next up previous
Next: Approximation of the Test-Statistic Up: Approximation of distribution function Previous: Problem Formulation

The Exact Analytical Expression for the Moments of the Test-Statistic

In order to evaluate $\L _{th}$ for a given $P_{FA}$ it is necessary to know the PDF (or the CDF) of the test-statistic $\L $. While the PDF of the random variable $\L $ is not known in analytical closed form, the exact analytical expression of the statistical moments of any order for the test-statistic monotonic function $V = \L ^{1/N}$ can be found by using the well-known complex Wishart distribution for the matrix ${\bf A}$ given in [3]

$\displaystyle \begin{array}{c}
W({\it A},\S ,N)=\frac{\vert A\vert^{N-p}}{I(\S )}e^{-Sp(\S ^{-1}A)},
\end{array}$

    (7)


where $I(\S )=\frac{1}{K(\S )}
=\pi^{\frac{1}{2}p(p-1)}\Gamma(n)\ldots\Gamma(N-p+1)\vert\S \vert^N$

If the null hypothesis $H_0$ is true, the matrix ${\bf A}$ is $W({\it A},\S _0,N)$ distributed.

The $h$-th order moment of the random variable

$V=\L ^{\frac{1}{N}}=\frac{\vert A\vert}{(\frac{SpA}{p})^p}$

can be straightforward determined from multidimensional integral


$\displaystyle \begin{array}{l}
M[V^h]=\int...\int\vert A\vert^h\frac{1}{\biggl(...
... _0,N)\times \  \times
\vert A\vert^{N-p}exp{[-Sp(\S _0^{-1}A)]dA}
\end{array}$

    (8)

The integration procedure is carry out here over area $dA=da_{11}da_{12}...da_{p-1,p}da_{pp}$, where A are all non-negative defined Hermitian $[p \times p]$ matrices. The integral (8) can be transformed into


$\displaystyle \begin{array}{l}
M[V^h]=\frac {p^{hp}K(\S _0,N)} {K(\S _0,N+h)}\i...
...xp[-Sp(\S _0^{-1}A)]\}
\frac{1}{(\sum\limits_{i=1}^pa_{ii})^{ph}}dA
\end{array}$

    (9)

The element of integration (9) in figured brackets is also Wishart distribution (7) with $N+h$ degrees of freedom $W(A,\S _0,N+h)$. After integration of (9) over all nondiagonal $W(A,\S _0,N+h)$ distribution elements $\it {a_{ij}}$ $(i\ne
j)$, the partial joint distribution of the diagonal elements $a'_{11}...a'_{pp}$ of the matrix ${\bf A}$ is obtained. For hypothesis $H_0$ this distribution is transformed into product of one-dimensional distributions $W({\it a}_{ii},{\bf\sigma}^2,N+h)$. The one-dimensional complex Wishart distribution $W({\it a}_{ii},{\bf\sigma}^2,N+h)$ is the distribution of the $i$-th diagonal element of the matrix ${\bf A}$. Note, that $W({\it a}_{ii},{\bf\sigma}^2,N+h)$ is $\chi^2$ distribution with $2(N+h)$ degrees of freedom:


$\displaystyle W(a_{ii},{\bf\sigma}^2,N+h)=
(\frac{a_{ii}^{N+h-1}}{\Gamma(N+h){\bf\sigma}^{2(N+h)}})exp{(\frac{-a_{ii}}{{\bf\sigma}^2})}$

    (10)

Thus, the expression for the $h$-th order moment of the test-statistic $V$ can be reduced to the integral

$M[V^h]=\underbrace{p^{hp}\frac{K(\S _0,N)}{K(\S _0,N+h)}}_{g}\int...\int
\frac{...
...}{{\bf\sigma}^2}}}{{\bf\sigma}^{2(N+h)}\Gamma(N+h)}
da_{11}da_{22}\dots da_{pp}$

and can be represented as

$M[V^h]=g<\frac{1} {(\sum\limits_{i=1}^pa_{ii}')^{ph}}> $,

where $<.>$ denotes expectation value.

Diagonal elements $a_{ii}'$ of the matrix ${\bf A}$ are independent and identically $\chi^2$ distributed random values with $2(N+h)$ degrees of freedom and variances ${\bf\sigma}^2/2$. Therefore, the random variable $B=\sum a_{ii}'$ is also $\chi^2$ distributed with $2(N+h)p$ degrees of freedom. Thereby, $h$-th order moment of the random variable $V$ is proportional to the $r$-th order moment $(r=-ph)$ of the random variable $B$ which can be expressed through known one-dimensional integral


$\displaystyle \begin{array}{l}
M[B^r]=\int\limits_0^{\infty}B^r\frac{B^{(N+h)p-...
...c{{\bf\sigma}^{2Np}\Gamma(Np)}{{\bf\sigma}^{2(N+h)p}\Gamma((N+h)p)}
\end{array}$

    (11)

For practical calculations of the high order moments of the test-statistic $V$ it is conveniently to transform the last expression into product using properties of gamma-function:

$M[V^h]=g\frac{{\bf\sigma}^{2pN}\Gamma(pN)}{{\bf\sigma}^{2p(N+h)}\Gamma(p(N+h))}...
...\times\ \times\frac
{\Gamma(N+h)...\Gamma(N+h-p+1)}{\Gamma(N)...\Gamma(N-p+1)}$

So, the exact analytical expression for the $h$-th order moment of the test-statistic $V$ is


$\displaystyle M[V^h]=
p^{hp}\frac{\prod\limits_{i=1}^{p}\prod\limits_{j=1}^{h}(j+N-i)}{\prod\limits_{i=1}^{ph}(i+pN-1)}$

    (12)

It is easy to see that the distribution of the random variable $V$ is concentrated in interval [0,1]. So, there are moments of any order (12) for $V$, and they completely determine its distribution function [5]. It is worth to notice that the characteristic function of the random variable $V$ is expressed through moments as a converging Taylor series.


next up previous
Next: Approximation of the Test-Statistic Up: Approximation of distribution function Previous: Problem Formulation