How Mathematica and Maple Get

Meijer's G-function into Problem 9

Folkmar Bornemann, Feb 8, 2005
in memoriam John Boersma (1937-2004)
Asking Wolfram Research, Inc.
Stan's inquiry to the WRI team as to how Mathematica manages to get Meijer's G-function into play, was
answered by Daniel Lichtblau on J uly 27, 2003, in the following rather cryptic way:
I ' l l be di scussi ng aspect s of Mei j er G i ssues at ACA next
week. I t i s basi cal l y a l ookup t hat conver t s var i ous
f unct i ons t o Mei j er G, t hen f i gur es out t he i nt egr al of a
pr oduct of 2 Mei j er G' s vi a Sl at er convol ut i on.
This was not detailed enough to help our understanding of how this works.
Mathematica Reference Guide
The Notes on Internal Implementation reveal the following knowledge:
Many ot her def i ni t e i nt egr al s ar e done usi ng Mar i chev- -
Adamchi k Mel l i n t r ansf or m met hods. The r esul t s ar e of t en
i ni t i al l y expr essed i n t er ms of Mei j er G f unct i ons, whi ch
ar e conver t ed i nt o hyper geomet r i c f unct i ons usi ng Sl at er ' s
Theor emand t hen si mpl i f i ed.
The problem is similar as above: these keywords are not sufficient to get reasonable pointers to some
literature. But there are some names mentioned which could be traced.
The Relevant Literature
An extensive search with Google Scholar (great stuff indeed) has led me to the following relevant items:
[Mari83] Oleg Marichev, Handbook of Integral Transforms of Higher
Transcendental Functions: Theory and Algorithmic Tables, Ellis Horwood Ltd.,
Chichester, 1983.
[Adam96] Victor Adamchik, Definite Integration in Mathematica, Mathematica
in Education and Research, no. 3, 5 (1996), 1622.
According to [Adam96] the ideas of [Mari83] are implemented in Mathematica. Marichev's involvement
with Maple suggests that the same is true for this CAS, too.
The General Idea
Marichev's method is based on the observation that any definite integral can be written in the form of
Mellin's convolution of two functions, that is
f HzL =

0

f
1
HxL f
2
J
z

x
N
x

x
.
Now, this is a convolution since the Mellin transforms of these functions, that is,
f
*
HsL =

0

f HxL x
s-1
x, f
j
*
HsL =

0

f
j
HxL x
s-1
x, j = 1, 2,
simply multiply
f
*
HsL = f
1
*
HsL f
2
*
HsL.
Of course, generally this observation will be of not much use. However, if the functions are of
hypergeometric type, which is true for many elementary functions and the majority of special functions, the
inverse Mellin transform of f ,
f HzL =
1

2p

f
*
HsL z
-s
s,
turns out to be a Mellin-Barnes integral. (One has to exercise some care in choosing a path of integration.)
Depending on the involved coefficients, this may represent Fox's H-function, or in simpler cases, Meijer's
G-function. Finally, Lucy Slater's theorem of 1966 (see [Mari83, p. 57]) states precise conditions on those
coefficients such that Meijer's G-function reduces to a hypergeometric series, and therefore, in many cases
can be simplified to more familiar special functions. In fact, this situation of a reduction to a hypergeometric
series turns out to be the generic case.
In the non-reducible, non-generic case, one can still evaluate the Mellin-Barnes integral by means of the
residue theorem yielding an infinite series involving polygamma functions. This way Maple 9.5 completely
avoids delivering Meijer's G-functions.
Let's Do It: Problem 9
We want to use this information to understand what Mathematica is doing when performing the following
evaluation:
J = SimplifyA

0
2
H2 - xL
a
sinJ
a

x
N x, 8a > 0<E
!!!!
p GHa + 1L G
2,4
3,0
i
k
j
j
j
j
jj
a
2

16

a+2

2
,
a+3

2
1

2
,
1

2
, 1, 0
y
{
z
z
z
z
zz
With z = 1 and
f
1
Hx_L = x H2 - xL
a
If@0 x 2, 1, 0D; f
2
Hx_L = sinHa xL;
the integral is obviously of the required convolution type. The Mellin transforms of f
1
and f
2
are given by
2 MeijerG.nb
f
1
*
Hs_L = SimplifyA

f
1
HxL x
s-1
x, 8a > 0, ReHsL > -1<E
2
s+a+1
GHs + 1L GHa + 1L

GHs + a + 2L
andby the reflection formula of the gamma function
f
2
*
Hs_L = SimplifyA

f
2
HxL x
s-1
x .
i
k
j
jjsinHp t_L
p

GHtL GH1 - tL
y
{
z
zz, 8a > 0, 0 < ReHsL < 1<E
p a
-s
GHsL

GH1-
s

2
L GH
s

2
L
Both transforms, which could also have been obtained by a short calculation by hand, are of the required
hypergeometric form.
Thus, the inverse Mellin transform yields the Mellin-Barnes integral
J = f H1L =
1

2p

p 2
a+1
GHa + 1L GHs + 1L GHsL

GHs + a + 2L GHs 2L GH1- s 2L
J
a

2
N
-s
s .
This is not yet recognizable as a Meijer's G-function, since in the definition of these functions the variable s
can only enter the gamma function terms with a coefficient 1. Therefore we apply the transform s 2s
and obtainby using the duplication formula for the gamma functionthe transformed integrand as:
2 f
1
*
HsL f
2
*
HsL . Hs 2 sL .
i
k
j
j
j
j
j
GHt : 2 g_ +d_: 0L
1

!!!!
p
2
t-1
G
i
k
j
j
t

2
y
{
z
z G
i
k
j
j
j
t +1

2
y
{
z
z
z
y
{
z
z
z
z
z
Simplify
16
s
!!!!
p a
-2s
GHs +
1

2
L
2
GHs + 1L GHa + 1L

GH1- sL GHs +
a

2
+ 1L GH
1

2
H2s + a + 3LL
Summarizing we get
J = f H1L =
!!!!
p G Ha + 1L
i
k
j
j
j
j
j
jj
1

2p

GHs +
1

2
L
2
GHs + 1L

GHs +
a+2

2
L GHs +
a+3

2
L GH1- sL

i
k
j
j
j
a
2

16
y
{
z
z
z
-s
s
y
{
z
z
z
z
z
zz
,
where the large parenthesis is nothing but the definition of Meijer's G-function,
J =
!!!!
p G Ha + 1L G
2,4
3,0
i
k
j
j
j
j
j
a
2

16

a+2

2
,
a+3

2
1

2
,
1

2
, 1, 0
y
{
z
z
z
z
z
.
(It does not reduce to a hypergeometric series because of the double appearance of the coefficient
1

2
.)
Thus, the approach is very systematic while involving a minimumof calculation.
MeijerG.nb 3