You are on page 1of 13

3670 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO.

7, JULY 2012

On the Performance of Covariance Based Spectrum


Sensing for Cognitive Radio
Ming Jin, Member, IEEE, Youming Li, and Heung-Gyoon Ryu, Member, IEEE

Abstract—In this paper, the distribution of the test-statistic for scaled identity matrix. However, in some applications, for ex-
the covariance based detection (CBD) is studied in order to obtain ample, in multiple antenna case, the noise power among an-
the mathematical expressions of false alarm probability and tennas may not be coincident after some array calibration. Then
detection probability . The formulation of decision threshold
for any is also presented. The expression of allows to the performance of these detections will degrade seriously. The
evaluate the performance of the CBD technology. Finally, the de- CBD, as well as the STBD, the GLRT and the EBD, exploits
cision threshold, as well as the derived and , is verified with time and/or spatial correlation of primary user signals without
Monte-Carlo simulations. requiring prior information of noise power. Hence, they are
Index Terms—Cognitive radio, covariance based detection, per- nonsensitive to noise uncertainty which always exists in prac-
formance analysis, spectrum sensing. tice [20], [21]. In contrast with the STBD and the EBD, the
CBD maintains high performance even when there is non-co-
incidence of noise power among antennas. It should be noticed
I. INTRODUCTION that the STBD and the EBD have relatively higher complexity

C OGNITIVE RADIO (CR) [1]–[3] is identified as one of than the CBD due to the determinant and the eigenvalue decom-
the promising technologies for alleviating the problem of position, respectively [29].
wireless spectrum resource shortage which is caused by current The expression for the decision threshold of the EBD tech-
fixed spectrum allocation policies. In cognitive radio networks, nology has been analyzed based on random matrix theory. It
in order to prob available spectrum resources and to avoid un- is mentioned in [22] and [23] that the largest and the smallest
acceptable interference to primary users, one of the most impor- eigenvalues of Wishart matrices are approximated to determin-
tant tasks is spectrum sensing [4]. istic values. With large sample number of the received signal,
Several spectrum sensing algorithms such as the matched under the largest eigenvalue follows Tracy-Widom distri-
filtering detection (MFD) [5], [6], the energy detection (ED) bution [24]. Based on the theorems in [22]–[24], the asymptotic
[7]–[9], [34], the cyclostationary detection (CSD) [10]–[12], the decision threshold is given in [15]. The exact decision threshold
sphericity test based detection (STBD) [35]–[38], the general- is derived in [25] by exploiting the expressions of the joint distri-
ized likelihood ratio test (GLRT) [42], the eigenvalue based de- butions of an arbitrary subset of ordered eigenvalues in Wishart
tection (EBD) [13]–[15], [39] and the covariance based detec- matrices [26]. In addition, Kortun et al. obtain the non-asymp-
tion (CBD) [16], [17] have been proposed in the literature. Pros totic analytical decision threshold using the distribution of the
and cons for most of these detectors are elaborated in many condition number of Wishart matrices in [27]. Besides false
studies (e.g., see [18] and [19]). Both the STBD and the GLRT, alarm probability , which is used to set decision threshold,
as well as the EBD, implement spectrum sensing via deter- [28] gives the detection probability for the EBD algorithm
mining whether the eigenvalues of the population covariance using the theory of spiked population models. The theoretical
matrix from received signal are all equal to each other, i.e., performance of the STBD have also been derived recently [38].
they determine whether the population covariance matrix is a To the best of our knowledge, there is only one theoretical
analysis [17] for the CBD algorithm. And because of some ap-
proximations, as the authors of [17] claimed, the theoretical re-
Manuscript received August 14, 2011; revised December 19, 2011 and De-
cember 19, 2011; accepted April 04, 2012. Date of publication April 16, 2012; sults do not exactly match the simulated results even when the
date of current version June 12, 2012. The associate editor coordinating the re- number of samples is 50 000. In this paper, we derive the distri-
view of this manuscript and approving it for publication was Prof. Francesco
bution of the test-statistic of the CBD technology to obtain the
Verde. This work was partially supported by the National Science Foundation
of China under Grant 61071119, the Natural Science Foundation of Zhejiang expressions of and . The expression of decision threshold
Province under Grants Y1110657 and Y1091155, the Scientific and Techno- for any is also derived. These analytical results match Monte
logical Innovations Teams of Zhejiang Province under Grant 2010R50009, the
Carlo simulation results well. It should be mentioned that al-
Natural Science Foundation of Ningbo Municipality under Grant 2011A610184,
and the K. C. Wong Magna Fund in Ningbo University. though the theoretical results are derived based on identity noise
M. Jin and Y. Li are with the College of Information Science and Engineering, power, the numerical examples given in Section VI show that
Ningbo University, 315211 Ningbo, China (e-mail: {jinming@nbu.edu.cn; liy-
the analytical results still match the simulation results as long
ouming@nbu.edu.cn).
H.-G. Ryu is with the Department of Electrical, Electronic and Computer as the nonidentity of noise power is not very serious.
Engineering, Chungbuk National University, 361-763 Cheongju, Republic of It is noted that the analytical expressions of CBD for com-
Korea (e-mail: ecomm@cbu.ac.kr).
plex signal case are difficult to be obtained. In complex signal
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. case, we implement spectrum sensing by using improved
Digital Object Identifier 10.1109/TSP.2012.2194708 CBD (iCBD) [41], which converts the complex signal problem

1053-587X/$31.00 © 2012 IEEE


JIN et al.: ON THE PERFORMANCE OF COVARIANCE BASED SPECTRUM SENSING FOR COGNITIVE RADIO 3671

to be a real signal problem and has a better detection per-


formance over CBD. The analytical expressions for iCBD in
complex signal case are obtained by extending the expressions
for CBD in real signal case.
The rest of this paper is organized as follows. Section II
introduces the signal model and reviews the covariance based
detection, also gives description of iCBD. Section III derives
the analytical results for the detection performance and some
statistical parameters required in the analytical results are given
in Section IV. After discussing the decision threshold setting
issue in Section V, we present numerical results in Section VI
to verify the analysis. And some conclusions are made in
Section VII.
Notational Remark: Upper-case boldface letters denote ma-
trices hereafter, and lower-case boldface letters indicate vectors.
The superscript , , and denote the conjugate, the Fig. 1. Cross-correlation coefficient of primary signal over two antennas with
versus angular spread.
transpose and the conjugate transpose operators, respectively.
represents expectation operator, is the corresponding
variance, is the magnitude operator, denotes Kronecker
product, and represents convolution operator. The symbols is the primary signal power on each receiving antenna, and
and represent the real and imaginary parts of , is spatial auto- or cross-correlation
respectively. coefficient between the primary signals from the th and the
th antennas. The time domain cross-correlation coefficient
II. COVARIANCE BASED DETECTION of the primary signal from one antenna is assumed to be
with . In most
A. Signal Model and Covariance Based Detection practical cases, the real and imaginary parts of communication
Either time or spatial correlation (or both of them) can be ex- signals are approximately independent of each other. Here, it is
ploited by the CBD technology for sensing the primary users. assumed that the real and imaginary parts of are indepen-
Time correlation is guaranteed for several reasons [17]: 1) over- dent of each other. Hence, both and are real number.
sampled signal; 2) propagation channel with time dispersion; In addition, it is assumed that and are independent
and 3) correlated primary signal. While closely placed antennas of each other.
are deployed, there is always spatial (or antenna) correlation es- The time correlation coefficient has been described in detail
pecially in low population density areas, typical of rural envi- in [17]. The spatial cross-correlation coefficient between the th
ronments [30]. and the th antennas is [40]
Consider a cognitive user with antennas being deployed.
Let be the discrete-time received signal from the th (2)
antenna. With the binary hypotheses, the received signal from
the antennas can be described in vector form by the following
where with being angular spread,
expression [31]:
and represent spacing between the antennas and wave-
length of carrier, respectively. It is shown in Fig. 1 that the
(1) spatial cross-correlation coefficient of primary signal over
two antennas versus angular spread under the constraint of
where represents the absence of primary users, i.e., the . It shows that the spatial cross-correlation coef-
received signal con- ficient reaches peak (i.e., ) when the angular spread
tains only independent and identically distributed (i.i.d.) is 0 rad. With the increasing of the angular spread, the spatial
additive zero-mean white circularly symmetric complex cross-correlation coefficient decreases. However, even the
Gaussian noise vector , angular spread increases to 1.5 rad, the spatial cross-correlation
and with and being the coefficient is still over 0.75. This illuminates that the spatial
noise power and an identify matrix of size , respec- cross-correlation coefficient is high while smart antenna is
tively. represents the presence of primary users, i.e., deployed with small spacing between antennas.
consists of primary user signal corrupted by noise. Considering consecutive time samples from the an-
denotes the received tennas, the recorded signal is written in vector form as
signal, and , where
denotes the primary signal from azimuthal angle and
represents its corresponding impulse response of channel on (3)
..
the th antenna. It is assumed that , with .
being the th element of . The term
3672 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 7, JULY 2012

When , it gives where


which means only time correlation can be ex-
ploited. When , implies one can exploit
(11)
spatial correlation.
The statistical covariance matrix of the received signal is de-
fined as and

(4)
(12)
Then
Considering that the real part of primary signal is indepen-
(5) dent of its imaginary part in complex signal case, an improved
version of the CBD has been given in [41]. The iCBD utilizes
where [17] only the real part of , i.e., , where

.. .. .. .. (6) (13)
. . . .
Then the test-statistic of the improved CBD is given by
and
(14)

.. .. .. .. (7) which converts the complex signal problem to be a real signal


. . . . problem.
It has be shown in [41] that better detection performance can
be achieved by using (14) than (10) in complex signal case. And
It should be noted that and in both real
in real signal case, the test-statistics in (14) and (10) become the
and complex signal cases. And , .
same.
Hence, is a multivariate random vector with the gen-
In [41], we have derived the decision threshold expression of
eral statistical properties
the improved CBD for only two antennas case. In this paper,
both the decision threshold (or false alarm probability, ) and
(8) detection probability expressions are obtained for any an-
. tennas case.
One can select and to exploit both spatial
B. Sample Covariance Matrix
and time correlation in multiple antenna case. Also, one can ex-
ploit time correlation by setting and . To simplify In practice, the statistical covariance matrix is estimated
the notion, in the rest of this paper, we consider the CBD ex- through a sample covariance matrix. Introduce as the number
ploiting only spatial correlation in multiple antenna case, i.e., of samples collected by each receiver during the sensing period.
we set and . Hence, In real signal case and , the received signal
is a real vector. Then the sample covariance matrix
(9) is defined as
.

It will show by simulation in Section VI that the obtained (15)


analytical results are still good also in case of time correlated
received signals.
While in complex signal case, the sample covariance
If primary user is not present, then . Hence, the off-
matrix is defined as
diagonal elements of are all zeros. Otherwise, because of the
spatial-time correlation of primary signals, is not a diagonal
matrix. Therefore, the ratio of the non-diagonal elements to the
diagonal elements of can be used to detect the presence of
primary users. With these observations, the test-statistic of CBD (16)
is given by [17] Then, the test-statistic of (improved) CBD is given by

(10) (17)
JIN et al.: ON THE PERFORMANCE OF COVARIANCE BASED SPECTRUM SENSING FOR COGNITIVE RADIO 3673

where Then we can write as

(18) (26)

and where , ,

and .
(19)
Let , it can be easily verified that

Based on the description above, we can get that once the de- (27)
tection performance expressions of CBD for real signal case
based on in (15) is derived, the results can be easily ex- Hence, is independent of . So we have
tended to complex signal case for iCBD based on in (16).
I.e., as for iCBD in the complex signal case, the analytical ex-
(28)
pressions can be obtained by just replacing , , with ,
and , respectively, in the expressions for the real signal
case. In the following, we derive the analytical expressions of Furthermore, (28) has the same distribution as
CBD based on in (15).
Denoting with the decision threshold of CBD, such that we (29)
can make decision as
Combining (28) and (29) gives
decision (20)
.
(30)
The and are given by

(21) Hence, we have

and

(22)

respectively.
These probabilities depend on the distribution of under the
two hypotheses. Thus, (21) and (22) can be written as

(23)

and

(24)

where and are the cumulative distribution


functions (CDFs) of under and hypotheses, respec-
tively.
If the CDFs of can be obtained with analytical expressions,
one can set the decision threshold for any given , and easily
evaluate the detection performance. In the next sections, the dis- (31)
tribution of is derived and the formulae of detection perfor-
mance and decision threshold are obtained. It can be easily obtained that the cross-correlation coefficient
of and is
III. DISTRIBUTION OF
(32)
With central limit theorem, and approach Gaussian dis-
tributions [17]. Assumed that and have the means
and , and the variances and . Their cross-correla- Then we can achieve
tion coefficient is . Let and be independent standard
normal variables, i.e.,

(25)
(33)
3674 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 7, JULY 2012

where is 2D Gaussian -function with Lemma 2: If , we have

(34)

and (45)

(35)
Proof: follows normal distribution, so follows
(36) folded normal distribution [32]. Hence, (45) can be easily ob-
tained with the result in [32].
With Lemmas 1 and 2, the following theorem is given.
When is large enough that we can ignore the possibility Theorem 1:
of , (33) can be reduced to
(46)

with given in (45), and


(37)
(47)
where

(38)
Lemma 3: We have the following cross-correlation coeffi-
cients:
Both (33) and (37) show that we can obtain the probability
of as long as we have the statistical parameters of ,
(48)
, , and . In the next section, we will derive these
parameters.
(49)
IV. STATISTICAL PARAMETERS OF
In this section, we will derive the statistical parameters of .
First of all, these parameters are derived under hypothesis. (50)
The parameters can be forwardly obtained under hypothesis
by setting . Let . We have the following
lemmas. (51)
Lemma 1:

with , , , and being different from each other.


Proof: See Appendix A.
(39) Lemma 4: This result can be found in [33]. If
and . The cross-correlation coeffi-
(40) cient of and is . Then we have

Proof: It can be easily obtained that (52)

where
(41)

(42)
(43)

(44)

Based on the central limit theorem, we have (39) and (40). (53)
JIN et al.: ON THE PERFORMANCE OF COVARIANCE BASED SPECTRUM SENSING FOR COGNITIVE RADIO 3675

and Theorem 3: The cross-correlation coefficient of and is

(61)

where , , , and are given in Lemma 5,


Theorems 1 and 2.
(54)
V. THE DECISION THRESHOLD
(55) In this Section, we derive the expression of the decision
threshold for any given . Exploiting the results in Section III
and Section IV and setting , we can obtain the following
results:

(62)
(56) (63)

(64)
Theorem 2: The variances of and are given by
(65)

(66)
(57)
and

and
(67)
(58)

which is independent of the noise power .


respectively, with being different from each other. And
Here . In practice, the sample number is
and can be calculated using Lemmas 1,
3, and 4. always larger than the antenna number . So it is reasonable
Proof: See Appendix B. to assume that is large enough that we can use (37) to
Lemma 5: elevate the .
We have

(59) (68)
Proof: It can be easily obtained that

Combining (37) and (68), we can obtain the decision threshold


for any via finding the solution of

(69)

where is the inverse function of .


(60) The (69) can be transformed to a quadratic equation of that

(70)
3676 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 7, JULY 2012

where primary signals from different antenna elements have the same
gain, one should calibrate the channels via applying weighting
(71) factors which will produce the nonidentity noise power among
antenna elements. The noise power on the th antenna after
calibration is assumed to be , and

(76)

where is a random variable which has a uniform distribu-


tion between 0.5 and 0.5. means that the noise power
among antennas are identity. Otherwise, the antennas have non-
(72) identity noise power.

and A. Versus
Fig. 2 shows the of CBD for fixed and different
in real signal case. Fig. 2(a)–(c) is obtained with for
different , and Fig. 2(d)–(f) is obtained by increasing to
(73) 1000. Fig. 2 shows that the analytical given in [17] does
not fit the empirical results well. While the proposed analytical
with given in (67). matches the empirical results quite well even when there is
Considering that , the decision non-coincidence of noise power among antennas.
threshold is given by With different values of , the curves are given in
Fig. 3(a)–(c) and Fig. 3(d)–(f) for and ,
respectively. In Fig. 3, only the proposed analytical and the
(74)
empirical results are given, for considering that the analytical
for and of [17] does not fit the empirical results well. Fig. 3 illu-
minates that the proposed analytical is consistent with the
empirical results while is large enough.
(75)
In complex signal case, the numerical results are given in
for . Figs. 4 and 5. Considering that the analytical expressions of
The decision threshold is independent of the noise power . CBD for complex signal case are difficult to be obtained, the
Hence, the CBD is nonsensitive to noise uncertainty. results of iCBD for complex signal is given. In addition, the
results of STBD are also considered. Fig. 4 shows the analyt-
VI. NUMERICAL RESULTS ical and empirical versus for different with given
and . It indicates that the analytical of both iCBD and
In this section, the obtained analytical expressions of the de- STBD match their empirical results when , which means
cision threshold , and are validated by comparing them the noise power among antennas is equal. When , how-
with empirical results, obtained from Monte-Carlo simulations ever, the analytical of STBD will deviates its empirical re-
with independent trails. The decision threshold is calcu- sult. And the analytical of iCBD still matches its empirical
lated with (74) and (75), and and are calculated with result. Fig. 5 evaluates the convergency of the analytical of
(37). For simplification, here, we assume that the sensing array, iCBD. Fig. 5 illuminates that the analytical of iCBD conver-
whose antennas are placed in a line, is deployed. And the spatial gence to its empirical one even when .
correlation coefficients of the received signals between the th
and the th receiving antennas is . In the following B. Versus SNR
simulations, the empirical and analytical results of [17] and this
paper of CBD are compared in real signal case. While in com- In this subsection, we will investigate the detection proba-
plex signal case, the empirical result of CBD, the empirical and bility for different signal-to-noise ratio (SNR). The SNR is de-
analytical results of iCBD and STBD are compared. The analyt- fined as SNR . The and are selected as 4 and 1000,
ical expressions of iCBD are obtained by replacing , , respectively. The is set to 0.1, and the decision threshold is
with , and , respectively, in the analytical expres- calculated with (74) for the given . Fig. 6(a) and (b) gives the
sions of CBD. results in real and complex signal cases, respectively. In both
It should be mentioned that the test-statistic given in [17] cases, we set . Fig. 6(a) shows that the proposed ana-
is instead of . Hence, we should use the decision lytical for CBD matches its corresponding empirical results,
threshold instead of in (74) and (77) given in [17] while the analytical given in [17] does not. Partial enlarged
to calculate the and . drawing of Fig. 2(d) is also placed in Fig. 6(a). The partial en-
In practice, there is always amplitude and phase mismatch larged one shows that the decision threshold calculated with
between antenna elements [43]. To guarantee that the received [17] for produces far below 0.1 of actual . This
JIN et al.: ON THE PERFORMANCE OF COVARIANCE BASED SPECTRUM SENSING FOR COGNITIVE RADIO 3677

Fig. 2. False-alarm probability versus decision threshold for CBD in real signal case. Circle: empirical of CBD; Dash-dot line: analytical given in [17];
Solid line: the proposed analytical . (a) , ; (b) , ; (c) , ; (d) , ; (e) ,
; (f) , .

Fig. 3. False-alarm probability versus decision threshold for CBD in real signal case. Point-dash-dot line: empirical of CBD; Solid line: the proposed analytical
. (a) , ; (b) , ; (c) , ; (d) , ; (e) , ; (f) , .

results into the being far below 0.1 when SNR is very small the effect of nonidentity noise power on STBD will be given in
as shown in Fig. 6(a). the next subsection.
In the complex case, Fig. 6(b) shows that the proposed an-
alytical for iCBD also matches its corresponding empirical C. Receiver Operating Characteristics
results. In addition, iCBD has better detection performance than The receiver operating characteristics (ROC) curves of the
both CBD and STBD. Fig. 6(b) also shows that the detection CBD for real signal and iCBD, STBD for complex signal are
performance of STBD degrades seriously when there is non- shown in Fig. 7(a) and (b), respectively. The ROC curves of
identity noise power among antennas. We can see from Fig. 6(b) GLRT and EG (energy detection with exact noise level knowl-
that STBD is sensitive to nonidentity noise power among an- edge) are given also. The curves are obtained with ,
tennas, while both CBD and iCBD are not. The explanation for , and SNR 12 dB. The curves with
3678 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 7, JULY 2012

Fig. 4. False-alarm probability versus decision threshold for iCBD and STBD in complex signal case. Circle: empirical of iCBD; Solid line: the proposed
analytical of iCBD; Point-dash-dot line: empirical of STBD; Dash-dot line: analytical of STBD. (a) , ; (b) , ; (c)
, ; (d) , ; (e) , ; (f) , .

Fig. 5. False-alarm probability versus decision threshold for iCBD in complex signal case. Point-dash-dot line: empirical of iCBD; Solid line: the proposed
analytical of iCBD. (a) , ; (b) , ; (c) , ; (d) , ; (e) , ; (f) ,
.

, , and are presented. From Fig. 7, we sumption of identity noise power among antennas. Essentially,
get the same conclusions as in Fig. 6. In addition, Fig. 7 shows they make decision by determining whether the eigenvalues of
that GLRT has better performance over STBD, although there the population covariance matrix from received signal are all
is angular spread about 1.4 rad. With the knowledge of noise equal to each other. If the population covariance matrix is not
power, it is also shown that the EG has the best performance a scaled identity matrix (or the eigenvalues are not equal to
among these detectors. each other), they make decision of that there are primary users.
It is given here that why GLRT and STBD degrade when Hence, when the noise powers among antennas are not equal
as shown in Fig. 6(b) and Fig. 7(b). GLRT and STBD to each other, GLRT and STBD will make wrong decision in
implement spectrum sensing via determining whether the popu- primary use absence case. And then the performances of GLRT
lation covariance matrix is a scaled identity matrix with the as- and STBD degrade.
JIN et al.: ON THE PERFORMANCE OF COVARIANCE BASED SPECTRUM SENSING FOR COGNITIVE RADIO 3679

Fig. 6. Detection probability versus SNR with , and . (a) real signal case; (b) complex signal case. (a) SNR (dB); (b) SNR (dB).

Fig. 7. Detection probability versus false alarm probability with , , and SNR . (a) real signal case; (b) complex signal
case. In this figure, the same line specifications is used for each detector with different . (a) False alarm probability ; (b) false alarm probability .

Fig. 8. Detection probability versus false alarm probability with , , and SNR . (a) real signal case; (b) complex signal case.
(a) False alarm probability ; (b) false alarm probability .

D. Time Correlation Case and . The sampled primary user signal has
time correlation as , , and .
In this subsection, the time correlation case is considered. As- With SNR 12 dB, the ROC curves for real and complex
sume that only one antenna is deployed. We select signal cases are given in Fig. 8(a) and (b), respectively. Fig. 8
3680 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 7, JULY 2012

depicts that the proposed analytical expressions can be applied and


to time correlation case. Also, the GLRT and EG are compared.
As the same as the previous case, here GLRT is better than
STBD.

VII. CONCLUSION

In this paper, analytical formulae have been founded for


the distribution of the test-statistic of the CBD. Mathematical
expressions are derived for the false alarm probability, the
detection probability and the decision threshold for CBD in (80)
real signal case. These expressions can be extended to complex Hence, based on the Lemmas 1 and 2, the cross-correlation
signal case when using iCBD to implement spectrum sensing. coefficients are
Finally, these expressions are verified by the corresponding
Monte Carlo simulations. (81)

APPENDIX A
PROOF OF LEMMA 3
With being different from each other, the expecta- (82)
tions of , , and are, respectively,

(83)

and

(84)

(77)
APPENDIX B
PROOF OF THEOREM 2
It is easy to verify that
(85)
and
(86)
where are different from each other.
Hence, we have
(78)

(87)
(79)
JIN et al.: ON THE PERFORMANCE OF COVARIANCE BASED SPECTRUM SENSING FOR COGNITIVE RADIO 3681

and based on Lemma 1, we have [11] A. Tkachenko, D. Cabric, and R. W. Brodersan, “Cyclostationary fea-
ture detector experiments using reconfigurable BEE2,” in Proc. IEEE
Int. Symp. New Frontiers in Dynamic Spectrum Access Netw., Dublin,
Ireland, Apr. 2007, pp. 216–219.
[12] Y. Xiao and F. Hu, Cognitive Radio Networks. Boca Raton, FL:
(88) Taylor & Francis, CRC, 2009.
[13] Y. Zeng, C. L. Koh, and Y. C. Liang, “Maximum eigenvalue detection:
Combining (46), (87), and (88) gives (57). Theory and application,” in Proc. IEEE Int. Conf. Commun., May 2008,
pp. 4160–4164.
[14] M. Maida, J. Najim, P. Bianchi, and M. Debbah, “Performance anal-
ysis of some eigen-based hypothesis tests for collaborative sensing,”
in Proc. IEEE Workshop Datatistical Signal Process., Cardiff, U.K.,
2009, pp. 5–8.
[15] Y. Zeng and Y. C. Liang, “Eigenvalue-based spectrum sensing algo-
rithms for cognitive radio,” IEEE Trans. Commun., vol. 57, no. 6, pp.
1784–1793, Jun. 2009.
[16] Y. Zeng and Y. C. Liang, “Covariance based signal detections for cog-
nitive radio,” in Proc. IEEE DySPAN, Dublin, Ireland, Apr. 2007, pp.
202–207.
[17] Y. Zeng and Y. C. Liang, “Spectrum sensing algorithms for cognitive
radio based on statistical covariance,” IEEE Trans. Veh. Technol., vol.
58, no. 4, pp. 1804–1815, May 2009.
[18] Y. Zeng, Y. Liang, A. T. Hoang, and R. Zhang, “A review on spec-
trumsensing for cognitive radio: Challenges and solutions,” EURASIP
J. Adv. Signal Process., vol. 2010, p. 15, 10.1155/2010/381465, Article
ID 381465.
[19] T. Yücek and H. Arslan, “A survey of spectrum sensing algorithms for
cognitive radio applications,” IEEE Commun. Surv. Tutorials, vol. 11,
(89) no. 1, pp. 116–130, 2009.
[20] R. Tandra and A. Sahai, “Fundamental limits on detection in low SNR
under noise uncertainty,” in Proc. Int. Conf. Wireless Netw., Commun.,
Combining (47) and (89) gives (58). Mobile Comput., Jun. 2005, vol. 1, pp. 464–469.
[21] D. Cabric, A. Tkachenko, and R. W. Brodersen, “Experimental study of
spectrum sensing based on energy detection and network cooperation,”
ACKNOWLEDGMENT presented at the 1st Int. Workshop Technol. Policy Accessing Spectrum
(TAPAS), New York, Aug. 2006.
The authors would like to thank Dr. H. Li, Z. Zhang, and Dr. [22] Z. D. Bai, “On the distributuion of largest eigenvalue in principle com-
Y. Gao for their helpful suggestions. The authors are grateful ponents analysis,” Anal. Statist., vol. 29, no. 2, pp. 295–327, 2001.
to the anonymous reviewers for providing them with a large [23] Z. D. Bai, “Methodologies in spectral analysis of large dimensional
random matrices, a review,” Statist. Sinica, vol. 9, pp. 611–677, 1999.
number of detailed suggestions for improving the submitted [24] C. A. Tracy and H. Widom, “The distribution of the largest eigenvalue
manuscript. in the Gaussian ensembles,” in Calogero-Moser-Sutherland Models.
New York: Springer, 2000, pp. 641–472.
[25] F. Penna, R. Garello, D. Figliloli, and M. A. Spirito, “Exact non-asymp-
REFERENCES totic threshold for eigenvalue based spectrum sensing,” presented at
[1] J. Mitola, “Cognitive radio for flexible mobile multimedia communi- the 4th Int. Conf. Cognitive Radio Oriented Wireless Netw. Commun.,
cations,” in Proc. IEEE Int. Workshop Mobile Multimedia Commun. Hannover, Germany, Jun. 2009.
(MoMuC), San Diego, CA, Nov. 1999, pp. 3–10. [26] M. Chiani and A. Zanella, “Joint distribution of an arbitrary subset of
[2] S. Haykin, “Cognitive radio: Brain-empowered wireless communica- the ordered eigenvalues of Wishart matrices,” in Proc. IEEE 19th Int.
tions,” IEEE J. Sel. Areas Commun., vol. 23, no. 2, pp. 201–220, Feb. Symp. Pers., Indoor, Mobile Radio Commun. (PIMRC), Sept. 2008, pp.
2005. 1–6.
[3] S. J. Shellhammer, “Spectrum sensing in IEEE 802.22,” presented at [27] A. Kortun, T. Ratnarajah, M. Sellathurai, C. Zhong, and C. B. Papa-
the IAPR Workshop Cognitive Inf. Process., Santorini, Greece, Jun. dias, “On the performance of eigenvalue-based cooperative spectrum
2008. sensing for cognitive radio,” IEEE J. Sel. Topics Signal Process., vol.
[4] D. Cabric, S. M. Mishra, and R. W. Brodersen, “Implementation issues 5, no. 1, pp. 49–55, Feb. 2011.
in spectrum sensing for cognitive radio,” in Proc. Asilomar Conf. Sig- [28] F. Penna and R. Garello, “Theoretical performance analysis of eigen-
nals, Syst., Comput., 2004, pp. 772–776. value-based detection,” [Online]. Available: http://arxiv.org/PS_cache/
[5] A. Sahai and D. Cabric, “Spectrum sensing: Fundamental limits and arxiv/pdf/0907/0907.1523v2.pdf
practical challenges,” in Proc. IEEE Int. Symp. New Frontiers Dynamic [29] Y. Zeng, Y. Liang, E. Peh, and A. T. Hoang, “Cooperative covariance
Spectrum Access Netw. (DySPAN), Baltimore, MD, Nov. 2005. and eigenvalue based detections for robust sensing,” in Proc. IEEE
[6] H. S. Chen, W. Gao, and D. G. Daut, “Signature based spectrum Global Telecommun. Conf. (IEEE GLOBECOM), Nov. 2009, pp. 1–6.
sensing algorithms for IEEE 802.22 WRAN,” in Proc. IEEE Int. Conf. [30] R. M. Buehrer, “Generalized equations for spatial correlation for low
Commun. (ICC), Jun. 2007, pp. 6487–6492. to moderate angle-spread,” in Proc. 10th Annu. Virginia Tech Symp.
[7] A. Sahai, N. Hoven, and R. Tandra, “Some fundamental limits on Wireless Pers. Commun., Jul. 2000, pp. 70–77.
cognitive radio,” presented at the Allerton Conf. Commun., Control, [31] A. Taherpour, M. Nasiri-Kenari, and S. Gazor, “Multiple antenna spec-
Comput., Monticello, IL, Oct. 2004. trum sensing in cognitive radios,” IEEE Trans. Wireless Commun., vol.
[8] A. Ghasemi and E. S. Sousa, “Collaborative spectrum sensing for 9, no. 2, pp. 814–823, Feb. 2010.
opportunistic access in fading environments,” in Proc. IEEE Symp. [32] F. C. Leone, R. B. Nottingham, and L. S. Nelson, “The folded normal
New Frontiers Dynamic Spectrum Access Netw., Baltimore, MD, Nov. distribution,” Technometrics, vol. 3, no. 4, pp. 543–550, 1961.
2005, pp. 131–136. [33] M. Krumin and S. Shoham, “Generation of spike trains with controlled
[9] D. C. Oh and Y. H. Lee, “Energy detection based spectrum sensing for auto- and cross-correlation functions,” Neural Comput., vol. 21, no. 6,
sensing error minimization in cognitive radio networks,” in Proc. Int. pp. 1642–1664, 2009.
J. Commun. Netw. Inf. Security, Apr. 2009, vol. 1, no. 1, pp. 1–5. [34] E. H. Gismalla and E. Alsusa, “Performance analysis of the peri-
[10] M. Ghozzi, F. Markx, M. Dohler, and J. Palicot, “Cyclostationarity- odogram-based energy detector in fading channels,” IEEE Trans.
based signal detection of vacant frequency bands,” in Proc. IEEE Int. Signal Process., vol. 59, no. 8, pp. 3712–3721, Aug. 2011.
Conf. Cognitive Radio Oriented Wireless Netw. Commun. (Crowncom), [35] S. John, “Some optimal multivariate tests,” Biometrika, vol. 58, no. 1,
Mykonos Island, Greece, Jun. 2006, pp. 1–5. pp. 123–127, Apr. 1971.
3682 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 60, NO. 7, JULY 2012

[36] D. B. Williams and D. H. Johnson, “Using the sphericity test for source Youming Li received the B.S. degree in computa-
detection with narrow-band passive arrays,” IEEE Trans. Acoust., tional mathematics from Lanzhou University, China,
Speech, Signal Process., vol. 38, no. 1, pp. 2008–2014, Nov. 1990. in 1985, the M.S. degree in computationtional
[37] R. Zhang, T. J. Lim, Y. C. Liang, and Y. Zeng, “Multi-antenna based mathematics from Xi’an Jiaotong University, Xi’an,
spectrum sensing for cognitive radios: A GLRT approach,” IEEE China, in 1988, and the Ph.D. degree in electrical
Trans. Commun., vol. 58, no. 1, pp. 84–88, Jan. 2010. engineering from Xidian University, Xi’an, China,
[38] L. Wei and O. Tirkkonen, “Spectrum sensing in the presence of in 1995.
multiple primary users,” IEEE Trans. Commun., vol. 60, no. 5, pp. From 1988 to 1998, he worked in the Department
1268–1277, 2012. of Applied Mathematics, Xidian University, where he
[39] P. Wang, J. Fang, N. Han, and H. Li, “Multiantenna-assisted spectrum was an Associate Professor. From 1999 to 2004, he
sensing for cognitive radio,” IEEE Trans. Veh. Technol., vol. 59, no. 4, worked in the School of Electrical and Electronics
pp. 1791–1800, May 2010. Engineering, Nanyang Technological University; DSO National Laboratories,
[40] G. D. Durgin and T. S. Rappaport, “Effects of multipath angular spread Singapore; and the School of Engineering, Bar-Ilan University, Israel, respec-
on the spatial cross-correlation of received voltage envelopes,” in Proc. tively. Since 2005, he has been with Ningbo University, Ningbo, China, where
IEEE 49th Conf. Veh. Technol., 1999, vol. 2, pp. 996–1000. he is currently a Professor. His research interests are in cognitive radio, and
[41] M. Jin, Y. Li, Z. Zhang, and R. Wang, “A new spectrum sensing algo- wireless communications.
rithm based on antenna correlation for cognitive radio systems,” Wire-
less Pers. Commun. May 2011, 10.1007/s11277-011-0349-9 [Online].
Available: http://www.springerlink.com/content/l7h500667j677955/
fulltext.pdf Heung-Gyoon Ryu (M’88) was born in Seoul,
[42] P. Bianchi, M. Debbah, M. Maida, and J. Najim, “Performance of sta- Republic of Korea, in 1959. He received the B.S.,
tistical tests for single-source detection using random matrix theory,” M.S., and Ph.D. degrees in electronic engineering
IEEE Trans. Inf. Theory, vol. 57, no. 4, pp. 2400–2419, Apr. 2011. from Seoul National University, Seoul, Korea, in
[43] K. R. Dandekar, H. Ling, and G. Xu, “Smart antenna array calibration 1982, 1984, and 1989.
procedure including amplitude and phase mismatch and mutual cou- Since 1988, he has been with Chungbuk National
pling effects,” in Proc. IEEE Int. Conf. Pers. Wireless Commun., 2000, University, Korea, where he is currently Professor of
pp. 293–297. the Department of Electrical, Electronic and Com-
puter Engineering in Chungbuk National University,
Korea. He worked as Chief of RICIC (the Research
Institute of Computer, Information Communication
Ming Jin (M’12) received the B.S. and Ph.D. de- Center) in Chungbuk National University from March 2002 to February 2004.
grees in electrical engineering from Xidian Univer- His main research interests are digital communication systems, communication
sity, Xi’an, China, in 2005 and 2010, respectively. circuit design, spread spectrum system, and communication signal processing.
He is currently working in the College of Infor- Since 1999, he has worked as reviewer of the IEEE transaction papers. He
mation Science and Engineering, Ningbo University, was a winner of the 2002 Academy Award from the Korea Electromagnetic En-
Zhejiang, China. His current research interests in- gineering Society, Korea. He received the Best Paper Award at the Fourth Inter-
clude cognitive radio and smart antenna techniques. national Conference on Wireless Mobile Communications (ICWMC) Athens,
Greece, July 27–August 1, 2008. He also received the Best Paper Award at the
International Conference on Advances in Satellite and Space Communications
(SPACOMM), Colmar France, July 20–25, 2009.

You might also like