Manuscript

Click here to view linked References

1
1
2
3
4
Kurtosis Based Spectrum Sensing in Cognitive
5
6
7
8
9
Radio
10
11
12 Shanthan Suresh, Shankar Prakriya, and Manav R. Bhatnagar.
13
14
15
16
17
18
19
20
21 Abstract
22
23
In this paper, we consider spectrum sensing in cognitive radio networks based on higher order statistics. The
24
25
26 kurtosis, a fourth order statistic, which is a measure of deviation from Gaussianity, is used as a detection statistic.
27
28 An optimum threshold is set up based on the Neyman-Pearson criterion and analytical expression for upper bound
29
30 on probability of miss is derived for a single pair of primary and secondary users. Further, we also propose a
31
32 collaborative spectrum sensing scheme for more than one secondary user and it is shown by simulations that the
33
34 proposed kurtosis based method outperforms the energy based spectrum hole detection method significantly.
35
36
37 Index Terms
38
39
40 Cognitive radio, cooperative spectrum sensing, kurtosis, Neyman Pearson test, spectrum sensing.
41
42
43
44 I. I NTRODUCTION
45
46
47 Evolution of the wireless systems and technologies have made a huge impact on the society. However, at
48
49 the same time, it has caused an acute scarcity of frequency bands for new services. This calls for efficient
50
51
52 management of the spectrum. In this regard, the Federal Communication Commission (FCC) published a
53
54 report on the spectrum management and utilization [1], which revealed, (as is now well known) that the
55
56
57 The authors are with the Department of Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, 110016, India (e-mails:
58
59 shanthan.n2@gmail.com,{shankar,manav}@ee.iitd.ac.in).
60
61 This work is supported by UKIERI Project SA08-080.
62
63
64
65
 2
1
2
3 allocated spectrum is grossly under-utilized. This problem is the motivation behind the development of
4
5 the cognitive radio (CR) technology. Cognitive radio is based on the concept of software defined radio
6
7
8 and provides techniques for spectrum sensing, spectrum management, and spectrum access for unlicensed
9
10 secondary users when the primary users are not utilizing the spectrum. A function of the CR is to support
11
12
13 intelligent and efficient dynamic spectrum access. Therefore, reliable spectrum sensing is an important
14
15 aspect of cognitive radio technology [2].
16
17
18 In the literature, a number of spectrum sensing algorithms have been suggested [3]-[6]. These can be
19
20 broadly classified based on their requirements for sensing [7], as, (A) Methods requiring both noise and
21
22
23 signal source information, like matched filter based spectrum sensing, cyclostationary based technique,
24
25 (B) Methods requiring only noise power information, which is termed as semiblind detection and includes
26
27
energy based detection, and, (C) Methods requiring no information about source signal or noise, which
28
29
30 is termed as totally blind detection (covariance based approach is an example).
31
32 The matched filter based sensing is an optimum spectrum sensing technique [3][6]- provided - knowl-
33
34
35 edge about the primary user signal, modulation characteristics, transmitted power, pulse shape and the
36
37 packet format are known at the secondary user terminal. This implies that, the suitability of this method is
38
39
40 limited to cases where the secondary user terminal has knowledge about the pilot signals and preambles
41
42 used by the primary user transmitter. Also the matched filter techniques performance degrades with
43
44
45 timing and carrier frequency offset between the primary transmitter and the secondary user terminal.
46
47 The cyclostationary based technique is a feature based technique [5], which exploits the fact that the man
48
49
50
made signals exhibit second order cyclostationary property in the wide sense [8] . This has the advantage of
51
52 being robust to uncertainty in the noise and also the ability to differentiate between the primary user signal
53
54 and interference signal, based on their cyclic properties. However this method requires the knowledge of
55
56
57 the signal phase and one or multiple cyclic frequencies [5] of the primary user signal. The disadvantages
58
59 of the cyclostationary based approach is that it requires a high sampling rate and it is computationally
60
61
62 complex.
63
64
65
 3
1
2
3 Another spectrum sensing method is the energy based detector, which is the simplest of the existing
4
5 spectrum sensing methods [3] [9], in terms of the computation and implementation complexity. As stated
6
7
8 above, energy based detector is a semi blind technique, which is the optimum detector for detecting a
9
10 random uncorrelated Gaussian signal [3]. The energy detector does not require any prior knowledge of
11
12
13 the primary user signals. However this technique is susceptible to uncertainty in the noise variance [3].
14
15 For fading channels, the energy based detector is not necessarily optimal. We show in this paper that the
16
17
18 kurtosis based method outperforms the popular energy based scheme at low SNRs.
19
20 The covariance based technique exploits the fact that the received data vector of the primary user has a
21
22
23 non-diagonal covariance matrix [10]. Hence, use of the non-diagonal elements of the covariance matrix as
24
25 a test statistic is the essence of the covariance based spectrum sensing method. Another method which uses
26
27
the higher order statistics (HOS) is the entropy based spectrum sensing method [11], which is based on the
28
29
30 principle that the entropy is greater for noise samples from additive white gaussian noise (AWGN) process
31
32 than that for the primary user’s signal. This method is based on the histogram approach and likelihood
33
34
35 ratio test is applied to make the decision. However, the authors consider the situation where the secondary
36
37 user has knowledge about the primary transmitter signal, which is not feasible in practical situation.
38
39
40 In [12], the author proposes an algorithm for sensing the Digital Television signals in Gaussian noise
41
42 using Higher Order Statistics (HOS). The algorithm performs non-Gaussianity check in the frequency
43
44
45 domain in the vicinity of the pilot of the DTV, to make a decision about the presence or absence of
46
47 the primary user signal. The scheme of [12] is backed by simulation results under different conditions.
48
49
50
However, no theoretical analysis has been carried out to obtain an analytical expressions for probability
51
52 of miss and false alarm in [12]. On similar lines, the author in [13] gives a sensing technique based on
53
54 higher order statistics, in the frequency domain, using a 2048 point fast fourier transform. The bispectrum,
55
56
57 which is a third order statistic, is used as a decision statistic in [14]. The performance of the scheme is
58
59 illustrated by simulations only and no theoretical analysis has been conducted by the authors in [14].
60
61
62 In this paper, we have chosen kurtosis-the fourth order cumulant at zero lag as a test statistic for spectrum
63
64
65
 4
1
2
3 sensing. Since it is difficult to obtain an accurate estimate of higher order statistics and prohibitively high
4
5 number of observation samples are required to get a reliable estimate, we utilize only the fourth order
6
7
8 statistic in this paper. To the best of our knowledge, no analysis has been carried out in the existing
9
10 literature to obtain expressions of probability of miss and false alarm for spectrum sensing using fourth
11
12
13 order statistics.
14
15 The reasons for choosing kurtosis as a test statistic for semi-blind spectrum sensing are, (a) At low
16
17
18 SNR, when the noise variance is high, the energy estimate tends to accumulate the noise energy, whereas,
19
20 the kurtosis based scheme will suppress the effect of noise, since the kurtosis of noise, which is assumed
21
22
23 to be additive Gaussian, will be zero. This can be observed from Fig. 1 in Section VII, in the Simulation
24
25 results, wherein, at low SNR range of -15dB to 11dB, the kurtosis based scheme outperforms the energy
26
27
based scheme, significantly. (b) If there are multiple secondary users terminals experiencing independent
28
29
30 fading, then at least one of them will have a good link to the primary terminal. Since kurtosis involves
31
32 the fourth order absolute power of the channel, the kurtosis yields a better estimate, as compared to the
33
34
35 energy based scheme, which uses second order absolute power of the channel. This fact is illustrated in
36
37 Fig. 3 in Section VII, where, the kurtosis based method offers significant improvement over the energy
38
39
40 based method with 4 secondary users. (c) The energy detector cannot distinguish among the primary user
41
42 signals, and interference. But kurtosis based method has an advantage, that, in the presence of independent
43
44
45 interferers, even though the individual interference signals be non Gaussian, their sum tends to Gaussian
46
47 by central limit theorem and hence the kurtosis value of the interference goes to zero and helps in
48
49
50
differentiating between the primary user signal and interference . The disadvantage of the kurtosis based
51
52 method is that its performance advantage dimnishes when the interference is non-Gaussian. This is due
53
54 to the fact that in such a situation, the received signal will always follow a non-Gaussian distribution
55
56
57 irrespective of the presence or absence of the primary user signal. In a sense it becomes quite similar to
58
59 the energy based scheme - the interference noise simply adds to the metric. In the TV band where the
60
61
62 primary interference is the TV signal, this assumption may not be true. However, in a dynamic spectrum
63
64
65
 5
1
2
3 sharing scenario that is commonly studied today, this assumption is quite reasonable.
4
5 In this paper we assume that the secondary receiver terminals have a perfect knowledge of the noise
6
7
8 variance, for the operating SNR values and also the the primary signal samples are statistically uncor-
9
10 related. In [10], the authors have shown by simulations that with perfect noise variance knowledge and
11
12
13 uncorrelated symbols, the energy based detector performs better than the covariance based approach.
14
15 Therefore, in our simulations, we have compared our scheme with the optimal semiblind sensing scheme,
16
17
18 namely the energy based detector.
19
20 The contributions of this paper are as follows. 1) A kurtosis based spectrum sensing for single pair
21
22
23
of primary and secondary users has been proposed. 2) We derive an upper bound of the probability of
24
25 miss, i.e, probability of missing the presence of the primary user signal. 3) The proposed kurtosis based
26
27 scheme has been extended to the collaborative spectrum sensing scenario, with a single primary user and
28
29
30 multiple secondary users which provides significant improvement as compared to the conventional energy
31
32 based collaborative scheme.
33
34
35 Rest of the paper is organized as follows. Section II describes the system model and the problem
36
37 formulation. In Section III, brief description of the energy based spectrum sensing is done. Section IV
38
39
40 discusses the proposed spectrum sensing scheme based on kurtosis. An upper bound of the average
41
42 probability of miss has been obtained in Section V. An extension of the proposed spectrum sensing
43
44
45
problem for the case of constraining the interference caused by secondary user to the primary user is
46
47 discussed in Section VI. Section VII presents the simulation results. Some conclusions are drawn in
48
49 Section VIII. This paper also contains two Appendices. The variance of the kurtosis estimate has been
50
51
52 given in Appendix A. In Appendix B, the unbiased property of the kurtosis estimate has been proved.
53
54
55 II. P ROBLEM FORMULATION
56
57
58 We consider a model in which there is a single primary user and a single secondary user. We assume
59
60
61
that there is no interferer in the neighborhood of the primary user and the secondary user. Spectrum
62
63 sensing in CR is the following binary hypothesis testing problem,
64
65
 6
1
2
3 Hypothesis H0 : Absence of a primary user;
4
5 Hypothesis H1 : Presence of a primary user.
6
7
8 The secondary user has to perform spectrum sensing and decide between the following binary hypotheses
9 
10 
n(i), H0


11
12 r(i) = (1)

13  hs(i) + n(i), H1

14
15 for i = 1, 2, ..., Nobs , where, Nobs is the total number of observation samples in the observation interval,
16
17
18 during the spectrum sensing process, r(i) is the ith sample of the received signal, s(i) is the ith sample of
19
20 the signal from the primary user, h is the complex Gaussian channel gain, assumed to be fixed during the
21
22
23 observation interval, n(i) ∼ Nc (0, 2ρ2n ) is circular AWGN. We assume that s(i) to have discrete uniform
24
25 distribution with zero mean. The signal constellation can be BPSK, as in [3]. This is a valid assumption,
26
27
28 as, the signal constellation will, in general, have a finite set of symbols, and in typical communication
29
30 scenarios, it is reasonable to assume that each of the symbol has a equiprobable occurrence. Also, the
31
32
fading nature of the channel ensures that the overall signal under hyposthesis 1 is non-gaussian. Hence,
33
34
35 r(i) is non-Gaussian under hypothesis H1 and Gaussian under H0 .
36
37 The secondary user has to decide the correct hypothesis during the observation interval and make a
38
39
40 decision whether the primary user is transmitting or not. The fact that r(i) is Gaussian under hypothesis
41
42 H0 and non-Gaussian under hypothesis H1 makes kurtosis an appropriate choice as a decision statistic.
43
44
45
46 III. R EVIEW OF E NERGY BASED SPECTRUM SENSING
47
48
49 Energy based method is one of the oldest spectrum sensing schemes used in cognitive radio [3]. In this
50
51 scheme, the channel is sensed in the desired frequency band and the RF energy is measured to detect the
52
53
54 presence of a primary user. Once the energy estimate is obtained, it is compared with a pre-determined
55
56 threshold to make the binary decision.
57
58
59 This scheme is the best choice in the scenario, where knowledge of the primary user signal is totally
60
61 absent at the CR. The drawbacks of this approach is that it is not able to differentiate between primary
62
63
64
65
 7
1
2
3 users signal and interference and it also has a performance degradation in the low SNR range.
4
5 From (1), the energy estimate can be obtained as,
6
7 Nobs
8 1 X
9
ê = |r(i)|2 , (2)
Nobs i=1
10
11
12
which is modeled in [15] [16] as,
13
14 
15 
χ22Nobs , H0

16 
17 ê = (3)
 χ22Nobs (2γ), H1

18 
19
20
21
where χ22Nobs denotes a central chi-square distribution with 2Nobs degrees of freedom, χ22Nobs (2γ) denotes
22
23 a non-central chi-squared distribution with 2Nobs degrees of freedom and a non-centrality parameter 2γ.
24
2
E[|s| ]
25 Here γ = |h|2 E[|n| 2 ] is the instantaneous SNR. Expressions for the probability of false alarm and probability
26
27
28 of miss have been obtained in [15].
29
30
31
32
IV. P ROPOSED SPECTRUM SENSING SCHEME
33
34 The kurtosis, k, for a circular zero mean random variable r is defined as
35
36
37 k = E[|r|4 ] − 2(E[|r|2 ])2 . (4)
38
39
40 It can be established that the kurtosis of a Gaussian random variable is zero while it is non-zero for a
41
42
43 non-Gaussian random variable [17]. Therefore, the kurtosis is zero under hypothesis H0 , and non-zero
44
45 under hypothesis H1 . The magnitude of the kurtosis, therefore, serves as a good candidate for the decision
46
47
48 statistic of the binary hypothesis test.
49
50
51
52 A. Estimation of Kurtosis using Finite Observation Samples
53
54 The kurtosis can be estimated as [18]
55
56 Nobs  Nobs 2
57 1 X 4 1 X 2
k̂ = |r(i)| − 2 |r(i)| . (5)
58 Nobs i=1 Nobs i=1
59
60
61
It has been shown in [20], that asymptotically, k̂ tends to a Gaussian distribution. For medium SNR range
62
63 of around 15dB, the kurtosis estimate calculated using finite number of observation samples (Nobs = 100)
64
65
 8
1
2
3 exhibit this Gaussian property. Hence, we use the finite observation kurtsosis estimate, k̂, as the decison
4
5 statistic in the spectrum sensing problem.
6
7
8
9
10
B. Decision Statistic and Cost Function
11
12 Since (5) is an unbiased estimate 1 , under hypothesis H0 it has a zero mean and under hypothesis H1
13
14
15 it has a non-zero mean. The separation of the mean values can be used as a suitable decision statistic, as
16
17 follows,
18
19
20 |k̂| RH1
H0 λ, (6)
21
22
23 where, λ is a decision threshold which has to be chosen by the minimization of a cost function. The
24
25 kurtosis estimates have the following statistics: k̂|H0 ∼ N(µ0 , σ02 ) and k̂|H1 ∼ N(µ1 , σ12 ) 2 .
26
27
28 The probability of false alarm, Pf , indicates the under utilization of the frequency spectrum by the
29
30 secondary user terminal. On the other hand, the probability of miss, Pm , implies that the secondary user
31
32
33 has missed sensing the presence of primary user signal and starts transmitting it’s signal in the same
34
35 frequency band as that of the primary user signal, causing interference to the licensed primary user. The
36
37
38
main virtue of the secondary user is not to interfere at all with the primary transmission. This is critical
39
40 from the viewpoint of the primary user. The sum of Pm and Pf could have been chosen as a cost function.
41
42 However, this does not guarantee a desired Pf . Therefore, we choose we choose Pm as the cost function
43
44
45 and minimize it, by fixing the value of Pf . Formally,
46
47
48 min Pm (λ),
49 λ
50
51 such that Pf ≤ α00 (7)
52
53
54 where Pf is the probability of False alarm, the probability of deciding in favour of a primary user’s signal
55
56 when it is absent.
57
58 1
Unbiased property of k̂ has been proved in Appendix- B
59 2
The expressions for the mean and variance are given in the Appendix- A
60
61
62
63
64
65
 9
1
2
3 We can use the Neyman-Pearson criterion to find the optimized threshold, λ∗ [21]. From (7), setting
4
5 Pf = α0 , we have,
6
7
Z ∞ Z −λ∗
8 fk̂|H0 (z)dz + fk̂|H0 (z)dz = α0 , (8)
9 λ∗ −∞
10
11 which reduces to,
12
λ∗
 
13 2Q = α0 , (9)
14 σ0
15
16 where we have used the fact that µ0 = 0, since k̂|H0 is an unbiased estimate of noise kurtosis. Hence
17
18
19 λ∗ = σ0 Q−1 ( α20 ) is the optimized threshold. It is important to note that for deriving the optimum threshold,
20
21 no knowledge of the channel SNR is required.
22
23
24
25 C. Collaborative Spectrum Sensing
26
27 If the secondary users are far away from the primary users then there might be deep fading and/or
28
29
30 shadowing which might prevent the secondary users from detecting the primary user signal. In such a
31
32 case, the decision of a single secondary user will not be reliable. In such scenarios, collaborative spectrum
33
34
35 sensing techniques can be used [22]. The proposed kurtosis based scheme can be extended to the case
36
37 of a collaborative CR model. We shall consider the case where the participating secondary users make
38
39
40 a hard decision i.e., the remaining collaborating secondary users send their binary decisions about the
41
42 presence or absence of a primary user to a centralized fusion center and this fusion center makes a final
43
44
45 decision based on some voting rule [16]. In particular, one of the rule is the OR fusion rule, where, the
46
47 fusion center considers the binary decisions of the participating secondary users and decides H1 , if, at
48
49
least one of the binary decisions is 1. The probability of detection and false alarm under such a scenario
50
51
52 is given as [22],
53
54
55 Qm,avg = (Pm,avg )N ,
56
57
58 Qf,avg = 1 − (1 − Pf,avg )N , (10)
59
60
61
where Pm,avg and Pf,avg are the average probability of miss and the probability of false at the secondary
62
63 user, given in the Section V. Since Pm,avg < 1 and Pf,avg < 1, for N > 1, Qd,avg < Pm,avg and
64
65
 10
1
2
3 using binomial approximation [23, Eq. (1.110)] Qm,avg ≈ N Pf,avg . Hence, for collaborative scheme, the
4
5 probability of miss decreases, but at the cost of increase in the value of probability of false alarm. We
6
7
8 have chosen the OR-rule because the OR-rule gives more emphasis to minimizing the probability of miss.
9
10 This is attributed to the fact that the OR-rule is a minority rule for detecting the presence of a primary
11
12
13 user, in the sense, even if at least one of the participating secondary users detects the presence of the
14
15 primary user signal, the fusion center gives the decision in favor of hypothesis H1 .
16
17
18
19 V. D ERIVATION OF UPPER BOUND OF AVERAGE PROBABILITY OF MISS
20
21
22 A. Bound on Probability of Miss
23
24
25 The probability of miss, conditioned on the channel, is expressed as,
26
27
Z ∞ Z −λ∗
28 Pm|γ,θ = 1 − ˆ 1 (z)dz −
fk|H fk|H
ˆ 1 (z)dz,
29 λ∗ −∞
30
31 = 1 − Pm1 |γ,θ − Pm2 |γ,θ , (11)
32
33 R∞ R −λ∗ 2
2 E[|s| ]
34 where Pm1 |γ,θ = fk|H ˆ 1 (z)dz. Here γ = |h| E[|n|2 ] is the instantaneous
ˆ 1 (z)dz, and Pm2 |γ,θ = −∞ fk|H
λ∗
35  
36 −1 Im[h]
37
SNR, and θ = tan Re[h]
, where Im[X] and Re[X] denote the imaginary and real parts of X respectively,
38
39 is the instantaneous phase angle of the channel. Since the kurtosis estimate follows a Gaussian distribution
40
41 under hypothesis H1 , i.e., k̂ ∼ N(µ1 , σ12 (γ, θ)), (11) can be rewritten as,
42
43
44
45
Pm|γ,θ = 1 − Pm1 |γ,θ − Pm2 |γ,θ , (12)
46  ∗   ∗ 
47 1 λ − µ1 1 λ + µ1
Pm|γ,θ = 1 − erfc √ − erfc √ ,
48 2 2σ1 (γ, θ) 2 2σ1 (γ, θ)
49  ∗   ∗ 
50 1 λ −µ1 1 λ +µ1
where Pm1 |γ,θ = 2 erfc 2σ (γ,θ) , Pm2 |γ,θ = 2 erfc 2σ (γ,θ) , and erfc(.) is the complementary error
√ √
51 1 1
52
53 function [Eq. (8.253)][23]. The exact expression for σ12 (γ, θ)) in terms of γ and θ is given in the Appendix
54
55 A.
56
57
58 The conditional probability, Pm|γ,θ , can be averaged with respect to the random variables γ and θ. For a
59
60 circular Gaussian channel h, γ follows an exponential distribution, i.e., γ ∼ ξ( γ̄1 ), where γ̄ is the average
61
62
63
64
65
 11
1
2
3 SNR, and θ is uniformly distributed between [0, 2π]. Here, ξ(·) denotes the exponential distribution. The
4
5 average value, Pm,avg , after averaging Pm|γ,θ , with respect to γ and θ, can be written as,
6
7
8
9 Pm,avg = 1 − Pm1 ,avg − Pm2 ,avg , (13)
10
11
12 where,
13
14
Z ∞ Z 2π
15 Pm1 ,avg = Pm1 |γ,θ fγ,θ (γ, θ)dθdγ,
16 γ=0 θ=0
17 Z ∞ Z 2π   
1 K 1 − γγ̄
18 = erfc e dθdγ, (14)
19 γ=0 θ=0 2 σ1 (γ, θ) 2πγ̄
20 γ
21 (λ∗ −µ1 ) 1 − γ̄
where, K = √
2
. In (14), fγ,θ (γ, θ) = fγ (γ)fθ (θ) = 2πγ̄
e is the joint pdf of γ and θ, which are
22
23
24 statistically independent. We can simplify integral given in (14) as follows,
25 Z x0 Z 2π  
26 K γ
27 Pm1 ,avg =c erfc e− γ̄ dθdγ+
28 γ=0 θ=0 σ1 (γ, θ)
Z 1 Z 2π  
29 K γ
30 c erfc e− γ̄ dθdγ+
31 γ=x0 θ=0 σ1 (γ, θ)
32 Z ∞ Z 2π  
K γ
33 c erfc e− γ̄ dθdγ. (15)
34 γ=1 θ=0 σ1 (γ, θ)
35
36 = I1 + I2 + I3 ,
37
38 R x0 R 2π   γ R 1 R 2π   γ
− γ̄
39 where x0 < 1, I1 = c γ=0 θ=0
erfc K
σ1 (γ,θ)
e dθdγ, I2 = c γ=x0 θ=0
erfc K
σ1 (γ,θ)
e− γ̄ dθdγ, I3 =
40
41 R ∞ R 2π   γ
42
K
c γ=1 θ=0 erfc σ1 (γ,θ) e− γ̄ dθdγ, and c is a constant that is independent of the variables of integration.
43
44 1) Evaluation of I1 : The range of γ is 0 ≤ γ ≤ x0 . Using the series expansion of erf c(·) [Eq. (8.253)][23],
45
46 ∞
(−1)k+1 z 2k−1
 X
47 2
erfc(z) = 1 − √ , (16)
48 π k=1
(2k − 1)(k − 1)!
49
50
51
using (16) in I1 ,  2k−1
∞ k+1 K
52
c
Z x0 Z 2π
2 X (−1) σ1 γ
53 I1 = [1 − √ ]e− γ̄ dθdγ, (17)
54 4πγ̄ γ=0 θ=0 π k=1
(2k − 1)(k − 1)!
55
56 For 0 ≤ γ ≤ x0 , σ1 (γ, θ) in (39) may be approximated by
57
58
59 √ p
60 σ1 (γ, θ) ≈ a00 + a10 γ = a00 (1 + b0 γ) (18)
61
62
63
64
65
 12
1
2
3 where a00 and a10 are constants which depend on the modulation scheme being employed under hypothesis
4
5 H1 , and whose values are given in the Appendix A, for 16-QAM constellation, and b0 = a10
. The k-th
6 a00
7
8 term of the integrand, tk , of I1 in (17) is given by,
9  2k−1
10  ∞ (−1)k+1
X K
2 σ1 (γ,θ) γ
11 tk = −√ e− γ̄ , (19)
12 π k=1
(2k − 1)(k − 1)!
13
14 for k ≥ 1. Using (18), (19) can be rewritten as
15  2k−1
16  ∞ (−1)k+1
X √ K
√
17 2 a00 1+b0 γ γ
tk = −√ e− γ̄ (20)
18 π k=1
(2k − 1)(k − 1)!
19
20
21
Since |b0 γ| < 1 for the range of γ considered above, using the binomial expansion [Eq. (1.110)][23] for
22 1
23 (1 + b0 γ)−k+ 2 , (20) can be further simplified as,
24 (
2k−1
(−1)k+1
 
25 2 K
26 tk = − √ √ 1
27 π a00 (2k − 1)(k − 1)!
28 ∞
)
29
X (−1)p (2k − 1)(2k − 3)...(2k − (2p − 3))bp00 γ p − γγ̄
+ p p!
e (21)
30
p=1
2
31
32 Therefore, I1 in (17) is a double integration with a double summation. The (k, p)-th term of the double
33
34
35 summation, where k and p are the indices of the the first and second summation, respectively, after
36
37 integration yields,
38
39 Z x0 Z 2π
γ
40 I1,k,p = (γ)p exp(− )dθdγ
41 γ=0 θ=0 γ̄
42
43 x0
= 2πγL (p + 1, ), (22)
44 γ̄
45
46 where, γL (·, ·) is the lower incomplete Gamma function [Eq. (8.350.1)][23]. Therefore, I1 in (17) can be
47
48
49 written as,
50
51 ∞ X
X ∞
52 I1 = I1,k,p
53 k=1 p=1
54  2k−1 (
55 ∞ k+1 √K
56
X 2c (−1) a00 1 x0
= √ p
γL (p + 1, )δ(p)+
57
k=1
π (2k − 1)(k − 1)! 2γ̄ γ̄
58
∞
)
59 X (−1)p (2k − 1)(2k − 3)...(2k − (2p − 3))γ̄ p bp 0 x0
60 γL (p + 1, ) ,
61 p=1
2p p! γ̄
62
63 where δ(.) is the Kronecker delta function.
64
65
 13
1
2
3 2) Evaluation of I2 : The range of γ is x0 ≤ γ ≤ 1. In this range of γ, since a10 γ is the dominant
4
5 term in the expression for σ(γ, θ) in (39), it can be approximated as,
6
7
8 √
9
σ1 (γ, θ) = a10 γ, (23)
10
11
12
where a10 is a constant which depends on the modulation format and is given in the Appendix A. Using
13
14 the series expansion of erfc(·) in I2 in (17) and proceeding in similar lines as for I1 , the k-th term integrand
15
16 for I2 in (17) can be written as,
17
   2k−1
18
19 − √2π (−1) k+1 √K
a10 1 γ
20 tk = γ −k+( 2 ) exp(− ) (24)
(2k − 1)(k − 1)! γ̄
21
22
23 Using the series expansion for exp(− γγ̄ ) [Eq. (1.211.1)][23], (24) reduces to,
24    2k−1
25 − √2π (−1)k+1 √Ka10 ∞ γ p
1 X (− γ̄ )
26 tk = γ −k+( 2 ) (25)
27 (2k − 1)(k − 1)! p=0
p!
28
29 Therefore the (k, p)-th term in the integrand in I2 will be
30
31
   2k−1
√2 (−1) k+p+1 √K
32 π a10 p−k+ 12
33 tk,p = γ (26)
34
(2k − 1)(k − 1)!p!γ̄ p
35
36 Therefore, using (26), the expression for I2 becomes,
37  2k−1
p−k+ 32
c k+p+1 √K
38 ∞ ∞
X X π (−1)
√
a10
(1 − x 0 )
39 I2 = (27)
40 k=1 p=0
(2k − 1)(k − 1)!p!(p − k + 12 )γ̄ p+1
41
42 3) Evaluation of I3 : γ > 1 In this range of γ we can neglect the lower order terms retaining the highest
43
44
45 order term of γ in the expression for σ1 (γ, θ) in (39). Hence
46 q
47
48 σ1 (γ, θ) ≈ [a40 + a41 sin2 2θ − a41 sin4 4θ]γ 4
49 q
50 [a40 + a41 sin2 4θ 1 − sin2 2θ ]γ 4


≈
51
52 q
53 ≈ a40 [1 + b1 sin2 4θ]γ 4 (28)
54
55
56 a41
where b1 = 4a40
, and a40 a41 are constants, given in the Appendix A. Therefore, the k-th term in the
57
58
59 integrand of I3 can be written as
60  2k−1 1 γ
61 √2 (−1)k+1
π
√K
a40
[1 + b1 sin2 4θ]−k+ 2 γ −4k+2 e− γ̄
62 tk = (29)
63 (2k − 1)(k − 1)!
64
65
 14
1
2 1
3 Since |b1 | < 1 the term [1 + b1 sin2 4θ]−k+ 2 can be expanded as a binomial series,
4
5 1
6 [1 + b1 sin2 4θ]−k+ 2 = 1+
7 ∞
8 X (−1)p (2k − 1)(2k − 3)...(2k − (2p − 3))
9 }bp1 sin2p 4θ (30)
10 p=1
2p p!
11
12 The (k, p)-th term integral of I3 will be
13 Z ∞ Z 2π
14 γ
15 I3,p,k = sin2p 4θγ −4k+2 e− γ̄ dθdγ
16 γ=1 θ=0
√
2 πΓ(p + 12 )
 
17 1
18 = E4k−2 (31)
19 Γ(p + 1) γ̄
20
21 where Γ(·) is the Gamma function [Eq. (8.310.1)][23] and E(·) (·) is the Exponential integral [Eq. (12),
22
23 pp xxxv][23]. Therefore,
24  
25
 2k−1
c √K
26 ∞ 3 (−1)k+1 Γ(p + 12 ) 
a10
 
X γ̄π 2 1
27 I3 = E4k−2 δ(p)
28 k=1
(2k − 1)(k − 1)!Γ(p + 1) γ̄
29 ∞ 
30 1 X (−1)p (2k − 1)(2k − 3)...(2k − (2p − 3))
+ √ ×
31 2γ̄ π p=1 2p p!
32
bp Γ(p + 21 )
 
33 1
34 E4k−2 (32)
35 Γ(p + 1) γ̄
36
λ∗ +µ1
37 An expression for Pm2 ,avg can be derived by following the same analysis with K = 2
. Therefore, the
38
39 expression for Pm,avg can be written as,
40
41 ∞ 2k−1 2k−1 
λ∗ − µ1 λ∗ + µ1
X  
42 ∗
Pm,avg (λ ) = 1 − Ak + , (33)
43
k=1
2 2
44
45 where Ak is the term independent of λ∗ .
46
47
48
49 B. Probability of False Alarm
50
51
52 ˆ the probability of false alarm, Pf , is the probability of deciding in the
For the decision statistic |k|,
53
54 favor of presence of the primary user signal when it is actually not present. Since this probability does
55
56
57 not depend on the channel statistics, the conditional (on the channel) probability of false alarm, Pf |γ,θ ,
58
59 and the average probability of false alarm,Pf,avg , are identical. Hence,
60
61 Z ∞ Z −λ∗
62 Pf,avg = fk|H
ˆ 0 (z)dz + fk|H
ˆ 0 (z)dz, (34)
63 λ∗ −∞
64
65
 15
1
2
3 Since for hypothesis H0 , k̂ ∼ N(µ0 , σ02 ), (34) reduces to,
4
5
 ∗   ∗ 
λ − µ0 λ + µ0
6 Pf,avg = erf c √ + erf c √ (35)
7 2σ0 2σ0
8
9 In (35), the quantity σ0 depends on the noise variance ρ2n only.
10
11
12
13 VI. S PECTRUM S ENSING - THE D UAL P ROBLEM
14
15
16
In section IV, the spectrum sensing problem was formulated by constraining the probability of false
17
18 alarm, Pf,avg to α0 , and minimizing the probability of miss, Pm,avg . This implies that the secondary user
19
20 can use only a fraction(α0 ) of the times the primary user channel is free. However, in many scenarios,
21
22
23 it is required that the interference caused by the secondary user to the primary user should not exceed
24
25 a predetermined level. Motivated by this fact, we formulate the dual spectrum sensing problem namely
26
27
28 minimizing Pf,avg , putting a constraint on Pm,avg , i.e.,
29
30
31 min Pf,avg (λ),
λ
32
33
34 such that, Pm,avg ≤ β0 (36)
35
36
37 It is important to note that we are putting a constraint on the Pm,avg (averaged over the primary user
38
39 channel) and not the instantaneous Pm . With the limited processing capabilities of the secondary users,
40
41
42
it is not very practical to assume that they will have the instantaneous channel SNR. Hence, the optimal
43
44 threshols, λ∗dual will only depend on the average channel SNR and it is reasonable to assume the average
45
46
channel SNR knowledge at the secondary user terminal.
47
48
49 By NP criterion,
50
51 Pm,avg (λ∗dual ) = β0 (37)
52
53
54 Using the expression for Pm,avg derived in 33, we have,
55
56 ∞  ∗ 2k−1  ∗ 2k−1 
57 ∗
X λdual − µ1 λdual + µ1
Pm,avg (λdual ) = 1 − Ak + = β0 , (38)
58 2 2
59 k=1
60
61
where, µ1 is the kurtosis mean value. λ∗dual can be otained by solving (3) numerically. Comparing the
62
63 two thresholds, λ∗ and λ∗dual , it can be seen that, λ∗ , is totally independent of the instantaneous channel
64
65
 16
1
2
3 SNR(γ) and average channel SNR, (γavg ), whereas, λ∗dual , is dependent γavg . Hence, λ∗dual will be more
4
5 sensitive to the channel SNR (in an average sense), and we can choose β0 in a sensible manner depending
6
7
8 on γavg .
9
10
11 VII. S IMULATION R ESULTS
12
13
14 In our simulations, we have compared the kurtosis based semi-blind scheme with the known semiblind
15
16
17
detector, namely, the energy based detector [7]. The energy detector cannot distinguish among the primary
18
19 user signals, secondary user signals, and interference. However, kurtosis based method has an advantage
20
21 that in the presence of independent interferers, the sum of the interference signals tends to become Gaussian
22
23
24 distributed, by the central limit theorem. Hence, the kurtosis goes to zero and helps in differentiating
25
26 between the primary user signal and interference . The simulations have been done for the slow fading
27
28
29 Rayleigh channel, 16-QAM constellation and zero mean Gaussian noise with unit variance. The SNR is
30
31 given by the signal power. Fig. 1 compares the performance of the proposed kurtosis and the coventional
32
33
34 energy based scheme. It can be seen that, for the single secondary user case, under low SNR range of
35
36 -15dB to 11dB, kurtosis based scheme outperforms Energy based scheme. This is intuitively appealing as
37
38
39
in the low SNR region, noise dominates and increases the energy in the absence of a primary user signal
40
41 whereas the cumulant based kurtosis scheme tends to suppress the Gaussian noise and hence shows an
42
43 improved performance. For Pm,avg = 10−2 , the kurtosis based scheme gives an improvement of 1.5dB
44
45
46 over the energy based scheme.
47
48 The complementary Receiver Operating Characteristics (ROC) curve is depicted in Fig. 2 with the
49
50
51 average SNR as a parameter. For a fixed probability of false alarm, lower value of probability of miss
52
53 and therefore better performance is observed at a higher average SNR, for the proposed scheme.
54
55
56 The proposed spectrum sensing scheme is compared with the energy based scheme in a collaborative
57
58 system with 1,2, and 4 secondary users and a single primary user. It can be seen from Fig. 3 that increasing
59
60
61
the number of secondary users users from two to four leads to a considerable improvement in the kurtosis
62
63 based scheme compared to the Energy based scheme. This is based on the fact that kurtosis has absolute
64
65
 17
1
2
3 fourth power of random variables as compared to absolute second power for the energy method. If there
4
5 is at least one good channel path among the independent fading paths for the different secondary users,
6
7
8 kurtosis will be more reliable than the energy estimate and hence provides better performance. Fig. 4 is a
9
10 plot of the probability of miss curve obtained by simulation and also the analytical upper bound, derived
11
12
13 in Section V.
14
15 The fundamental problem in the proposed kurtosis method is that the kurtosis estimate requires more
16
17
18 number of observation samples to have a better estimate than that required for the energy based. Fig. 5 is
19
20 an illustration of this fact. To achieve a probability of false alarm of 0.1, 250 samples are required for the
21
22
23
energy based scheme as compared to 350 by the kurtosis based scheme. But the huge advantage that is
24
25 obtained by the proposed kurtosis method, in the collaborative scenario outweighs the fact that it requires
26
27 more samples for obtaining a good estimate.
28
29
30 The ROC curves for the dual spectrum sensing problem is plotted in Fig. 6. Here λdual is obtained
31
32 nemerically by fixing the value of Pm,avg to β0 . The ROC curves have been plotted for γavg = −5dB and
33
34
35 γavg = 0dB. If it is required that the intereference caused by the secondary user to the primary user be
36
37 as low as β0 = 0.005, it can be seen from the curves that, at -5dB, the secondary user will utilize the
38
39
40 free channel for 60% of the times and for the 0dB case it will use the free channel with 99.5%. Hence
41
42 depending on the average SNR, the utilization of the primaty channel and hence the secondary user traffic
43
44
45
can be determined in the dual spectrum sensing case.
46
47
48 VIII. C ONCLUSION
49
50
51 A spectrum sensing scheme based on the higher order statistics has been proposed which exploits the
52
53 fact that kurtosis of Gaussian noise is zero. An analytical expression is derived for an upper bound of the
54
55
56 probability of miss, assuming that the probability of false alarm is fixed to some acceptable level. The
57
58 dual problem is also addressed. It has been shown by simulations that the proposed algorithm outperforms
59
60
61
the energy based scheme in the low SNR region in the case of a single secondary user as well as in the
62
63 case of multiple secondary users.
64
65
 18
1
2
3 A PPENDIX A
4
5 K URTOSIS ESTIMATE STATISTICS
6
7
8 The expressions for the mean and variance of the cumulants of a complex random variable, r, has been
9
10
11 derived in [19]. Using those expressions, we have the following :
12
13
14 µ1 = E(|r|4 ) − 2{E(|r|2 )}2
15
16
17 σ12 (γ, θ) = a00 + a10 γ + [a20 + a21 sin2 (2θ)]γ 2 + [a30 +
18
19 +a31 sin2 (2θ)]γ 3 + [a40 + a41 sin2 (2θ) + a42 sin4 (2θ)]γ 4 ,
20
21
22 σ02 (γ, θ) = c00 (39)
23
24
25 where ai,j s are constants which depend on the type of modulation scheme being employed and number
26
27 2

28 of observation samples used to estimate the kurtosis. γ = σσ2s |h|2 is the instantaneous SNR and θ =
n
29  
30 tan−1 Im[h]
Re[h]
is the instantaneous phase angle of the channel. The values of ai,j s for 16-QAM constellation
31
32
33 are evaluated and given in the Table 1. It is important to note that unlike second order statistics which
34
35 requires only the instantaneous SNR, the higher order statistics depend on the phase information of the
36
37
38
channel also.
39
40
41
42 TABLE I
43
44 M ODULATION CONSTANTS FOR 16-QAM CONSTELLATION
45
46
47 Modulation Scheme Modulation Constants
48
49
24ρ8
n 96ρ8
n 46.08ρ8
n −48.96ρ8
n 33.28ρ8
n
50 16-QAM a00 = N
, a10 = N
, a20 = N
, a21 = N
, a30 = N
,
51
52 128.64ρ8 10.33ρ8 −1.93ρ8 1.74ρ8 24ρ8
n n n n n
53 a31 = N
, a40 = N
, a41 = N
, a42 = N
, c00 = N
54
55
56
57
58
59
60
61
62
63
64
65
 19
1
2
3 A PPENDIX B
4
5 U NBIASED PROPERTY OF THE K URTOSIS ESTIMATE
6
7
8 The kurtosis estimate in (5) is an unbiased estimate. This can be proved by taking the expectation on
9
10
11 both sides of (5),
12  Nobs  Nobs 2 
13 1 X 4 1 X 2
14 E[k̂] = E |r(i)| − 2 |r(i)|
15
Nobs i=1 Nobs i=1
16  Nobs  Nobs  Nobs 
17 1 X 4 1 X 2 1 X 2
=E |r(i)| − 2 |r(i)| |r(j)| ,
18 Nobs i=1 Nobs i=1 Nobs j=1
19
20  Nobs  Nobs N obs 
1 X 4 1 X X
2 2
21 =E |r(i)| − 2 2
|r(i)| |r(j)| , (40)
22 Nobs i=1 Nobs i=1 j=1
23
24 where, E[·] is the expectation operator. Taking the expectation operator inside in summation, in (40), and
25
26
27 combining the like terms and unlike terms, we obtain,
28
Nobs  
29 1 X 4 2 4 2 2
30 E[k̂] = E[|r(i)| ] − 2 Nobs E[|r(i)| ] + (Nobs − 1){E[|r(i)| ]} , (41)
31 Nobs i=1 Nobs
32
33 Since the expectation operator is independent of the index i, we can rewrite (41) as,
34
35    
2 4 1
36 E[k̂] = 1− E[|r| ] − 2 1 − {E[|r|2 ]}2 . (42)
37 Nobs Nobs
38
39 For large Nobs , (42) can be approximated as E[k̂] ≈ k, and hence, k̂ is an asymptotically unbiased estimate.
40
41
42 It has been verified by simulations that for low to medium SNR values, choosing around Nobs = 50 samples
43
44 are sufficient to produce an almost unbiased estimate, (E[k̂ = k).
45
46
47
48 R EFERENCES
49
50
51 [1] Federal Communications Commission, “Spectrum Policy Task Force,” Rep. ET docket no. 02-135, Nov. 2002.
52
53 [2] S. Haykin, “Cognitive Radio: Brain-Empowered Wireless Communications”, IEEE journal on Selected Areas in Communications, vol.
54
23, no. 2, pp. 201–220, Feb. 2005.
55
56 [3] A. Sahai, N. Hoven, and R. Tandra, “Some fundamental limits on Cognitive Radio”, Proc. Allerton Conf. on Commununictions, control,
57
58 and computing, Monticello, Oct. 2004.
59
60 [4] Y. Hur, J. Park, W. Woo, K. Lim, C.-H. Lee, H. S. Kim, and J. Laskar, “A wideband analog multi-resolution spectrum sensing (MRSS)
61
62 technique for cognitive radio (CR) systems”, in Proc. IEEE Int. Symp. Circuit and System, Greece, May 21–24, 2006, pp. 4090-4093.
63
64
65
 20
1
2
3 [5] J. Lunden, V. Koivunen, A. Huttunen, and H. V. Poor, “Spectrum sensing in cognitive radios based on multiple cyclic frequencies”, in
4
5 Proc. of the Second Int. Conf. on Cognitive Radio Oriented Wireless Networks and Communications (CrownCom), Orlando, FL, USA,
6
7 July 31–August 3, 2007.
8
[6] A. Sahai and D. Cabric, “Spectrum sensing: fundamental limits and practical challenges”, in IEEE International Symposium on New
9
10 Frontiers in Dynamic Spectrum Access Networks (DySPAN 05), Baltimore, Md, USA, Nov. 2005.
11
12 [7] Y. Zeng, Y.C. Liang, A.T. Hoang, and R. Zhang, “A Review on Spectrum Sensing for Cognitive Radio: Challenges and Solutions”,
13
14 EURASIP Journal on Advances in Signal Processing, vol. 2010, Article ID 381465.
15
16 [8] W. A. Gardner, “Exploitation of spectral redundancy in cyclostationary signals”, IEEE Signal Processing Magazine, vol. 8, no. 2, pp.
17
18 14-36, 1991.
19
20 [9] V. I. Kostylev, “Energy detection of a signal with random amplitude”, in Proc. IEEE Int. Conf. Commun., New York, Apr. 28–May 2,
21
2002, pp. 1606–1610.
22
23 [10] Y. Zeng and Y. C. Liang,“Covariance based signal detections for cognitive radio”, in Proceedings of the 2nd IEEE International
24
25 Symposium on New Frontiers in Dynamic Spectrum Access Networks (DySPAN 07), pp. 202-207, Dublin, Ireland, April 2007.
26
27 [11] X. Chen and S. Nagaraj, “Entropy Based Spectrum Sensing in Cognitive Radio”, IEEE Wireless Telecommunications Symposium, 2008,
28
29 pp. 57–61, April. 2008.
30
31 [12] A. Mody, “Spectrum Sensing of the DTV in the Vicinity of the Pilot Using Higher Order Statistics”, doc.: IEEE 802.22-07/0370r1,
32
Aug. 2007.
33
34 [13] S. J. Shellhammer, “Spectrum Sensing in IEEE 802.22”, IAPR Wksp. Cognitive Info. Processing, in Santorini, Greece, Jun. 2008.
35
36 [14] Y. Sun, Y. Liu, X. Tan, “Spectrum Sensing for Cognitive Radio Based on Higher-Order Statistics”, 4th International Conference on
37
38 Wireless Communications, Net- working and Mobile Computing, pp. 1–4, Oct. 2008.
39
40 [15] F. F. Digham, M.-S. Alouini and M. K. Simon, “On the energy detection of unknown signals over fading channels”, Proc. IEEE, Int.
41
42 Conf. Commun., Anchorage, AK, USA, pp. 3575–3579, May. 2003.
43
44 [16] W. Zhang, R. K. Mallik and K. B. Letaief, “Cooperative Spectrum Sensing Optimization in Cognitive Radio Networks”, Proc. IEEE
45
ICC, Beijing, China, pp. 3411–3415, May. 2008.
46
47 [17] J. M. Mendel, “Tutorial on Higher-Order Statistics(Spectra) in Signal Processing and System Theory: Theoretical Results and some
48
49 Applications”, Proc. IEEE, vol. 79, pp. 278–305, Mar. 1991.
50
51 [18] J. A. R. Fonollosa, “Sample Cumulants of Stationary Processes: Asymptotic Results”, IEEE transactions on Signal Processing, vol.
52
53 43, no. 4, April. 1995.
54
55 [19] A. Swami and B. M. Sadler, “Hierarchical Digital Modulation Classification Using Cumulants”, IEEE transactions on Communications,
56
vol. 48, no. 3, Mar. 2000.
57
58 [20] A. V. Dandawate and G. B. Giannakis, “Asymptotic Properties and covariance expressions of k-th order sample moments and cumulants”,
59
60 Twenty-Seventh Asilomar Conference on Signals, Systems and Computers, vol. 2, pp. 1186–1190 , Nov. 1993.
61
62 [21] H. V. Poor, An Introduction to Signal Detection and Estimation, 2nd ed., Springer Verlag, 1994.
63
64
65
 21
1
2
3 [22] A. Ghasemi and E. S. Sousa, “Collaborative spectrum sensing for opportunistic access in fading environments”, Proc. IEEE Symp.
4
5 New Frontiers in Dynamic Spectrum Access Networks (DySPAN05), Baltimore, USA, pp. 131–136, Nov. 2005.
6
7 [23] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series and Products, 7th ed., San Diego, CA: Academic, 2007.
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
 22
1
2
3
4
5
6
7
8
9
10
11 10
−1

12
13
Pm,avg

14
15
16
17
18 −2
19 10
20 Kurtosis based method
21 Energy based method
22
23 −15 −10 −5 0 5 10
24 SNR(dB)
25
26
27
Fig. 1. Comparison of energy and proposed kurtosis based spectrum sensing method in terms of probability of miss, for α0 =0.01 and
28
29 Nobs =350.
30
31
32
33
34
35
36
37 10
−1

38
39
40
41
42
43
44
45
Pm,avg

−2
46 10
47
48
49
50
51
52
Average SNR = 8.82dB
53
54 Average SNR = 7.32dB
−3
10
55 10
−1
10
0

56 Pf,avg
57
58
59
Fig. 2. Complementary ROC curves for two different average SNR values using Nobs = 500 samples for kurtosis based scheme.
60
61
62
63
64
65
 23
1
2
3
4
5 −1
6 10

7
8
9
−2
10 10
11
12
13
Pm,avg

14 10
−3

15
16
17 Kurtosis : 1 user
18 −4 Energy : 1 user
10
19 Kurtosis : 2 users
20 Energy : 2 users
21 Kurtosis : 4 users
22 Energy : 4 users
23 0 5 10 15
24 Average SNR(dB)
25
26
27
Fig. 3. Probability of miss vs average SNR for collaborative scheme using OR-fusion rule, for Nobs = 350, which depicts the considerable
28
29 improvement offered by kurtosis based scheme over energy based scheme for 4 collaborative CRs.
30
31
32
33
34
35
36
37 −1
10
38 Simulated
39 Analytical
40
41
42
43
44 −2
10
45
Pm,avg

46
47
48
49
50
−3
51 10
52
53
54
55 5 10 15 20 25 30
Average SNR(dB)
56
57
58
59
Fig. 4. Curves of analytical upper bound on probability of miss, derived in Section V and that obtained by simulation for Nobs =50.
60
61
62
63
64
65