-1-

A simple proof of the Gaussian correlation conjecture

extended to multivariate gamma distributions
T. Royen
University of applied sciences Bingen, Berlinstrasse 109, D-55411 Bingen, Germany,
e-mail: [email protected]

Abstract
An extension of the Gaussian correlation conjecture (GCC) is proved for multivariate gamma distributions (in the
sense of Krishnamoorthy and Parthasarathy). The classical GCC for Gaussian probability measures is obtained by
the special case with an integer degree of freedom   1.

Keywords and phrases: probability inequalities, Gaussian correlation conjecture, multivariate gamma distribution
2010 Mathematics Subject Classification: 60E15

1. Introduction

Let P be a probability measure on n ,n  1, given by a Gaussian density (2 ) n /2 | Σ |1/2 exp( 12 xT Σ1x) with
a non-singular covariance matrix Σ . The Gaussian correlation conjecture (GCC) asserts the inequality
P(C1 C2 )  P(C1 ) P(C ) (1.1)

for all convex and central-symmetric sets C1 ,C2  n , (see [2]). The bivariate case was proved in [8]. Further
milestones towards a complete proof were the papers [12] and [4]. In [12] the GCC is verified for all centered ellip-
soids and for all sufficiently small C1 , C2 . In [4] a proof is given if only C1 is a centered ellipsoid. Recently, a com-
plete very long proof has been put into the arXiv.org server by Y. Memarian [7], which seems to be still under
examination. The here presented proof is totally different and comparatively short. According to [12] the GCC is
equivalent to
 n
  k   n 
P Ai   P  Ai  P  Ai  (1.2)
 i 1   i 1   i k 1 
with Ai  | X i |xi  ,x1 ,..., xn  0,1  k  n, and any Nn (0, Σ) -Gaussian vector ( X1 ,..., X n ). The inequality
(1.2) was independently proved for k  1 in [5] and [13]. Here, (1.2) is proved for ( X1 ,..., X n ) with an n-variate
gamma distribution ( n ( , R) -distribution in the sense of Krishnamoorthy and Parthasarathy [6]), defined by its La-
place transform (Lt)
| I n  RT | (1.3)

with the identity matrix I n , a non-singular correlation matrix R  ( ij ) and T  Diag (t1 ,..., tn ),t1 ,..., tn  0. Ad-
missible values for  are 2  , all values 2  n  2,n  2, (for n  2  2  n  1 see [10]) and all   0 if
the Lt is infinitely divisible, for which sufficient and necessary conditions are found in [1] and [3]. The n ( , R) -
distribution was originally derived from the joint distribution of the diagonal elements of a Wn (2 , R) -Wishart matrix
with the Lt | I n  2 RT | . The classical GCC is obtained from the special case   1/ 2. A proof of it could be
easily extended inductively for increasing degrees of freedom   2 by means of convolution integrals. However,
the here provided proof includes also non-integer values of 2 .

We need the non-central gamma-pdf

 -2-
 k
y
g ( x, y)  e y  g  k ( x) , , x  0, y  0, (1.4)
k 0 k!
with the central gamma-pdf g ( x)  (( ))1 x 1e x and the corresponding non-central gamma-cdf
 x
yk
G ( x, y )  e y  G  k ( x) ,G ( x)   g ( )d . (1.5)
k 0 k! 0

If R is represented by
R   I n  AAT  1R  I n  BBT ,B   1/2 A, (1.6)

with the minimal eigenvalue  of R and an n  (n  1) -matrix A of rank m  n  1, then the n ( , R) - pdf and the
n ( , R) - cdf, 2  or 2  n  2, can be represented by

 n 
f ( x1 ,..., xn ; , R)  E    1 g ( 1 x j , 12 b j SbTj  (1.7)
 j 1 
and
 n 
F ( x1 ,..., xn ; , R)  E   G ( 1 x j , 12 b j SbTj  (1.8)
 j 1 
respectively with the rows b j from B and the expectation referring to the Wn1 (2 , In1 ) - Wishart matrix S . These
formulas are derived in a slightly more general form in [9] and [10]. They can be verified by Lt.

2. Proof of the Gaussian correlation conjecture for n ( , R) - distributions

R R12 
Theorem 1. Let R   11  be a non-singular partitioned n  n -correlation matrix with ni  ni -submatrices
 R21 R22 
Rii and a positive rank of R12 . Then, it holds for the n ( , R) - cdf

F ( x1 ,..., xn ; , R)  F ( x1,..., xn1 ;  , R11 ) F ( xn1 1,..., xn ;  , R22 ) (2.1)

for all positive numbers x1 ,..., xn and 2  or 2  n  2.

Proof. All the matrices

 R  R12 
R   11  ,0    1, (2.2)
  R21 R22 

are non-singular correlation matrices. Theorem 1 will be proved if we show


F ( x1 ,..., xn ;  , R )  0,0    1. (2.3)

By means of the Lt we shall represent the lhs of (2.3) by a sum of not identically vanishing non-negative terms.

Let AJ denote the submatrices with row and column indices i  J  1,..., n from any n  n - matrix A. The deter-
minant | I n  R T | is equal to 1   J | R , J || TJ |,  J  1,..., n. With J  J1 J 2 ,J1  J 1,..., n1  ,
 RJ1  RJ1 , J 2 
J2  J n1  1,..., n  , R , J 
 RJ , J  ,rJ1 , J 2  rank ( RJ1 , J 2 ) and the canonical correlations  J1 , J 2 ,i
 2 1
RJ 2 
 -3-

belonging to RJ1 ,RJ2 ,RJ1 , J2 ,i  1,..., rJ1 , J2 , we find

 
rJ1,J 2 rJ1 , J 2
 J2 , J

| R , J | RJ1 || RJ 2 |

  
1   2  J21 , J 2 ,i  2 | R , J | 
i 1 1 2
1 2,

2
i
and
i 1 J1 , J 2 ,i

 rJ1 ,J2  J2 , J ,i 

 | R ,J |  1

 i 1 1   2  2
2
  t j  cJ ( ) t j
| I n  R T | 2
J  J ,J 1 2 ,i  jJ  J jJ
, (2.4)
 | I n  R T |  1
| I n  R T | 1

which is the Lt of h* (t1 ,..., tn ; , R ) of

  
h(x1 ,..., x n ;  , R )   cJ ( )    f ( x1 ,..., xn ;   1, R ). (2.5)

J  jJ x j 
We have to verify that


h(x1 ,..., x n ;  , R )  f ( x1 ,..., xn ;  , R ). (2.6)

Integration of the lhs in (2.4) over [0, ] provides
n

  f ( x1,..., xn ; , R )  f ( x1,..., xn ; , R0 )  e dx j 
t x
| I n  R T |  | I n  R0T |  j j

n j 1

 
  n t j x j
 h (t1,..., tn ; , R )d  n  0 1 n      (2.7)
*
 h(x ,..., x ; , R ) d e dx j ,
0 j 1


where the change of the order of integration can be justified by Fubini’s criterion in the following way:

With (1.7), g 1 (x, y)  g (x, y)  g 1 (x, y), decompositions ( ( ))1 R  I n  B( )( B( ))T - similar as in
x
(1.6) - with rows b j ( ) in B( ) and indicator functions eJ of J we obtain

| h(x1 ,..., x n ;  , R ) | cJ ( ) | ( ) f ( x1,..., xn ;  1, R ) |
J jJ x j
cJ ( ) 
  
n
 E   ( ( ))1 eJ ( j) g (  (j ) , 12 b j ( )Sb j ( ))  g 1 (  (j ) , 12 b j ( ) Sb j ( )) ,
x T x T

J ( ( ))|J |  j 1

S Wn1  2(  1), I n1  . Therefore, the Lt of | h | is bounded by a linear combination of integrals of the form
 n
E   ( ( ))1 g e
 n
 (1   ( ) t i )
( j ) (  ( ) , 2 b j ( ) Sb j ( ))  e ,K  1,..., n.
x

t j x j
dx j  iK
c
j 1 T

n  j 1
K
 j 1 | I n  RT | 1


The coefficients of this linear combination are non-negative continuous functions of  [0, ]. Integration over
[0, ] yields a finite value, which implies (2.7) and (2.6).

Integration over x1 ,..., xn in (2.5) leads to

 
F ( x1 ,..., xn ;  , R )   cJ ( )( )F ( x1 ,..., xn ;   1, R )  0, (2.8)
 J jJ x j

since the canonical correlations in (2.4) do not identically vanish because of the positive rank of R12 .
 -4-

Remarks. By continuity, theorem 1 holds also for a singular R, at least with “  ” instead of “  ”.

A modified proof uses the char. function ĥ corresponding to (2.4). It is the continuous limit of the char. functions of
the signed measures defined by  1 ( F ( x1 ,..., xn ; , R  )  F ( x1 ,..., xn ; , R )),  0. Thus, ĥ is the char.

function of the signed measure, which is defined by F ( x1 ,..., xn ;  , R ) and which is the difference of two finite

positive measures on n
.
An approximation by a series of univariate integrals for the difference of both sides in (2.1) is proposed in [11] with
identical values xi  x under the additional conditions
2
 1 n1 n 
   
2 2
1  ij 0,2  ij 0,  
2
ij   12 .
n1 (n1  1) 1i  j  n1 n2 (n2  1) n1 1i  j  n  n1n2 i 1 j  n1 1 
The error of this approximation tends to zero with a decreasing variability of the correlations within the submatrices
R11 ,R22 and R12 . This approximation is recommended in particular for small values of 1  F ( x,..., x; ; R). In a mo-
n1 n
 
1
dified approximation  2 is replaced by 2
ij .
n1n2 i 1 j  n 
1 1

References
[1] Bapat, R.B. (1989). Infinite divisibility of multivariate gamma distributions and M-matrices, Sankhyā 51, 73-78.
[2] Das Gupta, S., Eaton, M.L., Olkin, I., Perlman, M.D., Savage, L.J. and Sobel, M. (1972). Inequalities on the
probability content of convex regions for elliptically contoured distributions, Proc. Sixth Berkeley Symp. Math.
Statist. Probab. 2, 241-264, Univ. of California Press, Berkeley.
[3] Griffiths, R.C. (1984). Characterization of infinitely divisible multivariate gamma distributions, J. Multivariate
Anal. 15, 13-20.
[4] Hargé, G. (1999). A particular case of correlation inequality for the Gaussian measure, Ann. Probab. 27,1939-
1951.
[5] Khatri, C.G. (1967). On certain inequalities for normal distributions and their application to simultaneous confi-
dence bounds, Ann. Math. Stat. 38, 1853-1867.
[6] Krishnamoorthy, A.S. and Parthasarathy, M. (1951). A multivariate gamma type distribution, Ann. Math. Stat.
22, 549-557.
[7] Memarian, Y. (2013). The Gaussian correlation conjecture proof, arXiv:1310.8099v1[mathPR].
[8] Pitt, L.D. (1977). A Gaussian correlation inequality for symmetric convex sets, Ann. Probab. 5, 470-474.
[9] Royen, T. (1995). On some central and non-central multivariate chi-square distributions, Statist. Sinica 5, 361-
371.
[10] Royen, T. (2007). Integral representations and approximations for multivariate gamma distributions. Ann. Inst.
Statist. Math. 59, 499-513.
[11] Royen, T. (2013). Some upper tail approximations for the distribution of the maximum of correlated chi-square
or gamma random variables, Far East J. Theor. Statist. 43, 27-56.
[12] Schechtman, G., Schlumprecht, T. and Zinn, J. (1998). On the Gaussian measure of the intersection of sym-
metric convex sets, Ann. Probab. 26, 346-357.
[13] Šidák, Z. (1967). Rectangular confidence regions for the means of multivariate normal distributions, J. Amer.
Statist. Assoc. 62, 626-633.