You are on page 1of 5

On the Energy Detection of Unknown Signals

over Fading Channels


+
Fadel F. Digham

, Mohamed-Slim Alouini

, and Marvin K. Simon

Department of Electrical and Computer Engineering


University of Minnesota
Minneapolis, MN 55455, USA
E-mails: <fdigham,alouini@ece.umn.edu>

Jet Propulsion Laboratory (JPL)


California Institute of Technology (Caltech)
Pasadena, CA 91109-8099, USA.
E-mail: <marvin.k.simon@jpl.nasa.gov>
AbstractThis paper presents another look at the problem
of energy detection of unknown signals over different fading
channels. We start with the no diversity case and present some
alternative closed-form expressions for the probability of detec-
tion (P
d
) to those recently reported in [1]. We then investigate
the system performance when different diversity schemes are
employed. It is shown that there is not much improvement
in the probability of detection when either the probability of
false alarm (P
f
) exceeds 0.1 or the average signal-to-noise
ratio (SNR) exceeds 20 dB. In addition, receiver operating
characteristic (ROC) curves comparing the performance of equal-
gain combining (EGC), selection combining (SC), and switch
and stay combining (SSC) are presented. As an example, EGC
introduces a gain of two orders of magnitude from the probability
of miss perspective compared to the no diversity case while both
SC and SSC introduce a gain of about one order of magnitude.
I. INTRODUCTION
In many wireless applications, it is of great interest to check
the presence and availability of an active communication link
when the transmitted signal is unknown. In such scenarios,
one appropriate choice consists of using an energy detector
which measures the energy in the received waveform over an
observation time window. This interesting problem was studied
in the classic paper by Urkowitz [2] assuming deterministic
signals transmission over a at band-limited Gaussian noise
channel. More specically, relying on the sampling theorem
to approximate the received signal energy and on chi-square
statistics of the resulting sum of squared Gaussian random
variables, Urkowitz derived both the probability of detection
(P
d
) and the probability of false alarm (P
f
). The receiver op-
erating characteristic (ROC) [3] was then obtained by plotting
P
d
versus P
f
for different system parameters of interest.
This energy detection problem has been revisited recently
by Kostylev [1] for signals operating over a variety of fading
channels. While the probability of detection has been derived
in closed-form for Rayleigh fading channels, the results in-
volve a numerical integration and an innite summation for
Nakagami and Rician fading channels, respectively. In this
+
This work was supported by the ARL Communications and Networks
CTA, under cooperative agreement DAAD19-01-2-0011.
paper, we have another look at the problem. Our contribution
is two-fold. First, we present an alternative analytical approach
to the one presented in [1] and obtain closed-form expressions
for the probability of detection not only over Rayleigh fading
but also over Nakagami and Rician fading channels. Second,
and more importantly, we quantify the improvement in the
probability of detection when equal gain combining (EGC),
selection combining (SC), and switch and stay combining
(SSC) diversity schemes are used.
The rest of this paper is organized as follows. In Section II,
the system model under consideration is described and impor-
tant notations are listed. Section III evaluates the conditional
(on the fading) probabilities of detection and of false alarm
(or equivalently P
d
and P
f
over additive white Gaussian noise
(AWGN) channels). While Section IV deduces the detection
probability over various fading channels, Section V studies
the impact of diversity on this detection probability. Finally,
numerical examples and concluding remarks are offered in
Section VI.
II. SYSTEM MODEL AND NOTATIONS
Before describing the system model, we rst list the main
notations that are going to be used in this paper for additional
clarity and to avoid any kind of confusion when going back
to [2] and [1].
s(t) : signal waveform.
n(t) : noise waveform which is modelled as a zero-mean
white Gaussian random process.
N
01
: one-sided noise power spectral density, i.e., N
01

N
0
in [1].
N
02
=
N01
2
: two-sided noise power spectral density.
E
s
: signal energy=
_
T
0
s
2
(t)dt.
=
Es
N01
: signal-to-noise ratio (SNR), i.e.,

2
in [2].
: average SNR, i.e., d
2
in [1].
: energy threshold used by the energy detector.
T : observation time interval, seconds.
W : one-sided bandwidth (Hz), i.e. positive bandwidth of
the low-pass (LP) signal.
u=TW : time bandwidth product, i.e., u
B
2
1 in [1].
3575
0-7803-7802-4/03/$17.00 2003 IEEE
f
c
: carrier frequency.
P
d
: probability of detection.
P
f
: probability of false alarm.
P
m
= 1 P
d
: probability of missing.
H
0
: hypothesis 0 corresponding to no signal transmitted.
H
1
hypothesis 1 corresponding to signal transmitted.
N(,
2
) : a Gaussian variate with mean and variance

2
.

2

: a central chi-square variate with degrees of


freedom.

2

() : a noncentral chi-square variate with degrees


of freedom and noncentrality parameter .
The received signal r(t) takes the form
r(t) = h s(t) +n(t), (1)
where h=0 or 1 under hypotheses H
0
or H
1
, respectively. As
described in [2], the received signal is rst pre-ltered by an
ideal bandpass lter with transfer function
H(f) =
_
2

N01
, |f f
c
| W,
0, |f f
c
| > W,
(2)
to limit the average noise power and normalize the noise
variance. The output of this lter is then squared and integrated
over a time interval T to nally produce a measure of the
energy of the received waveform. The output of the integrator
denoted by Y will act as the test statistic to test the two
hypotheses H
0
and H
1
. Although this process is of band-pass
type, one can still deal with its low-pass equivalent form and
eventually translate it back to its band-pass type [4]. Besides,
it has been veried in [2] that both low-pass and band-pass
processes are equivalent from the decision statistics perspec-
tive which is our main concern. Therefore, for convenience,
we address in this paper the problem for a low-pass process.
According to the sampling theorem, the noise process can be
expressed as [5]
n(t) =

i=
n
i
sinc(2Wt i), (3)
where sinc(x) =
sin(x)
x
and n
i
= n
_
i
2W
_
. One can easily
check that
n
i
N(0, N
01
W), for all i. (4)
Over the time interval (0, T), the noise energy can be approx-
imated as [2]
_
T
0
n
2
(t) dt =
1
2W
2u

i=1
n
2
i
, (5)
where u = TW. We assume that T and W are chosen to
restrict u to integer values. If we dene
n
i
=
n
i

N
01
W
, (6)
then, the test or decision statistic Y can be written as [2]
Y =
2u

i=1
n
i
2
. (7)
Y can be viewed as the sum of the squares of 2u standard
Gaussian variates with zero mean and unit variance. Therefore,
Y follows a central chi-square (
2
) distribution with 2u
degrees of freedom.
The same approach is applied when the signal s(t) is
present with the replacement of each n
i
by n
i
+ s
i
where
s
i
= s
_
i
2W
_
. The decision statistic Y in this case will have
a noncentral
2
distribution with 2u degrees of freedom and
a non centrality parameter 2 [2]. Following the short-hand
notations mentioned in the beginning of this section, we can
describe the decision statistic as
Y
_

2
2u
, H
0
,

2
2u
(2), H
1
.
(8)
The probability density function (PDF) of Y can then be
written as
f
Y
(y) =
_
_
_
1
2
u
(u)
y
u1
e

y
2
, H
0
,
1
2
_
y
2
_
u1
2
e

2+y
2
I
u1
(

2y), H
1
,
(9)
where (.) is the gamma function [6, Section 8.31] and I

(.)
is the th-order modied Bessel function of the rst kind [6,
Section 8.43].
III. DETECTION AND FALSE ALARM PROBABILITIES OVER
AWGN CHANNELS
An approximate expression for P
d
over AWGN channels
was presented in [2]. In this section we present exact closed-
form expressions for both P
d
and P
f
.
The probability of detection and false alarm can be generally
computed by
P
d
= Pr(Y > |H
1
), (10)
P
f
= Pr(Y > |H
0
), (11)
where is the decision threshold. Using (9) to evaluate (11)
yields
P
f
=

_
u,

2
_
(u)
, (12)
where (., .) is the incomplete gamma function [6]. This result
matches the one obtained in [1, (19)] after replacing each B
by B/21 since each B should be equal to 2u in our notation.
Furthermore, the parameter in [1] should be equal to B or
equivalently 2u.
On the other hand, the probability of detection can be
obtained from (9) to evaluate (10). More specically making
use of [4, Eq. (2.1-124)], the cumulative distribution function
(CDF) of Y can be evaluated (for an even number of degrees
of freedom which is 2u in our case) as
F
Y
(y) = 1 Q
u
(
_
2,

y), (13)
where Q
u
(a, b) is the generalized Marcum Q-function [7].
Hence,
P
d
= Q
u
(
_
2,

). (14)
3576
IV. AVERAGE DETECTION PROBABILITY OVER FADING
CHANNELS WITH NO DIVERSITY
In this section, we derive the average detection probability
over Rayleigh, Nakagami, and Rician fading channels. We
provide alternative expressions to those obtained in [1]. Our
expressions are in closed form and are based on a different
approach by averaging the conditional P
d
in the AWGN case
as given by (14) over the SNR fading distribution. Of course,
P
f
of (12) will remain the same under any fading channel
since P
f
is considered for the case of no signal transmission
and as such is independent of SNR.
A. Rayleigh Channels
If the signal amplitude follows a Rayleigh distribution, then
the SNR follows an exponential PDF given by
f() =
1

exp
_

_
, 0, (15)
The average P
d
in this case, P
dRay
, can now be evaluated by
averaging (14) over (15) while making the change of variable
x =

2 and making use of [7, Eq. (30)] yielding


P
dRay
=e

2
u2

n=0
1
n!
_

2
_
n
+
_
1 +

_
u1
_
e


2(1+)
e

2
u2

n=0
1
n!

2(1 +)
_
.(16)
This expression is equivalent to that given in [1, Eq. (20)]
taking into account the following observations and corrections
in [1, Eq. (20)]: the normalized incomplete gamma function
P(a, b) = (a, b)/(a) can be expressed in its series form [6,
Eq. (8.352.1)], setting d
2
= , and correcting two typos: each
B should be replaced by B/2 1 and a minus sign in the
exponential argument should be inserted.
B. Nakagami Channels
If the signal amplitude follows a Nakagami distribution,
then the PDF of follows a gamma PDF given by
f() =
1
(m)
_
m

_
m

m1
exp
_


_
, 0, (17)
where m is the Nakagami parameter. The average P
d
in the
case of Nakagami channels P
dNak
can now be obtained by
averaging (14) over (17) and then using again the change of
variable x =

2 yielding
P
dNak
=
_

0
x
2m1
exp
_

mx
2
2
_
Q
u
_
x,

_
dx, (18)
where
=
1
(m)2
m1
_
m

_
m
. (19)
Evaluating the integral in (18) as described in Appendix A,
P
dNak
can be written as
P
dNak
=
_
G
1
+
u1

n=1
(/2)
n
2(n!)
1
F
1
_
m; n + 1;

2

m+
_
_
,(20)
where
1
F
1
(.; .; .) is the conuent hypergeometric function (
(., .; .)) [6, Section 9.2],
= (m)
_
2
m+
_
m
e
/2
, (21)
and
G
1
=
_

0
x
2m1
exp
_

mx
2
2
_
Q
_
x,

_
dx, (22)
where Q(., .) =Q
1
(., .) is the rst-order Marcum Q-function.
G
1
can be evaluated for integer m with the aid of [7, Eq. (25)]
as
G
1
=
2
m1
(m1)!
_
m

_
m

m+
e

2
m
m+
_
_
1 +
m

__
m
m+
_
m1
L
m1
_

m+
_
+
m2

n=0
_
m
m+
_
n
L
n
_

m+
_
_
,
(23)
where L
n
(.) is the Laguerre polynomial of degree n [6,
8.970]. As a byproduct, we obtain an alternative expression
for P
dRay
when setting m = 1 in (20) and this expression
is numerically equivalent to the one obtained in (16). When
comparing our expression in (20) with the one including a
numerical integration in [1, Eq. (33)], it seems that there is an
error in the Kostylevs expression since it does not reduce to
the special case of Rayleigh. Moreover, it is equal to zero for
the special case of u = 1 (equivalently B = 0) and m = 1.
C. Rician Channel
If the signal strength follows a Rician distribution, the PDF
of will be
f()=
K + 1

exp
_
K
(K + 1)

_
I
0
_
2

K(K + 1)

_
,
0, (24)
where K is the Rician factor. The average P
d
in the case of
a Rician channel, P
dRic
, is then obtained by averaging (14)
over (24) and substituting x for

2. The resulting expression


can be solved for u = 1 using [8, Eq. (45)] to yield
P
dRic
|
u=1
= Q
_
2K
K + 1 +
,

(K + 1)
K + 1 +
_
. (25)
For K = 0, this expression reduces to the Rayleigh expression
with u = 1.
V. AVERAGE DETECTION PROBABILITY WITH DIVERSITY
RECEPTION
In this section, we address the energy detection performance
when EGC, SC, or dual SSC diversity schemes are employed.
For simplicity, we focus on the case in which the diversity
paths are independent and identically distributed (IID) and are
subject to Rayleigh fading.
3577
A. Equal Gain Combining
The output SNR,
t
, of the EGC combiner is the sum of
the SNRs on all branches, i.e.,
t
=

L
l=1

l
where L is
the number of diversity branches. Adding L IID noncentral

2
variates with 2u degrees of freedom and non-centrality
parameter 2
l
each results in another noncentral
2
variate
with 2Lu degrees of freedom and non-centrality parameter

L
l=1
2
l
[4, Eq. (2.1-117)]. Hence, the P
d
at the EGC output
for AWGN channels can be evaluated by analogy to (14) as
P
dEGC
= Q
Lu
(
_
2
t
,

). (26)
The PDF of
t
for IID Rayleigh branches is known to be given
by
f(
t
) =
1
(L 1)!
L

L1
t
exp (
t
/). (27)
The average P
d
for the EGC diversity scheme, P
dEGC
, can
then be obtained by averaging (26) over (27). One can notice
that the PDF in (27) is similar to that in (17) when replacing
each m by L and each by L. This is intuitively correct since
the Nakagami parameter m can be viewed as a diversity order.
Hence, P
dEGC
is equivalent to P
dNak
in (20) after replacing
each m by L, each by L, and each u by Lu.
B. Selection Combining
In the SC diversity scheme, the branch with maximum SNR,

max
, is to be selected. The PDF of
max
for IID Rayleigh
branches is known to be given by
f
max
() =
L

_
1 e
/
_
L1
e
/
. (28)
This PDF can be rewritten as
f
max
() = L
L1

i=0
(1)
i
i + 1
_
L 1
i
_
1
/(i + 1)
e


/(i+1)
. (29)
The PDF in (29) represents a weighted sum of exponential
variates each with parameter

i+1
. Hence, the average P
d
for
the SC diversity scheme, P
dSC
, can be evaluated as
P
dSC
= L
L1

i=0
(1)
i
i + 1
_
L 1
i
_
P
dRay
_

i + 1
_
, (30)
where P
dRay
_

i+1
_
is the P
dRay
obtained in (16) with the
replacement of each by

i+1
.
C. Switch and Stay Combining
We address in this section the evaluation of the average P
d
for the dual SSC diversity scheme [9]. The PDF of the SNR at
the output of the SSC with dual IID Rayleigh branches,
SSC
,
is [9]
f
ssc
() =
_
1

e
/
_
1 e
T /
_
, <
T
,
1

e
/
_
2 e
T /
_
,
T
,
(31)
where
T
is the switching threshold. The average P
d
for
the dual SSC diversity scheme, P
dSSC
, can be obtained by
averaging (14) over the PDF in (31) yielding
P
dSSC
=
_
1 e
T /
_
P
dRay
+
_

T
Q
u
(
_
2,

)
1

e
/
d, (32)
where P
dRay
is given in (16). Relying one more time on the
change of variable x =

2 in the integral part of (32) and
making use of [7, Eq. (32)], P
dSSC
can be obtained in closed
form as
P
dSSC
=
_
1 e
T /
_
P
dRay
+e
T /
Q
u
(
_
2
T
,

)
+
_
1 +

_
u1
e


2(1+)

_
1 Q
u
_
_
2
T
1 +


1 +
__
. (33)
The optimal threshold

T
which maximizes the P
dSSC
can be
obtained by solving
P
dSSC
T
= 0 in (32) yielding

T
=
1
2
_
Q
1
u
(P
dRay
,

)
_
2
, (34)
where Q
1
u
(., .) denotes the inverse uth-order Marcum Q-
function with respect to its rst argument.
VI. NUMERICAL EXAMPLES AND CONCLUSIONS
We are interested in describing the receiver performance
through its ROC curves (P
d
versus P
f
) or complementary
ROC curves (P
m
versus P
f
) for different situations of interest.
Figure 1 illustrates the complementary ROC over a Rayleigh
channel for different average SNR, , values and a time-
bandwidth product u equal to 5. In Figure 1, one can generally
infer that the P
m
P
f
curves have low slopes for P
f
< 0.1.
One can also notice that there is an improvement of about a
half order of magnitude from the P
m
perspective with each
step of 5 dB increase in values starting from 10 to 25 dB.
Figure 2 shows the effect of the Nakagami parameter on
the complementary ROC curves for a Nakagami channel with
= 20 dB and u = 5. There is an improvement of roughly one
order of magnitude from the probability of miss perspective for
m = 2 compared to the Rayleigh case and this improvement
starts to diminish as m increases.
The performance of different diversity systems with dual
branches over IID Rayleigh channels is illustrated in Figure 3
for = 20 dB and u = 5. There is an obvious gain
from the P
m
perspective in the case of diversity systems
compared to the no diversity one. This gain can be quantied
as approximately one order of magnitude for both SC and SSC
schemes, and approximately two orders of magnitude for the
EGC scheme.
Finally, the effect of number of diversity branches is illus-
trated in Figure 4 wherein the SC diversity is considered since
it represents a good compromise between system performance
and receiver complexity. It seems that there is a gain of
roughly one order of magnitude from the P
m
perspective with
a diminishing effect as L increases.
3578
10
4
10
3
10
2
10
1
10
0
10
5
10
4
10
3
10
2
10
1
10
0
Probability of False Alarm P
f
P
r
o
b
a
b
b
ilit
y

o
f

M
is
s
in
g

P
m
<SNR>=10 dB
<SNR>=15 dB
<SNR>=20 dB
<SNR>=25 dB
Fig. 1. Complementary ROC curves for Rayleigh channel at different
values and u = 5.
APPENDIX A
EVALUATION OF G
M
=
_

0
x

e
p
2
x
2
/2
Q
M
(ax, b)dx
With the aid of [7, Eq. (29)], G
M
can be recursively
evaluated as
G
M
= G
M1
+C
M1
F
M
, for > 1, (35)
where
C
M1
=

_
+1
2
_
_
b
2
2
_
M1
e
b
2
/2
2(M 1)!
_
p
2
+a
2
2
_
+1
2
, (36)
and
F
M
=
1
F
1
_
+ 1
2
; M;
b
2
2
a
2
p
2
+a
2
_
. (37)
One can evaluate G
M
iteratively as follows
G
M
= G
M1
+C
M1
F
M
= G
M2
+C
M2
F
M2
+C
M1
F
M1
.
.
.
= G
1
+
M1

n=1
C
n
F
n+1
. (38)
REFERENCES
[1] V. I. Kostylev, Energy detection of a signal with random amplitude, in
Proc. IEEE Int. Conf. on Commun. (ICC02), New York City, New York,
pp. 16061610, May 2002.
[2] H. Urkowitz, Energy detection of unknown deterministic signals, Proc.
IEEE, vol. 55, pp. 523531, April 1967.
[3] H. V. Trees, Detection, Estimation, and Modulation Theory. New York,
NY: John Wiley & Sons, 1968.
[4] J. G. Proakis, Digital Communications. McGraw-Hill, fourth ed., 2001.
[5] C. E. Shannon, Communication in the presence of noise, Proc. IRE,
vol. 37, pp. 1021, January 1949.
[6] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
Products. San Diego, CA: Academic Press, sixth ed., 2000.
[7] A. H. Nuttall, Some integrals involving the Q
M
-function, Naval
Underwater Systems Center (NUSC) technical report, May 1974.
[8] A. H. Nuttall, Some integrals involving the Q-function, Naval Under-
water Systems Center (NUSC) technical report, April 1972.
[9] G. L. St uber, Principles of Mobile Communications. Norwell, MA:
Kluwer Academic Publishers, second ed., 2000.
10
4
10
3
10
2
10
1
10
0
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Probability of False Alarm P
f
P
r
o
b
a
b
b
ilit
y

o
f

M
is
s

P
m
m=3
m=2
m=1
Fig. 2. Complementary ROC curves for Nakagami channel at different m
values ( = 20 dB and u = 5).
10
4
10
3
10
2
10
1
10
0
10
8
10
7
10
6
10
5
10
4
10
3
10
2
10
1
10
0
Probability of False Alarm P
f
P
r
o
b
a
b
b
ilit
y

o
f

M
is
s

P
m
No Diversity
SSC
SC
EGC
Fig. 3. Complementary ROC curves for dual-branch diversity systems over
Rayleigh channel (u = 5 and = 20 dB).
10
4
10
3
10
2
10
1
10
0
10
12
10
10
10
8
10
6
10
4
10
2
10
0
Probability of False Alarm P
f
P
r
o
b
a
b
b
ilit
y

o
f

M
is
s

P
m
L=1
L=2
L=3
L=4
L=5
Fig. 4. Complementary ROC curves for L-branch SC diversity reception
over a Rayleigh channel ( = 20 dB and u = 5).
3579

You might also like