194 IRE TRANSL4CTIONX ON INFORMATION THEORY April

On a Moment Theorem for Complex Gaussian Processes*

MAGIC TERM I. S. REEDt
E[abs(u^2(m+1)] = (m+1)! P_n^m+1

Summary-A general theorem is provided for the moments of a (apparently not well known) moment theorem for the
complex Gaussian video process. This theorem is analogous to the process :
well-known property of the multivariate normal distribution for
real variables, which states that an nth order central product
Theorem: Suppose Z, = z(tJ for (n = 1, 2, .. . N)
moment is zero if n is odd and is equal to a sum of products of are samples from zero-mean, complex Gaussian video
covariances when n is even. process x(t).
a) If s # t, then
HE PROPERTIES of a Gaussian, zero-mean,
E.&z,, . . . z,,x,,z,, . . . x,, = 0(2)
stationary, complex video process z(t) are developed
by Arens’ and Kelly, Reed’ and Root.3,4 The where ??%kand n, are integers from set (1, 2, 3, . . . , NJ.
properties of interest here are b) If s = t, then

Ex(t’)z(t) = R(t - t’) E.&z,~ -0. zmtxPLl”rLl

9 . . * &I
= c (Ezrn.(~)~n~)(E~~.(.,x,,)
. +. @%n~~t,~w) (3)
E&‘)&) = 0 T
for all t and t’ where E is LLexpected value” and R(7) is where x is a permutation of the set of integers
the (Hermitian-symmetric) covariance function of the (1, 2, 3, .*. ) 1).
process. With these two properties one may show’,’ that Proof: Expand both sides of (1) with the exponential
the joint probability density of N samples, expansion. Consider the (1 + s)-th term in the expansion
of the left side,
z, = X(&J for (n = 1, 2, . . . , N)
is given by

P(Zl ) . . . , 2%)= (2~)~” 1j H / 1-l exp - $[ g z#Z-~)&] ‘(.?k$k, + bk,Zk,) ’ . ’ (ikt+.zkr+. + i-kt+a2kt+r). (4

If the product in (4) is expanded it is evident that the

with the 2N-dimensional volume element,
left side of (2) appears as a factor of the coefficient of
dv = dx, dxz ) . ‘. dXN cly1 dya ) ‘. . dy,jr, term j-,><,, * . * ln, Fnlfn, . . . fn,. The t-t,h term of the
exponential expansion of the right side of (1) is
where x, = Re (x,), yn = Im (x,) ; the bar denotes con-
jugation; and H is the Hermitian-symmetric, ‘(moment”
matrix
H = (H,J = (+E&) = (+R(tj - tJ)
Thus, the {,, . . . {,, fn, . . . fnc term in the Taylor’s
with 11 H 11, its determinant. It is further possible to expansion in Ii, fk for (j, 1c = 1, 2, . . . N) of the right
demonstrate that the 2N-dimensional characteristic side of (1) is missing if s # t. Hence, by equating coeffi-
function’ of this distribution is given by the identity cients, part a) of the theorem is proved.
f r. ~7 l\ In (4) let s = t. We now wish to equate the coefficient
of i-m, . . * i-,n, .?“A>. * + f%, in (4) with the coefficient of
the same term in (5). Multiplying the factors of (4) and
= exp using part a), the only terms that remain are those
-i k$ .tkHkjfi>* (1)
11 terms with precisely t barred x’s and t unbarred x’s
The purpose of this note is to use (1) to prove the following There are exactly 2tt
such terms and they are identical
0
* Received by the PGIT, May 9, 1961.
to one another. Consequently for s = t expression (-1)
- . t . Electronics Dept., The RAND Corporation, Santa Monica, reduces to
1 R. Arens, “Complex processes for envelopes of normal noise,”
IRE TRANS. ON INFORMATION THEORY, vol. IT-3, pp. 204-207;
September, 1957.
* E. J. Kelley and I. S. Reed, “Some Properties of Stationary
Gaussian Processes,” M.I.T. Lincoln Lab., Lexington, Mass.,
Tech. Rept. TR-157; June 5, 1957.
3 I. S. Reed! E. J. Kelly, and W. L. Root, “The Detection of .E(E~, . . . ~klxj~ . . ’ ~,,}m, ‘. . j-~~scii. ‘. j-;l.
Radar Echoes m Noise,” M.I.T. Lincoln Lab., Lexington, Mass.,
Tech. Rept. TR-158, June 20, 1957. In this multiple sum the particular term f,, . . . j-,,
4 E. J. Kelly, I. S. Reed, and W. L. Root, “The detection of
radar echoes in noise, I,” J. SIAAT, vol. 8, pp. 309-341; June, 1960. fn, 0-s fn, and its coefficient is repeated (t!)” times. Thus,
Authorized licensed use limited to: Universita degli Studi di Parma. Downloaded on October 29,2023 at 22:04:42 UTC from IEEE Xplore. Restrictions apply.
1962 Reed: Moment Theorem for Complex Gaussian Processes 195
the coefficient of cm, The latter identity has been useful ____in signal-to-noise
of (1) is calculations of the product process’ w(t - 7) w(t) where
w(t) is a signal-plus-noise complex Gaussian process. The
above theorem is analogous to the known property6 of
the multivariate normal distribution for real random
variables, which states that an n-th order central product
moment is zero if n is odd and is equal to a sum of products
of covariances when n is even.
In conclusion it should be mentioned that the joint
Let rr be a permutation of the set of subindices probability density for N samples of a complex Gaussian
i&2, -** 7 t) ; there are t! such r’s Then (5) evidently can video process has been usefull.’ in making Rice-type
be expressed as probability density calculations for envelope and phase.
The main advantage of the complex formalism in this
application lies in the fact that the moment matrix for
the complex process has half t#he dimensionality of the
corresponding real moment matrix for the process when
regarded as a joint real process. For example, Kelly and
Reed’ require only the inversion of a 2 X 2 Hermitian
The coefficient of j-,, . 1 . pm, fn, . . . in, in this sum is symmetric matrix in the computation of the probability
repeated t! times. Hence, the coefficient of lm, * . . crnl density of the envelope, sampled at two points in time,
Lb, *** fn, on the right side of (1) is whereas Rice’ required the inversion of a 4 X 4 real
moment matrix. A more complicated problem, for ex-
ample, that of computing the probability density of in-
stantaneous frequency, sampled at two points in time,
Equating coefficients and using identity would require either an inversion of a 4 X 4 Hermitian
matrix or, by Rice’s method, an inversion of an 8 X 8
Hkj = $E(Q.J
real moment matrix. For obvious reasons, E. J. Kelly has
the theorem is proved. recently called a complex Gaussian process with the
One may obtain the following useful identities from special property Ez(t)x(t’) = 0 for all t and t’ a circular
this theorem : complex Gaussian process.

E / x jzn = n! (E 1z 1’)”
5 I. S. Reed, “Semi-Coherent Detection,” The RAND Corp.,
E(@Jn = n! [E.z~zJ Santa Monica, Calif., Paper No. P-2106; September 19, 1960.
6 M. C. Wang and G. E. Uhlenbeck, “On the theory of the
Rrownian Motion II,” Revs. Mode. Phys., vol. 17; p. 332; April-
EzAz~x~ = E(z&E(@J + E(.z&E(&) July, 1945.
7 S. 0. Rice, “Mathematical analysis of random noise,” Bell
and so forth. Sys. Tech. J., vol. 24, p. 83; 1944.

Authorized licensed use limited to: Universita degli Studi di Parma. Downloaded on October 29,2023 at 22:04:42 UTC from IEEE Xplore. Restrictions apply.