 ≤ b1 ≤ b2 ≤ · · · ≤ b p ,

k 
k
(4)
and ai ≤ bi for k = 1, 2 . . . , p.
i=1 i=1

These inequalities are known as weak supermajorization [12, Def. A.2] and are abbre-
viated as b ≺W a, where a = (a1 , . . . , a p ), b = (b1 , . . . , b p ). In their 2006 paper,
[5] Grinshpan and Ismail found new sufficient conditions for U (x) to be l.c.m. when

123
 Author's personal copy
Completely Monotonic Gamma Ratio and Infinitely Divisible… 137

p = 2n or p = n!/2, n ≥ 1. We will explain and slightly generalize their results

in the next section. Finally, in 2009 Guo and Qi [6] used another approach to inves-
tigate logarithmic complete monotonicity of U (x) for arbitrary real values of ai , bi .
Their results, however, lead to complete monotonicity of U (x) on some subinterval
of (0, ∞) of the form (γ , ∞), where γ > 0.
The purpose of the present paper is to investigate complete monotonicity of the
ratio p
(Ai x + ai )
W (x) = qi=1 , (5)
j=1 (B j x + b j )
where A = (A1 , . . . , A p ) and B = (B1 , . . . , Bq ) are strictly positive scaling factors,
while a = (a1 , . . . , a p ) and b = (b1 , . . . , bq ) are non-negative. We find a necessary
and sufficient condition for W to be l.c.m. on (0, ∞) in terms of non-negativity of some
function which, unfortunately, is not easy to verify. Further, we supply several simple
sufficient conditions for such non-negativity in terms of the vectors A, B, a, b as well as
some necessary conditions. When W is completely monotonic, we proceed by deriving
its representing measure in the Bernstein theorem which leads us to studying some new
properties of Fox’s H -function. We begin, however, by revisiting the ratio defined in
(3) which we call the unweighted case. In the following section, we refine some of the
known results for U (x) and discuss its representing measure in Bernstein’s theorem.

2 The Unweighted Case Revisited

All proofs of complete monotonicity of U (x) we are aware of are based on the integral
representation of the logarithmic derivative of the gamma function [11, p. 16]
 ∞
  (z) e−t − e−t z
ψ(z) = = −γ + dt, (6)
(z) 0 1 − e−t

where γ is Euler–Mascheroni constant. The next two properties of the digamma func-
tion ψ will be repeatedly used in the sequel

∞
 te−xt
ψ (x) = dt, x > 0, (7)
1 − e−t
0

and
1
ψ(z) = log(z) − + O(z −2 ) as z → ∞, (8)
2z
see [2, Thm. 1.6.1,Cor. 1.4.5].
Representation
 p (6) lead Grinshpanand Ismail to formulate [5, Lem. 2.1] stating
that q(x) = k=1  σk (x + λk ) with k=1 σk = 0 and λk ≥ 0 is l.c.m. on (0, ∞) if
p

and only if


p
v(t) = σk t λk ≥ 0 for t ∈ (0, 1].
k=1

123
 Author's personal copy
138 D. B. Karp, E. G. Prilepkina

In fact, looking at the asymptotic expansion of −(log q(x)) as x → ∞ it is easy to

sharpen
p this lemma as follows: q(x) is l.c.m. on (0, ∞) iff v(t) ≥ 0 on (0, 1] and
k=1 σ k = 0. When q(x) = U (x), where U is defined as in (3), we have

p
U (x) is l.c.m. on (0, ∞) ⇔ v(t) = (t ak − t bk ) ≥ 0 on (0, 1]. (9)
k=1

As mentioned above Alzer noticed that v(t) ≥ 0 on (0, 1] under majorization condi-
tions (4) which follows from the 1949 result of Tomić [12, Prop. 4.B.2]. Let us remark
that for p = 2 conditions 0 ≤ min(a1 , a2 ) ≤ min(b1 , b2 ) and a1 + a2 ≤ b1 + b2 ,
equivalent to (4), are necessary and sufficient for non-negativity of v(t). Necessity
follows by considering the asymptotics of v(t) as t → 0 and t → 1. A proof of a
more general result is given in Corollary 1 below.
Grinshpan and Ismail took another path in [5] and considered two types of factor-
ization
 
n
v(t) = (t α j − t αi ) and v(t) = (1 − t αi ).
1≤i< j≤n i=1

It clear that for α1 > α2 > · · · > αn > 0 for the first factorization and αi > 0 for
the second we get expressions non-negative on (0, 1]. The corresponding value of p
is n!/2 for the first factorization and 2n−1 for the second. In both cases, the authors of
[5] found explicit combinatorial descriptions of the vectors a, b in (3) which lead to
the above factorizations. We would like to remark here that taking v in the form


n
β 
v(t) = t i − t αi
i=1

with αi ≥ βi ≥ 0 for i = 1, . . . , n certainly results in non-negative v on (0, 1].

This, of course,
can be reduced to the second factorization of Grinshpan and Ismail by
factoring out t βi . Also one can get the first type of factorization by properly choosing
βi . The vector a (b) is expressed in [5] in terms of sum over even (odd) permutations
of the numbers αi for the first factorization, and in terms of sums of the numbers αi
indexed by components of increasing integer vectors for the second factorization. In
the theorem below which generalizes both factorizations, the description of the vectors
a and b in terms of the numbers αi and βi turns out to be different and simpler.
Theorem 1 Let I = {1, 2, . . . , n} and suppose Iodd ( Ieven ) comprises all subsets of
I having odd (even) number of elements, ∅ ∈ Ieven . Suppose that αi ≥ βi ≥ 0 for
i = 1, . . . , n. Then
  

J ∈Ieven  x + i∈J αi + i∈I \J βi

U (x) =   

J ∈Iodd  x + i∈J αi + i∈I \J βi

is l.c.m. on (0, ∞).

123
 Author's personal copy
Completely Monotonic Gamma Ratio and Infinitely Divisible… 139

Proof We will write I = In , Iodd = Iodd (n), Ieven = Ieven (n) to emphasize the
dependence on n. According to (9) it suffices to verify that v(t) ≥ 0 on (0, 1], where
   
 αi + βi  αi + βi
v(t) = t i∈J i∈In \J
− t i∈J i∈In \J
.
J ∈Ieven (n) J ∈Iodd (n)

We will demonstrate that


n
v(t) = (t βi − t αi )
i=1

by induction in n. Indeed, the case n = 1 is obvious. The induction step now follows
from the identity
   
 αi + βi  αi + βi +βn
i∈In−1 \J
t i∈J i∈In \J
= t i∈J
J ∈Ieven(n) J ∈Ieven (n−1)
 
 αi +αn + βi
i∈In−1 \J
+ t i∈J
,
J ∈Iodd (n−1)

and a similar identity for the negative term in v(t). Non-negativity is now obvious
from αi ≥ βi ≥ 0 for i = 1, . . . , n.

The next example generalizes [5, Cor. 1.4].

Example 1 For αi ≥ βi ≥ 0 for i = 1, 2, 3 the function

U (x)
(x + β1 + β2 + β3 )(x + β1 + α2 + α3 )(x + α1 + β2 + α3 )(x + α1 + α2 + β3 )
=
(x + α1 + β2 + β3 )(x + β1 + α2 + β3 )(x + β1 + β2 + α3 )(x + α1 + α2 + α3 )

is l.c.m. on (0, ∞).

Further, changing variable in [16, 2.24.2.1] or in particular case of [9, Thm. 2.2]
we get
p   
(x + ai ) −t x p,0 −t b
= e G p, p e dt (10)
(x + bi ) a
i=1 (0,∞)
p p,0
provided that i=1 (bi − ai ) > 0. Here G p, p is Meijer’s G-function defined by the
contour integral
  
p,0 b 1 (a1 +s) . . . (a p +s) −s
G p, p z = z ds.
a 2πi (b1 +s) . . . (b p +s)
L

123
 Author's personal copy
140 D. B. Karp, E. G. Prilepkina

This function is a particular case of Fox’s H -function defined in (21) below on setting
p = q and Ai = Bi = 1, i = 1, . . . , p in that definition. See the paragraph following
ofp the contour L. It follows from an expansion due to Nørlund
(21) for a description
[14, (2.28)] that for i=1 (bi − ai ) = 0 formula (10) must be modified as


p   
(x + ai ) b
e−t x δ0 + G p, p e−t
p,0
= dt, (11)
(x + bi ) a
i=1 [0,∞)

where δ0 denotes the unit mass concentrated at zero. The representation (10) was
previously observed by us in [8].
The l.c.m ⇒ c.m. implication mentioned above translates into the next assertion:

p  
a  p,0 b
v(t) = t k − t bk ≥ 0 on [0, 1) ⇒ G p, p x ≥ 0 on (0, 1)
a
k=1

p
under the additional conditions ai ≥ 0, i = 1, . . . , p, and i=1 (bi − ai ) > 0. The
following much stronger assertion is supported by numerical evidence:

p
Conjecture 1 If ai ≥ 0, i = 1, . . . , p and i=1 (bi − ai ) > 0 then

   
p,0 b
# zeros of G p, p x on (0, 1) ≤ # zeros of v(t) on (0, 1) .
a

We think that this conjecture is encoded in the following integral equation for Meijer’s
G-function which we believe to be new.


p
Theorem 2 Suppose a, b ≥ 0 and (bi −ai ) > 0. Then Meijer’s G-function satisfies
i=1
the following integral equation:

     p  
p,0 b 1
p,0 b  x ak x bk dt
log(1/x)G p, p x = G p, p t − b
a x a t a k t k t−x
k=1

for 0 < x < 1.

We postpone the proof of this result until the proof of Theorem 8, which contains a
more general statement. Using Theorem 2 we were also able to demonstrate that if
p,0
G p, p (x) has a zero on (0, 1) then v(t) also has at least one zero on this interval.

123
 Author's personal copy
Completely Monotonic Gamma Ratio and Infinitely Divisible… 141

3 Generalized Gamma Ratio

For the positive vectors A = (A1 , . . . , A p ), B = (B1 , . . . , Bq ) and non-negative

vectors a = (a1 , . . . , a p ), b = (b1 , . . . , bq ), consider the positive function
p
(Ai x + ai )
W (x) = qi=1
j=1 (B j x + bj)

defined on (0, ∞). Our goal is to study logarithmic complete monotonicity of W . It

is convenient to start with the following result.
Lemma 1 The function (log W ) is c.m. if and only if


p
e−ai u/Ai 
q
e−bi u/Bi
P(u) = −u/A
− ≥ 0 for all u > 0. (12)
1−e i 1 − e−u/Bi
i=1 i=1

In the affirmative case, we have

∞

(log W ) (x) = e−xu u P(u)du. (13)
0

Proof Differentiating (6) we have


p 
q
(log W ) (x) = Ai2 ψ  (Ai x + ai ) − B 2j ψ  (B j x + b j ).
i=1 j=1

Using (7) and making change of variable u = Ai t or u = Bi t in the appropriate

integrals we get

  
te−(Ai x+ai )t dt  2
p ∞ q ∞ te−(B j x+b j )t dt
(log W ) (x) = Ai2 − Bj
0 1 − e−t 0 1 − e−t
i=1 j=1
p  ∞ q  ∞

ue−xu−uai /Ai du ue−xu−ub j /B j du
= −
0 1 − e−u/Ai 1 − e−u/B j
i=1 j=1 0
⎧ ⎫
 ⎨  e−b j u/B j ⎬
∞ p q
e−ai u/Ai
= e−xu u − du
0 ⎩ 1 − e−u/Ai 1 − e−u/B j ⎭
i=1 j=1
 ∞
= e−xu u P(u)du.
0

The integral converges around zero due to asymptotic formula (19) below. Conver-
gence at infinity is obvious. Hence, (log W ) is c.m. iff P(u) ≥ 0 by Bernstein’s
theorem [18, Thm. 1.4].


123
 Author's personal copy
142 D. B. Karp, E. G. Prilepkina

Remark 1 By substituting t = e−u condition (12) can be also written in the form


p
t ai /Ai  t b j /B j
q
Q(t) = − ≥ 0 for t ∈ (0, 1). (14)
1 − t 1/Ai 1 − t 1/B j
i=1 j=1

If (log W ) is completely monotonic two interesting possibilities appear: (log W )

> 0 and (log W ) < 0. These two possibilities are handled in the next two theorems.
Recall that a non-negative function f on (0, ∞) is called Bernstein function if f  is
c.m. [18, Def. 3.1and below].

Theorem 3 Suppose A, B > 0 and a, b ≥ 0. Then the function


p 
q
(log W ) (x) = Ai ψ(Ai x + ai ) − B j ψ(B j x + b j )
i=1 j=1

is a Bernstein function iff condition (12) holds and

 p q 
lim Ai ψ(Ai x + ai ) − B j ψ(B j x + b j ) ≥ 0.
x→0 i=1 j=1

Proof The proof is immediate from Lemma 1 and the following equivalences:

f is Bernstein function ⇔ f  is c.m. and f ≥ 0 on (0, ∞)

⇔ f  is c.m. on (0, ∞) and lim f (x) ≥ 0. (15)
x→0

Theorem 4 The function W is l.c.m. if and only if


q 
p 
p 
q
−B j
Bj = Ai , ρ = AiAi Bj ≤1 (16)
j=1 i=1 i=1 j=1

and condition (12) holds. In the affirmative case

 ∞
− (log W ) (x) = e−xu P(u)du + log(1/ρ). (17)
0

Proof We have


q 
p
−(log W ) (x) = B j ψ(B j x + b j ) − Ai ψ(Ai x + ai ).
j=1 i=1

123
 Author's personal copy
Completely Monotonic Gamma Ratio and Infinitely Divisible… 143

Further it follows from (8) that

 
1 1
ψ(C z + c) = log z + log C + c − + O(z −2 ) as z → ∞.
2 Cz

Substituting the latter formula into the former, we get

⎧ ⎫ ⎧ ⎫
⎨ q p ⎬ ⎨ q p ⎬
− (log W ) (x) = log(x) Bj − Ai + B j log(B j ) − Ai log(Ai )
⎩ ⎭ ⎩ ⎭
j=1 i=1 j=1 i=1
⎧ ⎫
1 ⎨  p−q⎬
q p
+ bj − ai + + O(x −2 ) as x → ∞. (18)
x⎩ 2 ⎭
j=1 i=1

If conditions (12) and (16) hold then

lim (− log W (x)) = − log ρ ≥ 0.

x→∞

Then Lemma 1 and relations (2) imply that W is l.c.m. In the opposite direc-
tion, if W is l.c.m.
then −(logW ) is c.m. and hence decreasing which in view of
p q
(18) implies that i=1 Ai ≥ i=1 Bi . On the other hand, relations (2) show that
p q
lim x→∞ (− log W (x)) ≥ 0 so that i=1 Ai ≤ i=1 Bi again by (18). This proves
the first condition in (16). Once this condition is satisfied, non-negativity of the limit
leads to the second condition in (16). Finally, integrating (13) we get

 x ∞  x
 
(log W (x)) = (log W (x)) dx + log ρ = u P(u)du e−tu dt + log ρ
∞ ∞
0
∞
=− e−xu P(u)du − log(1/ρ).
0


p
Remark 2 Expression of the form SA = i=1 Ai log(Ai ) for positive numbers Ai is
known as Shannon’s entropy in information theory, so that the second condition in
(16) can be restated as “entropy of the vector A does not exceed that of the vector B.”
We collect the necessary conditions for logarithmic complete monotonicity of W
in the next corollary.
Corollary 1 The following conditions are necessary for W to be logarithmically com-
pletely monotone:
q p
(a) j=1 B j = i=1 Ai
(b) ρ ≤ 1 p
q
(c) μ = j=1 b j − i=1 ai + 21 ( p − q) ≥ 0

123
 Author's personal copy
144 D. B. Karp, E. G. Prilepkina

(d) min (ai /Ai ) ≤ min (b j /B j )

1≤i≤ p 1≤ j≤q

Proof Necessity of (a) and (b) has been demonstrated in Theorem 4. Next, straight-
forward computation shows that
⎧ ⎫ ⎧ ⎫
1 ⎨  ⎬ ⎨  ⎬
p q q p
1
P(u) = Ai − Bj + bj − ai + ( p − q) +O(u) as u → 0.
u⎩ ⎭ ⎩ 2 ⎭
i=1 j=1 j=1 i=1
(19)
Since the first sum is zero by (a) we conclude that (c) is necessary for non-negativity
of P(u). Further, using representation (14), we easily compute

Q(t) = t α (1 + o(1)) − t β (1 + o(1)) as t → 0,

where α = min (ai /Ai ) and β = min (b j /B j ). Hence, we conclude that condition
1≤i≤ p 1≤ j≤q
(d) is also necessary.

Given the vectors A, B, a, b condition (12) is not easy to verify. The next theorem
provides some practical sufficient conditions.
Theorem
q 5 Inequality
 p (12) is true if any of the following conditions holds:
(a) j=1 B j = i=1 Ai and max (ai /Ai ) ≤ min (b j − 1)/B j ;
p  p 1≤i≤ p 1≤ j≤q
(b) p = q, i=1 Bi = i=1 Ai with Ai ≥ Bi for i = 1, . . . , p − 1, and

max bk /Bk ≤ (b p − 1)/B p , ai /Ai ≤ (bi − 1)/Bi for i = 1, . . . , p;

1≤k≤ p−1

(c) p = q, 0 ≤ a1 /A1 ≤ a2 /A2 ≤ · · · ≤ a p /A p ,

0 ≤ b1 /B1 ≤ b2 /B2 ≤ · · · ≤ b p /B p ,
0 < 1/A1 ≤ 1/A2 ≤ · · · ≤ 1/A p ,
k k
0 < 1/B1 ≤ 1/B2 ≤ · · · ≤ 1/B p and i=1 ai /Ai ≤ i=1 bi /Bi ,
k k
i=1 1/Ai ≤ i=1 1/Bi for each integer k = 1, . . . , p.

Proof (a) Denote ai = ai /Ai , Ai = 1/Ai , bi = bi /Bi , Bi = 1/Bi , x = eu . Then by
the mean value theorem we have

p
e−ai u/Ai p
e−ai u p
1 p
1
= = =  x ci log x
1 − e−u/Ai −Ai u ai ai −Ai
i=1 1 − e i=1 x − x
A
i=1 i=1 i

1 
p
= Ai x −ci ,
log x
i=1

where ci ∈ (ai − Ai , ai ). Hence, for P(u) we get

⎧ ⎫
1 ⎨  ⎬
p q
P(u) = Ai e−ci u − B j e−d j u , (20)
u⎩ ⎭
i=1 j=1

123
 Author's personal copy
Completely Monotonic Gamma Ratio and Infinitely Divisible… 145

where d j is some point of the interval (bj − B j , B j ). Denote c = max ci , d = min d j .

Suppose max(ai /Ai ) ≤ min(b j − 1)/B j , which is equivalent to max(ai ) ≤ min(bj −
B j ). Then c < d and non-negativity of P(u) follows from the relations


p 
p 
p 
q 
q
Ai x −ci ≥ Ai x −c ≥ Ai x −d = B j x −d ≥ B j x −d j .
i=1 i=1 i=1 j=1 j=1


p 
p
(b) Denote S = Bj = Ai . Non-negativity of P(u) follows from represen-
j=1 i=1
tation (20) and the following chain

⎛ ⎞

p 
p 
p−1 
p−1
Ai e−ci u ≥ Ai e−di u = Ai e−di u + ⎝S − Ai⎠ e−d p u
i=1 i=1 i=1 i=1


p−1
= Ai (e−di u − e−d p u ) + Se−d p u
i=1

p−1
≥ Bi (e−di u − e−d p u ) + Se−d p u
i=1
p
= Bi e−di u .
i=1

The first inequality here is due to conditions ai /Ai ≤ (bi − 1)/Bi , i = 1, . . . , p

which imply that ci < di . The second inequality follows from conditions Ai ≥ Bi ,
i = 1, . . . , p−1, combined with inequality max1≤k≤ p−1 bk /Bk ≤ (b p −1)/B p which
ensures that each term e−di u − e−d p u is non-negative.
(c) Expanding each expression like (1−e−u/Ai )−1 in geometric series and exchang-
ing the order of summations we get

p
∞ 


P(u) = e−ai u/Ai −lu/Ai − e−bi u/Bi −lu/Bi .

l=0 i=1

Conditions (c) guarantee that for any l ≥ 0, u ≥ 0 we have

(b1 u/B1 + lu/B1 , . . . , b p u/B p + lu/B p )

≺W (a1 u/A1 + lu/A1 , . . . , a p u/A p + lu/A p )

123
 Author's personal copy
146 D. B. Karp, E. G. Prilepkina

(see (4) for the definition of weak supermajorization ≺W ). Hence, according to [12,
Prop. 4.B.2] we conclude that

p


e−ai u/Ai −lu/Ai − e−bi u/Bi −lu/Bi ≥ 0

i=1

since x → e−x is convex and decreasing.



Remark
p 3 Conditions
p (c) of Theorem 5 are only compatible with the condition
=  W 
A
j=1 j j=1 j if A = B (up to permutation). Indeed, majorization B ≺ A
B
(recall that Aj = A−1  −1
j , B j = B j ) forms a part of conditions (c). According to [12,
Prop. 4.B.2] this majorization implies that


p 
p
1 
p
1 
p
Bj = ≤ = Aj,
B j Aj
j=1 j=1 j=1 j=1

because the function x → 1/x is decreasing and convex. Further, [12, 3.A.6.a] says
that for continuous strictly decreasing functions, the above inequality is strict unless
A is a permutation of B . This brings us to the conclusion that for A = B (modulo
permutations) conditions (c) can be used to check whether (log W ) is a Bernstein
function using Theorem 3, but they cannot be used to check whether W is com-
pletely monotone. On the other hand, if A = B, conditions (c) reduce to checking
the majorization a ≺W b (a j = a j A−1  −1
j , b j = b j A j ). This majorization is weaker
than (a) or (b) in this particular situation. One reason why the case A = B might be
important is applications in probability as explained in Remark 4 below.

4 The Representing Measure

We will need a particular case of Fox’s H -function defined by

    p
 (B, b) 1 (Ak s + ak ) −s
p,0
Hq, p 
z = qk=1 z ds, (21)
(A, a)
j=1 (B j s + b j )
2πi
L

where Ak , B j > 0 and ak , b j are real. The contour L can be either the left loop L−
starting at −∞ + iα and ending at −∞ + iβ for some α < 0 < β such that all poles
of the integrand lie inside the loop; or the right loop L+ starting at ∞ + iα and ending
at ∞ + iβ and leaving all poles on the left; or the vertical line Lic , z = c, traversed
upward and leaving all poles of the integrand on the left. Denote the rightmost pole of
the integrand by γ :

γ = − min (ak /Ak ).

1≤k≤ p

123
 Author's personal copy
Completely Monotonic Gamma Ratio and Infinitely Divisible… 147

Recall the definition of ρ from (16) and the definition of μ from Corollary 1(c):


p 
q
−B 
q 
p
p−q
ρ= AkAk Bj j , μ= bj − ak + .
2
k=1 j=1 j=1 k=1

Existence conditions of H -function under each choice of the contour L have been
thoroughly considered in the book [9]. Under the restrictions (16) and x > 0 Theo-
p,0
rem 1.1 from [9] states that Hq, p (x) exists if we choose L = L+ or L = Lic under
the additional restriction μ > 1. Only the second choice of the contour ensures the
p,0
existence of the Mellin transform of Hq, p (x) as demonstrated in [9, Thm. 2.2]. In our
next theorem, we relax the condition μ > 1 to μ > 0 and demonstrate that the first
p,0
condition in (16) leads to the compact support of Hq, p (x).
p q
Theorem 6 Suppose μ > 0 and i=1 Ai = j=1 B j . Then the integral over the
p,0
contour Lic with c > γ in the definition of Hq, p (x) converges for x > 0 except
possibly for x = ρ and   
 (B, b)
Hq, p x 
p,0
=0 (22)
(A, a)
for x > ρ. Moreover, under these restrictions, the Mellin transform exists for s > γ
and
    p
ρ  (B, b) (Ak s + ak )
x 
p,0
Hq, p x s−1 dx = qk=1 .
0 (A, a) (B j s + b j )
j=1

p q
Proof By applying Stirling’s asymptotic formula and in view of k=1 Ak = j=1 Bj
it is not difficult to derive the formula [15, (2.2.4)]
p
(Ak s + ak )
qk=1 = Aρ s s −μ + ρ s g(s), (23)
j=1 (B j s + bj)

with g(s) = O(s −μ−1 ) as |s| → ∞ uniformly in | arg s| ≤ π −  for any 0 <  < π
and
p a −1/2
q 1/2−b j
A = (2π )( p−q)/2 Ak k Bk .
k=1 j=1

This asymptotic behavior implies that t → g(c + it) is an absolutely integrable

continuous function on the real line for any c > γ so that the integral v(x) =
 c+i∞ −s
2πi c−i∞ (x/ρ) g(s)ds exists and we are in the position to apply the Mellin inver-
1

sion theorem yielding

 ∞
x s−1 v(x)d x = ρ s g(s).
0

123
 Author's personal copy
148 D. B. Karp, E. G. Prilepkina

Denote


c+i∞
1
h(x) = Aρ s s −μ x −s ds.
2πi
c−i∞

An application of [3, (6),§12] after a simple rearrangement gives

⎧
+∞ ⎨ A  μ−1
A elog(ρ/x)(c+it) log ρx , 0 < x < ρ,
h(x) = μ
dt = (μ)
2π (c + it) ⎩ 0, x > ρ.
−∞

It then follows from (23) that

    p
c+i∞
 (B, b) 1 (Ak s + ak ) −s
x 
p,0
Hq, p = qk=1 x ds = h(x) + v(x)
(A, a)
j=1 (B j s + b j )
2πi
c−i∞

and the above integral exists for all x > 0, x = ρ. Direct computation gives

∞    ρ
 (B, b) A ρ
μ−1 s−1
x 
p,0
x s−1
Hq, p dx = log x dx + ρ s g(s)
(A, a) (μ) x
0 0
s −μ
= Aρ s + ρ g(s),
s

which is equivalent to

∞    p
 (B, b) (Ak s + ak )
x 
p,0
x s−1 Hq, p dx = qk=1 .
(A, a) (B j s + b j )
j=1
0

Next we show that (22) holds so that the integration in the above formula is in fact
over the interval (0, ρ). To this end, we apply an argument similar to Jordan’s lemma.
According to the definition (21)

    Rp
c+i
 (B, b) 1 (Ak s + ak ) −s
p,0
Hq, p 
z = lim qk=1 z ds.
(A, a)
j=1 (B j s + b j )
R→∞ 2πi
c−i R

By Cauchy’s theorem, the last integral is equal to the integral along the right semicircle
of radius R centered at c. Hence, we need to prove that

π/2  p
R (Ak c + Ak Reiϕ + ak ) −(c+R cos ϕ+i R sin ϕ) log x+iϕ
IR = qk=1 e dϕ
j=1 (B j c + B j Re + b j )
2π iϕ
−π/2

123
 Author's personal copy
Completely Monotonic Gamma Ratio and Infinitely Divisible… 149

goes to zero as R → ∞ for x > ρ. Setting s = Reiϕ in (23) we get

p
(Ak c + Ak Reiϕ + ak )
= Aρ Re (Reiϕ )−μ (1 + O(R −1 )) as R → ∞.
iϕ
qk=1
j=1 (B j c + Bj Reiϕ + bj)

Hence, for sufficiently large R and some constant M1

π/2
M1 e−c log x x
|I R | ≤ R −μ+1 e−R log ρ cos ϕ dϕ.
2π
−π/2

Employing the elementary inequality cos ϕ ≥ 1 − π2 ϕ, 0 < ϕ < π/2, we obtain the
following estimate with some constant M2 (recall that x > ρ):

π/2 1 − e −R log ρx
2 x
|I R | ≤ 2M2 R −μ+1 e−R(1− π ϕ) log ρ dϕ = π M2 R −μ .
log ρx
0

Hence, lim R→∞ I R = 0 which completes the proof of the theorem.



Theorem 7 Suppose that μ > 0 and

p
(Ai x + ai )
W (x) = qi=1
j=1 (B j x + bj)

is l.c.m. Then
   
∞
−t x p,0 
−t  (B, b)
W (x) = e Hq, p e  dt,
log(1/ρ) (A, a)

and the H -function in the integrand is non-negative. In particular, the conclusion is

true if (16) and the hypotheses of Theorem 5 are satisfied.

 p q to Theorem 4 logarithmic complete monotonicity of W (x) implies

Proof According
A
k=1 k = j=1 B j , so that we are in the position to apply Theorem 6 yielding the
formula
p    
(Ai x + ai ) ρ  (B, b)
W (x) = qi=1 =
p,0
Hq, p 
u u x−1 du.
(A, a)
j=1 (B j x + bj) 0

The claimed Laplace transform representation for W (x) follows by substitution u =

e−t . Non-negativity of Hq, p (u) on (0, ρ) follows from the Bernstein theorem in view
p,0

of the uniqueness of the measure with given Laplace transform, see [18, Prop. 1.2] or
[19, Thm. 6.3]. 

123
 Author's personal copy
150 D. B. Karp, E. G. Prilepkina

Theorem 7 requires μ to be strictly positive, while the necessary conditions from

Corollary 1 allow for μ = 0. In analogy with (11) in that case we believe the following
conjecture is valid.

Conjecture 2 For μ = 0 the representing measure is given by

   
−t x p,0 
−t  (B, b)
W (x) = e Aδlog(1/ρ) + Hq, p e  dt,
[log(1/ρ),∞) (A, a)

where A is a positive constant and δlog(1/ρ) denotes the unit mass concentrated at the
point log(1/ρ).

Unlike the case of the G-function (11), no expansion similar to Nørlund’s is avail-
able for the H -function. We plan to study its behavior in the neighborhood of the
singular point log(1/ρ) in our future work.

Corollary 2 Suppose μ > 0, a, b > 0 and conditions (12) and (16) are satisfied (in
particular, (12) is true under Theorem 5). Then the function
q   
j=1 (b j )  (B, b)
Hq, p e−t 
p,0
t → p (24)
(A, a)
i=1 (ai )

represents an infinitely divisible probability density on (0, ∞).

Proof The result follows from Theorem 7 combined with [18, Def. 5.6] and [18,
Thm. 5.9]. See also [10, Thm. 7].


Remark 4 The case A = B is also connected to probability as follows. Suppose ζk ,

k = 1, . . . , p are independent beta-distributed random variables, so that ζk has the
density

(αk + βk ) αk
t (1 − t)βk , t ∈ (0, 1).
(αk )(βk )
p Ak
Then the random variable u = k=1 ζk with Ak > 0 has the following moments


p
(αk + βk )  (Ak x + αk − Ak )
p
E(u x−1 ) = , x > 0.
(αk ) (Ak x + αk + βk − Ak )
k=1 k=1

The probability density of u is expressed via Fox’s H -function:


p   
(αk + βk ) p,0  (A j , α j + β j − A j ), j = 1, . . . , p

H p, p z 
(αk ) (A j , α j − A j ), j = 1, . . . , p
k=1

See [13, Sec. 4.2.1].

123
 Author's personal copy
Completely Monotonic Gamma Ratio and Infinitely Divisible… 151

p q
Theorem 8 Suppose A, B > 0, a, b ≥ 0, μ > 0, ρ ≤ 1 and i=1 Ai = j=1 Bj.
Then for all x ∈ (0, ρ) the following identity holds:
    1   
 (B, b) 1 p,0 x  (B, b) Q(u)
x  
p,0
Hq, p = Hq, p  (A, a) du, (25)
(A, a) log(ρ/x) x/ρ u u

where Q(u) is defined in (14).

Proof Since − log(ρ) = SB − SA , where SA (SB ) stands for the entropy of A (B), we
adopt the notation S = SB − SA = − log(ρ). Further, denote

1, t ≥ S
I (t) = ,
0, t < S

and define the (signed) measure

dν(u) = Sδ0 + P(u)du,

where δ0 is a unit mass at zero. It follows from the proof of Theorem 4 and the
asymptotic formula (19) that the representation (17) is true under the conditions of the
theorem. In terms of the measure dν (17) takes the form:

W  (s) = W (log W ) = −W e−su dν(u).
[0,∞)

On the other hand, under the hypotheses of Theorem 6 we have the representation
   
∞  (B, b)
W (s) = e −ts −t
H (e )I (t)dt, where H (x) =
p,0
Hq, p 
x .
0 (A, a)

Hence,
 ∞
W  (s) = − te−ts H (e−t )I (t)dt.
0

Putting these representations together yields

 ∞  ∞ 
te−ts H (e−t )I (t)dt = e−ts H (e−t )I (t)dt e−su dν(u).
0 0 [0,∞)

The convolution theorem for the Laplace transform [19, Thm. 11.4] then leads to the
formula

t H (e−t )I (t) = H (e−(t−τ ) )I (t − τ )dν(τ ) for t ∈ (S, ∞).
[0,t)

123
 Author's personal copy
152 D. B. Karp, E. G. Prilepkina

Recalling the definition of dν, we can rewrite the above relation as

 t−s
−t −t
t H (e ) = S H (e ) + H (e−t+τ )P(τ )dτ
0

or
 t−S
1
H (e−t ) = H (e−t+τ )P(τ )dτ.
t − S 0

Substituting x = e−t and u = e−τ yields identity (25).



Theorem 4 shows that under additional restrictions (16) we have the implication
  
 (B, b)
x 
p,0
Q(t) ≥ 0 on [0, 1) ⇒ Hq, p ≥ 0 on (0, ρ).
(A, a)

The following much stronger assertion is supported by numerical evidence:

p q
Conjecture 3 Suppose A, B > 0, a, b ≥ 0, μ ≥ 0, ρ ≤ 1 and i=1 Ai = j=1 B j .
Then
    
 (B, b)
# zeros of Hq, p x 
p,0
on (0, ρ) ≤ # zeros of Q(t) on (0, 1) .
(A, a)

Using Theorem 8, we were able to demonstrate that this conjecture is true for the case
when the left-hand side is equal to one.

Acknowledgments We thank the anonymous referee for correcting a mistake in the original version of
the paper and numerous useful suggestions that helped to improve the exposition substantially. This work
has been supported by the Russian Science Foundation under project 14-11-00022.

References
1. Alzer, H.: On some inequalities for the gamma and psi functions. Math. Comput. 66(217), 373–389
(1997)
2. Andrews, G.E., Askey, R., Roy, R.: Special Functions. Cambridge University Press, Cambridge (1999)
3. Bochner, S.: Lectures on Fourier Integrals. Princeton University Press, Princeton (1959)
4. Bustoz, J., Ismail, M.E.H.: On gamma function inequalities. Math. Comp. 47, 659–667 (1986)
5. Grinshpan, A.Z., Ismail, M.E.H.: Completely monotonic functions involving the gamma and q-gamma
functions. Proc. Am. Math. Soc. 134(4), 1153–1160 (2005)
6. Guo, B.N., Qi, F.: Properties and applications of a function involving exponential functions. Commun.
Pure Appl. Anal. 8(4), 1231–1249 (2009)
7. Ismail, M.E.H., Muldoon, M.E.: Inequalities and monotonicity properties for gamma and q-gamma
functions. In: Zahar, R.V.M. (ed.) Approximation and Computation: A Festschrift in Honor of
Walter Gautschi, ISNM, vol. 119, pp. 309–323. Birkhäuser, Boston (1994). Corrected version:
arXiv:1301.1749v1
8. Karp, D., Prilepkina, E.: Hypergeometric functions as generalized Stieltjes transforms. J. Math. Anal.
Appl. 393(2), 348–359 (2012)
9. Kilbas, A.A., Saigo, M.: H-Transforms and Applications, Analytical Methods and Special Functions,
vol. 9. Chapman & Hall/CRC, Boca Raton (2004)

123
 Author's personal copy
Completely Monotonic Gamma Ratio and Infinitely Divisible… 153

10. Koumandos, S.: On completely monotonic and related functions. In: Rassias, T.M., Pardalos, P.M.
(eds.) Mathematics Without Boundaries, pp. 285–321. Springer, New York (2014)
11. Magnus, W., Oberhettinger, F., Soni, R.P.: Formulas and Theorems for the Special Functions of Math-
ematical Physics. Springer, Berlin (1966)
12. Marshall, A.W., Olkin, I., Arnold, B.C.: Inequalities: Theory of Majorization and its Applications, 2nd
edn. Springer, New York (2011)
13. Mathai, A.M., Saxena, R.K., Haubold, H.J.: The H-Function: Theory and Applications. Springer, New
York (2010)
14. Nørlund, N.E.: Hypergeometric functions. Acta Math. 94, 289–349 (1955)
15. Paris, R.B., Kaminski, D.: Asymptotics and Mellin-Barnes Integrals. Cambridge University Press,
Cambridge (2001)
16. Prudnikov, A.P., Brychkov, YuA, Marichev, O.I.: Integrals and Series, Volume 3: More Special Func-
tions. Gordon and Breach Science Publishers, Philadelphia (1990)
17. Sato, K.: Lévy processes and infinitely divisible distributions. Cambridge University Press, Cambridge
(1999)
18. Schilling, R.L., Song, R., Vondraček, Z.: Bernstein Functions. Theory and Applications, Studies in
Mathematics, vol. 37. Walter de Gruyter, Berlin (2010)
19. Widder, D.W.: The Laplace Transform. Princeton University Press, Princeton (1946)

123