Sie sind auf Seite 1von 22

Wireless Pers Commun (2013) 68:553574

DOI 10.1007/s11277-011-0468-3

A Comparison of Energy Detectability Models


for Cognitive Radios in Fading Environments
Selami Ciftci Murat Torlak

Published online: 22 January 2012


Springer Science+Business Media, LLC. 2012

Abstract In this work, we first analyze the accuracy of different energy detector models in approximating the exact solution in AWGN. These models motivate us to develop
approximation analysis to address energy detection for fading channels. Our analysis develops approximation that has almost the same performance as the exact solution in Rayleigh
channels. Our new model is simple enough to derive the relationship between the required
number of samples (N ) and the signal-to-noise ratio for a single Rayleigh channel similar to
the one obtained for AWGN channels. We also define a fading margin for link budget calculations that relates N in fading channels to AWGN channels. Furthermore, we analyze the
impact of multiple antennas for cognitive radios considering two receiver diversity schemes
and quantify the improvement in performance regarding this margin. All the analytical results
derived in this paper are verified by simulations. Finally, we have implemented and verified
energy detection models in our multiple antenna wireless testbed.
Keywords Cognitive radios Spectrum sensing Energy detector Rayleigh fading
Receiver diversity
1 Introduction
The scarcity of spectrum and its underutilization have led to the development of cognitive
radios [1] for wireless networks. Cognitive radios sense their environment and determine
the most appropriate communication parameters. The first step of this cognitive capability is
to determine whether there exists a user or not at a given time interval in a specified bandwidth. Spectrum sensing techniques in cognitive radios can be classified as matched filter
detection, energy based detection and cyclostationary based detection [4,9]. Matched filter

S. Ciftci
R & D Department, Turk Telekom, AS, Ankara, Turkey
M. Torlak (B)
Department of Electrical Engineering, The University of Texas at Dallas, Richardson, TX 75080, USA
e-mail: torlak@utdallas.edu

123

554

S. Ciftci, M. Torlak

detection is a coherent detection method that maximizes SNR but, a priori knowledge of the
signal to be detected is necessary in this method. Cyclostationary based detection makes use
of the periodicity of the signals statistical characteristics; however, it is a computationally
expensive method. Although energy detectors suffer from noise level ambiguities and require
more samples than matched filter detectors, their easy implementation makes it an attractive
method in spectrum sensing.
Some of the well-known energy detector models to simplify the performance analysis are
considered in this paper for AWGN channel analysis. For instance, Edells [3] and Torrieris [11] models have been developed for low probability of intercept (LPI) radars. Their
performance in approximating the exact energy detection solution is examined in [2]. In
addition, the approximation of energy detectors for cognitive radios in [9] is considered as
the Berkeley model in this paper.
All of the aforementioned models have been used for analyzing the performance of energy
detectors in AWGN channels due to the complexity of the exact solution. These models give
us the connection between the required number of samples, N , and SNR at a given probability of false-alarm, P f , and detection, Pd to have useful engineering insight into the
energy detection problem. We define the number of (real) independent samples N = 2T W
where T and W are the observation time interval and bandwidth of the primary user signal,
respectively. We also define SNR of the primary user signal at the cognitive radio receiver
location as SNR = Ps /n2 where Ps is the average primary user signal power over N samples and n2 is the noise variance of the real additive Gaussian noise samples. Note that
knowing this relationship provides cognitive radio energy detector to determine their sensing
time at a given bandwidth interval. In fading environments, on the other hand, deriving the
closed form solutions for P f and Pd with and without diversity is the main focus in [6,7,13]
and in [12]. Among these references, the first three obtain the closed form Pd solutions by
integrating the Pd expression in AWGN over SNR distribution of the fading environment.
In [12], closed form solutions are obtained via characteristic function of the decision statistic
in fading channels. The main idea in [12] is to find the distribution of the decision statistic in
fading channels. In addition to these, analytical solutions for dual antenna system cognitive
radios are given in [8]. However, these analytical results are not simple enough to derive the
N SNR relationship in fading channels with and without diversity as in AWGN channels.
In this paper, we first derive the probability density function (pdf) of the decision statistic
without making any cumbersome computations as in [12]. Specifically, we propose a new
model based on approximating this pdf using both convolution property and Central Limit
Theorem (CLT). Our model yields simple N SNR relationship that approximates the exact
solution in fading channels. Secondly, we introduce fade margin metric to quantify the effect
of the fading environment on the performance of energy detectors. The square of the fade
margin relates N per antenna in Rayleigh fading channels to N in AWGN to achieve the same
performance level, Pd and P f , at the same SNR in both environments. This margin can be
regarded as a design rule that is used to simplify the performance analysis of energy detector
based cognitive radios. Moreover, we also consider square-law combining and square-law
selection schemes in [7] with multiple antennas for cognitive radios to quantify the improvement in performance regarding this margin.
In Sect. 2, the exact energy detector solution and the models are introduced. A new
approach in finding the pdf of the decision statistic in fading environments and its approximation are presented in Sect. 3 and in Sect. 4, respectively. In Sect. 5, we derive N SNR for
single antenna cognitive radios. The multiple antenna version of this N SNR connection
considering two different diversity schemes is in Sect. 6. Fade margin and its approximation

123

Energy Detectability Models

555

Fig. 1 Block diagram of energy detector

are given in Sect. 7. Section 8 is dedicated to verify our theoretical results via simulation and
experimental results and Sect. 9 concludes this work.

2 Review: Energy Based Spectrum Sensing in AWGN Channels


The spectrum sensing problem can be viewed as a binary detection problem. The binary
hypothesis testing considered in spectrum sensing can be given by

y(i) =

n(i),
H0
s(i) + n(i), H1

(1)

where H0 is null hypothesis and H1 is alternative hypothesis. The small bold letters represent
the vector of length N in (1) which is equal to twice the time-bandwidth product (T W ).
The noise samples n(i) is assumed to be additive white Gaussian noise with zero mean and
variance of n2 . The primary signal samples s = [s(1), s(2), . . . , s(N )] are assumed to uncorrelated to noise samples. The average primary signal power over N samples is Ps = s2 /N .
In this paper, we focus on energy detection based method as illustrated in Fig. 1. Given N
received samples, the energy based detection solution for this problem is to find a decision
statistic
=

N


|y(i)|2

(2)

i=1

and to compare it with a threshold . If  , the detector declares the presence of a primary
user in the spectrum.
2.1 Exact Solution [6]
According to the exact solution, which is based on the work in [5], the received samples are
Gaussian random variables. When the signal exists, the decision statistic, , is a noncentral 2 random variable with N degrees of freedom and a noncentrality parameter, . This
parameter is equal to the ratio of signal energy measured over the sensing time to two-sided
noise spectral density. Throughout this paper, SNR is defined as the ratio of average primary
signal power to noise variance, SNR = Ps /n2 . Due to this, the relationship between and
SNR is given by = N SNR. When there is no signal in the energy detectors input,  has a
central 2 distribution with N degrees of freedom. The decision statistic in each hypotheses
can be summarized as
 2
H0
N ,
(3)
=
N2 (), H1

123

556

S. Ciftci, M. Torlak

where m2 () is a 2 variate with m degrees of freedom and a non-centrality parameter .


The probability density function of the test statistic is

1
(N 2)/2 e/2 ,
H0
N /2
/2)
f  () = 21  (N
(4)


(N 2)/4 (+N SNR)/2

e
I(N 2)/2 ( N SNR), H1
2 N SNR
where (.) is the gamma function and Im (.) is the mth order modified Bessel function of the
first kind.
Closed form expressions for probability of false-alarm, P f , and probability of detection,
Pd , in AWGN channels are computed as in [6,7] for the performance analysis of this energy
detector. They are given by
(N /2, /2)
(N /2)

Pd = Q N /2 ( N SNR, )

Pf =
and

(5)

where Q m (.) is the generalized Marcum Q-function [10].


2.2 Gaussian Approximation with Low SNR (Edell [3])
As the number of samples increases, N 20, the distribution of the test statistic under both
hypothesis converges to Normal distribution due to the Central Limit Theorem.
Edell [3] has exploited this feature in the radiometer problem and computed the mean, ,
and variance, 2 , of the test statistic under both hypothesis. Thus, for large N , the statistics
of the decision variable, , is modeled as

 
N  N n2 , 2N n4
 under H0
=
(6)
N N n2 (1 + SNR), 2N n4 under H1
For small SNR values (SNR << 1), the variances of the test statistic under both hypothesis
are assumed to be equal to each other. According to this model, P f and Pd can be written as




N
N (1 + SNR)
Pf = Q
and Pd = Q
(7)

2N
2N
where Q(.) is the well known tail probability of a zero-mean unit-variance Gaussian random
variable; is the normalized threshold = /n2 to satisfy the given P f and Pd values. The
relationship between the number of samples, N , and SNR in AWGN channels for a given
P f and Pd is given by
N = 2(Q 1 (P f ) Q 1 (Pd ))2 SNR2 .

(8)

This result is exactly the same with the one given in [2] except that the definition of SNR is
defined as the ratio of the total energy over sensing time and the noise variance. The required
SNR can readily obtained from (8) as


SNRreq = 2 Q 1 (P f ) Q 1 (Pd ) / N
(9)

123

Energy Detectability Models

557

2.3 Regular Gaussian Approximation


Regular Gaussian approximation as originally stated in [11] consider mean and variance of
the test statistics without low SNR restriction:

 
N  N n2 , 2N n4
 under H0
(10)
=
N N n2 (1 + SNR), 2N n4 (1 + 2SNR) under H1
Using these values, P f and Pd can be expressed as in (11) and (12).


N
Pf = Q
2N (1 + 2SNR)


N (1 + SNR)
Pd = Q
2N (1 + 2SNR)

(11)

(12)

The relationship between the required SNR and the number of samples N given Pd and P f
be found as

2
1
SNRreq =
( ) + (, , N )
(13)
N
N

2 2
4
(14)
1
+
(, , N ) = 2 2 2N 1 +
N
2N
where, = Q 1 (P f ) and = Q 1 (P f ). The N SNR relationship in (13) is the same as
the one in [2] except that the SNR is defined as in Edells model.
2.4 Extended Berkeley Model
In this model, we build upon the approach taken in [9] which is referred as Berkeley Model
in this paper. Our extended approach may be the closest analytical approximation to the exact
N SNR relation established in 2 among the methods published in literature [2]. Berkeley model applies Gaussian approximation to the log-likelihood ratio rather than signal and
noise samples directly. The details of this derivation can be found in [9]. Below we extend the
Berkeley derivation. The log-likelihood ratio in the Neyman-Pearson test for the detection
problem is given by



Pr(y(i)|H1 )
LLR(y(i)) =
log
Pr(y(i)|H0 )
i=1



2
(y(i)x(n))
N
2
1
2
n
n=1 pn 2 2 e

N




=
log

y(i)2

i=1
1 2 e 2n2
N


2 n

:=

N

i=1


f

y(i)
n2


(15)

123

558

S. Ciftci, M. Torlak

where f (t) := log


decision strategy


N
n=1


2
pn exp( x(n)

x(n)t)
as defined in [9]. Based on the optimal
2 2
n

H1

LLR(Y)
H0

and using CLT P f and Pd are approximately evaluated as follows:

N
m

N
m
0
1
Pd = Q 

Pf = Q 
N 02
N 12
where

 


y(i)
m0 = E f
|H 0
2
  n 

y(i)
H
m1 = E f
|
1
n2

 


y(i)
= Var f
|H 0
2
  n 

y(i)
H
12 = Var f
|
1
n2

02

(16)

(17)
(18)

These equations are exactly the same to the one given in [9]. Since we need to find N SNR
2
relation, we need to compute the quantities m 0 , m 1 , 02 , and
 1 . In order to find these quantities as a function SNR, the Berkeley model expands f
(due to low SNR assumption) to get
! N
"
2

x(n)2 x(n)t
f (t) = log
pn e 2n

y(i)
n2

using the Taylor series at 0

(19)

n=1

N

n=1



x(n)2 2 x(n)3 3 x(n)4 4
pn x(n)t +
t +
t +
t +
2
3!
4!

(20)


. Different from [9], we include terms up to the forth order in t to evaluate
where t = y(i)
n2
the mean and variance quantities above more precisely as a function of SNR. Furthermore,
for evaluating mean and variance, we only include terms up to SNR2 and SNR3 , respectively.
The higher order terms of SNR are neglected. We need the first four raw (sample) moments
of the primary signal to generate the approximated values above:
N


pn x(n) = 0 (zero mean)

n=1
N


N


pn x(n)2 = Ps (average power)

n=1

pn x(n)3 0

n=1

N


pn x(n)4 3Ps2

(21)

n=1

The third and fourth sample moments are approximated by assuming that the primary signal follows Gaussian distribution up to fourth moment. If these moments are known for the
primary signal, they can be used to fine tune the approximation. Here, we now provide the
key results:
m1 m0 =

123

SNR2
2

02 =

SNR2
SNR2
12 =
(5SNR + 1)
2
2

(22)

Energy Detectability Models

559

Error (dB) in required SNR from exact results

0.1

0.2

0.3
GaussianLow SNR
Berkeley/Gaussian
Modified Berkeley

0.4

0.5

0.6

0.7

0.8
2
10

10

10

10

Number of samples, N
Fig. 2 Comparison of energy detector models in AWGNthis shows the required SNR deviation of models
from the exact solution in dB to meet Pd = 0.9 and P f = 106 for a given number of samples

The derivations can be found in the Appendix. Thus, the required SNR with the modified
Berkeley model can be expressed as

SNRreq =

2
1
( ) + e (, , N )
N
N

(23)

where


2

25
10
1
+
e (, , N ) = 5 2 2N 1 +
2N
2N

(24)

which is slightly different from the regular Gaussian approximation result in Sect. 2.3. Figure 2 compares the extended Berkeley model to the other approximation and the exact results.
The extended model is within 0.2 dB of the exact results.

3 Energy Detection in Rayleigh Fading Channels


In this section, we derive the probability density function (pdf) of the decision statistic in
Rayleigh fading channels without diving into any cumbersome computations. To do that, our
approach is to find the decision statistics characteristic function of whose Fourier Transform
gives us the pdf of the decision statistic. Understanding the Rayleigh channel characteristics
help us find the characteristic function, so the Rayleigh channel model in (25) is the starting
point of our approach.

y=

n,
H0
hs + n, H1

(25)

123

560

S. Ciftci, M. Torlak

In this model, bold letters represent a complex vector of length u that is half the total number
of samples, N . Here, the vector n contains the noise samples whereas s is the deterministic
signal vector. The complex variable h is the channel coefficient and it is assumed to be a
complex Gaussian variable with zero mean and unit variance for Rayleigh channels.
In AWGN channels, the decision statistics characteristic function is determined intuitively
by the elements of received y vector. Since y is multivariate complex Gaussian distributed
because of the statistics of the received samples, the distribution of the absolute square of
them is either central or non-central 2 . In this way, we can easily write the characteristic
function of the decision statistic. In any fading channels, we do not need to find the characteristic function of the decision statistic under null hypothesis because it is just noise. Thus,
it is the characteristic function of the summation of random variables each of which has a
central 2 distribution of degree 2. To determine it under alternative hypothesis in Rayleigh
fading channels intuitively, we apply a Householder Transformation to y under alternative
hypothesis. Due to linearity of this transformation, transforming y vector means to transform
hs and n. According to the Rayleigh channel model in (25), the samples of deterministic s are
in every dimension of signal space. There exists a Householder transformation that makes
all the deterministic samples lie in one dimension. This transformation matrix is of the form
in (26) as given in [15].
U=I

zz H
zH a

(26)

where a, b C u . The z vector is given by z = a b where a and b in our case are given by
a = hs and b = a[1 0 . . . 0]T

(27)

The circularly n vector exposed to this transformation preserves its characteristics. The illustration of this process is in the following.
y = Uy = hUs + Un

1
.
0 ..
#

n
= |h|2 E s . +
.. .
..
0

(28)

(29)

The transformed y vector is now composed of u 1 circularly Gaussian distributed noise


samples with zero mean and unit variance and one circularly Gaussian distributed sample
with zero mean and variance equal to 2n2 + E s . Since the Householder Transformation is a
norm preserving unitary transformation, the decision statistic is nothing but the sum of the
absolute square of the elements of y . Hence, the decision statistic is the sum of u 1 central
2 distributed variables of degree 2 of 2n2 variance and one central 2 distributed variable
of degree 2 whose variance is 2n2 (1 + uSNR). To summarize, the characteristic function of
the decision statistic under both hypothesis is given in (30).

H0
(1 2 jwn2 )u ,
(1 2 jwn2 )1u
( jw) =
(30)

, H1
1 j2wn2 (1 + uSNR)
In this way, we obtain the characteristic function of the decision statistic without applying any
cumbersome computations as in [12]. Similarly, the characteristic function of the decision
statistic in AWGN channels is also obtained by assuming h is a deterministic. In this case,

123

Energy Detectability Models

561

it is either the characteristic function of central 2 or that of noncentral 2 as summarized


in (31).

(1 2 jwn2 )u , 
H0


2
jw | |h| =
(31)
jwu|h|2 SNR
2
u
(1 2 jwn ) exp 1/(2 2 ) jw , H1
n

The pdf of the decision statistic in Rayleigh channels is obtained by taking the Fourier
Transform of the characteristic function in (30). False alarm and detection probabilities in
Rayleigh channels are computed by integrating above pdfs of the decision statistic from a
threshold value to infinity:
(u, /(2n2 ))
P f =
(32)
(u)

 2
 u2


u1
n + SNR
n
1


+
Pd = exp 2
2n
(n + 1) 2n2
SNR
n=0
!
 u2



n "


SNR

1
exp 2
exp
2(n2 + SNR)
2n
(n + 1) 2(n2 + SNR)
n=0

(33)
where u = N /2 and is the threshold to satisfy the given P f and Pd values. Since linear
functions of N are in the boundaries of the summation terms in (32), it is not possible to
derive a simple N SNR relationship for Rayleigh channels using the exact energy detector
solution. Due to this, we propose an approximate energy detector model for Rayleigh fading
channels based on approximating the pdf of the decision statistic.

4 Approximation of the Energy Detection Statistics in Fading


Since all those energy detector models in AWGN channels are based on approximating the
distribution of the decision statistic via Central Limit Theorem (CLT), we need to invoke
CLT for the decision statistic in Rayleigh fading channels to see the applicability of these
models to Rayleigh fading channels. To do that, we consider the sum of absolute square of
each element in y . Classical CLT states that the sum of a large number of independent and
identically distributed random variables is approximated by normal distribution if they have
finite variances. The sum of independent variables is approximated by CLT if they satisfy
Lyapunovs condition [14]. However, the random variables summed in the equation of the
decision statistic do not satisfy this condition. However, we rewrite the decision statistic in
terms of y as
r = y H y = | y1 |2 +

u

i=2

| yi |2 = Y1 +

u


Yi .

(34)

i=2

where u = N /2. The mean and variance of 2 distributed Yi s are


 2
2n (1 + uSNR), i = 1
i =
and
2n2 ,
i = 2, . . . , u
 4
4n (1 + uSNR)2 , i = 1
Y2i =
4n4 ,
i = 2, . . . , u

(35)
(36)

123

562

S. Ciftci, M. Torlak

If we consider the sum of Yi s for i = 2, . . . , u as a single 2 distributed random variable,


its mean and variance is
Z=

u


Yi Z = 2(u 1)n2 and Z2 = 4(u 1)n4 .

(37)

i=2

The variance of Y1 is much much greater than that of Z for large sample size values because
Y21 is proportional to u 2 whereas Z2 is proportional to u.
The probability density function and hence, Pd equation is very complicated. Our aim is to
find a simplified version of Pd . For large sample sizes, Central Limit Theorem is the very first
solution to this kind of problems but, it doesnt work in this case as shown in Appendix. The
reason is that the variance of one of the summed terms dominates the variance of the decision
statistic. Our approach for this case is based on the convolution property. Since the decision
statistic is nothing but the sum of independent random variables, its pdf is the convolution
of the summed terms pdfs. Since Y21 >> Z2 for large sample sizes, the variance of the
decision statistic is approximated to be equal to the variance of Y1 . The mean value, however,
is the sum of the summed terms mean values. This idea can be illustrated by considering the
convolution of any continuous function with delta dirac function. Thus, our approximated
pdf of the decision statistic for a single antenna case becomes


r /n2 2(u 1)
1
f r (r ) =
exp

for r 2(u 1)n2 (38)


2n2 (1 + uSNR)
2(1 + uSNR)
5 N-SNR Relation in Fading Channels
Our goal is to derive an approximated N SNR relation under Rayleigh fading. First, recall
P f which is already approximated in 7. Then, Pd for a single antenna cognitive radios can
be computed by integrating (38) from a threshold to infinity as
Pd =

'
,2(u1)n2



1
r /n2 2(u 1)
exp

dr
2n2 (1 + uSNR)
2(1 + uSNR)

(39)

We rewrite the threshold value in (39) from (7) as

= 2u + 2 u Q 1 (P f ).
n2
Then, substituting this value into (39) yields


u + 1

where = Q 1 (P f ).
Pd = exp
1 + uSNR
From (41), we obtain the following quadratic equation.
 
 

log Pd SNR u u (1 + log Pd ) = 0

(40)

(41)

(42)

By neglecting non-dominating terms (i.e., O(1/SNR)) in the exact solution of (42), we obtain
our desired N SNR expression for Nr = 2u as in (49).

2

2
 
(43)
Nr =
log Pd
SNR2

123

Energy Detectability Models

563

The beauty of the expression in (43) is that it depends only on P f and Pd , which are the
predetermined values for system design.
6 N SNR Relationship in Multi-Antenna Energy Detection
In this section, we consider that there are multiple antennas in cognitive radios and each of
them sense the same spectrum based on energy detection in fading environments. Our focus
is to understand how multiple antennas in cognitive radios, each of which employs energy
detector, advance their spectrum sensing performance. We consider two different combining
schemes for cognitive radios in this section. The first one is the square-law combining and
the other one is square-law selection. The exact expression both methods are already derived
in [7].
As with the single antenna case, we derive the N SNR relationship at a specified Pd
and P f for any number of antennas for these techniques.
6.1 Square-Law Combining (SLC)
The first combining technique is Square-Law Combining technique in which the output
of each antennas energy detector is added. The cognitive radio decides whether there is
a primary user or not by comparing the output of this summation to a threshold value. In
this way, the decision statistic is nothing but the sum of the outputs of each antenna. The
approximated pdf of the decision statistic for L antennas in this scheme is of the form

 L1


 S LC /n2 2L(u 1)
 S LC /n2 2L(u 1)
f  S LC ( S LC ) =
. (44)
exp
2(1 + uSNR)
(2(1 + uSNR)) L (L 1)!
Pd,S LC is found to be



/n2 2L(u 1)
1
 L,
(45)
(L)
2(1 + uSNR)
where is the threshold which is again calculated from a specified P f,S LC . Using the
definition of regularized incomplete gamma function, we can also express Pd,S LC as


(a, z)
/n2 2L(u 1)

where P(a, z) =
.
(46)
Pd,S LC = P L ,
2(1 + uSNR)
(a)
Pd,S LC =

With CLT approximation for the distribution of the decision statistic for L antennas, P f,S LC
is approximated as


/n2 2Lu

(47)
=
Q
P f,S LC

2 Lu
Obtaining the threshold value from (47) and substituting it to (46) yields the quadratic
equation in (48).





P 1 L , Pd,S LC SNR u L u (L P 1 L , Pd,S LC ) = 0
(48)
The desired N SNR expression for L i.i.d. antennas with SLC scheme is in (49).

2

L
2


N S LC =
P 1 L , Pd,S LC
SNR2

(49)

As expected, (49) is equivalent to (43) for L = 1.

123

564

S. Ciftci, M. Torlak

6.2 Square-Law Selection (SLS)


In this scheme, the decision statistic is the output of one of the energy detectors that has the
highest energy. The outputs of each antennas energy detector is compared and the highest
one is selected via a switch. The decision whether there is a primary user or not is made by
comparing the selected antennas output with a threshold value.
Similar to the analysis for SLC scheme, the probability of false alarm is represented by
P f,S L S and the probability of detection is represented by Pd,S L S for multiple antennas in
Rayleigh fading channels. As in SLC scheme, we have the same i.i.d channel assumption in
the analysis here.
Let L be the number of antennas, P f,S L S is given by

L
P f,S L S = 1 1 P f

(50)

where P f is defined in (7). Assuming that the average SNR is equal for each branch of L
Rayleigh channels, Pd,S L S is given by
L

(51)
Pd,S L S = 1 1 Pd .
With little manipulation of (51) and using (46), we obtain



/n2 2(u 1)
L
1 1 Pd,S L S = P 1,
.
2(1 + uSNR)

(52)

This equation yields a quadratic equation. Making some assumptions, which we have done for
both single antenna and multi-antenna with SLC scheme cases, helps us solve that quadratic
form for S L S . This solution gives us the desired N SNR relationship as

NS L S =

where = Q 1


1


L



SNR2
L
log 1, 1 1 Pd,S L S

1 P f

(53)


.

7 Fade Margin
We introduce the fade margin to quantify the effect of the fading environment on the performance of energy detectors. The fade margin for SNR is the factor that relates the required
average SNR per antenna in Rayleigh fading channels to the required SNR in AWGN to
achieve the same performance level by taking the same number of samples in both environments. Let us consider the following scenario. Assume that we know the required SNR in
AWGN to achieve some given P f and Pd values by taking N samples. Our aim is to find the
required average SNR in Rayleigh fading channels to achieve the same P f and Pd with the
same number of samples.
In order to obtain the fade margin, we make use of N SNR relationship given in (43).
The assumption of fixing P f (P f in AWGN), Pd (Pd in AWGN) and N values and N =
Nawgn = Nray . These assumptions allows us to replace the term Nr in (43) by the right
hand side of the equation in (8). Thus, we define fade margin (F) as the ratio between the

123

Energy Detectability Models

565

required SNRs in Rayleigh and AWGN channels for energy detection with fixed observation
intervals:


1
SNR R

F=
=
.
(54)
SNR A
( )) log ( Pd )
where SNR R and SNR A are the required SNRs in Rayleigh and AWGN channels, respectively. Here, our fade margin is nothing but
 

1

.
(55)
F =

k
where k = ln (Pd ). Similarly, we can also define a factor for some number of samples
that relates the required number of samples per antenna in Rayleigh fading channels to the
required number of samples in AWGN to achieve the same performance level at the same
SNR values in both environments. In this case, one can calculate how many samples he needs
to achieve the same performance in Rayleigh fading environment as compared to AWGN
environment. Using a similar approach to fade margin, we start from (43) and replace the
average SNR term with the one in (8) based on our P f and Pd . In this way, we obtain the
factor which is nothing but the square of our fade margin given in (56).

NF =

2 

1
log(Pd )

2
(56)

A more generalized version of this margin can be obtained by considering cognitive radios
with multiple antennas. Following the same steps we take for single antenna case, the fade
margin for multi-antenna SLC case is obtained by using (8) and (49) as



1

F,slc =
(57)

 1 (L , Pd )
Similarly, the fade margin for multi-antenna SLS case is obtained by using (8) and (53) as




1


F,sls =
(58)


log 1, 1 L 1 Pd
After finding this margin for multiple antenna systems, we will try to demonstrate how useful
this margin is from a system designer perspective. Assume that we know the required number
of samples per antenna, N1 , for a single antenna in Rayleigh fading environment for given
Pd , P f and average SNR values. We wonder how few number of samples per antenna is
required when we increase the number of antennas from 1 to L. Assume that the required
number of samples for L antennas is N L . So,
N L = N1

fading factor for L antennas


fading factor for a single antenna

(59)

Consider using SLC scheme for multiple antenna case. Thus, (59) becomes

N L = N1

1
 1 (L ,Pd )

1
 1 (1,Pd )

2
= N1

 1 (1, Pd )
 1 (L , Pd )

2
(60)

123

566

S. Ciftci, M. Torlak

where = erfc1 (2P f ) and = erfc1 (2Pd ). Using SLS scheme, a similar version of (60)
is obtained:


 2

 1 1, 1 1 1 Pd


(61)
N L = N1

 1 1, 1 L 1 Pd
We make the following definitions for our purposes in this section:
N F,slc : The ratio between the required number of samples (Nr eq ) per antenna for L
antennas and that for a single antenna with SLC
scheme at fixed Pd , P f and SNR.
N F,sls : The ratio between the Nr eq per antenna for L antennas and that for a single
antenna with SLS
scheme at fixed Pd , P f and SNR.
F,slc : The ratio between the required SNR per antenna for L antennas and that for a
single antenna with SLC
scheme at fixed Pd , P f and number of samples.
F,sls : The ratio between the required SNR per antenna for L antennas and that for a
single one with SLS
scheme at fixed Pd , P f and number of samples.
The mathematical expressions for N F,slc and N F,sls are given in (62).
2
 1
P (1, Pd )
N F,slc =
P 1 (L , Pd )


 2

P 1 1, 1 1 1 Pd


(62)
N F,sls =

P 1 1, 1 L 1 Pd
The expressions for F,slc and F,sls is the square root of the ones in (62).



P 1 1, 1 1 1 Pd
P 1 (1, Pd )

F,slc = 1
and F,sls = 1 

P (L , Pd )
1, 1 L 1 Pd
P

(63)

All of these expressions are nothing but a ratio of two inverse of upper incomplete gamma
functions. Since this function is not used as frequently as any other function like rational or
power functions, it is really difficult to get an intuition how the value of N F,slc or the other
ones change. Therefore, a simplified expression for inverse of upper incomplete gamma function can be obtained by using the findings of [16]. In [16], the connection of lower incomplete
gamma function with confluent hypergeometric function [17] is established as
xa
(64)
1 F1 (a, 1 + a, x)
a
where g(., .) is the lower incomplete gamma function and 1 F1 (a, 1 + a, x) is the confluent hypergeometric function. As x 0,1 F1 (a, 1 + a, x) 1 and the inverse of lower
incomplete gamma function is given by

g 1 (a, x) = a ax
(65)
g(a, x) =

Benefiting from the above approximation, we express the upper incomplete gamma function
in our expressions in terms of lower incomplete gamma function. By definition, upper and
lower incomplete gamma functions satisfy
(a, x) = (a) g(a, x).

123

(66)

Energy Detectability Models

567

Using (66), we rewrite (45) and (52) as



2 slc Nslc 1
1
g L,
(L)
(slc )2 Nslc



2 (sls ) Nsls 1
= g 1,
Nsls (sls )2

1 Pd,slc =
#
L

1 Pd,sls

(67)

With these representations, the expressions for the new definitions in this section can be
rewritten as
2
 1
g (1, 1 Pd )
N F,slc =
g 1 (L , 1 Pd )


 2
g 1 1, 1 1 Pd



(68)
N F,sls =
g 1 1, L 1 Pd
g 1 (1, 1 Pd )
g 1 (L , 1 Pd )


g 1 1, 1 1 Pd

.
= 1

g
1, L 1 Pd

F,slc =
F,sls

(69)

As Pd 1, (1 Pd ) 0 and these inverse functions can be approximated by (65).


Hence, (68) and (69) can be expressed as
N F,slc

1
(L + 1)

2
L

(1 Pd )2(1 L )

N F,sls (1 Pd )2(1 L )
1
1
F,slc
(1 Pd )1 L
1
(L + 1) L
1

F,sls (1 Pd )1 L .

(70)

(71)

8 Simulation and Experimental Results


In the simulations, we compare our proposed energy detector model with the exact solution
and figure out how well it approximates the exact solution in Rayleigh channels. All the simulations are evaluated in Mathematica to get accurate results. We also use the fading factor
for SNR and required number of samples to find the required SNR and N in Rayleigh fading
channels and compare these values with exact values. We also show how our fading factor
behaves as Pd varies.
Figure 2 plots SNR error on y-axis and shows how SNR difference between the approximated models and the exact solution varies with respect to number of samples, N . On average
extended Berkeley model is approximately 0.1 dB better than the second best model, which
is the regular Gaussian approximation. As can be easily seen in Fig. 3, SNR error decreases
as N increases.
Since some assumptions are made in our derivation of (43), we need to verify how well it
approximates our exact model in Rayleigh fading channels. To do that, we compute required
number of samples needed to be taken in Rayleigh channels to meet Pd = 0.9 and P f = 106
for some given SNR values using (43). In order to obtain our exact models N values to

123

568

S. Ciftci, M. Torlak
10
1 antenna

2 antennas
3 antennas

Fading Factor

4
0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Pd
Fig. 3 Rayleigh channels with SLS schemethis shows how the fading factor varies with probability of
detection for different number of antennas in SLS scheme
Table 1 Estimated and exact N
values of the approximated model
in Rayleigh fading channels

SNR (dB)

Estimated N

Exact N

10,228

10,178

25,721

25,566

64,567

64,218

162,074

161,307

10

407,160

405,183

meet the above specifications, we use (33). As can be seen in Table 1, the error between our
approximated and exact values of our model is 0.5%.
In Table 2, we compute the required SNR values in AWGN and Rayleigh channels to
meet Pd = 0.9 and P f = 106 for some given N values using (8) and (43). The fading effect
can be seen by looking at the differences in required SNR values in the first two columns of
the table for a fixed number of samples. We also put the required SNR values for the above
specifications obtained by the exact energy detector solution in a Rayleigh fading channel
using (33) as the third column. All the SNR values are in dB. In this way, we can compare
how well (43) estimates the exact solution. As expected, the difference between the SNR
values obtained by our derivation and the ones obtained by the exact energy detector solution
decreases as the number of samples increases.
Tables 3 and 4 give us a chance to make a comparison between our estimated SNR values using the fade margin for 2 antennas in a Rayleigh channels that deploys either SLS or
SLC diversity schemes and the exact SNR values. As in the previous simulations, the given
performance measures in these simulations are Pd = 0.9 and P f = 106 . As expected, the
difference between the estimates and the exact results decrease as N increases.
In all the previous simulations, we consider Pd = 0.9 and P f = 106 because these
are the desired values for these performance measures. Having the Pd value as close to 1

123

Energy Detectability Models


Table 2 Required SNR values
in AWGN and Rayleigh channels
for a single antenna

Table 3 Required SNR values


in AWGN and Rayleigh channels
with 2 SLS antennas

Table 4 Required SNR values


in AWGN and Rayleigh channels
with 2 SLC antennas

569
N

SNR using (8)

SNR using (57)

SNR using (33)

200

2.1932

6.5433

7.2333

500

4.1829

4.5536

4.9959

1,000

5.6881

3.0484

3.3617

1,500

6.5686

2.168

2.4232

2,000

7.1932

1.5433

1.7636

SNR using (8)

SNR using (58)

SNR using (51)

200

2.1932

500

4.1829

1.01897

0.6005

1,000

5.6881

2.57096

2.248

1,500

6.5686

3.40456

3.1902

2,000

7.1932

4.02926

3.841

SNR using (8)

SNR using (57)

SNR using (45)

200

2.1932

1.0176

1.615

500

4.1829

0.9721

0.655

1,000

5.6881

2.4772

2.11

1,500

6.5686

3.3577

3.061

2,000

7.1932

3.9824

3.6415

0.970733

1.6202

provides the secondary user with cognitive radios not to interfere the primary users by high
detection probability. However, we consider the fading factor for different Pd values in order
to see how it is affected by Pd values. The simulation results to exploit how fading factor
varies with Pd values are given in Fig. 3 and (57) depending on the diversity scheme. Not
surprisingly, fading factor increases as Pd increases. Similarly, the variation in fading factor
values decreases as the number of antennas increases. This is expected because the fading
factor is calculated with respect to SNR values in AWGN and as the number of antennas
increase, the fading effect on system decreases and gets closer to AWGN system. It is also
seen that the variation in the SLC results are more than those in SLS results.
We demonstrate how the exact ratio of the required number of samples per antenna between
L antennas and a single antenna varies for both SLC and SLS schemes in (5). As given in
Sect. 7, the exact expression is the ratio of two inverse-incomplete gamma functions. The
approximated ratios, which are given in (70), are also plotted in Figs 4 and 5 to show how
close they approximate the exact solution. We also plot the variation of the same ratio for
SNR in Fig. 6. As expected, the differences between consecutive number of antennas with
SLC and that with SLS decrease as the number of antennas in the system increases. As the
collaborated antennas increase, the strength of SLC scheme over SLS scheme becomes more
obvious. On the other hand, the differences between the exact ratio and approximated ratio
does not seem to get closer as the number of antennas increases. The reason behind this is

123

570

S. Ciftci, M. Torlak
10
1 antenna
8

2 antennas
3 antennas

Fading Factor

6
0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

Pd

Fig. 4 Rayleigh channels with SLC schemethis shows how the fading factor varies with probability of
detection for different number of antennas in SLC scheme
0
Exact Result for SLS in Rayleigh Channels
Approximated Result SLS in Rayleigh Channels
Exact Result for SLC in Rayleigh Channels
Approximated Result for SLC in Rayleigh Channels

Ratio for Number of Samples (dB)

5
10
15
20
25
30
35
40
45

Number of Antennas (L)

Fig. 5 The difference of required number of samples difference with respect to number of antennasthis
shows how the exact ratio of the required number of samples between L antennas and a single antenna in
Rayleigh channels with SLC and SLS schemes

that our approximation is valid as Pd 1 and (1 Pd ) 0. If we choose Pd = 1 106 ,


the differences become smaller and smaller.
We also conduct some experiments to see the applicability of the AWGN and Rayleigh
channel models using our multiple antenna testbed. A detailed description of the testbed is
provided in [18]. Our first experimental result shows the relationship between probability of

123

Energy Detectability Models

571

0
Exact Result for SLS in Rayleigh Channels
Approximated Result SLS in Rayleigh Channels
Exact Result for SLC in Rayleigh Channels
Approximated Result for SLC in Rayleigh Channels

Ratio for SNR (dB)

10

15

20

25

Number of Antennas (L)


Fig. 6 The difference of required SNR with respect to the number of antennasthis shows how the exact
and approximated ratio of SNR between L antennas and a single antenna in Rayleigh channels with SLC and
SLS schemes

detection, Pd , and signal-to-noise ratio, SNR, at a fixed number of samples, N . To do that,


different experiments were held at different locations to obtain different SNR values. For our
analysis purposes, we preferred to take N = 200 for each experiments in order to calculate
the probability of detection. As in the previous results, the threshold is determined via fixing
the false-alarm probability to 0.05. As expected, Pd value increases as SNR increases for
fixed P f and N values. As shown in Fig. 7, the experimental results match with the theoretical results. The deviation of the experimental results from the theoretical ones is in an
acceptable level. We also conduct the same experiments in a fading environment to see the
applicability of our Rayleigh channel model assumption for our theoretical analysis. We,
again, focus on Pd versus SNR relationship by fixing P f and N for each experiment. Different than the experiments for AWGN channels, 2,700 experiments are considered to calculate
the probability of detection. As shown in Fig. 8, the experimental results are between the
theoretical results for AWGN and Rayleigh fading channel models. That is, our experimental
channel is not as good as AWGN and also is not as bad as the Rayleigh channel that the
model assumes. Since the transmitter and the receiver in our experiments do not directly see
each other, it is expected to see that the experimental results are close to the ones we obtain
theoretically for Raylegh channels. Figure 8 also verifies the validity of our analytical study.
In our analytical work, we assume the channel to be Rayleigh. This assumption gives us the
worst case scenario for fading environments as can be seen in our experimental results.

9 Conclusions
In this work, we have compared various energy detector models to simplify the performance
analysis of exact energy based detection. We have presented a modified Berkeley model that
agrees to within 0.2 dB for the sample sizes N 100. These models provide the relationship

123

572

S. Ciftci, M. Torlak
1

Probability of Detection (Pd)

0.95
0.9
0.85
0.8
0.75
0.7
0.65
Experimental results
Theoretical results

0.6
0.55
7

6.5

5.5

4.5

3.5

SignaltoNoiseRatio (SNR in dB)

Fig. 7 The Pd versus SNR for AWGN channels


1

0.95

Experimental result
Theoretical result by our model in Rayleigh
Theoretical result by Edell in AWGN

Probability of detection

0.9

0.85

0.8

0.75

0.7

0.65

4.5

3.5

Signaltonoise ratio (SNR in dB)

Fig. 8 The Pd versus SNR for indoor fading environment

between the required number of samples (N ) and signal-to-noise ratio (SNR) in AWGN
channels. The complicated false alarm, P f and detection, Pd , probabilities in Rayleigh channels do not yield such simple N SNR relationship in Rayleigh channels as in AWGN
channels. Therefore, we proposed a new model for fading channel analysis to obtain simple
N SNR relationship as in AWGN channels. Based on N SNR in AWGN and Rayleigh
channels, we introduced a fade margin whose square relates the required number of samples
in AWGN to the one in Rayleigh fading channels for the same SNR to achieve the required
performance level. This factor quantifies the performance degradation in Rayleigh channels

123

Energy Detectability Models

573

as compared to AWGN channels. The improvement in the performance of cognitive radios


with multiple antennas is measured via fade margin. Simulation results coincide with experimental results done with the experimental multiple antenna testbed in indoor environment.
The closeness of the simulation and experimental results shows that this margin is promising
in reducing the time spent on cognitive radio design.
Appendix: Derivation of (22)
Following the derivation steps in [9], we approximate f (t) at t = y(i)/n2 0 using
Eqs. (20) and (21):


y(i)
3SNR2 y(i)4
SNR y(i)2
+
(72)
f

2
2
n
2
n
4!
n4
Using the above approximation and ignoring the last term for m 0 and m 1 , we have
SNR E(y(i)2 |H0 )
SNR
=
.
2
2
n
2

(73)

SNR E(y(i)2 |H1 )


SNR(1 + SNR)
.
=
2
2
n
2

(74)

m 0 = E( f (t)|H0 )
Similarly, we have
m 1 = E( f (t)|H1 )

Similar calculations for variances are obtained


SNR2
2 .
2
E(( f (t)2 |H1 ) E(( f (t)|H1 )2 SNR (1+5SNR)
2

02 = E( f (t)2 |H0 ) m 20

12 =

(75)

using
E(( f (t)2 |H1 ) =

SNR2
3SNR3
4
E
(y(i)
|
H
)
+
E(y(i)6 |H1 ) + h.o.t.
1
4n4
4!n6

SNR2 (3n2 + 6Ps n2 + h.o.t) 3SNR3 (15n2 + h.o.t.)


+
4n4
24n6

SNR2 (3 + 6SNR) 15SNR3


+
4
8

and


E(( f (t)|H1 ) =
2

SNR(1 + SNR) 3SNR3 (3n4 + h.o.t.)


+
2
4!n4

2

SNR2 (1 + 2SNR) 9SNR3


+
4
24

References
1. Mitola, J., III. (1999). Cognitive radio for flexible mobile multimedia communication. In Proceedings
of IEEE International Workshop on Mobile Multimedia Communications (pp. 310).
2. Mills, R. F., & Prescott, G. E. (1996). A comparison of various radiometer detection models. IEEE
Transactions on Aerospace and Electronic Systems, 32(1), 467473.

123

574

S. Ciftci, M. Torlak

3. Edell, J. (1976). Wideband, noncoherent, frequency-hopped waveforms and their hybrids in low-probability of intercept communications. Report Naval Research Laboratory (NRL) 8025.
4. Akyildiz, I. F., Lee, W. Y., Vuran, M. C., & Mohanty, S. (2006). NeXt generation/dynamic spectrum access/cognitive radio wireless networks: A survey. Computer Networks Journal (Elsevier), 50,
21272159.
5. Urkowitz, H. (1967). Energy detection of unknown deterministic signals. Proceedings of IEEE, 55,
223231.
6. Digham, F. F., Alouini, M. S. & Simon, M. K. (2003). On the energy detection of unknown signals
over fading channels. In Proceedings of IEEE (pp. 35753579).
7. Digham, F. F., Alouini, M. S. & Simon M. K. (2007). On the energy detection of unknown signals
over fading channels. IEEE Transactions on Communications, 55(1), 2124.
8. Pandharipande, A. & Linnartz, J.-P. M. G. (2007). Performance analysis of primary user detection in a multiple antenna cognitive radio. In IEEE International Conference on Communications
(pp. 64826486).
9. Cabric, D., Tkachenko, A. & Brodersen, R. W. (2006). Experimental study of spectrum sensing beased
on energy detection and network cooperation. In Proceedings on ACM International workshop on
Technology and Policy for Accessing Spectrum.
10. Nuttall, A. H. (1975). Some integrals involving the Q M function. IEEE Transactions on Information
Theory, 21, 9596.
11. Torrieri, D. J. (1992). Principles of secure communication systems. Boston, MA: Artech House.
12. Kostylev, V. I. (2002). Energy detection of a signal with random amplitude. IEEE International
Conference on Communications, ICC 2002, 3, 16061610.
13. Ghasemi, A. & Sousa, E. S. (2005) Colloborative spectrum sensing for opportunistic access in fading
environments. In Proceedings of IEEE Dynamic Spectrum Access Networks (pp. 131136).
14. Bhattacharya, R., & Waymire, E. C. (2007). A basic course in probability theory. New York,
NY: Springer.
15. Chung, K.-L. & Yan, W.-M. (1997). The complex householder transform. IEEE Transactions on
Signal Processing, 45(9), 23742376.
16. Dohler, M., & Arndt, M. (1997). Inverse incomplete gamma function and its application. IEEE
Electronics Letters, 42(1), 3536.
17. Gradshteyn, I. S., & Ryzhik, I. M. (2007). Table of integrals, series, and products. San Diego,
CA: Academic Press.
18. Kim, D. & Torlak, M. (2008). Rapid prototyping of a cost effective and flexible 44 MIMO testbed. In
Proceedings of IEEE Sensor Array and Multichannel Signal Processing Workshop (pp. 58), 2123
July 2008. Darmstadt, Germany.

Author Biographies
Selami Ciftci received M.S. degree in electrical engineering from The University of Texas at Dallas in 2008.
Upon graduation, he worked as a network software engineer at Intel until 2009. He is currently an R&D
specialist with Turk Telekom, Turkey.

Murat Torlak received M.S. and Ph.D. degrees in electrical engineering from The University of Texas at Austin in 1995 and 1999,
respectively. He spent the summers of 1997 and 1998 in Cwill Telecommunications, Inc., Austin, TX, where he participated in the design
of a smart antenna SCDMA system. In the Fall of 1999, he joined
the Department of Electrical Engineering, The University of Texas at
Dallas, where he is currently an Associate Professor. He held a visiting
position at University of California Berkeley in 2008. He has been an
active contributor in the areas of smart antennas and multiuser detection. His current research focus is on interference alignment, experimental platforms for multiuser MIMO antenna systems, millimeter
wave systems, and wireless communications with health care applications. He is an Associate Editor of IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS. He was the Program Chair of IEEE
Signal Processing Society Dallas Chapter during 20032005.

123

Das könnte Ihnen auch gefallen