353 CORRESPONDENCE

probability of error receiver, and since finally p = l/2 when

x = 0, we have the identity
from which (5) follows.
Using the inequality (4) and the identity (5) in (1) gives
the asserted inequality (3).
ROBERT 0. HARGER
Dept. Elec. Engrg.
University of Maryland
College Park, Md. 20742
REFERENCES
[l] J. N. Pierce, "Theoretical diversity improvement in frequency-shift keying,
Proc. IRE, vol. 46. pp. 903-910. May 1958.
[2] J. M. Wozencraft and I. M. Jacobs, Principles of Communication Engineering.
New York: Wiley, 1965, sec. 7.4.
A Gaussian Channel With Slow Fading
Abstract-An interleaved fading channel whose state is known
to the receiver is analyzed. The reliability function E(n) is obtained
for rates R in the range R, 5 R 5 C. The capacity is shown to be
C = EA { 4 ln (1 + A%)) where A is a factor describing the fading
mechanism and u is the signal-to-noise ratio per dimension.
INTRODUCTION
This correspondence is concerned with the analysis of a
fading channel that can be described as a Gaussian channel
with a slowly varying signal level. The following assumptions
are made.
The fading is slow enough to make it possible for the re-
ceiver to observe the instantaneous value of the signal
level.
Scrambling units are used at both ends of the channel. The
scramblers, which are regarded as parts of the channel,
are constructed to make it possible to neglect the effect
of memory in the channel.
The channel is band limited, making a vector notation
natural.
As a consequence, our channel is a time-discrete amplitude-
continuous memoryless channel. For each transmitted symbol X,
the receiver observes a pair of symbols Y and A, where A is the
momentary signal level. The channel-transition probability
function p(y, a ] Z) is given by
p(y, a 1 z) = p(u)(2~)-~ exp ( -*(y - a$}.
0)
Where explicit assumptions about the probability distribution
of the fading variable A are used, we assume A to be Rayleigh
distributed,
p(u) = 2ae-a a 1 0.
(59
GENERAL EXPRESSION FOR THE RELIABILITY FUNCTION
Suppose each input signal vector x = (x1, x2, **. , 2,) is
constrained to satisfy an energy constraint of the form
xv f(x,) < 0, where f(z) = ~2 - 7. For R, 2 R 2 C, the
Manuscript received Mmoh 3, 1969; r&Bed July 18, 1969. Thia work w&s done * R. G. Gallager, Information Theory and Reliable Communication. New York:
at the Research Institute of National Defence, Sweden. Wiley, 1968, pp. 323-333.
reliability function E(R) is given by*
E(R) = max ~-UP, p(x):), 4 - PRI,
where in our case,
J-UP, Pc4, 4
S(s )
l+P
= -1n p(x)p(y, a 1 x)~(+) exp rj(x) dx
du da,
(3)
and where the maximum is over all input probability distri-
butions p(x), all p E [0, I], and all r 2 0. For the Gaussian
channel with this input constraint, the maximum over p(x)
is known to be given by
pa(x) = (2~17)~~ exp - $ *
1 1
Our channel is a Gaussian channel with a varying signal level.
As the distribution (4) is optimum for each value of the signal-
to-noise ratio in a Gaussian channel, it follows that it is optimum
in our channel too. With the distributions given in (1) and (4)
the integrals in (3) can be evaluated. The result is
&(P, Pdd, 4 = ~(1 + ~>?t + + In (1 - 24
- In E,((l - 2q + A27j(1 + p)-y).
At this point let us introduce the new variable /3 = 1 - 2rq,
and let us define
dP, PI = Eo(P, P&l, 4,
6, P> = EAiW + A2170 + PY-11-p21.
In our new notation E(R) takes the form
(5)
E(R) = max { -PR -I- m;x dp, P>l ,
P
(6)
dp, P> = $41 + p)(l - P) + 3 ln P - In NP, PI.
In order to guarantee (5) to be meaningful, we must restrict
p to satisfy /3 > 0, so the maximum in (6) is over all p, p E
[0, I]. For each p let g(p, 0) be maximized over p by /3* = p*(p).
Using the general relation dlS/dR = -p, we alternatively have
E(R) = gb, P*> - P
C
ag dP* as7
ap + dp ap
II
8D8e>
(7)
.
8-P
In Appendix I we show that g(p, p) is a convex-upward
function of /I for each p E [O, 11. Thus, if the equation ag/afl= 0
has a solution in the interval 0 5 /3 5 1, it is unique and
equals fl*.
RAYLEIGH-DISTRIBUTED SIGNAL LEVEL
We now turn to the case of Rayleigh fading, i.e., we assume
p(u) to be given by (2). From (5) we have
h(p, p) = s,- 2ue-[p + u2z-1]-p2 da,
354
where we have introduced z = q-1(1 + p). Integrating by
parts and rearranging the terms, we obtain the differential
equation ah/a/3 - zh = -sZp+*, which has the solution
IEEE TRANSACTIONS ON INFORMATION THEORY, MAY 1970
hb, P> = ezB
[ h(p, 0) - 2 sup z-p2e-zz dz] .
(8)
By direct computation we have h(p, 0) = &2I(l - p/2)
which inserted in (8) gives
s
m
h(p, 0) = eza~pz tmpemt dt.
(9)
4
We thus see that for /3 2 0 and for each p E [0, l] h(p, 0) is a
decreasing function of /3 taking values in the bounded interval
[O, ZNY(l - p/2)].
In order to find fi* we differentiate g(p, /3) as given in (6)
Fig. 1. h(p, 8) and k@, ,9) 88 funotions of 6.
with respect to /3,
Observing that cYg/a/3 = 0 for p = 0, p = /3*(O), we thus have
at7
-=
w
-$(I + p) + +/j- _ h-$.
(10)
C = Ml - P> - EAI-4 ln (P + A2d11p=I
Let us define
= EA(3 ln(1 -I- Av)}.
k(p, p> = p1-P2[(1 + +,)/I - +z-1-l.
(11)
DISCUSSION
With the aid of (10) and from the fact that g(p, 0) is convex
upward in p, we realize that if the equation
~XP, i-9 = hh PI (12)
has a solution in the interval 0 2 0 < 1, it is unique and
equals fl*.
The functions k and h are shown in Fig. 1. The value PO,
defined in the figure, is obtained from (11) and is &, =
[2(2 + q)]-l 5 1, which shows that for each q 2 0 and each
p E [0, l] there exists a ,8 E [0, l] such that k(p, /3) > h(p, ,6?).
Thus, in order to show that (12) has a solution in the interval
0 5 /3 5 1, it is now sufficient to show that h(p, 1) 2 k(p, 1)
for all p E [0, 11. This is shown in Appendix II, and thus we
have shown that when the fading is of the Rayleigh type, the
reliability function E(R) can be represented in the parametric
form (7), where /3* is the solution to (12).
CAPACITY
For the capacity it is possible to find a simple expression
without making the special assumption that the fading is of
the Rayleigh type. Generally we have
ag/a/3 is given in (10) and from (5) we realize that :
Thus we have
ag
ap p-o
= -4 + ;p-1,
which shows that p*(O) = 1. From (6) we have
at7
-=
3P
30
- 0) - h-!$
and from (5) straightforward computations give
ah
I ep p-o
= Ed{-+ In@ + A2rl)l.
In order to visualize how E(R) depends on the signal-to-noise
ratio 17, we have plotted in Fig. 2 its two related parameters C
and Ro, the capacity and the exponential error-bound parameter,
as functions of g for the case of Rayleigh fading. R. is defined
as R. = g(1, p*(l)). For comparison we have also plotted the
corresponding functions for a nonfading channel, i.e., a channel
with p(a) = 6(a - 1). As shown, the degradation due to fading
may be described as a power loss of between 1 and 3 dB. We also
see that the degradation is almost the same whetherwe use C
or R. as the measure of performance.
APPENDIX I
We are going to show that g(p, ,L?) as given in (6) is a convex-
upward function of ,!I for each p E [0, I]. The logarithm is an
increasing convex-upward function, and thus it is sufficient to
show that h-1 is a convex-upward function of 0. This is trivially
true for p = 0, so let us assume 0 < p 5 1. We have
$ (h-l) = h-f2($) - h $1
and as h(p, ,8) 2 0, it is sufficient to show that
Let us define the random variables
X2 = [/j + AZZ-1]--p2
Y2 = $p(+p + I)[@ + A2.z-1]-p2-2 .
From Schwarzs inequality we have
f (; + 1)($)2
= [EAXY-j2 5 EAX2EA I = h $ ;
h$ > (1 + $$$$2 > 2($$2. Q.E.D.
APPENDIX II
From (9) and (11) we have h(p, 1) - k(p, 1) = e~.&~j(z)
where we have introduced
f(z) = Jm fY2e- dt - [ez(l + +pz-)I-.
CORRESPONDENCE
A C.R,(nats/dim)
355
interval length) are obtained from this waveform to form the
vector sample
Fig. 2. Channel capacity and exponential bound parameter z+s functions of
signal-to-noise ratio. @ Capacity. Gaussian channel; @ capacity, fading
channel; 0 exponential bound, Gaussian channel: @ exponential bound,
fading channel.
Straightforward computations give
4f
-=
dz
-2p.z -e-(p + 2~)~ < 0
p; j(Z) = 0
from which we conclude that f(z) 2 0. This implies h(p, 1) 2
MP, 1). Q.E.D.
ACKNOWLEDGMENT
The author wishes to thank his teacher Prof. L. H. Zetterberg,
who read the manuscript and with whom he had many helpful
discussions.
THOMAS ERTCSON
Res. Inst. National Defence
104 50 Stockholm, Sweden
Nonparametric Detection Using Dependent Samples
Abstract-A new general approach to the formulation of a non-
parametric detector using dependent samples is introduced and
applied to a space-diversity system employing dc signaling. A
comparison based on a form of asymptotic relative efficiency is
made between the new detector and a Mann-Whitney detector.
Under certain conditions the new procedure demonstrates an
improvement in transmission efficiency.
INTRODUCTION
A binary signal detector using a reference noise waveform
can be formulated in terms of a multivariate nonparametric
hypothesis test to allow for the use of dependent samples.
Assume that the detector has available a sample waveform N(t)
from the stationary noise process and that a group of p equally
spaced dependent samples over a time period T (decision-
Manuscript received May 9, 1968; revised March 12, 1969. This work w&8
supported in part by the Doctoral Support Plan of Bell Telephone Laboratories.
Inc., and in part by the National Science Foundation, under Grant GK-1075.
This paper is based on part of & thesis submitted to the Graduate Division of the
School of Engineerin and Science of New York Uni\versity in partial fulfillment
of the requirements or the degree of Doctor of Philosophy. f
( yy, y;y . . . ) yi) = y,
This process is repeated n times with sufficient delays between
sampling groups so that finally n independent vector samples
yw, . . . ) yw, . . . ,
Y(n) are obtained. The vector independence
is not a severe limitation since the noise-vector samples are
taken on a periodic reference interval that comprises a small
part of the total transmission period. During normal trans-
mission, p equally spaced dependent samples (X1, . . . , X,) = X
are obtained on the decision interval T. The following hypothesis-
testing problem!results:
H: noise: Y(l), . * . , Yen, X E F(zl, . * * , z,)
K: signal and noise:
yl
, s. . , Y E F(+ . . . ,:z,j
X E Gh, - . - , 4,
where F(zl, . . . , s,) and G(zl, . . . , 2,) are p-variate cumulative
distribution functions. The detector is then equivalent to a
a-sample p-variate nonparametric hypothesis test. During the
decision interval, m independent vector samples X(l), * . * ,
X(i), . . . )
X(m) can be obtained if an m-channel space- or time-
diversity system is used. In a space-diversity system a reference
noise sample Y(i) could be obtained separately on each channel,
thereby eliminating the spacing problem in obtaining the
independent reference vector samples when n = m. For some
situations, however, such as in an active sonar array, space
diversity does not insure independent noise processes on each
channel.
This general formulation has been introduced to stimulate
further study of nonparametric methods with dependence by
following a multivariate approach. Very little work has been
done on the problem of dependence [l] in signal detection.
Some multivariate nonparametric tests have been considered
in the mathematical literature [2], [3]. However, the forms of
these tests are such that their implementation for a signal-
detection system would be quite complicated and, possibly, of
limited use. In the present work a simple procedure that trans-
forms the multivariate data to univariate data and then applies
a univariate nonparametric test is applied to signal detection.
CONVERSION TO UNIVARIATE DATA
Let L be a fixed transformation from p-dimensional space
to one-dimensional space. Then
x
= q X(i))
and
yci)
= q Y)
are all univariate random variables. Since L is a fixed trans-
formation, if
XCi and Y( E F(zl, . . . , 3,
then X(i) and Y(t) have the same univariate distribution, say
J&). After the transformation, the multivariate problem
reduces to a univariate a-sample nonparametric problem with
independent samples, and a suitable 2-sample detector can be
used. The overall detector then consists of a transformation
or predetector L that combines the dependent samples on the
decision interval and a nonparametric detector that uses the
combined samples. The overall operation is nonparametric with
respect to the false-alarm rate 01. The transformation L reduces
the data, and in some cases this lowers the efficiency. The
proper selection of L should minimize this effect. A reasonable
choice for the predetector L is that device that is optimum
when the noise is Gaussian and white, for then there is at