Download to read offline
![JOURNAL OF ELECTROMAGNETIC ENGINEERING AND SCIENCE, VOL. 12, NO. 3, 203～209, SEP. 2012
http://dx.doi.org/10.5515/JKIEES.2012.12.3.203
ISSN 2234-8395 (Online)․ISSN 2234-8409 (Print)
203
Performance of Spiked Population Models for Spectrum Sensing
Tan-Thanh Le․Hyung Yun Kong
Abstract
In order to improve sensing performance when the noise variance is not known, this paper considers a so-called
blind spectrum sensing technique that is based on eigenvalue models. In this paper, we employed the spiked population
models in order to identify the miss detection probability. At first, we try to estimate the unknown noise variance
based on the blind measurements at a secondary location. We then investigate the performance of detection, in terms
of both theoretical and empirical aspects, after applying this estimated noise variance result. In addition, we study the
effects of the number of SUs and the number of samples on the spectrum sensing performance.
Key words: Random Matrix Theory, Cognitive Radio Networks, Spiked Population Model, Spectrum Sensing Te-
chniques, Eigenvalue-Based Spectrum Sensing, Noise Estimation.
Manuscript received December 5, 2011 ; Revised February 16, 2012 ; Accepted February 27, 2012. (ID No. 20111205-039J)
Dept. of Electrical. Engineering, University of Ulsan, Ulsan, Korea.
Corresponding Author : Hyung Yun Kong (e-mail : hkong@mail.ulsan.ac.kr)
This is an Open-Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/
by-nc/3.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
Ⅰ. Introduction
Much recent work has focused on eigenvalue-based
spectrum sensing methods for cognitive radio networks
(CRNs) [1], [2], where researchers have applied inno-
vations from random matrix theory to calculate the pro-
bability of a false alarm as a function of a threshold.
This work has also employed many kinds of test sta-
tistics, such as the ratio of the largest and the smallest
eigenvalues or the ratio of the average and the smallest
eigenvalues of the sample covariance matrix. Yonghong
and Ying-chang’s model in [2] requires a number of
samples, N, and a number of secondary users (SUs), K,
approaching infinity. A significant gap exists between
the simulations and analytical results.
Kortun et al. [1] improved the model by finding an
exact threshold and an approximate closed-form per-
formance that agreed well with the empirical results.
Independently, Penna et al. [7] derived work similar to
[1] and then further analyzed the probability of a miss
in [10]. However, the method used in [10] still requires
knowledge of noise level and the parameters of primary
user (PU) signals and channels, which is impractical in
CRN scenarios. To our knowledge, not much research
has been carried out regarding the relationship between
the threshold and the probability of a miss when using
blind spectrum sensing techniques.
This paper considers a novel blind spectrum sensing
technique for CRNs, where we analyze a threshold with
a constraint on the probability of a miss. CRNs are com-
posed of a specific bandwidth, including multiple PUs
and the number of SUs. Based on a variety of research
related to the distribution of eigenvalues [3], [4], we ap-
ply the results of random matrix theory to the spectrum
sensing model. In fact, the practical model is related to
a finite number of samples and SUs, while these para-
meters are infinite in random matrix theory. Therefore,
we exploit the limit distribution for approximate results
in our applied spectrum sensing model. Furthermore, the
observations referred to in this spectrum sensing model
undergo interference with by complex Gaussian noise and
are affected by fading channels. Performance still de-
pends on the channel parameters, noise powers, and the
PU signal powers. However, these parameters are un-
known in practice. Therefore, noise estimation schemes
must be combined with analytical approaches. This work
also focuses on improving the performance of the noise
estimation scheme in order to provide an efficient sol-
ution to the problem with unknown information.
The rest of this paper is organized as follows. Section
2 presents the sensing model. The analysis of the proba-
bility of a miss as a function of threshold is considered
in Section 3. Section 4 demonstrates simulation results
and discussions. Finally, concluding remarks are given
ⓒ Copyright The Korean Institute of Electromagnetic Engineering and Science. All Rights Reserved.](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/85/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-1-320.jpg)
![JOURNAL OF ELECTROMAGNETIC ENGINEERING AND SCIENCE, VOL. 12, NO. 3, SEP. 2012
204
in Section 5.
Ⅱ. System Model
The system model can be presented with some assum-
ptions. At the PU side, there are P sub-bands of interest
transmitted by P entities. Cognitive radio networks have
K SUs, which can share their information through high
speed transmissions. The high speed requirement for ex-
change of data is the problem of interest in the model
design. Fortunately, these cognitive radio nodes can be
linked by wired networks, which help to significantly
improve the speed.
The observable signals at the SUs can be expressed
as
Y HX V,= + (1)
where the K×N matrix Y is the measurement at the SUs.
The K×P matrix H represents the channels between the
PUs and SUs. The P×N matrix X denotes the trans-
mitted signals from the PUs. In fact, the transmitted sig-
nals can be expressed as X=[x(n)]T, n=1, …, N, where
each x(n) is the P´1 vector whose elements are obtained
from samples of PU signals. Therefore, the covarian-
ce matrix of these independent PUs can be written as
{ } 2 2
1E xx C diag( ,..., )H
x Ps s= = , where 2
,is [ ]1,i PÎ is
the variance of the i-th PU signal. The K×N matrix V
is the complex Gaussian noise with zero mean and var-
iances .
The hypotheses can be expressed as
00 : Y | V,=HH (2)
11 :Y| HX V,= +HH (3)
Ⅲ. Performance Analysis
3-1 Probability of Miss Analysis
In the case of 1H , the measurements at the SUs are
rewritten as
(4)
where the covariance matrix of the PU signals are Cx
{ } ( )H 2 2
1 Pxx diag σ , ,σ ,Cx E = ¼= 1/21
J HC I
σ
x K
é ù
= ê ú
ë û
,
1/2
σC X
R
V
x
-
é ù
= ê ú
ë û
.
The test statistic can be found by calculating the ei-
genvalues 2
1
JJ HC H I
σ
H
x
H
K= + as follows:
( )
( )
( )
( )( )
( )
H
2
2 ( )
~ ~
2 ( )
K
det JJ I 0,
1
det HC H I I 0,
det HC H 1 I 0,
det(HC H I ) 0,? ,
J
K
JH
x K K
H J
x K
H J
x i
t
t
t
t t t
s
s
s
- =
æ ö
Û + - =ç ÷
è ø
Û - - =
Û - = = -
(5)
Using the generalized matrix determinant lemma, we
have
det detdet
⇔det
det
⇔
≤ ≤
det
≤ ≤
(6)
Hence,
~
( )
2
1.iJ
it
t
s
= +
(7)
It is easily observed from Eqs. (5) and (6) that the
power of each PU and the channel information, repre-
sented by Cx and H, respectively, must be known. We
first consider the case of known channel state infor-
mation and unknown PU signal powers. The noise vari-
ances must be estimated, and then we can separate into
two groups of noises or signals with variances equal to
or larger than the estimated one. This technique will be
presented and analyzed in Subsection 3.1. Furthermore,
SUs do not know channel state information in practice;
hence, we approximate ( )J
it for a fully blind spectrum sen-
sing model in the Simulation section.
3-2 Methods for Noise Estimation
In Subsection 3.2, the final expression of miss proba-
bility requires knowledge of the sampled covariance ma-
trix of PU signals, the channel matrix, and the noise
level. In order to obtain these values, we must estimate
the number of PUs and the noise power. Based on pre-
vious research [8], we present and analyze the perfor-
mance of noise estimation, as follows.
The distribution of the largest eigenvalue 1l is
( )2
2
1
,
( , )
lim Pr
/
,
( , )
W
N p
N p
x F x
N p
l s m
s®¥
ì ü-
< =í ý
î þ (8)
where](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/85/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-2-320.jpg)
![LE and KONG : PERFORMANCE OF SPIKED POPULATION MODELS FOR SPECTRUM SENSING
205
21
( , ) ( 1/ 2 1/ 2) ,N p p
N
Nm = - + -
1/3
1 1 1
( , ) ( 1/ 2 1/ 2)
1/ 2 1/ 2
,N p N p
N N p
s
æ ö
= - + - +ç ÷ç ÷- -è ø
and ( )2WF x is the Tracy-Widom distribution with order
2 ( 2W ) [9].
In order to determine the total number of active Pus,
^
P , we apply a series of tests. In other words, we will
estimate the noise level of the communication environ-
ment at the p-th eigenvalue, and then test the likelihood
of this p-th eigenvalue, pl , to determine whether it has
arisen from PU signals or noises. In each test, there are
two hypotheses, 0H and 1H , where the former represents
at least p eigenvalues satisfying (9). The latter repre-
sents, at most, (p—1) eigenvalues not satisfying the con-
dition in (9), due to the fact that the p-th eigenvalue is
used as the edges of the two hypotheses, the eigenvalues
il , [ ]1,i p KÎ + are assumed to be arisen from noises.
0H has at least p eigenvalues satisfying (9).
0H has at most (p-1) eigenvalues not satisfying (9).
The condition used for testing is
( ) ( ) ( ) ( )( )2
, , ,p PN p N K p x b N K pl s m s> - + - (9)
where ( )2Wb F x= is the confidence level, and ( ) 2
1
Wx b F b-
=
( )2
1
Wx b F b-
is the value determined from inverting the second-
order Tracy Widom distribution. In our work, b is va-
ried from 0.1 % to 0.5 %.
The operation can be implemented as described be-
low. If the measurements satisfy hypothesis 0H , we in-
crease p and perform the test again. Otherwise, when
the hypothesis 1H is satisfied, we stop testing and record
the final result
^
1P p= - . Based on the above operation,
^
P can be defined as the solution to the following opti-
mization problem
( ) ( ) ( )( )
2
( )
arg min -1.
, - , -
p P N
p
p
P
N K p x b N K p
l s
m s
ì ü´ï ï
= í ý
£
+ï ïî þ
$
(10)
For Eq. (10), the noise level must be known. There-
fore, we next present the method to estimate 2
P Ns .
3-2-1 A Simple Method for Noise Estimation
Equation (1) can be rewritten as
1
y h x ,
P
j j
j
sx
=
= +å
(11)
The covariance matrix can be modified to a diagonal
matrix as
(12)
where U is now an unknown K×K matrix whose col-
umns are the eigenvectors of Cy , ..
1
11
2
λ
λ
.A
é ù
ê
´
ú
= ê
ê
ë û´
ú
ú
, and
( )22A 0 K P= - . It is clear that, with a very large num-
ber of samples ( N ® ¥ ), the first P eigenvalues of the
covariance matrix are 2
il s+ , i=1, …, P, and the re-
maining (K—P) eigenvalues are 2
s . The eigenvalues are
separated into two groups, the group of signals with es-
timated eigenvalues larger than ^
2
s and the group of
noises with the remaining eigenvalues. Unfortunately,
the noise eigenvalues are widely spread in the case of
a finite number of samples. In this setting, we define the
covariance matrix for a finite number of samples and
observe that
1
C y y .
N
H
yN j j
j=
= å
(13)
In practice, the sampled covariance matrix can be
written as
1τ 0
U C U
τ 0
T
yN
K
´ ´æ ö æ ö
ç ÷ ç ÷
= +ç ÷ ç ÷
ç ÷ ç ÷´ ´è ø è ø
O O
(14)
where
. It is easily seen that the diagonal
elements from (P+1) to K are due to noise. By avera-
ging these elements, we derive the noise variance as
2
1 1 1
1 1
1 1
1
( )
K K P
i i i
i P i i
K P
i i i
i P i
K P K P
K P
s t l t
l l t
= + = =
= + =
æ ö
= = -ç ÷
- - è ø
æ ö
= + -ç ÷
- è ø
å å å
å å
(15)
However, the matrix U is unknown. Hence, we must
estimate it . In the simple form, we set i it l= and get
^
2
simple
1
1
.
K
i
i PK P
s l
= +
æ ö
= ç ÷
- è ø
å
(16)
3-2-2 Improvement in Noise Estimation
In order to improve the estimated noise variance, we
now diagonalize the upper left submatrix 1 N 1U C UT
y , as](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/85/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-3-320.jpg)
![,
JOURNAL OF ELECTROMAGNETIC ENGINEERING AND SCIENCE, VOL. 12, NO. 3, SEP. 2012
206
in [8]
11 12
1 1
21 22
B B
U C U ,
B B
T
yN
æ ö
= ç ÷
è ø (17)
where
1
0
0
P
f
f
æ ö
ç ÷
ç ÷
ç ÷
è ø
O , 12 21B BT
Z= = and
1
22
0
0
B
P
K
t
t
+æ ö
ç ÷
= ç ÷
ç ÷
è ø
O .
The eigenvalues of 1 N 1U C UT
y are
≈
∈
(18)
where { }2 2 21
ij ij jz E z
N
f s» = . Therefore, Eq. (18) can be
derived as
2 2
2 2
1
1
,
K
j j
j j j
i P j j
K P
N N
f f
l f s f s
f s f s= +
-
» + = +
- -
å
(19)
where [ ]1,j PÎ . To this end, we iterate Eqs. (15) and
(19) to estimate the noise variance
^
2
1 1
^ ^ ^ ^
2 2 2
i
1
( )
,
λ 1 0
K P
iim i i
i P i
i i im i im
K P
K P
N
s l l f
f f s l s
= + =
ì æ ö
= + -ï ç ÷
- è øï
í
æ - öæ öï - + - + =ç ÷ç ÷ï è øè øî
å å
$
(20)
with the initial value
^ ^
2 2
0 simple / 1
P
N
s s
æ ö
= -ç ÷
è ø
for better per-
formance [5].
After estimating
^
2
ims , which is used for missed prob-
ability analysis with known channel state information
and PU signal powers, we also use ∈ to calcu-
late the eigenvalue ( )J
it as in Section 4 for the case of
blind spectrum sensing models.
3-3 The Probability of Miss
Based on the estimated result of noise variance,
^
2
ims ,
we perform the calculation of Pm. Note that the follow-
ing expressions use 2
s instead of
^
2
ims
.
From [6], we
have the following theorem with
( )( ) 2 ( )
1 1 ( )
1
1 ,
1
J J
J
c
t t
t
m s
æ ö
= +ç ÷
-è ø (21)
( )
( )
( ) 2 ( )
1 1 2( )
1
1 ,
1
J J
J
c
t t
t
s s= -
-
(22)
where c=K/N.
3-3-1 Theorem 1
Ref. [3] Consider the sequence of a complex Wishart
matrix with Cx. Let 1 m P£ £ , we have the case of iden-
tifiable signals; i.e., the spiked eigenvalues are above the
critical limit 1/2
1 c+ . If
…
. Let
( )
( )( )
11 1( )
1
,J
J
N
t
t
e l m
s
æ ö
= -ç ÷
è ø
$
(23)
then
( ) ( )1
,
lim Pr ,m
N P
e g g
¥
£ = G
™ (24)
where ( )m gG is the finite Gaussian unitary ensemble
(GUE) distribution given as
×
∞
⋯ ∞
≤ ≺ ≤
exp
⋯
(25)
Based on Eq. (24), the distribution of the largest ei-
genvalue is lim
→∞
.
In order to determine the distribution of the smallest
eigenvalues, which are the remaining (K-P) eigenvalues,
we adhere to Theorem 2 [4]. Let
~
,
K P
c
N
-
=
(26)
~ ~
1/2
2
2
1 ,c cm sæ ö æ ö
= -ç ÷ ç ÷
è ø è ø (27)
~ ~ ~
1/2
1/
1/2
3
2
1 1 ,c c cs s -æ ö æ öæ ö
= - -ç ÷ ç ÷ç ÷
è ø è øè ø (28)
3-3-2 Theorem 2
Consider the sequence of a complex Wishart matrix
~
Cx with size (K − P)×(K − P) (Note that
~
Cx is the
lower right submatrix of Cx ). Kl is the smallest ei-
genvalue of
~
Cx . The variable
2/3 ^ ~
~
KK
N
c
c
e l m
s
æ öæ ö= - ç ÷ç ÷
æ ö è øè ø
ç ÷
è ø
will be
satisfied
( )2,( )
lim Pr( ) .K W
N K P
Fe g g
- ¥
£ =
™ (29)
Based on Eq. (29), the distribution of the smallest ei-
genvalue is 2,
lim ( ) ( )K W
N P
f x F xe
®¥
= .
Now we can establish the ratio test as follows,](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/85/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-4-320.jpg)
![LE and KONG : PERFORMANCE OF SPIKED POPULATION MODELS FOR SPECTRUM SENSING
207
( )1/2 ( )
1 1 1
2
^
/
1
^ ~
3
~
( ) ( )
.
JJ
K K
N t t
T
c cN
sl
l
e m
e s m
-
-
+
= =
æ ö æ ö
+ç ÷ ç ÷
è ø è ø (30)
Following [7], the joint distribution of the largest and
smallest eigenvalues is
( ) ( ) ( )1 1|
0
,KT Hf t xf tx f x dxl l
¥
= ò
(31)
where the distributions of the largest and smallest ei-
genvalues are derived from ( )1
f xe and ( )K
f xe , respecti-
vely, and are given as
( ) ( )
( )
( )( )1
1/21/2
( )
( )
,
( ) m
J
GJ J
N N
f x f x t
t t
l m
s s
æ ö
ç ÷= -
ç ÷
è ø (32)
( ) 2
2
2/3 3 ~
~ ~
|
.
|
K w
N N
c
c
f x f x
c
l m
s s
æ ö
ç ÷
æ öæ öç ÷= - ç ÷ç ÷ç ÷æ ö è øæ ö è ø
ç ÷ ç ÷ç ÷
è ø è øè ø (33)
Therefore, the probability of a miss can be calculated
as
( ) 11 |Pr | ( )m T HP T H Fg g= < = (34)
and the threshold is the inversion of Eq. (34)
( ) ( )1
-1
| .m T H mP F Pg = (35)
It is easily observed that the threshold derived from
Eq. (35) depends on the channel matrix and the PU sig-
nal powers. This is due to the fact that 1| ( )T HF g , which
includes H and Cx. Information on the PU signal powers
is included in Subsection 3.1, and the model considered
here is a CSI case. In the next section, we approximate
( )
1
J
t in order to obtain the results in the case of the fully
blind spectrum sensing model.
Ⅳ. Simulation Results and Discussions
After noise estimation, the results are the noise level
2
PNs and the number of PU signals, Pest, with the proba-
bility of the correct number of PU signals Pr(Pest=P).
The covariance matrix of the measurements at K of the
SUs can then be diagonalized as in Eq. (14) and the
first P eigenvalues [ ]
^
, 1,i i Pl Î represented as P PU sig-
nal powers are approximated as Eq. (19). Substituting
these eigenvalues to Eq. (7) at i=1, we have
≈
(36)
Proof: When N is very large, the approximation is as
follows:
HC H ,H
yN x vC C» + (37)
where Cv is the sampled covariance matrix of noise.
From [11], Weyl’s inequality theorem reads as
( )
~ ^ ~
max 1 1 maxC (C ),p v vt tl l l+ £ £ + (38)
where ( )max Cvl is the estimated noise eigenvalue,
2
P Ns ,
which is calculated as in Subsection 3.1,
~ ~ ~
1 2, ,..., Pt t t are P
eigenvalues of HC HH
x , and
^
1l is the largest eigenvalue
of CyN . Moreover, at very large N, the simulation shows
that the following approximation performs in a manner
that is in good agreement with the empirical results
( )
^
2
1 max
~ ~
1 1λ Cv P Nt tl s» + = + (39)
Substituting Eq. (39) into Eq. (36), we have the final
result.
We now perform spectrum sensing without using any
information from the channels or the noise level, in this
case. We note that noise estimation relates to the correct
number of PUs Pr(Pest=P), so the final probability of a
miss is ( )
^
estPr .mP Pm P P= ´ =
Now, we analyze the effects of parameters such as K,
N, and P on the noise estimation performance and there-
fore the spectrum sensing performance. In Fig. 1, the
miss probabilities for the unknown model, the ideal
model (known model), and the empirical simulation are
calculated using various numbers of samples and SUs.
As illustrated in this figure, the gap between the ideal
and the unknown schemes is slightly too small. In par-
ticular, when the number of SUs and samples are in-
creased ((N, K)=(2,000, 40), (N, K)=(2,000, 50) and (N,
K)=(800, 40)), the results of both cases are the same. To
confirm a good match between the theoretical analysis
and the empirical experiment, the model uses only N=
2,000 or N=800 samples, which are appropriate for the
practical model.
In order to completely understand the estimation sche-
me, Fig. 2 shows the performance when estimating PU
signal powers. Even though a pair of samples and the
number of SUs (N, K) are chosen to be (1,000, 50) (i.e.,
a not very large N), the estimated power densities agree
well with the original power densities. A larger number
of samples gives a more exact result for the estimate of
the powers. Hence, if we want to obtain a more precise
estimate in order to reduce the difference between the
theoretical and empirical results, we slightly increase the
number of samples N.](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/85/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-5-320.jpg)
![JOURNAL OF ELECTROMAGNETIC ENGINEERING AND SCIENCE, VOL. 12, NO. 3, SEP. 2012
208
Fig. 1. Probability of miss vs. the threshold in the case of
P=4 and variable K and N.
Fig. 2. Density of the ratio of signal eigenvalues to noise ei-
genvalues in the case of P=3, K =50 and N=1,000.
Fig. 3 shows the receiver operating characteristic
(ROC), which describes the sensitivity of the spectrum
sensing model. We also note that, in the case of H0, the
probability of a false alarm, faP , is calculated based on
the results in [7], which include the analysis for faP as
a function of a decision threshold.
Fig. 4 shows the convergence of our proposed scheme
at P=4 with the constraint c=0.1. It is clearly observed
that the increases in both K and N confirm the con-
vergence of the curve represented by the relation be-
tween Pm and the threshold. In the case of N=80 and
K=8, the estimated scheme no longer agrees with the
known scheme due to the fact that noise estimation per-
formance is heavily degraded when N is too low.
Moreover, after the noise estimation stage, the perfor-
mance of the spectrum sensing technique is also affected
by very low K and N. Increases in both N and K with
the constraint N/K=c ensure that the performance of the
analytical unknown case matches well with that of the
empirical known case.
Finally, we did not refer to the energy detection tech-
nique (ED) in our work; our focus was on comparison
of our proposed method with state-of-the-art eigenvalue-
based approaches. Several other studies [1], [2], [7],
Fig. 3. Probability of miss vs. probability of a false alarm
in the case of P=1, K=50 and N =1,000.
Fig. 4. Miss probability vs. threshold in the case of P=4 and
variable K and N with the constraint c=K/N=0.1.
[10] have confirmed that the eigenvalue-based methods
outperform the ED.
Ⅴ. Conclusion
Our work considered a new sensing model for cogni-
tive radio networks, where multiple SUs blindly sense
the band of interest, including multiple PUs. A novel
analysis is presented to determine the threshold based on
the probability of a miss, which is a new method that
utilizes the best knowledge from spectrum sensing. Mo-
reover, the proposed hybrid analytical simulation method
enhances the blind features of spectrum sensing models
since the unknown noise level, and hence the parameters
of PU signals and channels, are estimated with small
errors. We also show the effects of the estimation sche-
me and attempt to improve its performance. As a con-
sequence, the empirical simulation results illustrate small
differences between the performances of the ideal (kno-
wn information) and estimate (blind) models.
References
[1] A. Kortun, T. Ratnarajah, M. Sellathurai, and C. Zhong,
"On the performance of eigenvalue-based spectrum
sensing for cognitive radio," in IEEE DySPAN 2010,](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/85/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-6-320.jpg)
![LE and KONG : PERFORMANCE OF SPIKED POPULATION MODELS FOR SPECTRUM SENSING
209
Tan-Thanh Le Hyung Yun Kong
received the B.S. degree in Telecommuni-
cation Engineering from Poly-technique
University of Hochiminh, Vietnam in 2002.
In 2005, he got the degree of Master
from Poly-technique University of Hochi-
minh, Vietnam in major of Electrical and
Electronics Engineering. Since 2009, he
has been studying Ph.D. program at Uni-
versity of Ulsan, Korea. His major research area is Cognitive
Radio Network, Cooperative Communication.
received the M.E. and Ph.D. degrees in
electrical engineering from Polytechnic
University, Brooklyn, New York, USA, in
1991 and 1996, respectively, He received
a BE in electrical engineering from New
York Institute of Technology, New York,
in 1989. Since 1996, he has been with
LG Electronics Co., Ltd., in the multi-
media research lab developing PCS mobile phone systems,
and from 1997 the LG chairman's office planning future satel-
lite communication systems. Currently he is a Professor in
electrical engineering at the University of Ulsan, Korea. His
research area includes channel coding, detection and estima-
tion, cooperative communications, cognitive radio and sensor
networks. He is a member of IEEK, KICS, KIPS, IEEE, and
IEICE.
2010.
[2] Z. Yonghong, L. Ying-chang, "Eigenvalue-based spec-
trum sensing algorithms for cognitive radio," IEEE
Transactions on Communications, vol. 57, no. 6, pp.
1784-1793, 2009.
[3] S. Peche, "Universality results for the largest eigen-
values of some sample covariance matrix ensem-
bles," Probability Theory and Related Fields, vol.
143, no. 3, pp. 481-516, 2009.
[4] O. Feldheim, S. Sodin, "A universality result for the
smallest eigenvalues of certain sample covariance ma-
trices," Geometric and Functional Analysis, vol. 20,
no. 1, pp. 88-123, 2010.
[5] N. M. Faber, L. M. C. Buydens, and G. Kateman,
"Aspects of pseudorank estimation methods based on
the eigenvalues of principal component analysis of
random matrices," Chemometrics and Intelligent La-
boratory Systems, vol. 25, no. 2, pp. 203-226, 1994.
[6] D. Feral, S. Peche, "The largest eigenvalues of sam-
ple covariance matrices for a spiked population: Dia-
gonal case," Journal of Mathematical Physics, vol.
50, no. 7, pp. 073302-073333, 2009.
[7] F. Penna, R. Garello, and M. A. Spirito, "Cooperati-
ve spectrum sensing based on the limiting eigenval-
ue ratio distribution in Wishart matrices," IEEE Co-
mmunications Letters, vol. 13, no. 7, pp. 507-509,
2009.
[8] S. Kritchman, B. Nadler, "Non-parametric detection
of the number of signals: Hypothesis testing and
random matrix theory," IEEE Transactions on Sig-
nal Processing, vol. 57, no. 10, pp. 3930-3941,
2009.
[9] I. M. Johnstone, "On the distribution of the largest
eigenvalue in principal components analysis," An-
nals of Statistics, vol. 29, 2001.
[10] F. Penna, R. Garello, "Theoretical performance ana-
lysis of eigenvalue-based detection," in Preprint, 2009.
[11] R. Bhatia, Matrix Analysis, Springer, 1997.](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/85/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-7-320.jpg)
![JOURNAL OF ELECTROMAGNETIC ENGINEERING AND SCIENCE, VOL. 12, NO. 3, 203～209, SEP. 2012
http://dx.doi.org/10.5515/JKIEES.2012.12.3.203
ISSN 2234-8395 (Online)․ISSN 2234-8409 (Print)
203
Performance of Spiked Population Models for Spectrum Sensing
Tan-Thanh Le․Hyung Yun Kong
Abstract
In order to improve sensing performance when the noise variance is not known, this paper considers a so-called
blind spectrum sensing technique that is based on eigenvalue models. In this paper, we employed the spiked population
models in order to identify the miss detection probability. At first, we try to estimate the unknown noise variance
based on the blind measurements at a secondary location. We then investigate the performance of detection, in terms
of both theoretical and empirical aspects, after applying this estimated noise variance result. In addition, we study the
effects of the number of SUs and the number of samples on the spectrum sensing performance.
Key words: Random Matrix Theory, Cognitive Radio Networks, Spiked Population Model, Spectrum Sensing Te-
chniques, Eigenvalue-Based Spectrum Sensing, Noise Estimation.
Manuscript received December 5, 2011 ; Revised February 16, 2012 ; Accepted February 27, 2012. (ID No. 20111205-039J)
Dept. of Electrical. Engineering, University of Ulsan, Ulsan, Korea.
Corresponding Author : Hyung Yun Kong (e-mail : hkong@mail.ulsan.ac.kr)
This is an Open-Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/
by-nc/3.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
Ⅰ. Introduction
Much recent work has focused on eigenvalue-based
spectrum sensing methods for cognitive radio networks
(CRNs) [1], [2], where researchers have applied inno-
vations from random matrix theory to calculate the pro-
bability of a false alarm as a function of a threshold.
This work has also employed many kinds of test sta-
tistics, such as the ratio of the largest and the smallest
eigenvalues or the ratio of the average and the smallest
eigenvalues of the sample covariance matrix. Yonghong
and Ying-chang’s model in [2] requires a number of
samples, N, and a number of secondary users (SUs), K,
approaching infinity. A significant gap exists between
the simulations and analytical results.
Kortun et al. [1] improved the model by finding an
exact threshold and an approximate closed-form per-
formance that agreed well with the empirical results.
Independently, Penna et al. [7] derived work similar to
[1] and then further analyzed the probability of a miss
in [10]. However, the method used in [10] still requires
knowledge of noise level and the parameters of primary
user (PU) signals and channels, which is impractical in
CRN scenarios. To our knowledge, not much research
has been carried out regarding the relationship between
the threshold and the probability of a miss when using
blind spectrum sensing techniques.
This paper considers a novel blind spectrum sensing
technique for CRNs, where we analyze a threshold with
a constraint on the probability of a miss. CRNs are com-
posed of a specific bandwidth, including multiple PUs
and the number of SUs. Based on a variety of research
related to the distribution of eigenvalues [3], [4], we ap-
ply the results of random matrix theory to the spectrum
sensing model. In fact, the practical model is related to
a finite number of samples and SUs, while these para-
meters are infinite in random matrix theory. Therefore,
we exploit the limit distribution for approximate results
in our applied spectrum sensing model. Furthermore, the
observations referred to in this spectrum sensing model
undergo interference with by complex Gaussian noise and
are affected by fading channels. Performance still de-
pends on the channel parameters, noise powers, and the
PU signal powers. However, these parameters are un-
known in practice. Therefore, noise estimation schemes
must be combined with analytical approaches. This work
also focuses on improving the performance of the noise
estimation scheme in order to provide an efficient sol-
ution to the problem with unknown information.
The rest of this paper is organized as follows. Section
2 presents the sensing model. The analysis of the proba-
bility of a miss as a function of threshold is considered
in Section 3. Section 4 demonstrates simulation results
and discussions. Finally, concluding remarks are given
ⓒ Copyright The Korean Institute of Electromagnetic Engineering and Science. All Rights Reserved.](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/75/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-1-2048.jpg)
![JOURNAL OF ELECTROMAGNETIC ENGINEERING AND SCIENCE, VOL. 12, NO. 3, SEP. 2012
204
in Section 5.
Ⅱ. System Model
The system model can be presented with some assum-
ptions. At the PU side, there are P sub-bands of interest
transmitted by P entities. Cognitive radio networks have
K SUs, which can share their information through high
speed transmissions. The high speed requirement for ex-
change of data is the problem of interest in the model
design. Fortunately, these cognitive radio nodes can be
linked by wired networks, which help to significantly
improve the speed.
The observable signals at the SUs can be expressed
as
Y HX V,= + (1)
where the K×N matrix Y is the measurement at the SUs.
The K×P matrix H represents the channels between the
PUs and SUs. The P×N matrix X denotes the trans-
mitted signals from the PUs. In fact, the transmitted sig-
nals can be expressed as X=[x(n)]T, n=1, …, N, where
each x(n) is the P´1 vector whose elements are obtained
from samples of PU signals. Therefore, the covarian-
ce matrix of these independent PUs can be written as
{ } 2 2
1E xx C diag( ,..., )H
x Ps s= = , where 2
,is [ ]1,i PÎ is
the variance of the i-th PU signal. The K×N matrix V
is the complex Gaussian noise with zero mean and var-
iances .
The hypotheses can be expressed as
00 : Y | V,=HH (2)
11 :Y| HX V,= +HH (3)
Ⅲ. Performance Analysis
3-1 Probability of Miss Analysis
In the case of 1H , the measurements at the SUs are
rewritten as
(4)
where the covariance matrix of the PU signals are Cx
{ } ( )H 2 2
1 Pxx diag σ , ,σ ,Cx E = ¼= 1/21
J HC I
σ
x K
é ù
= ê ú
ë û
,
1/2
σC X
R
V
x
-
é ù
= ê ú
ë û
.
The test statistic can be found by calculating the ei-
genvalues 2
1
JJ HC H I
σ
H
x
H
K= + as follows:
( )
( )
( )
( )( )
( )
H
2
2 ( )
~ ~
2 ( )
K
det JJ I 0,
1
det HC H I I 0,
det HC H 1 I 0,
det(HC H I ) 0,? ,
J
K
JH
x K K
H J
x K
H J
x i
t
t
t
t t t
s
s
s
- =
æ ö
Û + - =ç ÷
è ø
Û - - =
Û - = = -
(5)
Using the generalized matrix determinant lemma, we
have
det detdet
⇔det
det
⇔
≤ ≤
det
≤ ≤
(6)
Hence,
~
( )
2
1.iJ
it
t
s
= +
(7)
It is easily observed from Eqs. (5) and (6) that the
power of each PU and the channel information, repre-
sented by Cx and H, respectively, must be known. We
first consider the case of known channel state infor-
mation and unknown PU signal powers. The noise vari-
ances must be estimated, and then we can separate into
two groups of noises or signals with variances equal to
or larger than the estimated one. This technique will be
presented and analyzed in Subsection 3.1. Furthermore,
SUs do not know channel state information in practice;
hence, we approximate ( )J
it for a fully blind spectrum sen-
sing model in the Simulation section.
3-2 Methods for Noise Estimation
In Subsection 3.2, the final expression of miss proba-
bility requires knowledge of the sampled covariance ma-
trix of PU signals, the channel matrix, and the noise
level. In order to obtain these values, we must estimate
the number of PUs and the noise power. Based on pre-
vious research [8], we present and analyze the perfor-
mance of noise estimation, as follows.
The distribution of the largest eigenvalue 1l is
( )2
2
1
,
( , )
lim Pr
/
,
( , )
W
N p
N p
x F x
N p
l s m
s®¥
ì ü-
< =í ý
î þ (8)
where](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/75/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-2-2048.jpg)
![LE and KONG : PERFORMANCE OF SPIKED POPULATION MODELS FOR SPECTRUM SENSING
205
21
( , ) ( 1/ 2 1/ 2) ,N p p
N
Nm = - + -
1/3
1 1 1
( , ) ( 1/ 2 1/ 2)
1/ 2 1/ 2
,N p N p
N N p
s
æ ö
= - + - +ç ÷ç ÷- -è ø
and ( )2WF x is the Tracy-Widom distribution with order
2 ( 2W ) [9].
In order to determine the total number of active Pus,
^
P , we apply a series of tests. In other words, we will
estimate the noise level of the communication environ-
ment at the p-th eigenvalue, and then test the likelihood
of this p-th eigenvalue, pl , to determine whether it has
arisen from PU signals or noises. In each test, there are
two hypotheses, 0H and 1H , where the former represents
at least p eigenvalues satisfying (9). The latter repre-
sents, at most, (p—1) eigenvalues not satisfying the con-
dition in (9), due to the fact that the p-th eigenvalue is
used as the edges of the two hypotheses, the eigenvalues
il , [ ]1,i p KÎ + are assumed to be arisen from noises.
0H has at least p eigenvalues satisfying (9).
0H has at most (p-1) eigenvalues not satisfying (9).
The condition used for testing is
( ) ( ) ( ) ( )( )2
, , ,p PN p N K p x b N K pl s m s> - + - (9)
where ( )2Wb F x= is the confidence level, and ( ) 2
1
Wx b F b-
=
( )2
1
Wx b F b-
is the value determined from inverting the second-
order Tracy Widom distribution. In our work, b is va-
ried from 0.1 % to 0.5 %.
The operation can be implemented as described be-
low. If the measurements satisfy hypothesis 0H , we in-
crease p and perform the test again. Otherwise, when
the hypothesis 1H is satisfied, we stop testing and record
the final result
^
1P p= - . Based on the above operation,
^
P can be defined as the solution to the following opti-
mization problem
( ) ( ) ( )( )
2
( )
arg min -1.
, - , -
p P N
p
p
P
N K p x b N K p
l s
m s
ì ü´ï ï
= í ý
£
+ï ïî þ
$
(10)
For Eq. (10), the noise level must be known. There-
fore, we next present the method to estimate 2
P Ns .
3-2-1 A Simple Method for Noise Estimation
Equation (1) can be rewritten as
1
y h x ,
P
j j
j
sx
=
= +å
(11)
The covariance matrix can be modified to a diagonal
matrix as
(12)
where U is now an unknown K×K matrix whose col-
umns are the eigenvectors of Cy , ..
1
11
2
λ
λ
.A
é ù
ê
´
ú
= ê
ê
ë û´
ú
ú
, and
( )22A 0 K P= - . It is clear that, with a very large num-
ber of samples ( N ® ¥ ), the first P eigenvalues of the
covariance matrix are 2
il s+ , i=1, …, P, and the re-
maining (K—P) eigenvalues are 2
s . The eigenvalues are
separated into two groups, the group of signals with es-
timated eigenvalues larger than ^
2
s and the group of
noises with the remaining eigenvalues. Unfortunately,
the noise eigenvalues are widely spread in the case of
a finite number of samples. In this setting, we define the
covariance matrix for a finite number of samples and
observe that
1
C y y .
N
H
yN j j
j=
= å
(13)
In practice, the sampled covariance matrix can be
written as
1τ 0
U C U
τ 0
T
yN
K
´ ´æ ö æ ö
ç ÷ ç ÷
= +ç ÷ ç ÷
ç ÷ ç ÷´ ´è ø è ø
O O
(14)
where
. It is easily seen that the diagonal
elements from (P+1) to K are due to noise. By avera-
ging these elements, we derive the noise variance as
2
1 1 1
1 1
1 1
1
( )
K K P
i i i
i P i i
K P
i i i
i P i
K P K P
K P
s t l t
l l t
= + = =
= + =
æ ö
= = -ç ÷
- - è ø
æ ö
= + -ç ÷
- è ø
å å å
å å
(15)
However, the matrix U is unknown. Hence, we must
estimate it . In the simple form, we set i it l= and get
^
2
simple
1
1
.
K
i
i PK P
s l
= +
æ ö
= ç ÷
- è ø
å
(16)
3-2-2 Improvement in Noise Estimation
In order to improve the estimated noise variance, we
now diagonalize the upper left submatrix 1 N 1U C UT
y , as](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/75/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-3-2048.jpg)
![,
JOURNAL OF ELECTROMAGNETIC ENGINEERING AND SCIENCE, VOL. 12, NO. 3, SEP. 2012
206
in [8]
11 12
1 1
21 22
B B
U C U ,
B B
T
yN
æ ö
= ç ÷
è ø (17)
where
1
0
0
P
f
f
æ ö
ç ÷
ç ÷
ç ÷
è ø
O , 12 21B BT
Z= = and
1
22
0
0
B
P
K
t
t
+æ ö
ç ÷
= ç ÷
ç ÷
è ø
O .
The eigenvalues of 1 N 1U C UT
y are
≈
∈
(18)
where { }2 2 21
ij ij jz E z
N
f s» = . Therefore, Eq. (18) can be
derived as
2 2
2 2
1
1
,
K
j j
j j j
i P j j
K P
N N
f f
l f s f s
f s f s= +
-
» + = +
- -
å
(19)
where [ ]1,j PÎ . To this end, we iterate Eqs. (15) and
(19) to estimate the noise variance
^
2
1 1
^ ^ ^ ^
2 2 2
i
1
( )
,
λ 1 0
K P
iim i i
i P i
i i im i im
K P
K P
N
s l l f
f f s l s
= + =
ì æ ö
= + -ï ç ÷
- è øï
í
æ - öæ öï - + - + =ç ÷ç ÷ï è øè øî
å å
$
(20)
with the initial value
^ ^
2 2
0 simple / 1
P
N
s s
æ ö
= -ç ÷
è ø
for better per-
formance [5].
After estimating
^
2
ims , which is used for missed prob-
ability analysis with known channel state information
and PU signal powers, we also use ∈ to calcu-
late the eigenvalue ( )J
it as in Section 4 for the case of
blind spectrum sensing models.
3-3 The Probability of Miss
Based on the estimated result of noise variance,
^
2
ims ,
we perform the calculation of Pm. Note that the follow-
ing expressions use 2
s instead of
^
2
ims
.
From [6], we
have the following theorem with
( )( ) 2 ( )
1 1 ( )
1
1 ,
1
J J
J
c
t t
t
m s
æ ö
= +ç ÷
-è ø (21)
( )
( )
( ) 2 ( )
1 1 2( )
1
1 ,
1
J J
J
c
t t
t
s s= -
-
(22)
where c=K/N.
3-3-1 Theorem 1
Ref. [3] Consider the sequence of a complex Wishart
matrix with Cx. Let 1 m P£ £ , we have the case of iden-
tifiable signals; i.e., the spiked eigenvalues are above the
critical limit 1/2
1 c+ . If
…
. Let
( )
( )( )
11 1( )
1
,J
J
N
t
t
e l m
s
æ ö
= -ç ÷
è ø
$
(23)
then
( ) ( )1
,
lim Pr ,m
N P
e g g
¥
£ = G
™ (24)
where ( )m gG is the finite Gaussian unitary ensemble
(GUE) distribution given as
×
∞
⋯ ∞
≤ ≺ ≤
exp
⋯
(25)
Based on Eq. (24), the distribution of the largest ei-
genvalue is lim
→∞
.
In order to determine the distribution of the smallest
eigenvalues, which are the remaining (K-P) eigenvalues,
we adhere to Theorem 2 [4]. Let
~
,
K P
c
N
-
=
(26)
~ ~
1/2
2
2
1 ,c cm sæ ö æ ö
= -ç ÷ ç ÷
è ø è ø (27)
~ ~ ~
1/2
1/
1/2
3
2
1 1 ,c c cs s -æ ö æ öæ ö
= - -ç ÷ ç ÷ç ÷
è ø è øè ø (28)
3-3-2 Theorem 2
Consider the sequence of a complex Wishart matrix
~
Cx with size (K − P)×(K − P) (Note that
~
Cx is the
lower right submatrix of Cx ). Kl is the smallest ei-
genvalue of
~
Cx . The variable
2/3 ^ ~
~
KK
N
c
c
e l m
s
æ öæ ö= - ç ÷ç ÷
æ ö è øè ø
ç ÷
è ø
will be
satisfied
( )2,( )
lim Pr( ) .K W
N K P
Fe g g
- ¥
£ =
™ (29)
Based on Eq. (29), the distribution of the smallest ei-
genvalue is 2,
lim ( ) ( )K W
N P
f x F xe
®¥
= .
Now we can establish the ratio test as follows,](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/75/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-4-2048.jpg)
![LE and KONG : PERFORMANCE OF SPIKED POPULATION MODELS FOR SPECTRUM SENSING
207
( )1/2 ( )
1 1 1
2
^
/
1
^ ~
3
~
( ) ( )
.
JJ
K K
N t t
T
c cN
sl
l
e m
e s m
-
-
+
= =
æ ö æ ö
+ç ÷ ç ÷
è ø è ø (30)
Following [7], the joint distribution of the largest and
smallest eigenvalues is
( ) ( ) ( )1 1|
0
,KT Hf t xf tx f x dxl l
¥
= ò
(31)
where the distributions of the largest and smallest ei-
genvalues are derived from ( )1
f xe and ( )K
f xe , respecti-
vely, and are given as
( ) ( )
( )
( )( )1
1/21/2
( )
( )
,
( ) m
J
GJ J
N N
f x f x t
t t
l m
s s
æ ö
ç ÷= -
ç ÷
è ø (32)
( ) 2
2
2/3 3 ~
~ ~
|
.
|
K w
N N
c
c
f x f x
c
l m
s s
æ ö
ç ÷
æ öæ öç ÷= - ç ÷ç ÷ç ÷æ ö è øæ ö è ø
ç ÷ ç ÷ç ÷
è ø è øè ø (33)
Therefore, the probability of a miss can be calculated
as
( ) 11 |Pr | ( )m T HP T H Fg g= < = (34)
and the threshold is the inversion of Eq. (34)
( ) ( )1
-1
| .m T H mP F Pg = (35)
It is easily observed that the threshold derived from
Eq. (35) depends on the channel matrix and the PU sig-
nal powers. This is due to the fact that 1| ( )T HF g , which
includes H and Cx. Information on the PU signal powers
is included in Subsection 3.1, and the model considered
here is a CSI case. In the next section, we approximate
( )
1
J
t in order to obtain the results in the case of the fully
blind spectrum sensing model.
Ⅳ. Simulation Results and Discussions
After noise estimation, the results are the noise level
2
PNs and the number of PU signals, Pest, with the proba-
bility of the correct number of PU signals Pr(Pest=P).
The covariance matrix of the measurements at K of the
SUs can then be diagonalized as in Eq. (14) and the
first P eigenvalues [ ]
^
, 1,i i Pl Î represented as P PU sig-
nal powers are approximated as Eq. (19). Substituting
these eigenvalues to Eq. (7) at i=1, we have
≈
(36)
Proof: When N is very large, the approximation is as
follows:
HC H ,H
yN x vC C» + (37)
where Cv is the sampled covariance matrix of noise.
From [11], Weyl’s inequality theorem reads as
( )
~ ^ ~
max 1 1 maxC (C ),p v vt tl l l+ £ £ + (38)
where ( )max Cvl is the estimated noise eigenvalue,
2
P Ns ,
which is calculated as in Subsection 3.1,
~ ~ ~
1 2, ,..., Pt t t are P
eigenvalues of HC HH
x , and
^
1l is the largest eigenvalue
of CyN . Moreover, at very large N, the simulation shows
that the following approximation performs in a manner
that is in good agreement with the empirical results
( )
^
2
1 max
~ ~
1 1λ Cv P Nt tl s» + = + (39)
Substituting Eq. (39) into Eq. (36), we have the final
result.
We now perform spectrum sensing without using any
information from the channels or the noise level, in this
case. We note that noise estimation relates to the correct
number of PUs Pr(Pest=P), so the final probability of a
miss is ( )
^
estPr .mP Pm P P= ´ =
Now, we analyze the effects of parameters such as K,
N, and P on the noise estimation performance and there-
fore the spectrum sensing performance. In Fig. 1, the
miss probabilities for the unknown model, the ideal
model (known model), and the empirical simulation are
calculated using various numbers of samples and SUs.
As illustrated in this figure, the gap between the ideal
and the unknown schemes is slightly too small. In par-
ticular, when the number of SUs and samples are in-
creased ((N, K)=(2,000, 40), (N, K)=(2,000, 50) and (N,
K)=(800, 40)), the results of both cases are the same. To
confirm a good match between the theoretical analysis
and the empirical experiment, the model uses only N=
2,000 or N=800 samples, which are appropriate for the
practical model.
In order to completely understand the estimation sche-
me, Fig. 2 shows the performance when estimating PU
signal powers. Even though a pair of samples and the
number of SUs (N, K) are chosen to be (1,000, 50) (i.e.,
a not very large N), the estimated power densities agree
well with the original power densities. A larger number
of samples gives a more exact result for the estimate of
the powers. Hence, if we want to obtain a more precise
estimate in order to reduce the difference between the
theoretical and empirical results, we slightly increase the
number of samples N.](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/75/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-5-2048.jpg)
![JOURNAL OF ELECTROMAGNETIC ENGINEERING AND SCIENCE, VOL. 12, NO. 3, SEP. 2012
208
Fig. 1. Probability of miss vs. the threshold in the case of
P=4 and variable K and N.
Fig. 2. Density of the ratio of signal eigenvalues to noise ei-
genvalues in the case of P=3, K =50 and N=1,000.
Fig. 3 shows the receiver operating characteristic
(ROC), which describes the sensitivity of the spectrum
sensing model. We also note that, in the case of H0, the
probability of a false alarm, faP , is calculated based on
the results in [7], which include the analysis for faP as
a function of a decision threshold.
Fig. 4 shows the convergence of our proposed scheme
at P=4 with the constraint c=0.1. It is clearly observed
that the increases in both K and N confirm the con-
vergence of the curve represented by the relation be-
tween Pm and the threshold. In the case of N=80 and
K=8, the estimated scheme no longer agrees with the
known scheme due to the fact that noise estimation per-
formance is heavily degraded when N is too low.
Moreover, after the noise estimation stage, the perfor-
mance of the spectrum sensing technique is also affected
by very low K and N. Increases in both N and K with
the constraint N/K=c ensure that the performance of the
analytical unknown case matches well with that of the
empirical known case.
Finally, we did not refer to the energy detection tech-
nique (ED) in our work; our focus was on comparison
of our proposed method with state-of-the-art eigenvalue-
based approaches. Several other studies [1], [2], [7],
Fig. 3. Probability of miss vs. probability of a false alarm
in the case of P=1, K=50 and N =1,000.
Fig. 4. Miss probability vs. threshold in the case of P=4 and
variable K and N with the constraint c=K/N=0.1.
[10] have confirmed that the eigenvalue-based methods
outperform the ED.
Ⅴ. Conclusion
Our work considered a new sensing model for cogni-
tive radio networks, where multiple SUs blindly sense
the band of interest, including multiple PUs. A novel
analysis is presented to determine the threshold based on
the probability of a miss, which is a new method that
utilizes the best knowledge from spectrum sensing. Mo-
reover, the proposed hybrid analytical simulation method
enhances the blind features of spectrum sensing models
since the unknown noise level, and hence the parameters
of PU signals and channels, are estimated with small
errors. We also show the effects of the estimation sche-
me and attempt to improve its performance. As a con-
sequence, the empirical simulation results illustrate small
differences between the performances of the ideal (kno-
wn information) and estimate (blind) models.
References
[1] A. Kortun, T. Ratnarajah, M. Sellathurai, and C. Zhong,
"On the performance of eigenvalue-based spectrum
sensing for cognitive radio," in IEEE DySPAN 2010,](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/75/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-6-2048.jpg)
![LE and KONG : PERFORMANCE OF SPIKED POPULATION MODELS FOR SPECTRUM SENSING
209
Tan-Thanh Le Hyung Yun Kong
received the B.S. degree in Telecommuni-
cation Engineering from Poly-technique
University of Hochiminh, Vietnam in 2002.
In 2005, he got the degree of Master
from Poly-technique University of Hochi-
minh, Vietnam in major of Electrical and
Electronics Engineering. Since 2009, he
has been studying Ph.D. program at Uni-
versity of Ulsan, Korea. His major research area is Cognitive
Radio Network, Cooperative Communication.
received the M.E. and Ph.D. degrees in
electrical engineering from Polytechnic
University, Brooklyn, New York, USA, in
1991 and 1996, respectively, He received
a BE in electrical engineering from New
York Institute of Technology, New York,
in 1989. Since 1996, he has been with
LG Electronics Co., Ltd., in the multi-
media research lab developing PCS mobile phone systems,
and from 1997 the LG chairman's office planning future satel-
lite communication systems. Currently he is a Professor in
electrical engineering at the University of Ulsan, Korea. His
research area includes channel coding, detection and estima-
tion, cooperative communications, cognitive radio and sensor
networks. He is a member of IEEK, KICS, KIPS, IEEE, and
IEICE.
2010.
[2] Z. Yonghong, L. Ying-chang, "Eigenvalue-based spec-
trum sensing algorithms for cognitive radio," IEEE
Transactions on Communications, vol. 57, no. 6, pp.
1784-1793, 2009.
[3] S. Peche, "Universality results for the largest eigen-
values of some sample covariance matrix ensem-
bles," Probability Theory and Related Fields, vol.
143, no. 3, pp. 481-516, 2009.
[4] O. Feldheim, S. Sodin, "A universality result for the
smallest eigenvalues of certain sample covariance ma-
trices," Geometric and Functional Analysis, vol. 20,
no. 1, pp. 88-123, 2010.
[5] N. M. Faber, L. M. C. Buydens, and G. Kateman,
"Aspects of pseudorank estimation methods based on
the eigenvalues of principal component analysis of
random matrices," Chemometrics and Intelligent La-
boratory Systems, vol. 25, no. 2, pp. 203-226, 1994.
[6] D. Feral, S. Peche, "The largest eigenvalues of sam-
ple covariance matrices for a spiked population: Dia-
gonal case," Journal of Mathematical Physics, vol.
50, no. 7, pp. 073302-073333, 2009.
[7] F. Penna, R. Garello, and M. A. Spirito, "Cooperati-
ve spectrum sensing based on the limiting eigenval-
ue ratio distribution in Wishart matrices," IEEE Co-
mmunications Letters, vol. 13, no. 7, pp. 507-509,
2009.
[8] S. Kritchman, B. Nadler, "Non-parametric detection
of the number of signals: Hypothesis testing and
random matrix theory," IEEE Transactions on Sig-
nal Processing, vol. 57, no. 10, pp. 3930-3941,
2009.
[9] I. M. Johnstone, "On the distribution of the largest
eigenvalue in principal components analysis," An-
nals of Statistics, vol. 29, 2001.
[10] F. Penna, R. Garello, "Theoretical performance ana-
lysis of eigenvalue-based detection," in Preprint, 2009.
[11] R. Bhatia, Matrix Analysis, Springer, 1997.](https://image.slidesharecdn.com/03tan-thanhle-130815002657-phpapp01/75/Performance-of-Spiked-Population-Models-for-Spectrum-Sensing-7-2048.jpg)
Uploaded byPolytechnique Montreal
Performance of Spiked Population Models for Spectrum Sensing
This paper discusses a novel blind spectrum sensing technique for cognitive radio networks (CRNs) that improves detection performance without prior knowledge of noise variance. It employs spiked population models to estimate unknown noise variances and analyzes the probability of miss detection. The study addresses the effects of the number of secondary users (SUs) and samples on spectrum sensing performance, combining theoretical and empirical approaches.
More Related Content
Similar to Performance of Spiked Population Models for Spectrum Sensing
Performance of Spiked Population Models for Spectrum Sensing
- 1. JOURNAL OF ELECTROMAGNETICENGINEERING AND SCIENCE, VOL. 12, NO. 3, 203～209, SEP. 2012 http://dx.doi.org/10.5515/JKIEES.2012.12.3.203 ISSN 2234-8395 (Online)․ISSN 2234-8409 (Print) 203 Performance of Spiked Population Models for Spectrum Sensing Tan-Thanh Le․Hyung Yun Kong Abstract In order to improve sensing performance when the noise variance is not known, this paper considers a so-called blind spectrum sensing technique that is based on eigenvalue models. In this paper, we employed the spiked population models in order to identify the miss detection probability. At first, we try to estimate the unknown noise variance based on the blind measurements at a secondary location. We then investigate the performance of detection, in terms of both theoretical and empirical aspects, after applying this estimated noise variance result. In addition, we study the effects of the number of SUs and the number of samples on the spectrum sensing performance. Key words: Random Matrix Theory, Cognitive Radio Networks, Spiked Population Model, Spectrum Sensing Te- chniques, Eigenvalue-Based Spectrum Sensing, Noise Estimation. Manuscript received December 5, 2011 ; Revised February 16, 2012 ; Accepted February 27, 2012. (ID No. 20111205-039J) Dept. of Electrical. Engineering, University of Ulsan, Ulsan, Korea. Corresponding Author : Hyung Yun Kong (e-mail : [email protected]) This is an Open-Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/ by-nc/3.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited. Ⅰ. Introduction Much recent work has focused on eigenvalue-based spectrum sensing methods for cognitive radio networks (CRNs) [1], [2], where researchers have applied inno- vations from random matrix theory to calculate the pro- bability of a false alarm as a function of a threshold. This work has also employed many kinds of test sta- tistics, such as the ratio of the largest and the smallest eigenvalues or the ratio of the average and the smallest eigenvalues of the sample covariance matrix. Yonghong and Ying-chang’s model in [2] requires a number of samples, N, and a number of secondary users (SUs), K, approaching infinity. A significant gap exists between the simulations and analytical results. Kortun et al. [1] improved the model by finding an exact threshold and an approximate closed-form per- formance that agreed well with the empirical results. Independently, Penna et al. [7] derived work similar to [1] and then further analyzed the probability of a miss in [10]. However, the method used in [10] still requires knowledge of noise level and the parameters of primary user (PU) signals and channels, which is impractical in CRN scenarios. To our knowledge, not much research has been carried out regarding the relationship between the threshold and the probability of a miss when using blind spectrum sensing techniques. This paper considers a novel blind spectrum sensing technique for CRNs, where we analyze a threshold with a constraint on the probability of a miss. CRNs are com- posed of a specific bandwidth, including multiple PUs and the number of SUs. Based on a variety of research related to the distribution of eigenvalues [3], [4], we ap- ply the results of random matrix theory to the spectrum sensing model. In fact, the practical model is related to a finite number of samples and SUs, while these para- meters are infinite in random matrix theory. Therefore, we exploit the limit distribution for approximate results in our applied spectrum sensing model. Furthermore, the observations referred to in this spectrum sensing model undergo interference with by complex Gaussian noise and are affected by fading channels. Performance still de- pends on the channel parameters, noise powers, and the PU signal powers. However, these parameters are un- known in practice. Therefore, noise estimation schemes must be combined with analytical approaches. This work also focuses on improving the performance of the noise estimation scheme in order to provide an efficient sol- ution to the problem with unknown information. The rest of this paper is organized as follows. Section 2 presents the sensing model. The analysis of the proba- bility of a miss as a function of threshold is considered in Section 3. Section 4 demonstrates simulation results and discussions. Finally, concluding remarks are given ⓒ Copyright The Korean Institute of Electromagnetic Engineering and Science. All Rights Reserved.
- 2. JOURNAL OF ELECTROMAGNETICENGINEERING AND SCIENCE, VOL. 12, NO. 3, SEP. 2012 204 in Section 5. Ⅱ. System Model The system model can be presented with some assum- ptions. At the PU side, there are P sub-bands of interest transmitted by P entities. Cognitive radio networks have K SUs, which can share their information through high speed transmissions. The high speed requirement for ex- change of data is the problem of interest in the model design. Fortunately, these cognitive radio nodes can be linked by wired networks, which help to significantly improve the speed. The observable signals at the SUs can be expressed as Y HX V,= + (1) where the K×N matrix Y is the measurement at the SUs. The K×P matrix H represents the channels between the PUs and SUs. The P×N matrix X denotes the trans- mitted signals from the PUs. In fact, the transmitted sig- nals can be expressed as X=[x(n)]T, n=1, …, N, where each x(n) is the P´1 vector whose elements are obtained from samples of PU signals. Therefore, the covarian- ce matrix of these independent PUs can be written as { } 2 2 1E xx C diag( ,..., )H x Ps s= = , where 2 ,is [ ]1,i PÎ is the variance of the i-th PU signal. The K×N matrix V is the complex Gaussian noise with zero mean and var- iances . The hypotheses can be expressed as 00 : Y | V,=HH (2) 11 :Y| HX V,= +HH (3) Ⅲ. Performance Analysis 3-1 Probability of Miss Analysis In the case of 1H , the measurements at the SUs are rewritten as (4) where the covariance matrix of the PU signals are Cx { } ( )H 2 2 1 Pxx diag σ , ,σ ,Cx E = ¼= 1/21 J HC I σ x K é ù = ê ú ë û , 1/2 σC X R V x - é ù = ê ú ë û . The test statistic can be found by calculating the ei- genvalues 2 1 JJ HC H I σ H x H K= + as follows: ( ) ( ) ( ) ( )( ) ( ) H 2 2 ( ) ~ ~ 2 ( ) K det JJ I 0, 1 det HC H I I 0, det HC H 1 I 0, det(HC H I ) 0,? , J K JH x K K H J x K H J x i t t t t t t s s s - = æ ö Û + - =ç ÷ è ø Û - - = Û - = = - (5) Using the generalized matrix determinant lemma, we have det detdet ⇔det det ⇔ ≤ ≤ det ≤ ≤ (6) Hence, ~ ( ) 2 1.iJ it t s = + (7) It is easily observed from Eqs. (5) and (6) that the power of each PU and the channel information, repre- sented by Cx and H, respectively, must be known. We first consider the case of known channel state infor- mation and unknown PU signal powers. The noise vari- ances must be estimated, and then we can separate into two groups of noises or signals with variances equal to or larger than the estimated one. This technique will be presented and analyzed in Subsection 3.1. Furthermore, SUs do not know channel state information in practice; hence, we approximate ( )J it for a fully blind spectrum sen- sing model in the Simulation section. 3-2 Methods for Noise Estimation In Subsection 3.2, the final expression of miss proba- bility requires knowledge of the sampled covariance ma- trix of PU signals, the channel matrix, and the noise level. In order to obtain these values, we must estimate the number of PUs and the noise power. Based on pre- vious research [8], we present and analyze the perfor- mance of noise estimation, as follows. The distribution of the largest eigenvalue 1l is ( )2 2 1 , ( , ) lim Pr / , ( , ) W N p N p x F x N p l s m s®¥ ì ü- < =í ý î þ (8) where
- 3. LE and KONG: PERFORMANCE OF SPIKED POPULATION MODELS FOR SPECTRUM SENSING 205 21 ( , ) ( 1/ 2 1/ 2) ,N p p N Nm = - + - 1/3 1 1 1 ( , ) ( 1/ 2 1/ 2) 1/ 2 1/ 2 ,N p N p N N p s æ ö = - + - +ç ÷ç ÷- -è ø and ( )2WF x is the Tracy-Widom distribution with order 2 ( 2W ) [9]. In order to determine the total number of active Pus, ^ P , we apply a series of tests. In other words, we will estimate the noise level of the communication environ- ment at the p-th eigenvalue, and then test the likelihood of this p-th eigenvalue, pl , to determine whether it has arisen from PU signals or noises. In each test, there are two hypotheses, 0H and 1H , where the former represents at least p eigenvalues satisfying (9). The latter repre- sents, at most, (p—1) eigenvalues not satisfying the con- dition in (9), due to the fact that the p-th eigenvalue is used as the edges of the two hypotheses, the eigenvalues il , [ ]1,i p KÎ + are assumed to be arisen from noises. 0H has at least p eigenvalues satisfying (9). 0H has at most (p-1) eigenvalues not satisfying (9). The condition used for testing is ( ) ( ) ( ) ( )( )2 , , ,p PN p N K p x b N K pl s m s> - + - (9) where ( )2Wb F x= is the confidence level, and ( ) 2 1 Wx b F b- = ( )2 1 Wx b F b- is the value determined from inverting the second- order Tracy Widom distribution. In our work, b is va- ried from 0.1 % to 0.5 %. The operation can be implemented as described be- low. If the measurements satisfy hypothesis 0H , we in- crease p and perform the test again. Otherwise, when the hypothesis 1H is satisfied, we stop testing and record the final result ^ 1P p= - . Based on the above operation, ^ P can be defined as the solution to the following opti- mization problem ( ) ( ) ( )( ) 2 ( ) arg min -1. , - , - p P N p p P N K p x b N K p l s m s ì ü´ï ï = í ý £ +ï ïî þ $ (10) For Eq. (10), the noise level must be known. There- fore, we next present the method to estimate 2 P Ns . 3-2-1 A Simple Method for Noise Estimation Equation (1) can be rewritten as 1 y h x , P j j j sx = = +å (11) The covariance matrix can be modified to a diagonal matrix as (12) where U is now an unknown K×K matrix whose col- umns are the eigenvectors of Cy , .. 1 11 2 λ λ .A é ù ê ´ ú = ê ê ë û´ ú ú , and ( )22A 0 K P= - . It is clear that, with a very large num- ber of samples ( N ® ¥ ), the first P eigenvalues of the covariance matrix are 2 il s+ , i=1, …, P, and the re- maining (K—P) eigenvalues are 2 s . The eigenvalues are separated into two groups, the group of signals with es- timated eigenvalues larger than ^ 2 s and the group of noises with the remaining eigenvalues. Unfortunately, the noise eigenvalues are widely spread in the case of a finite number of samples. In this setting, we define the covariance matrix for a finite number of samples and observe that 1 C y y . N H yN j j j= = å (13) In practice, the sampled covariance matrix can be written as 1τ 0 U C U τ 0 T yN K ´ ´æ ö æ ö ç ÷ ç ÷ = +ç ÷ ç ÷ ç ÷ ç ÷´ ´è ø è ø O O (14) where . It is easily seen that the diagonal elements from (P+1) to K are due to noise. By avera- ging these elements, we derive the noise variance as 2 1 1 1 1 1 1 1 1 ( ) K K P i i i i P i i K P i i i i P i K P K P K P s t l t l l t = + = = = + = æ ö = = -ç ÷ - - è ø æ ö = + -ç ÷ - è ø å å å å å (15) However, the matrix U is unknown. Hence, we must estimate it . In the simple form, we set i it l= and get ^ 2 simple 1 1 . K i i PK P s l = + æ ö = ç ÷ - è ø å (16) 3-2-2 Improvement in Noise Estimation In order to improve the estimated noise variance, we now diagonalize the upper left submatrix 1 N 1U C UT y , as
- 4. , JOURNAL OF ELECTROMAGNETICENGINEERING AND SCIENCE, VOL. 12, NO. 3, SEP. 2012 206 in [8] 11 12 1 1 21 22 B B U C U , B B T yN æ ö = ç ÷ è ø (17) where 1 0 0 P f f æ ö ç ÷ ç ÷ ç ÷ è ø O , 12 21B BT Z= = and 1 22 0 0 B P K t t +æ ö ç ÷ = ç ÷ ç ÷ è ø O . The eigenvalues of 1 N 1U C UT y are ≈ ∈ (18) where { }2 2 21 ij ij jz E z N f s» = . Therefore, Eq. (18) can be derived as 2 2 2 2 1 1 , K j j j j j i P j j K P N N f f l f s f s f s f s= + - » + = + - - å (19) where [ ]1,j PÎ . To this end, we iterate Eqs. (15) and (19) to estimate the noise variance ^ 2 1 1 ^ ^ ^ ^ 2 2 2 i 1 ( ) , λ 1 0 K P iim i i i P i i i im i im K P K P N s l l f f f s l s = + = ì æ ö = + -ï ç ÷ - è øï í æ - öæ öï - + - + =ç ÷ç ÷ï è øè øî å å $ (20) with the initial value ^ ^ 2 2 0 simple / 1 P N s s æ ö = -ç ÷ è ø for better per- formance [5]. After estimating ^ 2 ims , which is used for missed prob- ability analysis with known channel state information and PU signal powers, we also use ∈ to calcu- late the eigenvalue ( )J it as in Section 4 for the case of blind spectrum sensing models. 3-3 The Probability of Miss Based on the estimated result of noise variance, ^ 2 ims , we perform the calculation of Pm. Note that the follow- ing expressions use 2 s instead of ^ 2 ims . From [6], we have the following theorem with ( )( ) 2 ( ) 1 1 ( ) 1 1 , 1 J J J c t t t m s æ ö = +ç ÷ -è ø (21) ( ) ( ) ( ) 2 ( ) 1 1 2( ) 1 1 , 1 J J J c t t t s s= - - (22) where c=K/N. 3-3-1 Theorem 1 Ref. [3] Consider the sequence of a complex Wishart matrix with Cx. Let 1 m P£ £ , we have the case of iden- tifiable signals; i.e., the spiked eigenvalues are above the critical limit 1/2 1 c+ . If … . Let ( ) ( )( ) 11 1( ) 1 ,J J N t t e l m s æ ö = -ç ÷ è ø $ (23) then ( ) ( )1 , lim Pr ,m N P e g g ¥ £ = G ™ (24) where ( )m gG is the finite Gaussian unitary ensemble (GUE) distribution given as × ∞ ⋯ ∞ ≤ ≺ ≤ exp ⋯ (25) Based on Eq. (24), the distribution of the largest ei- genvalue is lim →∞ . In order to determine the distribution of the smallest eigenvalues, which are the remaining (K-P) eigenvalues, we adhere to Theorem 2 [4]. Let ~ , K P c N - = (26) ~ ~ 1/2 2 2 1 ,c cm sæ ö æ ö = -ç ÷ ç ÷ è ø è ø (27) ~ ~ ~ 1/2 1/ 1/2 3 2 1 1 ,c c cs s -æ ö æ öæ ö = - -ç ÷ ç ÷ç ÷ è ø è øè ø (28) 3-3-2 Theorem 2 Consider the sequence of a complex Wishart matrix ~ Cx with size (K − P)×(K − P) (Note that ~ Cx is the lower right submatrix of Cx ). Kl is the smallest ei- genvalue of ~ Cx . The variable 2/3 ^ ~ ~ KK N c c e l m s æ öæ ö= - ç ÷ç ÷ æ ö è øè ø ç ÷ è ø will be satisfied ( )2,( ) lim Pr( ) .K W N K P Fe g g - ¥ £ = ™ (29) Based on Eq. (29), the distribution of the smallest ei- genvalue is 2, lim ( ) ( )K W N P f x F xe ®¥ = . Now we can establish the ratio test as follows,
- 5. LE and KONG: PERFORMANCE OF SPIKED POPULATION MODELS FOR SPECTRUM SENSING 207 ( )1/2 ( ) 1 1 1 2 ^ / 1 ^ ~ 3 ~ ( ) ( ) . JJ K K N t t T c cN sl l e m e s m - - + = = æ ö æ ö +ç ÷ ç ÷ è ø è ø (30) Following [7], the joint distribution of the largest and smallest eigenvalues is ( ) ( ) ( )1 1| 0 ,KT Hf t xf tx f x dxl l ¥ = ò (31) where the distributions of the largest and smallest ei- genvalues are derived from ( )1 f xe and ( )K f xe , respecti- vely, and are given as ( ) ( ) ( ) ( )( )1 1/21/2 ( ) ( ) , ( ) m J GJ J N N f x f x t t t l m s s æ ö ç ÷= - ç ÷ è ø (32) ( ) 2 2 2/3 3 ~ ~ ~ | . | K w N N c c f x f x c l m s s æ ö ç ÷ æ öæ öç ÷= - ç ÷ç ÷ç ÷æ ö è øæ ö è ø ç ÷ ç ÷ç ÷ è ø è øè ø (33) Therefore, the probability of a miss can be calculated as ( ) 11 |Pr | ( )m T HP T H Fg g= < = (34) and the threshold is the inversion of Eq. (34) ( ) ( )1 -1 | .m T H mP F Pg = (35) It is easily observed that the threshold derived from Eq. (35) depends on the channel matrix and the PU sig- nal powers. This is due to the fact that 1| ( )T HF g , which includes H and Cx. Information on the PU signal powers is included in Subsection 3.1, and the model considered here is a CSI case. In the next section, we approximate ( ) 1 J t in order to obtain the results in the case of the fully blind spectrum sensing model. Ⅳ. Simulation Results and Discussions After noise estimation, the results are the noise level 2 PNs and the number of PU signals, Pest, with the proba- bility of the correct number of PU signals Pr(Pest=P). The covariance matrix of the measurements at K of the SUs can then be diagonalized as in Eq. (14) and the first P eigenvalues [ ] ^ , 1,i i Pl Î represented as P PU sig- nal powers are approximated as Eq. (19). Substituting these eigenvalues to Eq. (7) at i=1, we have ≈ (36) Proof: When N is very large, the approximation is as follows: HC H ,H yN x vC C» + (37) where Cv is the sampled covariance matrix of noise. From [11], Weyl’s inequality theorem reads as ( ) ~ ^ ~ max 1 1 maxC (C ),p v vt tl l l+ £ £ + (38) where ( )max Cvl is the estimated noise eigenvalue, 2 P Ns , which is calculated as in Subsection 3.1, ~ ~ ~ 1 2, ,..., Pt t t are P eigenvalues of HC HH x , and ^ 1l is the largest eigenvalue of CyN . Moreover, at very large N, the simulation shows that the following approximation performs in a manner that is in good agreement with the empirical results ( ) ^ 2 1 max ~ ~ 1 1λ Cv P Nt tl s» + = + (39) Substituting Eq. (39) into Eq. (36), we have the final result. We now perform spectrum sensing without using any information from the channels or the noise level, in this case. We note that noise estimation relates to the correct number of PUs Pr(Pest=P), so the final probability of a miss is ( ) ^ estPr .mP Pm P P= ´ = Now, we analyze the effects of parameters such as K, N, and P on the noise estimation performance and there- fore the spectrum sensing performance. In Fig. 1, the miss probabilities for the unknown model, the ideal model (known model), and the empirical simulation are calculated using various numbers of samples and SUs. As illustrated in this figure, the gap between the ideal and the unknown schemes is slightly too small. In par- ticular, when the number of SUs and samples are in- creased ((N, K)=(2,000, 40), (N, K)=(2,000, 50) and (N, K)=(800, 40)), the results of both cases are the same. To confirm a good match between the theoretical analysis and the empirical experiment, the model uses only N= 2,000 or N=800 samples, which are appropriate for the practical model. In order to completely understand the estimation sche- me, Fig. 2 shows the performance when estimating PU signal powers. Even though a pair of samples and the number of SUs (N, K) are chosen to be (1,000, 50) (i.e., a not very large N), the estimated power densities agree well with the original power densities. A larger number of samples gives a more exact result for the estimate of the powers. Hence, if we want to obtain a more precise estimate in order to reduce the difference between the theoretical and empirical results, we slightly increase the number of samples N.
- 6. JOURNAL OF ELECTROMAGNETICENGINEERING AND SCIENCE, VOL. 12, NO. 3, SEP. 2012 208 Fig. 1. Probability of miss vs. the threshold in the case of P=4 and variable K and N. Fig. 2. Density of the ratio of signal eigenvalues to noise ei- genvalues in the case of P=3, K =50 and N=1,000. Fig. 3 shows the receiver operating characteristic (ROC), which describes the sensitivity of the spectrum sensing model. We also note that, in the case of H0, the probability of a false alarm, faP , is calculated based on the results in [7], which include the analysis for faP as a function of a decision threshold. Fig. 4 shows the convergence of our proposed scheme at P=4 with the constraint c=0.1. It is clearly observed that the increases in both K and N confirm the con- vergence of the curve represented by the relation be- tween Pm and the threshold. In the case of N=80 and K=8, the estimated scheme no longer agrees with the known scheme due to the fact that noise estimation per- formance is heavily degraded when N is too low. Moreover, after the noise estimation stage, the perfor- mance of the spectrum sensing technique is also affected by very low K and N. Increases in both N and K with the constraint N/K=c ensure that the performance of the analytical unknown case matches well with that of the empirical known case. Finally, we did not refer to the energy detection tech- nique (ED) in our work; our focus was on comparison of our proposed method with state-of-the-art eigenvalue- based approaches. Several other studies [1], [2], [7], Fig. 3. Probability of miss vs. probability of a false alarm in the case of P=1, K=50 and N =1,000. Fig. 4. Miss probability vs. threshold in the case of P=4 and variable K and N with the constraint c=K/N=0.1. [10] have confirmed that the eigenvalue-based methods outperform the ED. Ⅴ. Conclusion Our work considered a new sensing model for cogni- tive radio networks, where multiple SUs blindly sense the band of interest, including multiple PUs. A novel analysis is presented to determine the threshold based on the probability of a miss, which is a new method that utilizes the best knowledge from spectrum sensing. Mo- reover, the proposed hybrid analytical simulation method enhances the blind features of spectrum sensing models since the unknown noise level, and hence the parameters of PU signals and channels, are estimated with small errors. We also show the effects of the estimation sche- me and attempt to improve its performance. As a con- sequence, the empirical simulation results illustrate small differences between the performances of the ideal (kno- wn information) and estimate (blind) models. References [1] A. Kortun, T. Ratnarajah, M. Sellathurai, and C. Zhong, "On the performance of eigenvalue-based spectrum sensing for cognitive radio," in IEEE DySPAN 2010,
- 7. LE and KONG: PERFORMANCE OF SPIKED POPULATION MODELS FOR SPECTRUM SENSING 209 Tan-Thanh Le Hyung Yun Kong received the B.S. degree in Telecommuni- cation Engineering from Poly-technique University of Hochiminh, Vietnam in 2002. In 2005, he got the degree of Master from Poly-technique University of Hochi- minh, Vietnam in major of Electrical and Electronics Engineering. Since 2009, he has been studying Ph.D. program at Uni- versity of Ulsan, Korea. His major research area is Cognitive Radio Network, Cooperative Communication. received the M.E. and Ph.D. degrees in electrical engineering from Polytechnic University, Brooklyn, New York, USA, in 1991 and 1996, respectively, He received a BE in electrical engineering from New York Institute of Technology, New York, in 1989. Since 1996, he has been with LG Electronics Co., Ltd., in the multi- media research lab developing PCS mobile phone systems, and from 1997 the LG chairman's office planning future satel- lite communication systems. Currently he is a Professor in electrical engineering at the University of Ulsan, Korea. His research area includes channel coding, detection and estima- tion, cooperative communications, cognitive radio and sensor networks. He is a member of IEEK, KICS, KIPS, IEEE, and IEICE. 2010. [2] Z. Yonghong, L. Ying-chang, "Eigenvalue-based spec- trum sensing algorithms for cognitive radio," IEEE Transactions on Communications, vol. 57, no. 6, pp. 1784-1793, 2009. [3] S. Peche, "Universality results for the largest eigen- values of some sample covariance matrix ensem- bles," Probability Theory and Related Fields, vol. 143, no. 3, pp. 481-516, 2009. [4] O. Feldheim, S. Sodin, "A universality result for the smallest eigenvalues of certain sample covariance ma- trices," Geometric and Functional Analysis, vol. 20, no. 1, pp. 88-123, 2010. [5] N. M. Faber, L. M. C. Buydens, and G. Kateman, "Aspects of pseudorank estimation methods based on the eigenvalues of principal component analysis of random matrices," Chemometrics and Intelligent La- boratory Systems, vol. 25, no. 2, pp. 203-226, 1994. [6] D. Feral, S. Peche, "The largest eigenvalues of sam- ple covariance matrices for a spiked population: Dia- gonal case," Journal of Mathematical Physics, vol. 50, no. 7, pp. 073302-073333, 2009. [7] F. Penna, R. Garello, and M. A. Spirito, "Cooperati- ve spectrum sensing based on the limiting eigenval- ue ratio distribution in Wishart matrices," IEEE Co- mmunications Letters, vol. 13, no. 7, pp. 507-509, 2009. [8] S. Kritchman, B. Nadler, "Non-parametric detection of the number of signals: Hypothesis testing and random matrix theory," IEEE Transactions on Sig- nal Processing, vol. 57, no. 10, pp. 3930-3941, 2009. [9] I. M. Johnstone, "On the distribution of the largest eigenvalue in principal components analysis," An- nals of Statistics, vol. 29, 2001. [10] F. Penna, R. Garello, "Theoretical performance ana- lysis of eigenvalue-based detection," in Preprint, 2009. [11] R. Bhatia, Matrix Analysis, Springer, 1997.