Cooperative Spectrum Sensing of OFDM Signals
Using Largest Eigenvalue Distributions
Lu Wei and Olav Tirkkonen
Department of Communications and Networking
Helsinki University of Technology (TKK)
P.O. Box 3000, FIN-02015 TKK, Finland
Email: {Lu.Wei, Olav.Tirkkonen}@tkk.fi
Abstract—Spectrum sensing is a key component in cognitive
radio networks. Recently there has been intense research interest
in eigenvalue based sensing. The results presented so far rely
on the distributions of infinite dimensional covariance matrices,
therefore these analyses are not accurate for a small sample size.
In this paper, we propose a new spectrum sensing method based
on the distribution of the largest eigenvalue of the covariance
matrices. Using distribution functions for finite dimensional
matrices, we conduct exact analysis on the performance of the
proposed detector. Essentially, the detection problem requires
characterizing the decision threshold as a function of various
parameters. The threshold optimization problem is characterized
by a weighted sum of false alarm and miss detection probabilities.
This detector outperforms the cooperative energy detector with
all sample sizes and in the whole SNR range considered. Our
proposed detection scheme has direct application in OFDM
systems.
I. I NTRODUCTION
Cognitive radio (CR) is a promising technique for future
wireless communication systems. In CR networks, dynamic
spectrum allocation of the frequency band is implemented to
mitigate spectrum scarcity. In dynamic spectrum allocation, a
secondary (unlicensed) user is allowed to utilize the spectrum
resources when it does not cause intolerable interference to
the primary (licensed) user. The level of interference depends
mainly on the secondary user’s ability to detect the primary
user. Therefore it becomes a critical issue in CR network that
the secondary user will make a quick and reliable decision
based on spectrum sensing.
Several spectrum sensing methods exist in the literature.
Energy detection has been considered in [1], [2], [3], [4],
matched filter detection in [2], [5], cyclostationary feature de-
tection in [6] and recently emerged eigenvalue-based detection
in [7], [8], [9]. Each of these techniques has its strengths
and weaknesses. For example, matched filter detection and
cyclostationary feature detection require knowledge on the
wave form of the primary user, which is impractical for certain
applications. Energy detection and eigenvalue-based detection
belong to the so-called blind detection methods which do not
need any a priori information of the signal. However, energy
detection suffers from the noise-uncertainty problem [3], in
which inaccurate estimation of noise power will lead to
the SNR wall [10]. The newly proposed eigenvalue-based
detection algorithms circumvent the need of knowledge about
noise power. Therefore they outperform energy detection in the
978-1-4244-5213-4/09/ $26.00 ©2009 IEEE 2295
case of noise uncertainty [7]. The eigenvalue-based detection
schemes proposed so far are based on asymptotical analysis.
Specifically, these schemes rely on the distribution of the ratio
between the largest and smallest eigenvalues of the received
covariance matrix. This distribution has only been derived
for the case when the number of received samples tend to
infinity [9]. An exact analysis of eigenvalue-based detection,
however, requires the knowledge of the ratio distribution of
finite dimensional covariance matrices, which is generically
mathematically intractable.
In this paper, we propose a new detection scheme based
on the largest eigenvalue distributions of covariance matrices.
Utilizing existing closed-form distributions [11], [12], we
perform exact analysis of this new detector. The analysis is
carried out in the setting where there is only one primary
user transmitting an OFDM signal. However, the applicability
of this new detector is not limited by these assumptions;
the analysis is valid as long as the samples of the primary
user’s signal are drawn from a Gaussian process. The detection
problem is basically a hypothesis test between two possible
hypotheses. By our assumptions, the covariance matrix dis-
tributions under both hypotheses can be formulated as central
and non-central Wishart distributions respectively. Empirically
we found that the largest eigenvalue calculated from these
covariance matrices can be considered as a test statistics to
discriminate between the two hypotheses. It turns out that
this test statistics is an efficient decision making quantity in
terms of sample size and SNR requirement. Based on this
result we propose a new detector where the largest eigenvalue
is compared to a pre-calculated threshold. The threshold is
obtained by minimizing a weighted sum of the false alarm
and miss detection probabilities. The weighting coefficients
can be considered as design parameters. The proposed detector
outperforms the conventional energy detector with all sample
sizes and in the full SNR range considered. The superior
performance is due to choosing the largest eigenvalue as the
decision variable. As a result we obtain a less heavy tailed
distribution of decision statistics compared with that of the
energy detector.
Finally, we would like to note that the largest based de-
tection has been proposed independently in [13]. In [13],
limiting law of the largest eigenvalue distribution (Tracy-
Widom distribution) is utilized to set a decision threshold.
 TABLE I
The threshold can be calculated for a given target false alarm S AMPLE MEAN OF THE ORDERED EIGENVALUES UNDER BOTH
probability. Their results characterize the detector performance HYPOTHESES .
for larger number of samples and cooperative sensors. In
K = 4, N = 30 E[λ1 ] E[λ2 ] E[λ3 ] E[λ4 ]
this paper, we perform non-asymptotic exact analysis on the H1 (SNR=-5dB) 73.215 38.385 27.523 18.820
largest eigenvalue based detection. This result is valid for H0 44.768 33.265 24.747 17.233
arbitrary number of samples and sensors. Besides, the decision
threshold is obtained by considering both false alarm and
miss detection probabilities. We also notice that making use where IK denotes the identity matrix of dimension K. Under
of the statistics of the largest eigenvalue as opposed to the hypothesis H1 , by our assumptions the covariance R follows
statistics of the sum of all eigenvalues is not new [14]. Again, the complex non-central Wishart distribution, denoted as [15]
compared with this result, our work might be considered as
the corresponding finite dimensional counterpart. R ∼ WK (N, IK , MMH ). (5)
The rest of this paper is organized as follows. In section II, Here MMH is called non-centrality parameter matrix with
we formulate the primary user detection problem after M = hk (s(n) )H . The complex vectors hk = [h1 , h2 , . . . , hK ]
introducing the signal model. Section III presents the and s(n) = [s(1) , s(2) , . . . , s(N ) ] are column vectors. From the
comparison of detector performance. Finally in section IV we assumptions it follows that the matrix M is rank one, because
conclude the main results of this paper and point out possible
further research directions. Rank(M) = Rank(MMH ) (6)
= Rank(hk (s(n) )H s(n) hH
k ) (7)
= Rank(||s||2 hk hH
k ) = 1. (8)
II. P ROBLEM F ORMULATION
A. Signal Model In other words, the Hermitian matrix MMH has only one
non-zero eigenvalue, which we denote by φ1 . Without loss of
Consider a primary signal detection problem with K col- generality we assume that the variance of the complex noise
laborating secondary base stations. Each base station collects equals 1 (σn2 = 1). The average SNR under hypothesis H1 is
N samples during the sensing time. The K base stations will defined as
collaborate in a way that they together form the received data σ2 σ2
SNR = s 2 h , (9)
matrix using all the samples collected. The received K × N σn
data matrix Y is represented as
where σs2 can be estimated by ||s||2 /N .
⎛ (1) (2) (N )
⎞ We want to discriminate between the two hypotheses based
y1 y1 ... y1
⎜ (1) (2) (N ) ⎟ on the eigenvalues λ1 ≥ λ2 ≥ . . . ≥ λK of the observed
⎜ y2 y2 ... y2 ⎟ covariance matrix R. The fact that M is of rank one leads
Y=⎜
⎜ .. .. .. .. ⎟.
⎟ (1)
⎝ . . . . ⎠ to a major difference on the numerical value of the largest
(1)
yK
(2)
yK ...
(N )
yK eigenvalue λ1 , but the impact on other eigenvalues is much
smaller. This fact is illustrated in Table I, which shows sample
The primary user detection problem is a hypothesis test mean of the ordered eigenvalues of the covariance matrix R
between the following two hypotheses: under both hypotheses, where we set the parameters K = 4,
N = 30 and SNR = −5 dB.
(n) (n)
H0 : yk = nk (2) From Table I we can see that the largest eigenvalue λ1
H1 :
(n)
yk =
(n) (n)
hk sk +
(n)
nk , (3) provides a most prominent quantity to discriminate H0 from
H1 . Specifically, the difference between the mean value of the
where k = 1, . . . , K and n = 1, . . . , N . The received largest eigenvalues under both hypothesis can be as large as
covariance matrix R is defined as R = YYH , where H 28.447, whereas the difference between the mean value of the
denotes the Hermitian conjugate operator. Throughout this smallest eigenvalues is only 1.587. Based on this property, we
paper, we assume the following two basic assumptions. propose a new detector by considering the largest eigenvalue
Assumption 1: there is at most one primary user transmit- of the covariance matrix R as the decision statistics. We will
ting, and its signal amplitude is drawn independently from a show that this detector can efficiently, in terms of sample size
Gaussian process every sample. and SNR required, infer the presence or absence of the primary
Assumption 2: there is no fading in the temporal domain, user by looking at the largest eigenvalue of matrix R only. The
i.e. the channel stays constant during the sensing time. detailed detection process will be presented in the following
As a result of the above assumptions, under hypothesis H0 , subsection.
the received covariance matrix R follows the complex central
B. The Detection Problem
Wishart distribution, denoted as [15]
Computable closed-form expressions for the largest eigen-
R ∼ WK (N, IK ), (4) value distributions of central and non-central (non-central

2296
matrix M being rank one) Wishart matrices exist in the litera-
ture [11], [12]. The cumulative distribution function (CDF) of
the largest eigenvalue of the central Wishart matrix R under
hypothesis H0 is [11]
Fc (x) := Fc (λ1 < x) (10)
det A
,
= (11)

K
Γ(N − k + 1)Γ(K − k + 1)
k=1
where det(·) denotes the matrix determinant operator and
Γ(·) is the Gamma function. The i, j th entry of matrix
A (K × xK) is given by the incomplete Gamma function,
Ai,j = 0 αN −K+i+j−2 e−α dα.
When matrix M is rank one, the CDF of the largest eigen-
value of the non-central Wishart matrix R under hypothesis
H1 is [12]
Fnc (x) := Fnc (λ1 < x) (12)
−φ1
e det B
= (13),

K−1
φ1K−1 Γ(N − K + 1) Γ(N − k)Γ(K − k)
k=1
x N −ithe−β first column of matrix B is Bi,1
where =
0
β e 0 F 1 (N − K + 1; φ 1 β)dβ
xand the column from
the second to the Kth is Bi,j = 0 β N +K−i−j e−β dβ,
j = 2, . . . , N . Here, 0 F1 (·; ·) is the hypergeometric function
of Bessel type. Recall that φ1 is the only non-zero eigenvalue
of the Hermitian matrix MMH .
Based on the distribution functions quoted above, we can
calculate the optimal decision threshold analytically. The
threshold is utilized for making decision between the two
hypotheses. This threshold is obtained through an optimiza-
tion process such that a weighted sum of the false alarm
probability Pf a and the miss detection probability Pm is
minimized. In our case, the false alarm probability Pf a (r)
equals 1 − Fc (r) and the miss detection probability Pm (r)
equals Fnc (r), therefore the probability of correct detection
Pd (r) is 1−Fnc (r). Mathematically the threshold optimization
problem is formulated as
γopt = arg min C0 Pf a (r) + C1 Pm (r), (14)
γ≥0
where C0 and C1 are called weighting coefficients or design
parameters which we can choose. Generally we would like to
choose C0 ≤ C1 in order to intentionally set the miss detection
probability Pm lower. Equivalently, we make it less probable
that the secondary user will commence transmission when the
primary user is present. In this way, the total interference to
the primary user is smaller compared to the case with equal
weighting (C0 = C1 = 1). One example of the optimization
problem is illustrated in Fig. 1 where the areas S0 and S1
denote the false alarm probability and the miss detection
probability respectively. In Fig. 1 we have chosen K = 4,
N = 20 and SNR = −5 dB. The optimal threshold γopt is
found at 38.71 with the coefficients C0 = 1, C1 = 1.5.
2297
0.15
Largest Eigenvalue Distributions Under Two Hypotheses
H0
0.1 H1 (SNR = −5dB)
Probality
0.05
S1
S0
0
20 30 40 50 60 70 80
X: 38.71
x
Fig. 1. Threshold optimization result with parameters: K = 4, N = 20,
C0 = 1, C1 = 1.5. The optimal threshold is found at x=38.71.
Assuming a worst case signal-to-noise ratio, a performance
bound can be calculated based on the analysis above. This
worst case SNR value can e.g. be set based on the receiver
sensitivity of the primary OFDM system of interest [4]. Notice
that the actual detection performance could be better or worse
depending on the validity of the assumed worst case SNR
value. Using this SNR value the optimal threshold could be
calculated from (14).
Once we have the threshold value, the detection procedure is
as follows. Firstly, the K cooperative secondary base stations
form the K × N received data matrix Y as in (1). Secondly,
the largest eigenvalue of the covariance matrix R = YYH
is calculated, which is denoted by λ1 . Finally, we make a
decision. If λ1 is larger than γopt obtained from (14) it is
decided that the primary user is present and if λ1 is smaller
than γopt it is decided that the primary user is absent. In the
next section, we will show that the detection performance
of this largest eigenvalue based detector is better than the
conventional energy detector. Heuristic reasons behind the
superior performance will be discussed there as well.
III. D ETECTOR P ERFORMANCE
In general the detection performance is a function of the
SNR and the sample size. Specifically, as the SNR value
and the number of samples increase we would expect to
see increased probability of correct detection. In this section,
we will show the impact of both SNR and sample size on
the performance of our proposed detector. As a comparison,
we also show the detection performance of a conventional
cooperative energy detector.
The cooperative energy detector will collaborate in the same
way as the proposed detector in that the decision statistics is a
function of the collaboratively formed the received data matrix
Y. In the case of the conventional energy detector the test
statistics z is based on the norm of received data matrix Y,
i.e. the Frobenius norm z = ||Y||2F . Therefore the decision
 Gaussian Approximation of Distribution in eq.(18) Detector Performance (SNR = −5dB)
0.04 1
Gaussian Approximation
0.98
0.035 Exact Distribution, eq.(18)
0.96
0.03
0.94

Probality of Detection
Probality Measure

0.025
0.92

0.02 0.9 Largest Eigenvalue Based Detector

Conventional Energy Detector
0.88
0.015
0.86
0.01
0.84
0.005
0.82

0 0.8
100 120 140 160 180 200 220 10 15 20 25 30 35 40 45 50
z Number of Samples

Fig. 2. Accuracy of Gaussian approximation of distribution in eq.(18) with Fig. 3. Performance comparison for fixed SNR (SNR=-5dB).
parameters: K=4, N=30 and SNR=-5dB.
Detector Performance (N = 50)
1
rule is 0.98

H0 : ||Y||2F ≤ γopt (15) 0.96

H1 : ||Y||2F > γopt , (16) 0.94

Probality of Detection

where under H0 , the decision statistics ||Y||2F follows the 0.92

central Chi-square distribution with 2KN degree of freedom 0.9

Largest Eigenvalue Based Detector
as [1], [16] Conventional Energy Detector
1 0.88
||Y||2F ∼ z KN−1 e−z . (17)
Γ(KN) 0.86

Under H1 , ||Y||2F
follows the non-central Chi-square dis- 0.84

tribution with 2KN degree of freedom and a non-centrality 0.82

parameter γ =KN·SNR as [1], [16]
z KN−1 √ −10
||Y||2F ∼ ( ) 2 e−(γ+z) IKN−1 (2 γz), (18)
γ
where Iν (·) is the νth-order modified Bessel function of the
first kind. The optimal threshold γopt can be calculated in the
same way from (14) considering the corresponding distribu-
tions under both hypotheses. Notice that in order to ease the
numerical burden of evaluating the function Iν (·), we approxi-
mate the distribution (18) by a Gaussian distribution. The mean
and variance of this Gaussian distribution are obtained from
the first two moments of the distribution (18) respectively [16].
We illustrate the accuracy of this approximation in Fig. 2 for
the parameter values K = 4, N = 30 and SNR = −5 dB.
The performance curves of the energy detector in Fig. 3 and
Fig. 4 are plotted by using the above approximation.
As we can see from Fig. 3 and Fig. 4, the proposed largest
eigenvalue based detector outperforms the conventional energy
detector. We have chosen the performance metric of interest
to be the probability of correct detection. In both figures, the
optimal threshold for each point is obtained by assuming the
coefficients C0 = 1, C1 = 1.5.
In Fig. 3, we plot the probability of detection for the largest
eigenvalue based detector and the conventional energy detector
for the fixed SNR case, where the SNR is set at −5 dB. The
0.8
−8 −6 −4 −2 0
SNR in dB
Fig. 4. Performance comparison for fixed sample size (N=50).
number of collaborating secondary base stations is K = 4,
and the sample size ranges from 10 to 50. We can see that
the detection probability is uniformly larger for our proposed
detector in all the samples size considered. The improvement
is around 3 percentage in detection probability.
In Fig. 4 we consider the case for the fixed sample size,
where the number of samples equals N = 50. As in Fig. 3,
there are K = 4 cooperating secondary base stations. We de-
pict the performance of both largest eigenvalue based detector
and the energy detector in the SNR range starting from −10
to 0 dB. The detection probability is uniformly larger for the
proposed detector in the considered SNR range. Notice that the
performance improvement is not the same for all SNR values.
The detection probability of both schemes will converge to 1
when the SNR is more than −2 dB. When the SNR is −7 dB,
the increase in detection probability is more than 3 percents.
The superior performance of the proposed detector can be
understood as follows. For cooperative energy detection, the

2298
test statistics ||Y||2F , by definition, equals [4] K. Ruttik, K. Koufos and R. Jantti, “Spectrum sensing with multiple
antennas,” accepted in Proc. 2009 IEEE International Conference on

K Systems, Man, and Cybernetics, Oct. 2009.

||Y||2F = tr{YYH } = tr{R} = λi ,
i=1
where λi is the ith eigenvalue of the received covariance
matrix R. Therefore, the test statistics of the energy detector
blindly sums up all the K eigenvalues from the covariance
matrix R. On the other hand, for the largest eigenvalue based
detection the test statistics only involves λ1 . In other words,
our proposed detector will intentionally pick up only the
largest eigenvalue as decision statistics. This is an optimal
statistical test when there is only one primary user present
(matrix M being rank one). Recall also the implication from
Table I that blindly adding more eigenvalues as test statistics
is unnecessary. When summing all eigenvalues one obtains
a more heavy tailed distribution than the largest eigenvalue
distribution. This heavy tailed distribution will lead to worse
detection performance of the energy detector, which is the
main motivation behind the proposed detector.
IV. C ONCLUSION AND F UTURE W ORK
In this paper, we have provided a new spectrum sensing
scheme based on the largest eigenvalue distributions of the
received covariance matrix. This scheme picks up the largest
eigenvalue as the test statistics. We have shown that this
test statistics is efficient in terms of sample size and SNR
requirement when at most one primary user is present. The
decision threshold is obtained by an optimization process,
which takes into account both false alarm and miss detection
probability. The proposed scheme outperforms the conven-
tional energy detector for all the sample sizes and in the whole
SNR range considered. When there are more than one primary
user, the detection performance will be suboptimal compared
with the case when only one primary user is present, but the
proposed detector is still likely to outperform conventional
energy detectors.
In future work, we plan to extend the investigation to the
case of more than one primary user. We conjecture that the
optimal decision statistics distribution in that casewill be a
K
weighted summation of some/all the eigenvalues, i=1 ai λi ,
where the set of coefficients ai are chosen according to an
assumed number of simultaneously transmitting primary users.
(19) [5] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
Theory. Englewood Cliffs, NJ: Prentice-Hall, 1993.
[6] W. A. Gardner, “Exploitation of spectral redundancy in cyclostationary
signals,” IEEE Sig. Proc. Mag., vol. 8, pp. 14-36, 1991.
[7] Y. Zeng and Y. C. Liang, “Eigenvalue based spectrum sens-
ing algorithms for cognitive radio,” Apr. 2008. [Online]. Available:
http://arxiv.org/PS_cache/arxiv/pdf/0804/0804.2960v1.pdf.
[8] L. S. Cardoso, M. Debbah, P. Bianchi and J. Najim, “Cooperative spec-
trum sensing using random matrix theory,” 3rd International Symposium
on Wireless Pervasive Computing, pp. 334-338, May 2008.
[9] F. Penna, R. Garello and M. A. Spirito, “Cooperative
spectrum sensing based on the limiting eigenvalue ratio
distribution in Wishart matrices,” Feb. 2008. [Online]. Available:
http://arxiv.org/PS_cache/arxiv/pdf/0902/0902.1947v1.pdf.
[10] R. Tandra and A. Sahai, “SNR walls for signal detection,” IEEE J.
Select. Topic in Sig. Proc., vol. 2, no. 1, Feb. 2008.
[11] C. G. Khatri, “Distribution of the largest or the smallest characteristic
root under null hyperthesis concerning complex multivariate normal
populations,” Ann. Math. Stat., 35, 1807-1810, 1964.
[12] M. Kang and M. S. Alouini, “Largest eigenvalue of complex wishart
matrices and performance analysis of MIMO MRC systems,” IEEE J.
Selec. Areas Commun., vol. 21, no. 3, pp. 418-426, Apr. 2003.
[13] Y. H. Zeng, C. L. Koh and Y.-C. Liang, “Maximum eigenvalue detection:
theory and application,” IEEE ICC’2008, May 2008.
[14] H. Luutkepohl, P. Saikkonen and C. Trenkler, “Maximum eigenvalue
versus trace tests for the cointegrating rank of a VAR process,” Econo-
metrics Journal, vol. 4, issue 2, pp. 287-310, 2002.
[15] A. K. Gupta and D. K. Nagar, Matrix Variate Distributions. CRC Press,
2000.
[16] J. G. Proakis, Digital Communications. Boston, MA: McGraw-Hill, 4th
edition, 2001.

ACKNOWLEDGMENT
The first author wishes to thank Kalle Ruttik for discussions
of this topic.
R EFERENCES
[1] F. F. Digham, M. Alouini, and M. K. Simon, “On the energy detection of
unknown signals over fading channels,” in Proc. 2003 IEEE Intl. Conf.
Commun., May 2003.
[2] A. Sahai and D. Cabric, “Spectrum sensing: fundamental limits and
practical challenges,” in Proc. IEEE International Symposium on New
Frontiers in Dynamic Spectrum Access Network, Nov. 2005.
[3] R. Tandra and A. Sahai, “Fundamental limits on detection in low SNR
under noise uncertainty,” in WirelessCom., Jun. 2005.

2299