You are on page 1of 18

4322 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO.

9, NOVEMBER 2010
A Class of Spectrum-Sensing Schemes for Cognitive
Radio Under Impulsive Noise Circumstances:
Structure and Performance in Nonfading
and Fading Environments
Hyun Gu Kang, Student Member, IEEE, Iickho Song, Fellow, IEEE, Seokho Yoon, Senior Member, IEEE, and
Yun Hee Kim, Senior Member, IEEE
AbstractIn this paper, we propose a class of spectrum-sensing
schemes for cognitive radio with receive diversity. By employing
the generalized likelihood ratio test (GLRT) in the detectors on the
antenna branches and exploiting a nonlinear diversity-combining
strategy, the proposed scheme exhibits better performance than
conventional schemes in various fading and noise environments.
Exact expressions of the detection and false-alarm probabilities of
the proposed scheme are derived in nonfading and Nakagami fad-
ing channels with Gaussian noise. Through computer simulations,
it is conrmed that the proposed scheme provides a signicant
performance gain over conventional schemes in impulsive noise
environments.
Index TermsCognitive radio (CR), nonlinear diversity com-
bining, receive diversity, spectrum sensing.
I. INTRODUCTION
O
NE of the most useful and interesting ways to efciently
utilize the spectrum without causing interference to the
primary spectrum user is to employ spectrum sensing in cogni-
tive radio (CR) [1], [2]. The techniques of spectrum sensing in
CR can generally be classied into three categories: 1) coherent
detection; 2) feature detection; and 3) noncoherent detection
[3]. Because the transmission scenario (including the modula-
tion scheme and pilot information for channel estimation) of the
spectrum user is unknown to the CR [3] in most cases, the CR
should normally resort to noncoherent detection. Among the
noncoherent detectors, the energy detector is used most widely
in spectrum sensing because it usually produces a reasonable
Manuscript received December 26, 2009; revised May 2, 2010, June 28,
2010, and August 3, 2010; accepted August 5, 2010. Date of publication
August 19, 2010; date of current version November 12, 2010. This work was
supported in part by the National Research Foundation of Korea, with funding
from the Ministry of Education, Science, and Technology, under Grant 2010-
0015175 and by the Ministry of Knowledge Economy through the Information
Technology Research Center Program of the National IT Industry Promotion
Agency under Grant NIPA-2010-C1090-1021-0007. The review of this paper
was coordinated by Prof. R. Schober.
H. G. Kang and I. Song are with the Department of Electrical Engineering,
Korea Advanced Institute of Science and Technology, Daejeon 305-701, Korea
(e-mail: khg@Sejong.kaist.ac.kr; i.song@ieee.org).
S. Yoon is with the School of Information and Communication Engineering,
Sungkyunkwan University, Suwon 440-746, Korea (e-mail: syoon@skku.edu).
Y. H. Kim (Corresponding author) is with the Department of Electronics
and Radio Engineering, Kyung Hee University, Yongin 446-701, Korea (e-mail:
yheekim@khu.ac.kr).
Digital Object Identier 10.1109/TVT.2010.2068066
detection performance in Gaussian noise and possesses a low
complexity suitable for real-time operations [3][6]. In the
meantime, a variety of diversity-combining techniques such
as equal gain combining (EGC) and selection combining (SC)
have been incorporated in the energy detection for spectrum
sensing [7][10], resulting in the energy-detector-based EGC
(EEGC) and energy-detector-based SC (ESC) schemes.
In most of the schemes designed for spectrum sensing in CR,
it is usually assumed that the noise is Gaussian. Although we
can commonly justify the Gaussian assumption on noise with
the central-limit theorem in several practical communication
systems, we also frequently deal with noise environments,
where the non-Gaussian (impulsive or heavy-tailed) nature of
noise prevails in the system [11][14]. For example, com-
munication receivers are often annoyed by impulsive noise
components from various sources of interference, including
buildings in radar clutter, lightning in the atmosphere, moving
vehicles, and reections from sea waves. Under such impulsive
noise circumstances, spectrum-sensing schemes employing the
EEGC and ESC may be highly susceptible to a severe degrada-
tion of performance.
In this paper, to employ spectrum sensing in the CR with
a multiple of receive antenna branches, we propose a class of
spectrum-sensing schemes called combining with order statis-
tics (COS). Based on the observation that nonlinear schemes
can successfully mitigate the effects of impulsive noise com-
ponents generated from many natural, as well as man-made,
sources of interference as conrmed in many studies on signal
detection [15], [16], the COS scheme exploits nonlinear com-
bining strategies together with the generalized likelihood ratio
test (GLRT) detector. In particular, the test statistic of the COS
scheme is the sum of one or more order statistics of the log-
likelihood ratios (LLRs) obtained at the branches of a receiver.
By reducing the inuence of observations with very large
amplitudes, the COS scheme provides performance stability
for spectrum sensing in impulsive noise environments. Such
an attempt of simultaneously exploiting a nonlinear combining
strategy and the GLRT detector has yet to be reported in other
published works for spectrum sensing, although the GLRT
detector and order statistics have previously been addressed in
several investigations of signal-detection problems. The contri-
butions and novelty of this paper are given as follows: 1) A
0018-9545/$26.00 2010 IEEE
KANG et al.: SPECTRUM SENSING FOR CR UNDER IMPULSIVE NOISE: STRUCTURE AND PERFORMANCE 4323
novel spectrum-sensing scheme, i.e., the COS, is proposed for
the CR with receive diversity; 2) the exact detection and false-
alarm probabilities of the COS scheme are derived in complex
Gaussian noise; and 3) the performance characteristics of the
COS scheme are analyzed and compared with those of the
conventional schemes in various environments.
This paper is organized as follows. Details of the system
model are delineated in Section II. The COS scheme is de-
scribed in detail in Section III, where some examples of the
LLR are also examined for a few representative distributions
of impulsive noise. The detection and false-alarm probabilities
of the COS scheme are derived in complex Gaussian noise
environments in Section IV. The performance characteristics
of the COS scheme for spectrum sensing are compared through
computer simulations to those of the EEGC and ESC schemes
in various noise environments in Section V.
II. SYSTEM MODEL
For a CR with L receive antenna branches, let us denote
the low-pass discrete-time observation vector on the lth receive
antenna branch by
R
l
=
_
R
1,l
, R
2,l
, . . . , RN
2
,l
_
(1)
where R
n,l
= R
I,n,l
+ jR
Q,n,l
, with the subscripts I and Q
denoting the in-phase and quadrature components, respectively,
and the positive integer (N/2) is the sample size per either
component. Here, assuming that B samples are acquired per
second when the signal bandwidth is B (in hertz) and the
sampling period of R
l
is T (in seconds), the sample size can be
determined as (N/2) BT from the timebandwidth product
[17]. In this paper, we also assume a single transmit antenna
without loss of generality because, as it is well known [18], the
technique of transmit diversity with multiple transmit antennas
does not provide performance enhancement when the total
transmitted signal power is xed and the transmission schemes
of the spectrum user is unknown to the CR.
Then, the observation vector can be expressed as
R
l
= W
l
(2)
for l = 1, 2, . . . , L under the null hypothesis H
0
that the
frequency band is not currently used by a spectrum user
and as
R
l
= H
l
S + W
l
(3)
for l = 1, 2, . . . , L under the alternative hypothesis H
1
that
the frequency band is currently used. In (2) and (3), H
l
=
H
I,l
+ jH
Q,l
is the complex channel gain on the lth branch,
which is assumed to be constant during a symbol time (block
fading), S = [S
1
, S
2
, . . . , S
(N/2)
] is the transmit signal vector
of the spectrum user, with S
n
= S
I,n
+ jS
Q,n
, and W
l
=
[W
1,l
, W
2,l
, . . . , W
(N/2),l
] is the vector of independent and
identically distributed (i.i.d.) complex additive noise compo-
nents W
n,l
= W
I,n,l
+ jW
Q,n,l
with the common joint prob-
ability density function (pdf) f
W
I
,W
Q
of (W
I,n,l
, W
Q,n,l
).
Note that H
l
S = [H
l
S
1
, H
l
S
2
, . . . , H
l
S
(N/2)
] denotes the
block-faded transmit signal vector on the lth branch. It is
well known [19] that the uncertainties in noise distribution
impose fundamental limitations on the detection performance
of noncoherent detectors, and this condition is called the signal-
to-noise ratio wall behavior. Because our concern does not
lie in reducing the uncertainties of noise distribution in this
paper, we assume that the noise distribution is known to the
receiver.
The problem of spectrum sensing in CR can now be for-
mulated as a statistical hypothesis-testing problem of choosing
between the null and alternative hypotheses with the deci-
sion rule
T
S
(R)
H
1
>
<
H
0
(4)
where the test statistic T
S
(R) is a function of the observation
matrix R = [R
1
, R
2
, . . . , R
L
], and the parameter denotes the
threshold determined to satisfy the requirement PrT
S
(R) >
[H
0
on the false-alarm probability.
III. SPECTRUM-SENSING SCHEMES
A. Conventional Schemes: EEGC and ESC
Consider the lth branch output
E(R
l
) =
N/2

n=1
[R
n,l
[
2
(5)
obtained as the sum of the received signal energy. When the
EEGC scheme is used in spectrum sensing, all the branch
outputs E(R
l
)
L
l=1
are linearly combined with equal weights,
producing the test statistic [7], [9]
T
EEGC
(R) =
L

l=1
E(R
l
). (6)
On the other hand, when the ESC scheme is used in spectrum
sensing, the largest energy E
(L)
(R) is used for spectrum sens-
ing so that we have the test statistic [7], [10]
T
ESC
(R) = E
(L)
(R) (7)
where E
(k)
(R) denotes the kth-order statistic in E(R
l
)
L
l=1
.
Although the spectrum-sensing schemes with the test statis-
tics (6) and (7) can clearly provide a satisfactory performance
in Gaussian noise, they could result in a severe degradation of
performance when the noise is heavier tailed than the Gaussian
noise. This condition implies that, in impulsive environments,
the energy-based branch output (5) is not appropriate and that
the combiners EGC and SC cannot effectively suppress the
inuence of noise components of very large amplitudes.
B. Proposed Spectrum-Sensing Scheme
1) Preliminaries: When the noise is impulsive, nonlinear
schemes are reported to be fruitful in mitigating the effects
of noise components in many signal processing applications
4324 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 9, NOVEMBER 2010
Fig. 1. Block diagram of the COS scheme, where M < L.
[15], [16]. Providing reasonable performance in both Gaussian
and impulsive noise environments, order statistics of the av-
erages of observations [20], [21] and averages of the order
statistics of observations [22] are used for various applica-
tions, including signal detection and signal restoration. Based
on such observations on the performance characteristics of
nonlinear schemes in various signal processing areas, it is
easily anticipated that nonlinear strategies could also be very
useful in spectrum-sensing schemes that operate in impulsive
noise environments. In particular, when the CR exploits receive
diversity, it is conceivable that nonlinear diversity-combining
techniques, if adequately designed, can guarantee performance
stability of spectrum sensing in the CR, even in impulsive noise
environments.
In addition, as a more general and useful detector in im-
pulsive noise environments, we consider the GLRT detector,
instead of the energy detector, in the branches of the receiver.
The GLRT detector usually produces a good detection perfor-
mance with a simple structure for implementation, furnishing
the special case of the energy detector when the noise is
Gaussian.
2) COS Scheme: Fig. 1 shows a schematic of the COS
scheme. In this gure, the output of the GLRT detector in the
lth branch can be expressed as the LLR
G(R
l
) = ln
_
f
W
(R
l


H
l
S)
f
W
(R
l
)
_
(8)
where ln denotes the natural logarithm, f
W
is the common
joint pdf of W
l
, and

H
l
S = [

H
l
S
1
,

H
l
S
2
, . . . ,

H
l
S
(N/2)
] is
the vector of maximum-likelihood estimates (MLEs)

H
l
S
n
of
H
l
S
n
. An ordering operation on the L branch outputs G =
G(R
l
)
L
l=l
will produce the order statistics G
(l)
(R)
L
l=1
,
where G
(k)
(R) is the kth-order statistic in G. Then, M-order
statistics G
(i
a
)
(R)
M
a=1
are linearly combined with equal
weights to produce the test statistic
T
COS
(i
1
, i
2
, . . . , i
M
; R) =
M

a=1
G
(i
a
)
(R) (9)
for spectrum sensing, where M 1, 2, . . . , L is the number
of branches employed in the combining, and 1 i
1
< i
2
<
< i
M
L. We will use the notation COS(i
1
, i
2
, . . . , i
M
)
to denote the COS scheme with the test statistic (9). For
example, COS(2,4) denotes the proposed scheme with the
test statistic G
(2)
(R) + G
(4)
(R) and decision rule G
(2)
(R) +
G
(4)
(R)
H
1
>
<
H
0
.
Depending on the number M of branches employed in the
combining and the set i
a

M
a=1
of indices, the COS scheme can
also signify various conventional spectrum-sensing schemes.
For example, COS(1, 2, . . . , L) and COS(L) are the GLRT
detector-based EGC and SC schemes, respectively. In addition,
because the GLRT detector is the same as the energy detector
when the noise is Gaussian, COS(1, 2, . . . , L) and COS(L) are
the same as the EEGC and ESC schemes, respectively, when
the noise is Gaussian.
When the values of L and M are given, the test statistic
(9) allows
L
C
M
= L!/M!(L M)! possible schemes, the
performance characteristics of which will certainly depend
on the heaviness of the tails of the noise. For example, in
Gaussian noise environments, COS(L) (which is the same
as the ESC) would exhibit the best performance for spec-
trum sensing among COS(i)
L
i=1
, because the largest LLR
G
(L)
(R) would, with a high probability, have resulted from
signal components that have passed through a favorable (large
gain) channel. On the other hand, when the noise is heavy
tailed, COS(L) would not exhibit the best performance because
the largest LLR would, with a high probability, have resulted
from noise components rather than from signal components;
in such a case, some COS(i), with i < L, would exhibit better
performance than COS(L).
3) Examples of LLRs: Based on (8), it is clear that the LLR
is affected by the joint pdf f
W
I
,W
Q
. In most cases of complex
noise environments, it is not quite feasible to mathematically
express the joint pdf in closed form. Among the exceptions
are the symmetric -stable (SS), contaminated Gaussian
(CG), and Gaussian noises. Fortunately, the SS, CG, and
Gaussian distributions can accurately model practical noise
processes in many cases, and as a consequence, they are widely
adopted in the modeling of important classes of impulsive
noises. Let us thus briey examine the LLRs for these noise
models.
For the complex SS noise model in this paper, the bivariate
isotropic SS (BISS) pdf [11]
f
W
I
,W
Q
(u, v) =
1
(2)
2

exp
_
(x
2
+ y
2
)

2
j(xu + yv) dxdy, < u, v < (10)
is employed as the common joint pdf f
W
I
,W
Q
. In (10), the
positive dispersion parameter is related to the spread of the
BISS pdf, and the characteristic exponent (0, 2) is related
to the heaviness of the tails of the BISS pdf. Asmaller value of
indicates a higher degree of impulsiveness, whereas a value
closer to 2 indicates a more Gaussian type of behavior. For the
pdf (10), no closed-form expression is known to exist, except
for the special cases of = 1 and 2. In particular, the pdf (10)
is the bivariate Cauchy pdf
f
W
I
,W
Q
(u, v) =

2(u
2
+ v
2
+
2
)
3
2
, < u, v <
(11)
KANG et al.: SPECTRUM SENSING FOR CR UNDER IMPULSIVE NOISE: STRUCTURE AND PERFORMANCE 4325
when = 1, and the bivariate Gaussian pdf
f
W
I
,W
Q
(u, v) =
1
4
e

u
2
+v
2
4
, < u, v < (12)
when = 2. Because the MLE

H
l
S
n
of H
l
S
n
is R
n,l
, the LLR
in (8) can be expressed as
G(R
l
) =
1
2
2
N/2

n=1
[R
n,l
[
2
(13)
in Gaussian environments, with
2
= 2, and as
G(R
l
) =
N/2

n=1
ln
_
1 +
[R
n,l
[
2

2
_
(14)
in Cauchy environments.
For the complex CG noise model in this paper, we assume
the bivariate CG pdf [23]
f
W
I
,W
Q
(u, v) =
1
2
2
1
exp
_

u
2
+ v
2
2
2
1
_
+

2
2
2
exp
_

u
2
+ v
2
2
2
2
_
, < u, v < (15)
as the common joint pdf f
W
I
,W
Q
, where
2
1

2
2
in general,
and [0, 1] is called the contamination constant. Noting that
the MLE

H
l
S
n
of H
l
S
n
is R
n,l
, the LLR in (8) can be
expressed as
G(R
l
) =
N/2

n=1
ln
_
(1 )
2
2
exp
_

[R
n,l
[
2
2
2
1
_
+
2
1
exp
_

[R
n,l
[
2
2
2
2
__
(16)
where = (N/2) ln(1 )
2
2
+
2
1
.
IV. PERFORMANCE ANALYSIS IN GAUSSIAN NOISE
WITH/WITHOUT FADING
In this section, we analytically derive the detection and
false-alarm probabilities of the COS scheme in Gaussian noise
environments with/without Nakagami fading. We will consider
only the classes of COS(i) and COS(i, k), because linear
combinations of one or two order statistics are anticipated to
be sufcient for reasonable stability of performance without an
inordinate increase of complexity, as shown in such schemes
of signal processing as the median, Wilcoxon, and rank-based
lters [20][22], [24].
A. Distributions of G(R __
l
)
Ignoring the constant term 1/2, the LLR (13) is the sum of
the squares of N unit-variance Gaussian random variables with
zero mean under H
0
and positive mean under H
1
. Then, under
the assumption of i.i.d. noise components, the distribution
of G(R
l
) is the central chi-square distribution
2
(N) under
H
0
and the noncentral chi-square distribution
2
(N,
l
) under
H
1
, where
l
= ([H
l
[
2
/
2
)

N/2
n=1
[S
n
[
2
. Here, by
2
(n) and

2
(n, ), we denote the chi-square and noncentral chi-square
distributions, respectively. The parameter n is called the degree
of freedom, and is called the noncentrality parameter. In the
nonfading channel, the complex channel gains H
l

L
l=1
are all
1, and the noncentrality parameters
l
= (1/
2
)

N/2
n=1
[S
n
[
2
are a constant.
As it is well known, the pdf (x; N, ) and cumulative
distribution function (cdf) (x; N, ) of
2
(N, ) are given by
(x; N, ) =
1
2
_
x

_N
4

1
2
e

x+
2
IN
2
1
(

x), x 0
(17)
(x; N, ) =
x
_
0
1
2
_
t

_N
4

1
2
e

t+
2
IN
2
1
(

t)dt
=1 QN
2
(

x), x 0. (18)
Similarly, we have the pdf
(x; N) =
_

_
N
2
_
2
N
2
_
1
x
N
2
1
e

x
2
, x 0 (19)
and cdf
(x; N) =
_

_
N
2
_
2
N
2
_
1
x
_
0
t
N
2
1
e

t
2
dt (20)
=1

_
N
2
,
x
2
_

_
N
2
_ (21)
=1 e

x
2

N
2
1

p=0
1
p!
_
x
2
_
p

, x 0 (22)
of
2
(N). Here, I
a
(x) =

s=0
((x/2)
a+2s
/s!(a + s + 1)),
x 0 is the ath-order modied Bessel function of the
rst kind, Q
b
(

u,

v) =
_

v
(1/2)(t/u)
(b1)/2
exp((t +
u)/2)I
b1
(

tu)dt, u, v > 0 is the generalized Marcums Q


function [25], (u) =
_

0
x
u1
e
x
dx, u > 0 is the gamma
function, and (u, v) =
_

v
x
u1
e
x
dx, u, v > 0 is the upper
incomplete gamma function.
B. Performance in Nonfading Channels
Performance of COS(i). Let us denote by
P
D
(i; ) (P
FA
(i; )) the detection (false-alarm) probability
of COS(i) in nonfading channels when the threshold is .
Because P
D
(i; ) (P
FA
(i; )) is the probability that the test
statistic T
COS
(i; R) in (9) exceeds the threshold under
H
1
(H
0
), we have
P
D
(i; ) =1 F
T
COS
(i;R)
([H
1
) (23)
P
FA
(i; ) =1 F
T
COS
(i;R)
([H
0
) (24)
with F
T
COS
(i;R)
(x[H
z
) being the cdf of the test statistic
T
COS
(i; R) under H
z
for z = 0 and 1. Recollecting that the
4326 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 9, NOVEMBER 2010
test statistic T
COS
(i; R) is the ith-order statistic G
(i)
(R) of
G(R
l
)
L
l=1
, the cdf F
T
COS
(i;R)
(x[H
z
) can be expressed as
[15], [26], [27]
F
T
COS
(i;R)
(x[H
z
) =
L

p=i
L
C
p
F
p
(x[H
z
) 1 F(x[H
z
)
Lp
(25)
where F(x[H
z
) is the common cdf of the LLRs G(R
l
) under
H
z
. Then, using (18) and (25) in (23), the detection probability
can be obtained as
P
D
(i; ) = 1
L

p=i
L
C
p
_
1 QN
2
(

)
_
p

_
QN
2
(

)
_
Lp
. (26)
Similarly, using (21) and (25) in (24), we have the false-alarm
probability
P
FA
(i; ) = 1
L

p=i
L
C
p
_
1

_
N
2
,

2
_

_
N
2
_
_
p

_
N
2
,

2
_

_
N
2
_
_
Lp
. (27)
Note that the detection probability (26) [false-alarm proba-
bility (27)], with i = L, should provide the detection (false-
alarm) probability of the ESC scheme in Gaussian noise
without fading; indeed, it is straightforward to conrm that the
results are the same as the results derived in [7] and [8].
Performance of COS(i, k). Let us denote by P
D
(i,
k; ) (P
FA
(i, k; )) the detection (false-alarm) probability of
COS(i, k) in nonfading channels when the threshold is . Then,
the detection probability P
D
(i, k; ) can be obtained as
P
D
(i, k; ) =1 F
T
COS
(i,k;R)
([H
1
)
=
_ _
/
1
f
G
(i)
,G
(k)
(x, y[H
1
)dxdy, i < k (28)
where /
1
= (x, y) : x + y > , 0 < x < y denotes the in-
tegral region, F
T
COS
(i,k;R)
([H
z
) is the cdf of the test statistic
T
COS
(i, k; R) under H
z
, and
f
G
(i)
,G
(k)
(x, y[H
z
)
=
L!
(i 1)!(k i 1)!(L k)!
F
i1
(x[H
z
) F(y[H
z
) F(x[H
z
)
ki1
1 F(y[H
z
)
Lk
f(x[H
z
)f(y[H
z
)
x < y, 1 i < k L (29)
is the joint pdf [26], [27] of the two order statistics G
(i)
(R) and
G
(k)
(R) under hypothesis H
z
. In (29), f([H
z
) is the common
pdf of the LLRs G(R
l
) under H
z
. In addition, the false-alarm
probability P
FA
(i, k; ) is evaluated as
P
FA
(i, k; ) =
_ _
/
1
f
G
(i)
,G
(k)
(x, y[H
0
)dxdy, i < k.
(30)
As shown in Appendices C and E, the detection probability (28)
and false-alarm probability (30) can be expressed as
P
D
(i, k; ) =1
Lk

p=0
ki+p

q=0
i1

r=0
(1)
ki+p+q+r
(L k p)!q!r!

L!
(k i + p q)!(i r 1)!

_

2
_
0
_
QN
2
(

x)
_
q+r

_
QN
2
(

x)
_
Liq
(x; N, )dx
_
(31)
P
FA
(i, k; ) =1
Lk

p=0
ki+p

q=0
i1

r=0
(1)
ki+p+q+r
(L k p)!q!r!

L!
(k i + p q)!(i r 1)!

_

2
_
0
_

_
N
2
,
x
2
_

_
N
2
_
_
q+r

_
N
2
,
x
2
_

_
N
2
_
_
Liq
(x; N)dx
_
.
(32)
In Appendices C and E, specic results of (31) and (32)
when L = 2 are obtained. The specic results (65) and (95) are
clearly the same as the detection and false-alarm probabilities
obtained in [7] and [9] for the EEGC scheme in Gaussian
environments without fading. For convenience in the evaluation
when L 2, the detection probability P
D
(i, k; ) (false-alarm
probability P
FA
(i, k; )) is approximated (expressed) using a
nite quadruple series in (70) in Appendix C [see (88) in
Appendix E].
C. Performance in Nakagami Fading
In modeling wireless multipath fading, Nakagami fading is
frequently used. In this section, we assume block Nakagami
fading channels in which H
l

L
l=1
are i.i.d. Then, the noncen-
trality parameters can be modeled [28] as a gamma random
variable with the common pdf
f

(t) =
1
(m)
_
m

_
m
t
m1
exp
_

t
_
, t 0 (33)
where m is called the Nakagami parameter, and = E(
l
) =
(E([H
l
[
2
)/
2
)

N/2
n=1
[S
n
[
2
is the average of the noncentrality
parameters, with E() being the expectation operator. The
KANG et al.: SPECTRUM SENSING FOR CR UNDER IMPULSIVE NOISE: STRUCTURE AND PERFORMANCE 4327
Nakagami parameter m determines how severe the fading is,
with a larger value denoting a less-severe fading; for example,
m = 1 denotes Rayleigh fading, and m = denotes no fading.
Performance of COS(i). Let us denote by
P
D,m
(i; ) (P
FA,m
(i; )) the detection (false-alarm)
probability of COS(i) when the Nakagami parameter is
m and the threshold is . Then, the detection probability
P
D,m
(i; ) can be obtained by regarding the noncentrality
parameter , for each of the generalized Marcums Q functions
Q
(N/2)
in (26), as a random variable with the pdf (33) and then
taking the expectation. In particular, we have
P
D,m
(i; ) =1
L

p=i
L
C
p

_
0
QN
2
(

t,

) f

(t)dt

_
0
QN
2
(

t,

)f

(t)dt

Lp
=1
L

p=i
L
C
p
_
1

Q(; )
_
p
_

Q(; )
_
Lp
(34)
where =N, , m, and

Q(; )=e
(/2)

s=0

s
()(/
2)
s
, 0 is the averaged generalized Marcums Q function,
with

s
() dened in Appendix B. It is straightforward to
see that the false-alarm probability P
FA,m
(i; ) of COS(i) in
Nakagami fading will be the same as the false-alarm probability
P
FA
(i; ) obtained in (27).
The detection probability (34), with i = L, clearly provides
an alternative expression of the detection probability of the
ESC scheme in Gaussian noise with Nakagami fading [10].
Let us also note that, as the Nakagami parameter m approaches
innity, the detection probability (34) converges to the detection
probability (26) obtained in nonfading channels. In particular,
we have
lim
m
P
D,m
(i; ) = P
D
(i; ) (35)
because as shown in (55) in Appendix B, the averaged gener-
alized Marcums Q function

Q(; ) in (34) converges to the
generalized Marcums Q function Q
(N/2)
(

) in (26) as
m approaches innity.
Performance of COS(i, k). Let us denote by
P
D,m
(i, k; ) (P
FA,m
(i, k; )) the detection (false-alarm)
probability of COS(i, k) when the Nakagami parameter is
m and the threshold is . As we have done in the derivation
of P
D,m
(i; ), the detection probability P
D,m
(i, k; ) can
be obtained by regarding the noncentrality parameter for
each of the generalized Marcums Q functions Q
(N/2)
and the
noncentral chi-square pdf in (31) as a random variable with
the pdf (33). Then, taking the expectation, we have
P
D,m
(i, k; ) =1
Lk

p=0
ki+p

q=0
i1

r=0
(1)
ki+p+q+r
(L k p)!q!r!

L!
(k i + p q)!(i r 1)!

_

2
_
0
_

Q(x; )
_
q+r

Q( x; )
_
Liq

(x; )dx
_
(36)
where

(x; ) = e
(x/2)

s=0

s
()(x/2)
(N/2)1+s
, x
0, with
s
() dened in Appendix A. It is straightforward to
see that the false-alarm probability P
FA,m
(i, k; ) of COS(i, k)
in Nakagami fading will be the same as the false-alarm proba-
bility P
FA
(i, k; ) in (32).
Because the evaluation of the detection probability
P
D,m
(i, k; ) using (36) is not convenient, alternative
expressions of P
D,m
(i, k; ) are derived in Appendix D. In
particular, we can obtain the result (80), shown later, from (36)
when L = 2, which is equivalent to the result obtained in [7]
and [9] for the EEGC scheme in Gaussian noise with Nakagami
fading. We have also obtained the approximate expression
(84) for an easier evaluation of the detection probability
P
D,m
(i, k; ) for L 2.
Again, we can clearly see that the detection probability
P
D,m
(i, k; ) of (36) in Nakagami fading converges to the
detection probability P
D
(i, k; ) of (31) in nonfading channels
as the Nakagami parameter m approaches innity, i.e.,
lim
m
P
D,m
(i, k; ) = P
D
(i, k; ) (37)
which can readily be shown using (45) and (55), shown later.
In Appendix D, we have also shown that, as m approaches
innity, the approximate detection probability (84), shown later,
obtained in Nakagami fading converges to the approximate
detection probability (70), shown later, obtained in nonfading
channels.
V. PERFORMANCE CHARACTERISTIC IN VARIOUS
NOISE ENVIRONMENTS
In the simulations herein, we consider the nonfading chan-
nel and i.i.d. slow varying Nakagami fading channels with
Nakagami parameter m = 10, 1 (Rayleigh fading channel), and
0.5, where the complex channel gains H
l
, with E([H
l
[
2
) =
1, may change at each symbol time. In addition, we have
assumed that S
1
= S
2
= = S
(N/2)
, with S
I,n
= S
Q,n
,
for simplicity and N = 10. It is also assumed that P =

N/2
n=1
[S
n
[
2
= 5 for complex Gaussian noise with variance

2
= 1, P = 3, for BISS noise with = 1.6, P = 9, for
BISS noise with = 1, and P = 10 for complex CG noise.
The performances of the COS, EEGC, and ESC schemes in
various noise and fading environments are measured in terms of
the receiver operating characteristic (ROC) and constant false-
alarm rate (CFAR) behavior. A summary of how the results
are obtained is depicted in Table I for easy reference; in the
simulations, each point is obtained by averaging the results
from 10
6
runs. We have chosen u
2
, u
4
= 15, 15 in (84)
for the evaluation of the detection probability of COS(i, k),
4328 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 9, NOVEMBER 2010
TABLE I
SUMMARY OF HOW THE RESULTS ARE OBTAINED
because the choice did not turn out to be unreasonable for the
approximate evaluations of (44) and (54), shown later.
We assume that the number L of receive antenna branches
is 4. Although additional antennas could imply better perfor-
mance, it is reported [28] that the largest diversity gain is
obtained when the value of L is increased from 1 to 2, with
diminishing returns for larger values of L. In addition, hardware
costs and system-design constraints [3] of the CR would limit
the value of L in practice. Therefore, from the viewpoint of the
performancecomplexity tradeoff, four antennas would be an
appropriate choice [29].
The notations COS
G
, COS
C
, and COS
CG
will be used to
denote the COS scheme with the LLR obtained by numerical
simulation using (13), (14), and (16), respectively. Because the
lack of a closed-form expression for the BISS pdf (except
for = 1 and 2) prohibits the exact evaluation of the LLR,
we have obtained the LLR by numerical simulation using (14)
when = 2, 1.6, and 1 and using (13) when = 2. Note that
Cauchy detectors have frequently been used as a useful alterna-
tive under the general impulsive noise circumstances because
of their acceptable performance in various impulsive environ-
ments [11], although they are only suboptimal when the noise is
not Cauchy. We would also like to note that, because the energy
detector is known to result in a severe performance degradation
in impulsive noise environments, COS
G
is considered only in
complex Gaussian noise environments. For the CG noise
environments, we have used (16) in obtaining the LLR.
A. Performance Comparison of the COS, EEGC, and ESC
Figs. 2 and 3 show the ROCs of the COS
C
, COS
G
, EEGC,
and ESC schemes in complex Gaussian noise with various
fading conditions. Note that COS
G
(4) is the same as ESC in
these gures and that the ROCs of COS
G
(i, k) are shown only
for (i, k) = (2, 4) and (3, 4) for better visibility of gures. It
is observed that the results with a reasonably large value of
Nakagami parameter (e.g., m = 10) are almost the same as the
results in the nonfading case, as shown in Fig. 2, for example;
therefore, we will no longer consider the Nakagami fading
channel with m = 10.
As it is easily anticipated, the performance of the EEGC is
superior to the performance of the other schemes, regardless of
the fading conditions in complex Gaussian noise. This result
is not unexpected from the viewpoint of diversity combining;
when the noise is Gaussian, the EGC (averaging all available
branch outputs) is simply the best scheme in terms of reducing
the effect of Gaussian noise and fading.
Fig. 2. ROCs of the proposed (COS), EEGC, and ESC schemes in the
complex Gaussian noise (COS
G
(4) is the same as ESC). (a) No fading.
(b) Nakagami fading with m = 10.
Next, let us restrict our attention to the combining strategy
of selecting one single branch. When the fading is Rayleigh,
COS
G
(4) (ESC) is clearly observed to exhibit better detection
KANG et al.: SPECTRUM SENSING FOR CR UNDER IMPULSIVE NOISE: STRUCTURE AND PERFORMANCE 4329
Fig. 3. ROCs of the proposed (COS), EEGC, and ESC schemes in the
complex Gaussian noise and Nakagami fading with m = 1 (COS
G
(4) is the
same as ESC).
performance than the other schemes; on the other hand, when
the fading is less severe than Rayleigh fading, it is interestingly
observed that COS
G
(2) and COS
G
(3) outperform COS
G
(4).
This observation implies that the median performs better than
the other order statistics in weak-fading environments as it does
in impulsive noise environments.
It is, in addition, shown in Figs. 2(b) and 3(a) that, as the
severity of fading gets higher (m smaller), the performance
of the ESC scheme gets closer to the performance of the
EEGC scheme in complex Gaussian noise, which is consistent
with the report in [10]. Note that the performance difference
between COS
C
(i
1
, i
2
, . . . , i
M
) and COS
G
(i
1
, i
2
, . . . , i
M
) for
the same set i
a

M
a=1
is insignicant, which is an attractive
characteristic of the COS scheme because COS
C
would incur
Fig. 4. Detection probability of the proposed (COS), EEGC, and ESC
schemes in the bivariate CG noise with = 0.05,
2
1
= 1, and
2
2
= 100
when the false-alarm probability is 0.001. (a) No fading. (b) Nakagami fading
with m = 1.
only an insignicant loss of performance compared with COS
G
even in complex Gaussian noise environments.
Figs. 49 show the ROCs and CFAR behaviors of the
COS
CG
, COS
C
, EEGC, and ESC schemes in various fading
conditions when the noise is heavier tailed. In particular, we
have assumed = 1.6 in (10), = 1 in (10), and = 0.05,
with
2
2
/
2
1
= 100, in (15). As clearly shown in these gures,
the EEGC and ESC schemes exhibit a severe degradation of
performance when the noise is impulsive; they become prac-
tically useless when the impulsiveness of noise gets higher.
On the other hand, COS
C
provides a signicant performance
gain over the EEGC and ESC schemes under all of the fading
conditions over a wide range of false-alarm probabilities in
impulsive noise environments. Note that, if the COS scheme
4330 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 9, NOVEMBER 2010
Fig. 5. ROCs of the proposed (COS), EEGC, and ESC schemes in the BISS
noise with = 1.6 and Nakagami fading with m = 1.
is obtained with the exact LLR at = 1.6, it would exhibit
better performance than COS
C
in the complex SS noise of
= 1.6. In addition, the general tendency of the performance
observed in Figs. 6 and 9 resembles, and could be anticipated
from, the general tendency of performance in Figs. 5(a) and 8,
respectively.
Although we have shown the results only for some values
of parameters, including N and P, we have conrmed that
other values of parameters result in the same tendency of the
performance.
B. Performance Characteristics of the COS Scheme
Let us now discuss the inuence of noise and fading on the
performances of COS(i) and COS(i, k). Because COS
C
Fig. 6. Detection probability of the proposed (COS), EEGC, and ESC
schemes in the BISS noise with = 1.6 and Nakagami fading with m = 1
when the false-alarm probability is 0.001.
Fig. 7. ROCs of the proposed (COS), EEGC, and ESC schemes in the BISS
noise with = 1.6 and Nakagami fading with m = 0.5.
and COS
G
have almost the same tendency of performance as
shown in Fig. 3, we will consider the ROC and CFAR behaviors
of only COS
C
in examining the inuence of noise and fading on
the performance of the COS scheme without loss of generality.
As observed in the gures that depict the performance of
COS
C
(i), the order statistic of LLR that leads to better
performance than the other schemes clearly depends on the op-
erating environments, including the fading condition, noise pdf,
and false-alarm probability. For example, although COS
C
(3)
exhibits the best performance among COS
C
(i) for a wide
range of false-alarm probabilities in complex Gaussian noise
with no fading [see Fig. 2(a)], COS
C
(1) exhibits the best
performance when the impulsiveness of noise is high ( small),
KANG et al.: SPECTRUM SENSING FOR CR UNDER IMPULSIVE NOISE: STRUCTURE AND PERFORMANCE 4331
Fig. 8. ROCs of the proposed (COS), EEGC, and ESC schemes in the BISS
noise with = 1. (a) No fading. (b) Nakagami fading with m = 1.
with the fading condition xed [see Fig. 8(a)], and COS
C
(4)
exhibits the best performance when the severity of fading is
high, with the impulsiveness of noise xed [see Fig. 3(a)].
Collecting such observations from the simulation results
shown in the gures, we have the following summary.
As the impulsiveness of noise gets higher, with the
fading condition xed, the performance of COS(i) with
a smaller i becomes better than that with a larger i, as
shown in Figs. 2(a), 4(a), and 8(a). This result is easily
explainable from the viewpoint of signal detection; it is
well known that an observation with a very large mag-
nitude in impulsive environments should be considered
not as a signal plus noise but just as a noise. As a result, a
large value of the LLR in impulsive environments could,
with a high probability, imply that it has originated
Fig. 9. Detection probability of the proposed (COS), EEGC, and ESC
schemes in the BISS noise with = 1 when the false-alarm probability is
0.001. (a) No fading. (b) Nakagami fading with m = 1.
in noise components rather than in signal components.
Thus, in such an environment, selecting a smaller LLR
(smaller i) generally leads to better performance than
selecting a larger LLR.
As the severity of fading gets higher with the impulsive-
ness of noise xed, on the other hand, the performance
of COS(i) with a larger i becomes better than that with a
smaller i, as shown in Fig. 4. This result can be justied
by noting that signal components S
n
that have passed
through a favorable channel would exhibit a large power
and, consequently, a large value of the LLR.
When the severity of fading and the impulsiveness of
noise are both high, the impulsiveness of noise is more
inuential on the performance of COS(i) than the
4332 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 9, NOVEMBER 2010
TABLE II
SUMMARY OF HOW THE VALUES OF i AND k FOR COS(i) AND COS(i, k) CAN BE SELECTED TO PRODUCE AN OPTIMUM PERFORMANCE
severity of fading, as shown in Figs. 8(b) and 9(b). Note
that, although an unfavorable (small gain) channel could
lead to signicant performance degradations of COS(1)
and COS(2), the detection probabilities of COS(1) and
COS(2) are still higher than the detection probabilities
of COS(3) and COS(4) for a wide range of false-alarm
probabilities (see Fig. 8). As aforementioned, although
COS(i) with a larger i is expected to exhibit higher
detection probability than that with a smaller i when
the severity of fading is high, COS(3) and COS(4), in
fact, exhibit a lower detection probability than the others
[see Fig. 8(b)]. This condition implies that the noise
environment is more inuential on the performance of
COS(i) than the fading condition when the severity of
fading and the impulsiveness of noise are both high.
Based on the discussions on the performance of COS(i),
it is conceivable that the performance of COS(i, k) would be
dependent on i and k. In particular, in selecting two LLRs, it
is a reasonable strategy to rst select larger (smaller) LLRs in
Gaussian (impulsive) noise, as shown in Figs. 3(b) and 5(b).
Note that the performance of COS
C
(i, k) is less sensitive
(more robust) to changes in the fading conditions than the
performance of COS
C
(i). In addition, the best (worst) per-
formance among COS
C
(i, k) is superior to the best (worst)
performance among COS
C
(i). These results stem from that
the sum of two LLRs increases the stability of the test statis-
tic and consequently reduces the adverse effect of fading. In
Table II, we have summarized how we can choose the values of
i and k for the COS scheme to produce optimum performance
in various noise and fading environments.
VI. CONCLUSION
In this paper, we have proposed a class of spectrum-sensing
schemes called the COS, providing reasonable performance
and performance stability for spectrum sensing under various
fading conditions over a wide range of false-alarmprobabilities,
particularly in impulsive noise environments. The test statistic
of the COS scheme is the sum-of-order statistics of LLRs.
By considering the GLRT detector and exploiting a nonlinear
diversity-combining strategy, the COS scheme exhibits reason-
able performance for spectrum sensing with a smaller variation
of performance under various fading conditions, even in heavy-
tailed noise environments.
The performances of the COS, EEGC, and ESC schemes
have been compared in a variety of fading environments when
the noise is Gaussian and heavy tailed through computer sim-
ulations. The simulation results have conrmed that, although
the EEGC scheme performs better than the other schemes when
the noise is Gaussian, the COS scheme provides a signicant
performance gain over the EEGC and ESC schemes over a wide
range of false-alarm probabilities when the noise is heavier
tailed. Among the COS schemes that employ only a single
order statistic, COS(i) with a smaller i performs better than
that with a larger i as the impulsiveness of noise gets higher,
with the fading condition xed. As the severity of fading gets
higher, with the impulsiveness of noise xed, on the other hand,
COS(i) with a larger i performs better than that with a smaller
i. When the severity of fading and impulsiveness of noise
are both high, the impulsiveness of noise is more inuential
on the performance of COS(i) than the severity of fading.
We have also observed that the performance of COS(i, k) is
dependent on i and k and is less sensitive to changes in the
fading condition than that of COS(i). In selecting two LLRs
for the COS scheme that employs two order statistics, it is
a reasonable strategy to rst select larger LLRs in Gaussian
noise environments and smaller LLRs in impulsive noise
environments.
APPENDIX A
APPROXIMATION OF THE NONCENTRAL
CHI-SQUARE PDF (x; N, )
Employing the innite-series expression of the modied
Bessel function of the rst kind, we can express the pdf
(17) as
(x; N, ) =

s=0
x
N
2
1+s
exp
_

x+
2
_
s!2
N
2
+s

_
N
2
+ s
_
_

2
_
s
=e

x
2

s=0

s
(N, )
_
x
2
_N
2
1+s
, x 0 (38)
KANG et al.: SPECTRUM SENSING FOR CR UNDER IMPULSIVE NOISE: STRUCTURE AND PERFORMANCE 4333
where
s
(N, ) = (/2)
s
e
(/2)
/(2s!((N/2) + s)). Noting
that
s
(N, ) is negligible when s is large, we can approximate
(38) as
(x; N, ) e

x
2
u
1

s=0

s
(N, )
_
x
2
_N
2
1+s
, x 0 (39)
where u
1
is a positive integer that determines the accuracy of
the approximation.
Averaging the pdf (x; N, t) when the pdf of t is given as
(33). Let

(x; ) =

_
0
(x; N, t)f

(t)dt, x 0. (40)
Employing the innite series representation (38) of the pdf
(x; N, ), we can express

(x; ) as

(x; ) =e

x
2

s=0

_
0

s
(N, t)f

(t)dt

_
x
2
_N
2
1+s
=e

x
2

s=0
_
m

_
m _
x
2
_N
2
1+s
2
s+1
s!(m)
_
N
2
+ s
_

_
0
t
s+m1
exp
_

+ 2m
2
t
_
dt

, x 0.
(41)
Calculating the integral in (41) using the relation [30]

_
0
t
s+m1
e

+2m
2
t
dt = (s + m)
_
2
+ 2m
_
s+m
(42)
we can obtain

(x; ) = e

x
2

s=0

s
()
_
x
2
_N
2
1+s
, x 0 (43)
where
s
() = (((s + m) (m/)
m
)/(2
s+1
((N/2) +
s)(m)s!))(2/( + 2m))
s+m
. Noting that
s
() is
negligible when s is large, we can approximate the averaged
pdf

(x; ) in (43) as

(x; ) e

x
2
u
2

s=0

s
()
_
x
2
_N
2
1+s
, x 0 (44)
where u
2
is a positive integer that determines the accuracy and
computational complexity of the approximation. Note that the
averaged pdf (43) can be regarded as a generalization of the pdf
(38). In particular, we have
lim
m

(x; ) = (x; N, ) (45)


because lim
m
((s + m) (m/)
m
/(m))(2/( +
2m))
s+m
=
s
e
(/2)
.
APPENDIX B
APPROXIMATION OF THE GENERALIZED MARCUMS Q
FUNCTION Q
(N/2)
(

x)
Again, using the innite-series expression of the modied
Bessel function of the rst kind, the generalized Marcums Q
function Q
(N/2)
(

x) can be expressed with an innite


series as
QN
2
(

x) = e

s=0
_

2
_
s
s!2
N
2
+s

_
N
2
+ s
_

_
x
t
N
2
1+s
e

t
2
dt, , x 0. (46)
Then, using (22) and the relation (1/(2
(N/2)+s
((N/2) +
s)))
_

x
t
(N/2)1+s
e
(t/2)
dt = 1 (x; N + 2s) obtainable
from (20), the generalized Marcums Q function (46) can be
expressed as
QN
2
(

x) =e

s=0
1
s!
_

2
_
s
1 (x; N + 2s)
= exp
_

x +
2
_

s=0
N
2
1+s

p=0

()
s
2
s
s!p!
_
x
2
_
p
, , x 0. (47)
Expanding the double summation in (47), it is straightforward
to obtain the coefcients

1
k!

s=0
1
s!
_

2
_
s
, k = 0, 1, . . . ,
N
2
1
1
k!
_

s=0
1
s!
_

2
_
s

k
N
2

s=0
1
s!
_

2
_
s
_
, k =
N
2
,
N
2
+ 1,
(48)
of (x/2)
k
. Thus, based on the relation

s=0
(1/s!)(/2)
s
=
e
(/2)
, we can rewrite (47) as
QN
2
(

x) = e

x
2

k=0

k
(N, )
_
x
2
_
k
, , x 0
(49)
where
s
(N, ) = 1/s! for 0 s (N/2) 1, and

s
(N, )=(1/s!)1

s(N/2)
p=0
(1/p!)(/2)
p
e
(/2)
for s
(N/2). Because
s
(N, ) is negligible when s is large, we can
approximate (49) as
QN
2
(

x) e

x
2
u
3

k=0

k
(N, )
_
x
2
_
k
, , x 0
(50)
where u
3
is a positive integer that determines the accuracy of
the approximation.
Averaging Q
(N/2)
(

t,

x) when the pdf of t is given as


(33). Let

Q(x; ) =

_
0
QN
2
(

t,

x)f

(t)dt, x 0. (51)
4334 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 9, NOVEMBER 2010
Using (49) in (51), we can express

Q(x; ) as

Q(x; ) = e

x
2

s=0

_
0

s
(N, t)f

(t)dt

_
x
2
_
s
. (52)
If we evaluate the integral in (52) using (42), we can obtain

Q(x; ) = e

x
2

s=0

s
()
_
x
2
_
s
, x 0 (53)
where

s
() = 1/s! for 0 s (N/2) 1, and

s
() =
(1/s!)1

s(N/2)
p=0
((p+m) (m/)
m
/(2
p
p!(m)))(2/
( + 2m))
p+m
for s (N/2). Noting that

s
() is neg-
ligible when s is large, we can approximate the averaged
generalized Marcums Q function (53) as

Q(x; ) e

x
2
u
4

s=0

s
()
_
x
2
_
s
, x 0 (54)
where u
4
is a positive integer that determines the accuracy
and computational complexity of the approximation. Similar to
(45), it is straightforward to show that
lim
m

Q(x; ) = QN
2
(

x) (55)
using lim
m
((s+m)(m/)
m
/(m))(2/(+2m))
s+m
=

s
e
(/2)
, which implies that the generalized Marcums
Q function Q
(N/2)
(

x) is a special case of the averaged


generalized Marcums Q function

Q(x; ).
APPENDIX C
DETECTION PROBABILITY OF COS(i, k) IN
NONFADING CHANNELS
Derivation of (31) based on (28). The detection probability
(28) can be rewritten as
P
D
(i, k; ) =1

2
_
0
x
_
x
f
G
(i)
,G
(k)
(x, y[H
1
)dydx
=1

2
_
0

i1
(x; N, )(x; N, )

x
_
x
(y; N, ) (x; N, )
ki1
1 (y; N, )
Lk
(y; N, )dydx (56)
where = L!/(i 1)!(k i 1)!(L k)!. Next, employ-
ing integration by parts, the integral over y in (56) can be
expressed as
x
_
x
(y; N, ) (x; N, )
ki1
1 (y; N, )
Lk
(y; N, )dy
=
Lk

p=0
(L k)!(k i 1)!
(L k p)!(k i + p)!
( x; N, ) (x; N, )
ki+p
1 ( x; N, )
Lkp
. (57)
Then, the detection probability P
D
(i, k; ) in (56) can be
rewritten as
P
D
(i, k; )
= 1
Lk

p=0
L!
(i 1)!(L k p)!(k i + p)!

2
_
0

i1
(x; N, ) ( x; N, ) (x; N, )
ki+p
1 ( x; N, )
Lkp
(x; N, )dx.
(58)
Now, binomially expanding the term in the integral in (58),
we have

i1
(x; N, ) ( x; N, ) (x; N, )
ki+p
1 ( x; N, )
Lkp
=
ki+p

q=0
i1

r=0
(1)
ki+p+q+r
(k i + p)!(i 1)!
(k i + p q)!q!(i r)!r!
1 (x; N, )
q+r
1 ( x; N, )
Liq
.
(59)
Using (18) and (59) in (58), we nally obtain (31).
Detection probability P
D
(i, k; ) when L = 2. Because
i = 1 and k = 2 when L = 2, we have
P
D
(1, 2; ) =1 + 2

2
_
0
(x; N, )(x; N, )dx
2

2
_
0
( x; N, )(x; N, )dx
=1 +
2
_

2
; N,
_
2

2
_
0
( x; N, )(x; N, )dx (60)
based on (18) and (31). Let us rst evaluate the integral in (60).
Noting the Fourier transform T(; N, ) (; N, ) =
exp(j2/(1 + j2))/(j(1 + j2)
N
) of the convolution
integral (; N, ) (; N, ) =
_

0
( x; N, )(x; N,
)dx, we have

_
0
( x; N, )(x; N, )dx =T
1

exp
_
j2
1+j2
_
j(1 + j2)
N

=(; 2N, 2). (61)


KANG et al.: SPECTRUM SENSING FOR CR UNDER IMPULSIVE NOISE: STRUCTURE AND PERFORMANCE 4335
On the other hand, integrating the integral in (60) by parts,
we get

2
_
0
( x; N, )(x; N, )dx =
2
_

2
; N,
_
+

2
( x; N, )(x; N, )dx. (62)
If we add
_
(/2)
0
( x; N, )(x; N, )dx to both sides of
(62), we obtain
2

2
_
0
( x; N, )(x; N, )dx
=
2
_

2
; N,
_
+

2
( x; N, )(x; N, )dx
+

2
_
0
( x; N, )(x; N, )dx (63)
which can be expressed as
2

2
_
0
( x; N, )(x; N, )dx
=
2
_

2
; N,
_
+ (; 2N, 2) (64)
using (61). Finally, using (18) and (64) in (60), we have
P
D
(1, 2; ) = Q
N
(
_
2,

). (65)
Approximation of the detection probability P
D
(i, k; )
when L 2. With the approximation (50), we have
QN
2
(

x) e
x
2
u
3

s=0
e

s
(N, )
_
x
2
_
s
=e
x
2
u
3

s=0

s
(N, , , u
3
)
_
x
2
_
s
0, x (66)
where
s
(N, , , u) = (1)
s
e
(/2)

u
p=s
p
C
s

p
(N, )(/
2)
ps
. Then, using (39), (50), and (66), the integral in (31) can
be approximated as

2
_
0
_
QN
2
(

x)
_
q+r
_
QN
2
(

x)
_
Liq
(x; N, )dx

2
_
0
e

x
2
(1L+i+2q+r)

s=0

s
(q + r, L i q; N, , , u
1
, u
3
)x
N
2
1+s
_
dx
=

s=0

s
(q + r, L i q; N, , , u
1
, u
3
)

s
(1 L + i + 2q + r; N, ) (67)
where = (L i + r)u
3
+ u
1
,
s
(a, b; N, , , c, d) is the co-
efcient of x
(N/2)1+s
in the expansion of
_
d

s=0

s
(N, )
_
x
2
_
s
_
a
_
d

s=0

s
(N, , , d)
_
x
2
_
s
_
b

_
c

s=0

s
(N, )
_
x
2
_N
2
1+s
_
(68)
and

s
(a; N, ) =

2
_
0
x
N
2
1+s
e

xa
2
dx
=

1
N
2
+s
_

2
_N
2
+s
, a = 0
_
2
a
_N
2
+s

_
N
2
+ s
_

N
2
1+s

t=0
(
2
a
)
t+1
(
N
2
+s)e

a
4
(
N
2
+st)

2
_N
2
1+st
, a ,= 0.
(69)
Using (67) in (31), we nally get the approximation
P
D
(i, k; ) 1
Lk

p=0
ki+p

q=0
i1

r=0
(1)
ki+p+q+r
(L k p)!q!r!

L!
(k i + p q)!(i r 1)!

s=0

s
(q + r, L i q; N, , , u
1
, u
3
)

s
(1 L + i + 2q + r; N, ). (70)
APPENDIX D
DETECTION PROBABILITY OF COS(i, k)
IN NAKAGAMI FADING
Detection probability P
D,m
(i, k; ) in Nakagami fading
when L = 2. Let

(x; ) =

_
0
(x; N, t)f

(t)dt (71)
=1

Q(x; ), x 0 (72)
where we have used (18) and (51). Using (72) in (36), with
i = 1 and k = 2, we have
P
D,m
(1, 2; ) =1 + 2

2
_
0

(x; )

(x; )dx
2

2
_
0

( x; )

(x; )dx
4336 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 9, NOVEMBER 2010
=1+

2
_

2
;
_
2

2
_
0

(x; )

(x; )dx.
(73)
Let us evaluate the integral in (73). First, using (40), (61),
and (71), we have

_
0

( x; )

(x; )dx
=

_
0

_
0
( x; N, t
1
)f

(t
1
)dt
1

_
0
(x; N, t
2
)f

(t
2
)dt
2

dx
=

_
0

_
0

_
0
( x; N, t
1
)(x; N, t
2
)dx

(t
1
)f

(t
2
)dt
1
dt
2
=

_
0

_
0
(; 2N, t
1
+ t
2
)f

(t
1
)f

(t
2
)dt
1
dt
2
. (74)
Changing the variables as t = t
1
+ t
2
in (74) gives

_
0
(; 2N, t)

_
0
f

(t t
2
)f

(t
2
)dt
2

dt
=

_
0
(; 2N, t)f

E
(t)dt (75)
where f

E
(t) =
_

0
f

(t t
2
)f

(t
2
)dt
2
is the pdf of the sum
of two i.i.d. random variables with the common pdf f

(t).
Note that the pdf f

E
(t) can directly be obtained from (33) by
replacing m with 2m and with 2. Then, using (18) and (51),
we can obtain

_
0

( x; )

(x; )dx = 1

Q(; 2). (76)
On the other hand, employing integration by parts in the integral
in (73), we get

2
_
0

( x; )

(x)dx
=

2
_

2
;
_
+

( x; )

(x; )dx. (77)
If we add
_
(/2)
0

( x; )

(x; )dx to both sides of (77),
we obtain
2

2
_
0

( x; )

(x; )dx
=

2
_

2
;
_
+

( x; )

(x; )dx
+

2
_
0

( x; )

(x; )dx (78)
which can be expressed as
2

2
_
0

( x; )

(x; )dx =

2
_

2
;
_
+ 1

Q(; 2)
(79)
using (76). Combining (73) and (79), we nally have
P
D,m
(1, 2; ) =

Q(; 2). (80)
Approximation of the detection probability P
D,m
(i, k; )
when L 2. Using (54), the averaged generalized Marcums
Q function

Q( x; ) can be approximated as

Q( x; ) e
x
2
u
4

s=0

s
(, , u
4
)
_
x
2
_
s
, x (81)
where
s
(, , u) = (1)
s
e
(/2)

u
p=s
p
C
s

p
()(/2)
ps
.
Next, after following the steps similar to (67), the integral in
(36) can be approximated as

2
_
0
_

Q(x; )
_
q+r
_

Q( x; )
_
Liq

(x; )dx

s=0

s
(q + r, L i q; , , u
2
, u
4
)

s
(1 L + i + 2q + r; N, ) (82)
using (44), (54), and (81), where = (L i + r)u
4
+ u
2
, and

s
(a, b; , , c, d) is the coefcient of x
(N/2)1+s
in the expan-
sion of
_
d

s=0

s
()
_
x
2
_
s
_
a
_
d

s=0

s
(, , d)
_
x
2
_
s
_
b

_
c

s=0

s
()
_
x
2
_N
2
1+s
_
. (83)
KANG et al.: SPECTRUM SENSING FOR CR UNDER IMPULSIVE NOISE: STRUCTURE AND PERFORMANCE 4337
Using (82) in (36), we nally get the approximation
P
D,m
(i, k; ) 1
Lk

p=0
ki+p

q=0
i1

r=0
(1)
ki+p+q+r
(L k p)!q!r!

L!
(k i + p q)!(i r 1)!

s=0

s
(q + r, L i q; , , u
2
, u
4
)

s
(1 L + i + 2q + r; N, ). (84)
Note that
s
(a, b; , , c, d) in (82) can be regarded as a
generalization of
s
(a, b; N, , , c, d) in (67). In particular,
because
s
(),

s
(), and
s
(, , d) in (83) converge to

s
(N, ),
s
(N, ), and
s
(N, , , d), respectively, we have
lim
m

s
(a, b; , , c, d) =
s
(a, b; N, , , c, d). (85)
Consequently, (83) converges to (68). Using (85), it is straight-
forward to see that the approximate detection probability (84)
in Nakagami fading converges to the approximate detection
probability (70) in nonfading channels as m approaches innity
when u
2
= u
1
and u
4
= u
3
.
APPENDIX E
FALSE-ALARM PROBABILITY OF COS(i, k)
Following the steps similar to (31) in the derivations of
the detection probability P
D
(i, k; ), it is straightforward to
obtain the false-alarm probability (32). For convenience in the
evaluation of the false-alarm probability (32), let us evaluate
its integral. Using (19) and (22), the integral in (32) can be
expressed as

2
_
0
_

_
N
2
,
x
2
_

_
N
2
_
_
q+r
_

_
N
2
,
x
2
_

_
N
2
_
_
Liq
(x; N)dx
=

s=0

s
(q + r, L i q; N, )

s
(1 L + i + 2q + r; N, ) (86)
where
/
= ((N/2) 1)(L i + r), and
s
(a, b; N, ) is the
coefcient of x
(N/2)1+s
in the expansion of

N
2
1

p=0
_
x
2
_
p
p!

N
2
1

p=0
( x)
p
e

2
2
p
p!

_
x
2
_N
2
1
2
_
N
2
_

. (87)
Then, using (86) in (32), we get
P
FA
(i, k; ) =1
Lk

p=0
ki+p

q=0
i1

r=0
(1)
ki+p+q+r
(L k p)!q!r!

L!
(k i + p q)!(i r 1)!

s=0

s
(q + r, L i q; N, )

s
(1 L + i + 2q + r; N, ). (88)
Now, let us obtainthe false-alarmprobability(88) whenL=2.
Because i = 1 and k = 2 when L = 2, we have
P
FA
(1, 2; ) = 1 + 2
N
2
1

s=0

s
(0, 1; N, )
s
(0; N, )
2
N
2
1

s=0

s
(1, 0; N, )
s
(2; N, ) (89)
based on (88). Then, based on (69) and (87), it is straightfor-
ward to obtain

s
(0, 1; N, )
s
(0; N, )
= e

2
N
2
1

p=s
_

2
_N
2
+p
(1)
s
s!(p s)!2
N
2
+s

_
N
2
_ _
N
2
+ s
_ (90)
and

s
(1, 0; N, )
s
(2; N, )
=

_
N
2
+ s
_
s!2
N
2
+s

_
N
2
_
e

2
N
2
1+s

t=0

_
N
2
+ s
_ _

2
_N
2
1+st
s!2
N
2
+s

_
N
2
_
(
N
2
+ s t)
. (91)
By employing (90) and (91) and recollecting the relation [30]

(N/2)1
s=0
((N/2) + s)/(s!2
(N/2)+s
(N/2)) = 1/2, we can
express (89) as
P
FA
(1, 2; ) = 2e

2
N
2
1

s=0

N
2
1

p=s
(1)
s
_

2
_N
2
+p
s!(p s)!2
N
2
+s

_
N
2
_

1
_
N
2
+ s
_ +
N
2
1+s

p=0

_
N
2
+ s
_ _

2
_N
2
1+sp
s!2
N
2
+s

_
N
2
_
(
N
2
+ s p)

. (92)
Expanding the double summations in (92), it is straightforward
to obtain the coefcients

1
k!
N
2
1

s=0
(
N
2
+s)
s!2
N
2
+s
(
N
2
)
, 0 k
N
2
1
k
N
2

s=0
(1)
s
s!(k
N
2
s)!2
N
2
+s
(
N
2
)(
N
2
+s)
+
1
k!
N
2
1

s=k
N
2
+1
(
N
2
+s)
s!2
N
2
+s
(
N
2
)
,
N
2
k N 2
N
2
1

s=0
(1)
s
s!(
N
2
1s)!2
N
2
+s
(
N
2
)(
N
2
+s)
, k = N 1
=
1
2k!
, k = 0, 1, . . . , N 1 (93)
of (/2)
k
by noting
k
N
2

s=0
(1)
s
s!
_
k
N
2
s
_
!2
N
2
+s

_
N
2
_ _
N
2
+ s
_
=
1
k!
k
N
2

s=0

_
N
2
+ s
_
s!2
N
2
+s

_
N
2
_, k >
N
2
. (94)
4338 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 59, NO. 9, NOVEMBER 2010
With (93), we can rewrite (92) as
P
FA
(1, 2; ) =
(N,

2
)
(N)
(95)
based on (21). Again, when L = 2 in Gaussian noise, the false-
alarm probability (95) of COS(i, k) would be the same as the
false-alarm probability of the EEGC scheme [7], [9], i.e.,
P
FA
(EGC; ) =

_
LN
2
,

2
_

_
LN
2
_ (96)
because COS(i, k) is the same as the EEGC scheme when L =
2 in Gaussian noise. Indeed, it is clearly observed that the result
(95) is the same as (96) when L = 2.
APPENDIX F
NEW DERIVATION OF THE DETECTION PROBABILITY
OF THE EEGC SCHEME IN GAUSSIAN NOISE
WITH NAKAGAMI FADING
Let us denote by P
D
(EGC; ) (P
D,m
(EGC; )) the detection
probability of the EEGC scheme in a nonfading channel (the
Nakagami fading channel with Nakagami parameter m) with
Gaussian noise when the threshold is . Then, the detection
probability P
D
(EGC; ) is well known to be [7], [9]
P
D
(EGC; ) = QLN
2
(

S
,

) (97)
where
S
=

L
l=1

l
is the sum of noncentrality parameters

L
l=1
.
The detection probability P
D,m
(EGC; ) can be obtained
by regarding the noncentrality parameter
S
in (97) as
a random variable with the pdf f

S
(t) and then taking
the expectation. In particular, we have P
D,m
(EGC; ) =
_

0
Q
(LN/2)
(

t,

)f

S
(t)dt, where f

S
(t) is the pdf of
s
and can directly be obtained from (33) by replacing m with Lm
and with L. Following the steps similar to (53) in the deriva-
tion of

Q(x; ), we can obtain the detection probability, i.e.,
P
D,m
(EGC; ) =

Q(; L). (98)
The result (98) is an alternative expression of the detection
probability of the EEGC scheme in Gaussian noise with
Nakagami fading and is equivalent to the results in [7] and [9].
When L = 2 in Gaussian noise, the detection probability
(98) of the EEGC scheme would be the same as the detection
probability (80) of COS(i, k), because the EEGC scheme is the
same as COS(i, k) when L = 2 in Gaussian noise. As clearly
observed, the detection probability (98) of the EEGC scheme
with L = 2 is the same as the detection probability (80) of
COS(i, k).
ACKNOWLEDGMENT
The authors would like to thank the Associate Editor and
anonymous reviewers for their invaluably constructive sugges-
tions and helpful comments.
REFERENCES
[1] S.-Y. Tu, K.-C. Chen, and R. Prasad, Spectrum sensing of OFDMA
systems for cognitive radio networks, IEEE Trans. Veh. Technol., vol. 58,
no. 7, pp. 34103425, Sep. 2009.
[2] E. C. Y. Peh, Y.-C. Liang, Y. L. Guan, and Y. Zeng, Optimization of
cooperative sensing in cognitive radio networks: A sensing-throughput
tradeoff view, IEEE Trans. Veh. Technol., vol. 58, no. 9, pp. 52945299,
Nov. 2009.
[3] T. Ycek and H. Arslan, A survey of spectrum-sensing algorithms for
cognitive radio applications, Commun. Surveys Tuts., vol. 11, no. 1,
pp. 116130, First Quarter, 2009.
[4] Z. Quan, S. Cui, A. H. Sayed, and H. V. Poor, Optimal multiband joint
detection for spectrum sensing in cognitive radio networks, IEEE Trans.
Signal Process., vol. 57, no. 3, pp. 11281140, Mar. 2009.
[5] D. Datla, A. M. Wyglinski, and G. J. Minden, A spectrum surveying
framework for dynamic spectrum access networks, IEEE Trans. Veh.
Technol., vol. 58, no. 8, pp. 41584168, Oct. 2009.
[6] P. Wang, J. Fang, N. Han, and H. Li, Multiantenna-assisted spectrum
sensing for cognitive radio, IEEE Trans. Veh. Technol., vol. 59, no. 4,
pp. 17911800, May 2010.
[7] F. F. Digham, M.-S. Alouini, and M. K. Simon, On the energy detection
of unknown signals over fading channels, IEEE Trans. Commun., vol. 55,
no. 1, pp. 2124, Jan. 2007.
[8] A. Pandharipande and J.-P. M. G. Linnartz, Performance analysis of
primary user detection in a multiple-antenna cognitive radio, in Proc.
IEEE Int. Conf. Commun., Glasgow, U.K., Jun. 2007, pp. 64826486.
[9] S. P. Herath and N. Rajatheva, Analysis of equal-gain combining in
energy detection for cognitive radio over Nakagami channels, in Proc.
IEEE Global Telecommun. Conf., New Orleans, LA, Nov. 2008, pp. 15.
[10] S. P. Herath, N. Rajatheva, and C. Tellambura, On the energy detection
of unknown deterministic signal over Nakagami channels with selection
combining, in Proc. IEEE Can. Conf. Elect. Comput. Eng., St. Johns,
NF, Canada, May 2009, pp. 745749.
[11] C. L. Nikias and M. Shao, Signal Processing With Alpha-Stable Distribu-
tions and Applications. Hoboken, NJ: Wiley, 1995.
[12] I. Song, J. Koo, H. Kwon, S. R. Park, S. R. Lee, and B.-H. Chung,
A novel detection criterion for weak M-ary signals and its application
to ultrawideband multiple-access systems, IEEE Trans. Veh. Technol.,
vol. 54, no. 6, pp. 20242036, Nov. 2005.
[13] I. J. Kim, S. R. Park, I. Song, J. Lee, H. Kwon, and S. Yoon, Detection
schemes for weak signals in rst-order moving average of impulsive
noise, IEEE Trans. Veh. Technol., vol. 56, no. 1, pp. 126133, Jan. 2007.
[14] P. Toro and M. G. Snchez, Astudy of the correlation between horizontal
and vertical polarizations of impulsive noise in UHF, IEEE Trans. Veh.
Technol., vol. 56, no. 5, pp. 28442849, Sep. 2007.
[15] I. Song, J. Bae, and S. Y. Kim, Advanced Theory of Signal Detection.
New York: Springer-Verlag, 2002.
[16] B. S. Kim, J. Bae, I. Song, S. Y. Kim, and H. Kwon, A comparative
analysis of optimum and suboptimum rake receivers in impulsive UWB
environment, IEEE Trans. Veh. Technol., vol. 55, no. 6, pp. 17971804,
Nov. 2006.
[17] Z. Quan, S. Cui, and A. H. Sayed, Optimal linear cooperation for spec-
trum sensing in cognitive radio networks, IEEE J. Sel. Topics Signal
Process., vol. 2, no. 1, pp. 2840, Feb. 2008.
[18] B. Vucetic and J. Yuan, SpaceTime Coding. Hoboken, NJ: Wiley, 2003.
[19] R. Tandra and A. Sahai, SNR walls for signal detection, IEEE J. Sel.
Topics Signal Process., vol. 2, no. 1, pp. 417, Feb. 2008.
[20] R. J. Crinon, The Wilcoxon lter: A robust ltering scheme, in Proc.
IEEE Int. Conf. Acoust., Speech, Signal Process., Tampa, FL, Apr. 1985,
pp. 668671.
[21] P. P. Gandhi, I. Song, and S. A. Kassam, Nonlinear smoothing lters
based on rank estimates of location, IEEE Trans. Acoust., Speech, Signal
Process., vol. 37, no. 9, pp. 13591379, Sep. 1989.
[22] A. C. Bovik, T. S. Huang, and D. C. Munson, A generalization of me-
dian ltering using linear combinations of order statistics, IEEE Trans.
Acoust., Speech, Signal Process., vol. ASSP-31, no. 6, pp. 13421350,
Dec. 1983.
[23] G. L. Shevlyakov and N. O. Vilchevski, Minimax variance estimation
of a correlation coefcient for -contaminated bivariate normal distribu-
tions, Statist. Probab. Lett., vol. 57, no. 1, pp. 91100, Mar. 2002.
[24] I. Pitas and A. N. Venetsanopoulos, Order statistics in digital image
processing, Proc. IEEE, vol. 80, no. 12, pp. 18931921, Dec. 1992.
[25] J. I. Marcum, A statistical theory of target detection by pulsed radar,
IRE Trans. Inf. Theory, vol. 6, no. 2, pp. 59267, Apr. 1960.
[26] H. A. David and H. N. Nagaraja, Order Statistics, 3rd ed. Hoboken, NJ:
Wiley, 2003.
KANG et al.: SPECTRUM SENSING FOR CR UNDER IMPULSIVE NOISE: STRUCTURE AND PERFORMANCE 4339
[27] I. Song, K. S. Kim, S. R. Park, and C. H. Park, Principles of Random
Processes. Seoul, Korea: Kyobo, 2009.
[28] G. L. Stber, Principles of Mobile Communication, 2nd ed. Norwell,
MA: Kluwer, 2000.
[29] S. N. Diggavi, N. Al-Dhahir, A. Stamoulis, and A. R. Calderbank, Great
expectations: The value of spatial diversity in wireless networks, Proc.
IEEE, vol. 92, no. 2, pp. 219270, Feb. 2004.
[30] I. S. Gradshteyn, I. M. Ryzhik, A. Jeffrey, and D. Zwillinger, Table of
Integrals, Series, and Products, 6th ed. New York: Academic, 2000.
Hyun Gu Kang (S06) received the B.S. degree in
electronics engineering in 2004 from Korea Univer-
sity, Seoul, Korea, and the M.S.E. degree in electrical
engineering in 2006 from Korea Advanced Insti-
tute of Science and Technology (KAIST), Daejeon,
Korea, where he is currently working toward the
Ph.D. degree.
Since August 2004, he has been a Teaching and
Research Assistant with the Department of Elec-
trical Engineering, KAIST. His research interests
include mobile communications, multi-inputmulti-
output systems, detection and estimation theory, and statistical signal
processing.
Iickho Song (S80M87SM96F09) received
the B.S.E. (magna cum laude) and M.S.E. degrees
in electronics engineering from Seoul National Uni-
versity, Seoul, Korea, in 1982 and 1984, respec-
tively, and the M.S.E. and Ph.D. degrees in electrical
engineering from the University of Pennsylvania,
Philadelphia, in 1985 and 1987, respectively.
In 1988, he joined the Department of Electrical
Engineering, Korea Advanced Institute of Science
and Technology (KAIST), Daejeon, where he is cur-
rently a Professor. He served as an Editor of the Jour-
nal of the Acoustical Society of Korea (ASK) and the Journal of the Institute
of Electronics Engineers of Korea (IEEK). Since 1995, he has been an Editor
of the Journal of Korea Information and Communications Society (KICS).
Since 1998, he has been an Editor of the Journal of Communications and
Networks. He is a coauthor of Advanced Theory of Signal Detection (Springer,
2002), Random Processes (in Korean; Saengneung, 2004), Signals and Systems
(Hongreung, 2008; Springer, 2009), and Principles of Random Processes (in
Korean; Kyobo, 2009). His research interests include detection and estimation
theory, statistical communication theory, and mobile communication.
Prof. Song was the Treasurer of the IEEE Korea Section in 1989. He is
a member of the Korean Academy of Science and Technology, ASK, IEEK,
KICS, the Korea Institute of Information, Electronics, and Communication
Technology, and the Institute of Electronics, Information, and Communication
Engineers. He is a Fellow of the Institution of Engineering and Technology.
Seokho Yoon (S99M02SM07) received the
B.S.E. (summa cum laude), M.S.E., and Ph.D. de-
grees in electrical engineering from Korea Advanced
Institute of Science and Technology, Daejeon, Korea,
in 1997, 1999, and 2002, respectively.
From March 2002 to June 2002, he was with the
Department of Electrical Engineering and Computer
Sciences, Massachusetts Institute of Technology,
Cambridge, and from July 2002 to February 2003, he
was with the Department of Electrical Engineering,
Harvard University, Cambridge, as a Postdoctoral
Research Fellow. In March 2003, he joined the School of Information and
Communication Engineering, Sungkyunkwan University, Suwon, Korea, where
he is currently an Associate Professor. His research interests include spread-
spectrum systems, mobile communications, detection and estimation theory,
and statistical signal processing.
Dr. Yoon is a member of the Institute of Electronics Engineers of Korea
and the Korea Information and Communications Society (KICS). He was the
recipient of a Bronze Prize at the Samsung Humantech Paper Contest in 2000
and the LG Academic Award from KICS in 2009.
Yun Hee Kim (S97M00SM05) received the
B.S.E. (summa cum laude), M.S.E., and Ph.D.
degrees in electrical engineering from the Korea
Advanced Institute of Science and Technology,
Daejeon, Korea, in 1995, 1997, and 2000,
respectively.
In 2000, she was with the Department of Electrical
and Computer Engineering, University of California
at San Diego, La Jolla, as a Visiting Researcher.
From 2001 to 2004, she was a Senior Member of
Research Staff with the Electronics and Telecom-
munications Research Institute, where she participated in the development of
the wideband code-division multiple access systems and the air interface for
fourth-generation mobile radio access. In 2004, she joined the Department of
Electronics and Radio Engineering, Kyung Hee University, Seoul, where she
is currently an Associate Professor. Since 2006, she has been an Editor of
the Journal of Korea Information and Communications Society (KICS). Her
research interests include communication theory, statistical signal processing,
and estimation and detection for wireless communications.
Prof. Kim is a member of KICS and the Institute of Electronics, Information,
and Communication Engineers. She has also served as a member of the
Technical ProgramCommittee of several IEEE conferences, including the IEEE
Wireless Communications and Networking Conference and the IEEE Vehicular
Technology Conference.

You might also like