Cyclostationary-based Architectures for

Spectrum Sensing in IEEE 802.22 WRAN

Deepa B

NPOL, DRDO
Kochi, India 682 021
Email: [email protected]
Anand Padmanabha Iyer

Microsoft Research India

Bangalore, India 560 080
Email: [email protected]
Chandra R. Murthy
Dept. of ECE, Indian Institute of Science
Bangalore, India 560 012
Email: [email protected]
AbstractThe well known noise rejection property of the cy-
clostationary spectrum makes it an ideal candidate for spectrum
sensing in low SNR environments such as the IEEE 802.22
WRAN, which stipulates detection of primary signals at 20.8dB.
In this paper, we propose two novel detector architectures
that exploit cyclostationary properties: the Spectral Correlation
Density (SCD), and the Magnitude Squared Coherence (MSC).
Through extensive simulations, both on generated data and real
world ATSC capture data, we show that our detector achieves
an improvement of 2.5dB compared to existing proposals. Addi-
tionally, we compare our proposal against two popular choices
for spectrum sensing in cognitive radio the matched lter
detection and energy detection, and show the superiority of
cyclostationary spectrum sensing in such low SNR environments.
I. INTRODUCTION
A Cognitive Radio (CR) [1] is a device that is aware of its
radio environment, and only transmits in frequency bands that
are not being used by a primary user. A crucial rst step in
enabling CR systems is Spectrum Sensing (SS), where the goal
is the detection of the presence or absence of a signal from a
primary transmitter by the CR. The main challenge in SS is
to quickly detect the signal in a very low SNR environment,
and with high reliability.
The algorithms used for SS can be broadly classied into
three types: Energy Detector (ED), Matched Filter Detector
(MFD) and Cyclostationary Feature Detector. ED, where the
received signal energy in a frequency band of interest is com-
pared against a threshold to detect the presence of a primary,
is the simplest and most popular detector. The MFD correlates
the received signal with a copy of the transmitted signal (e.g.,
known preamble data). Although it is computationally simple,
it assumes knowledge of the primarys signal, which may not
be feasible in general. Cyclostationary feature detectors rely
on the second order cyclostationary characteristics inherent in
all communication signals, i.e., pilot sequences, carrier tones,
etc, that are transmitted on a recurrent basis. One of the goals
of this work is a performance comparison of these three types
of detectors on a level-playing eld.
Spectrum sensing for cognitive radio is a well researched
topic; only a summary of relevant literature is provided here
due to lack of space. In [2], a survey of the SS techniques that
are included in the IEEE 802.22 draft standard is provided.

The authors were at IISc during the course of this work. This work was
supported in part by a research grant from Microsoft Research India.
Excellent overviews of SS and other physical layer issues of
CR can be found in [3], [4]. [21] discusses spectrum sensing
in wideband OFDM CR. SS using the SCD function and its
application to IEEE 802.22 WRAN is discussed in [5]. In [6],
cycle frequency domain prole is used for signal detection and
classication. Additional references on cyclostationarity-based
signal detection include [7] - [10].
The works mentioned above assume perfect knowledge of
the receiver noise level. The effect of noise uncertainty on the
performance of SS has also been explored by several authors,
although the focus has largely been on EDs. In [11] authors
showed that there exists an SNR wall for energy detection, i.e.,
an SNR value below which reliable detection is not possible
even if the sensing duration is increased indenitely.
In this paper, the SS problem is modeled as one of testing
hypothesis H
0
(primary signal is absent) versus H
1
(primary
signal is present). Under H
0
, the signal received at the CR is
modeled as Additive White Gaussian Noise (AWGN). Under
H
1
, the received signal is modeled as a pilot tone signal
modulated by a slowly-varying analog signal, corrupted by
AWGN. This model is the similar to the one in [5], and
is motivated by the primary signal in the 802.22 WRAN
standard, where the (primary) TV broadcast signal contains a
strong pilot tone. The main contributions of this paper are:
Two different cyclostationarity-based detector implemen-
tations are proposed. The rst detector is based on esti-
mating the SCD of the signal, and the second is based
on estimating the MSC. These implementations can be
viewed as improvements on, or alternatives to, existing
cyclostationarity-based detectors. To the best of our knowl-
edge, the MSC detector has not been studied in the context
of CR spectrum sensing.
We compare ED and MFD with the proposed detector. We
also compare the MSC and ED against existing detectors
using real-world ATSC DTV capture data, and demonstrate
that the MSC detector offers a performance advantage of
2.5dB compared to existing detectors.
Initial results on the above work have been published in
[12]. Here, our previous work is extended to account for
frequency at fading channels and the comparison using the
ATSC DTV capture data is more comprehensive, including the
performance with the ED.
978-1-4244-5638-3/10/$26.00 2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.
II. BACKGROUND AND SYSTEM MODEL
A. Introduction to Cyclostationarity
Spectral Correlation Density Function: The cyclic autocor-
relation function R

x
() is dened as [13]
R

x
() lim
T
_
T/2
T/2
x(t +/2)x

(t /2)e
i2t
dt. (1)
A signal exhibits second-order periodicity when its cyclic
autocorrelation function, R

x
() is not zero for some nonzero
frequency , called the cyclic frequency. The Fourier transform
of the cyclic autocorrelation is called the Spectral Correlation
Density (SCD) function:
S

x
(f) =
_

x
()e
i2f
d = S
uv
(f). (2)
Here, u(t) = x(t)e
it
and v(t) = x(t)e
+it
are frequency
shifted versions of x(t). Thus, S

x
(f) is the cross spectral
density of frequency shifted signals u(t) and v(t).
FFT Accumulation Method for estimating the SCD: The
FFT Accumulation method (FAM) [14] incorporates the idea
of time smoothing using a Fourier transform to arrive at a
computationally efcient digital implementation of the SCD
function using N samples from a nite observation interval
of duration t. The complex demodulates
1
X
N
(n, k +

2
)
and X
N
(n, k

2
) are estimated by means of a sliding
N

point FFT, followed by a down-shift in frequency to

baseband. Here, X
N
(n, k +

2
) is the (k +

2
)
th
component
of the N

point FFT output (in baseband) of the n

th
N

point window. That is, n is a time index corresponding to

consecutive N

point windows that are used in the FAM. The

N

point FFT is hopped over the data in blocks of K samples.

The value of K is generally chosen to be N

/4 (i.e., 75%
overlap between adjacent segments) as it allows for a good
compromise between computational efciency and minimizing
cyclic leakage and aliasing. Next, the element-wise product
between the sequences X
N
(n, k +

2
) and X

N
(n, k

2
)
is formed and time smoothed by a P-point second FFT.
The value of N

depends on the frequency resolution f

required, and is given by N

=
fs
f
. The value of P is
given by P =
fs
K
, where f
s
denotes sampling frequency
and f and denote the frequency resolution and cyclic
frequency resolution, respectively. A block diagram of the
FAM implementation is shown in Fig.1.
An Example: In this paper, an Amplitude Modulated (AM)
signal of the form z(t) = s(t) cos(2f
0
t +
0
) is considered.
If the message signal is a tone, i.e., s(t) = cos(2f
m
t), the
SCD function of z(t) is given by S

z
(f) =
Ps
4
[(f f
0
f
m
) +(f f
0
+f
m
)
+(f +f
0
f
m
) +(f +f
0
+f
m
)] = 0,
Ps
4
[(f +f
0
) +(f f
0
)] = 2f
m
,
Ps
4
[(f +f
m
) +(f f
m
)] e
i20
= 2f
0
,
Ps
4
(f)e
i20
= 2(f
0
f
m
).
(3)
1
This notation follows classical literature on the topic, e.g., [14]. Also, with
a slight abuse of notation, the same is used to represent the cyclic frequency
in both discrete and continuous frequency domains.
Input w
N
point
First
FFT
X
X
X
P
point
Sec-
ond
FFT
output
e
j2kn
N

e
j2kn
N

XN (n, k +

2
)
XN (n, k

2
)
x(n)
S

xN
(n, k)N

Fig. 1: FAM block diagram

Magnitude Squared Coherence: The Spectral Autocoher-
ence (also called Spectral Coherence (SC)) of x(t) at cyclic
frequency and spectrum frequency f is dened as [13]
C

x
(f)
S

x
(f)
[S
0
x
(f +/2)S
0
x
(f /2)]
1/2
. (4)
Note that |C

x
(f)| [0, 1]. The difference between the SCD
and the SC is that the SC gives a normalized measure of
cross-correlation between frequency shifted versions of x(t) at
frequencies f /2 and f +/2. It follows from the denition
that the SC is identically zero for all = 0 if and only if x(t)
contains no second order periodicity. Since u(t) = x(t)e
it
and v(t) = x(t)e
it
,
C

x
(f) =
S
uv
(f)
_
S
u
(f)S
v
(f)

uv
(f), (5)
where S
u
(f) and S
v
(f) are the power spectral densities
of u(t) and v(t), and S
uv
(f) denotes the cross spectral
density. The Magnitude Squared Coherence (MSC) is dened
as |
uv
(f)|
2
, and in practice, its estimate, |
uv
(f)|
2
, obtained
from a nite observation interval, is used for signal detection.
B. System Model and Problem Setup
Cognitive radio was rst proposed to be implemented on
TV broadcast service, where the signal is vestigial sideband
modulated. There is a strong pilot tone in the power spectral
density of the TV transmission signal. Taking cue from this,
the SS problem can be stated as testing H
0
versus H
1
, where
H
0
: x(t) = w(t)
H
1
: x(t) = 2
_
P
s
s(t) cos(2f
0
t +
0
) +w(t), (6)
where f
0
is the carrier frequency and
0
is the initial
phase of the carrier. The signal s(t) is modeled as s(t) =
a(t) cos(2f
m
t), with f
m
being the frequency the pilot tone
in the TV signal, and a(t) being an analog waveform. In
practice, the waveform a(t) is slowly varying compared to the
sinusoid, and for short sensing durations, can be considered to
be approximately constant. Therefore, in the sequel, a(t) will
be dropped and its effect included in P
s
, which also models
the transmitted power, path loss and frequency at fading. This
makes the problem analytically tractable, and the model is
similar to ones used in existing literature, e.g., [5]. Moreover,
the efcacy of detectors based on this relatively simple model
will be illustrated using 802.22 RF capture database.
Two effects that are not explicitly captured in this model are
interference and frequency-selective fading. While interference
978-1-4244-5638-3/10/$26.00 2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.
adversely affects the ED and the MFD, its effect on the
cyclostationarity-based detectors would be minimal, the in-
terfering signals unlikely to exhibit a second-order periodicity
at the same cyclic frequency as the primary signal. The effect
of frequency-selective fading is ignored here for simplicity,
but a detailed analysis is presented in [15]. Now, since the
noise does not exhibit any cyclostationarity, its cyclic auto-
correlation function is given by R

w
() =
2
w
() when
= 0, and 0 otherwise. When the signal is present, the cyclic
spectrum of x(t) is
S

x
(f) = S

z
(f) +S

w
(f) (7)
where S

z
(f) is the cyclic spectrum of z(t)
2

P
s
cos(2f
m
t) cos(2f
0
t +
0
). Thus, evaluating S

x
(f)
at an appropriately chosen = 0 helps separate the signal
from the purely stationary AWGN.
III. STATISTICAL CHARACTERIZATION OF THE
CYCLOSTATIONARY DETECTORS
A. Statistical Characterization of the SCD
The SCD estimated using the FAM is S

x
N

(n, k)
N
=
1
P
P1

l=0
_
1
N

X
N
(n +lK, k +

2
)X

N
(n +lK, k

2
)
_
,
where K = N

/J, and there are (1

1
J
)N

samples overlapped
between adjacent segments, and k = fN

/f
s
is the discrete
frequency corresponding to the frequency f. When J = 1,
there is no overlap between adjacent segments, and using
central limit theorem arguments, it can be shown (details
are omitted due to lack of space) that the distribution of
S

x
N

(n, k)
N
is given by
H
0
: S
0
x
N

(n, k
0
)
N
CN(0, 2
2
0
)
H
1
: S
0
x
N

(n, k
0
)
N
CN(S
0
z
N

(n, k
0
)
N
, 2
2
1
),
where CN(0, 2
2
0
) is the circularly symmetric complex Gaus-
sian noise with mean 0 and variance 2
2
0
,
2
0
=

4
w
2Ps
and

2
1
=

4
w
2Ps
_
1 +
Sz
N

(k0+

0
2
)+Sz
N

(k0

0
2
)

2
w
_
. The S
z
N

(k) is
the power spectral density of the N point snapshot of z(t)
evaluated at discrete frequency k, and k
0
is the frequency bin
of interest, determined from the SCD of a clean signal sample.
The decision rule with the SCD detector is

S
0
x
N

(n, k
0
)
N

H
1

H
0

SCD
. (8)
Since

S
0
x
N

(n, k
0
)
N

follows a Rayleigh distribution under

H
0
and Rician distribution under H
1
, the probabilities of false
alarm and detection are given by
P
FA,SCD
= e

2
SCD

2
0
P
D,SCD
= Q
1
_
|S
0
z
N

(n, k
0
)
N
|

1
,

SCD

1
_
. (9)
Here, Q
1
(a, b) is the Marcum Q function of a and b [16].
When J > 1, one needs to account for the correlation
between adjacent spectral demodulates e.g., X
N
(n+lk, k+

2
)
and X
N
(n + (l + 1)k, k +

2
). This is done by multiplying
the variance under H
0
and H
1
by a correction factor
0
and
1
respectively. The correction factor can be found by
moment matching between the theoretical expression and
experimental results. Through simulation, it has been veried
that this distribution approximation gives an excellent match
between the theoretical and experimental Receiver Operating
Characteristic (ROC) curves.
B. Statistical Characterization of the MSC
In the past, several works (e.g., [17], [18]) have dealt with
the statistical distribution of MSC estimates, although not
for the signal model considered here. Here, these results are
adapted to the problem at hand to derive the false alarm and
detection probability performance of the MSC-based detector.
Segmentation Without Overlapping: Let u[n] and v[n] de-
note Nlength complex sequences which are segmented to L
disjoint symbols u
l
[n] and v
l
[n], l = 1, . . . , L, each of length
M = N/L. Let U
l
(k) = F{u
l
[n]}, V
l
(k) = F{v
l
[n]}, where
F denotes the FFT operation and k is the discrete frequency.
The spectral densities are estimated as

S
u
(k) =
L

l=1
|U
l
(k)|
2
,

S
v
(k) =
L

l=1
|V
l
(k)|
2
,

S
uv
(k) =
L

l=1
U
l
(k)V

l
(k). (10)
From this, the MSC is estimated as |
uv
(k)|
2
=
|

Suv(k)|
2

Su(k)

Sv(k)
.
Thus, the decision rule is given by |
uv
(k)|
2
H
1

H
0

MSC
.
The probability density of |
2
| conditioned on the number
of disjoint segments L and the true MSC ||
2
, from [17], is
p(| |
2
| L, ||
2
) =
(L 1)(1 ||
2
)
L
(1 | |
2
)
L2
2
F
1
(L, L; 1; ||
2
| |
2
), (11)
where
2
F
1
is the hypergeometric function [16]. The cumulative
distribution function is
P
_
| |
2
| L, ||
2
_
= | |
2
_
(1 ||
2
)
(1 ||
2
| |
2
)
_
L
L2

l=0
_
(1 | |
2
)
(1 ||
2
| |
2
)
_
l
2
F
1
(l, 1 L; 1; ||
2
| |
2
). (12)
The probability distribution of | |
2
under H
0
is obtained by
simply setting ||
2
= 0 in the above pdf to get
p(| |
2
| L, ||
2
= 0) = (L 1)(1 | |
2
)
(L2)
(13)
Hence, the probability of false alarm is
P
FA,MSC
= (1
MSC
)
(L1)
. (14)
The probability of detection is obtained by integrating (11)
between
MSC
and 1, and is given by
P
D,MSC
= 1 P
_
| |
2

MSC
| L, ||
2
_
, (15)
978-1-4244-5638-3/10/$26.00 2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
FA
P
D


SCD Detector
SCD Detector (theoretical)
MSC Detector
MSC Detector (theoretical)
MSC Detector (50% overlap)
MSC Detector (50% overlap, theoretical)
(a) MSC vs SCD detector, SNR=24 dB
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
P
D
P
FA


MSC (no overlap)
MSC (75% overlap)
ED
ED (theoretical)
MFD
MFD (theoretical)
MFD (1.7 sample sync error)
(b) ED, MFD and MSC detectors at SNR=22dB
(no noise uncertainty)
30 28 26 24 22 20 18 16 14 12 10 8
10
4
10
3
10
2
10
1
10
0
SNR
P
M
D


MSC
MSC (75% overlap)
ED
MFD
MFD (1 sample sync error)
MFD (1.5 sample sync error)
(c) P
MD
vs SNR, noise uncertainty=0.5dB,
P
FA
=0.01
Fig. 2: Comparison of different detectors with and without noise uncertainty
where P(| |
2

MSC
|L, ||
2
) is given by (12).
Overlapped Segmentation: The case of overlapped segmen-
tation can be analyzed using the tools rst developed in [19]
to nd the effective number of degrees of freedom in the
decision statistic under H
1
. The details are omitted here,
but the performance with overlapped segmentation will be
illustrated in the next section.
IV. SIMULATION RESULTS
A. Generated Data
First, we present the simulation results on generated data.
For Figs.2 and 3, a dataset with N = 8192 samples generated
according to (6) was used, which corresponds to L = 16
non-overlapping segments with M = 512 point FFT
2
. The
sampling frequency was xed at 4KHz with f
0
and f
m
as
512Hz and 16Hz respectively. The pilot tone signal considered
here is coherent at = 2f
0
and f = f
m
. So, the MSC/SCD
was evaluated at a frequency f = f
m
and a cyclic frequency
= 2f
0
. It can be shown that the MSC at = 2f
0
and
f = f
m
is related to the transmit power, noise variance and
the FFT size via
||
2
=
(P
s
/4)
2
(P
s
/4 +
2
w
/M)
2
(16)
The above equation was used to nd the ||
2
in (12). Samples
were generated according to (6) with P
s
= 1 and w(t) being
an AWGN with variance equal to the inverse of the SNR.
In Fig.2(a), the ROC of the SCD and MSC detectors are
plotted. The excellent match between the theoretical curves
and the experimental results can be readily observed in the g-
ure. It can be seen that the MSC detector performs marginally
better than the SCD detector. The performance is further
improved by overlap processing. Hence, we select the MSC
detector as our detector of choice in all further comparisons.
Fig.2(b) compares the performance of the different detectors
under no noise uncertainty, i.e., when the noise variance
2
is
known. The matched-lter detector performs the best, followed
by the MSC detector with 75% overlap. It is also seen that
2
We also modeled s(t) as a(t) cos(2fmt) where a(t) is a slowly varying
signal compared to the sinusoid, and obtained very similar performance.
the MSC detector outperforms the energy detector. The MFD,
on the other hand, is sensitive to synchronization errors.
The effect of noise uncertainty is illustrated in Fig.2(c). It
is assumed that the noise uncertainty range is x = 0.5dB
(corresponding to, for example, a temperature change of about
35 degrees celsius around room temperature). Therefore, for
xing the threshold, it is assumed that under H
0
, the signal
is AWGN with variance
2
w
. For calculating the P
D
, it
is assumed that the primary signal is corrupted by AWGN
with variance (1/)
2
w
. The gure shows that ED is severely
degraded by the noise uncertainty, while the MSC detector is
virtually unaffected. The performance degradation of the MFD
due to synchronization errors is also seen in the gure.
Now we consider the detection performance in the case of a
Rayleigh fading signal against average Energy to Noise ratio
(ENR). With exact knowledge of the noise variance, the ED
and MSC have similar performance and both give better results
than the SCD detector (Fig.3(a)). However, in the presence of
noise uncertainty, the performance of ED degrades and the
MSC and SCD outperform the ED (Fig.3(b)). The MFD has
to be modied to account for fading, resulting in the so-called
Quadrature Matched Filter (QMF). The derivation of the QMF
detector is omitted due to lack of space, but its performance
is depicted in the gure. Clearly, an exact knowledge of the
primary waveform offers a signicant performance advantage;
however, this is not a reasonable assumption in practice.
B. ATSC DTV Capture
Finally, the proposed MSC detector was tested on the DTV
RF signal capture database (Table I) from the 802.22 WRAN
working group, and the performance was evaluated using the
procedure described in the standard. The result is shown in
Fig.4 for the twelve data sets labeled A through L. Averaged
over all the data sets, the MSC detector achieves a P
MD
of
0.1 at an SNR of about 27.5dB and P
FA
of 0.1 with a
19.03ms sensing time. In comparison, the detector proposed in
[5] achieves the same performance at an SNR of 25dB. More
details on this test can be found in [15]. For comparison, we
also provide the performance of ED on this dataset in Fig.4(c).
The ED implementation here has been optimized over the FFT
size and bin index for obtaining the best performance. It is
978-1-4244-5638-3/10/$26.00 2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.
0 5 10 15 20 25 30 35 40
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Average energytonoise ratio(dB)
P
D


QMF
QMF (theoretical)
SCD
MSC (75 % overlap)
ED
ED theoretical
(a) No noise uncertainty
0 5 10 15 20 25 30 35 40
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Average nominal energytonoise ratio(dB)
P
D


QMF
QMF theoretical
SCD (no overlap)
MSC (75 % overlap)
ED
ED theoretical (no noise uncertainty)
(b) Noise uncertainty of 0.5dB
Fig. 3: Comparison of detection performance of Rayleigh fading AM modulated sinusoid for different detectors (PFA = 0.1)
35 30 25 20
10
3
10
2
10
1
10
0
SNR (dB)
P
M
D


A
B
C
D
E
F
G
H
I
J
K
L
AVG
Est P
FA
(a) MSC detector, max across from the data
35 30 25 20
10
3
10
2
10
1
10
0
SNR (dB)
P
M
D


A
B
C
D
E
F
G
H
I
J
K
L
AVG
Est P
FA
(b) MSC detector, estimated from the data
30 25 20 15 10 5 0
10
5
10
4
10
3
10
2
10
1
10
0
SNR
P
M
D


A
B
C
D
E
F
G
H
I
J
K
L
AVG
Est P
FA
(c) Energy detector
Fig. 4: PMD vs SNR for different detectors PFA = 0.1, sensing time= 19.03 ms
A WAS 06 34 06092000 REF G WAS 51 35 05242000 REF
B WAS 3 27 06022000 REF H WAS 68 36 05232000 REF
C WAS 32 48 06012000 OPT I WAS 86 48 07122000 REF
D WAS 47 48 06132000 opt J WAS 311 35 06052000 REF
E WAS 49 34 06142000 opt K WAS 311 36 06052000 REF
F WAS 49 39 06142000 opt L WAS 311 48 06052000 REF
TABLE I: DTV capture le and its symbol in gures
interesting to note that the performance of the ED is within
4 5dB of the best MSC detector.
V. CONCLUSIONS
This paper considered the problem of detection of a pri-
mary transmitters signal using cyclostationary properties. Our
proposed scheme achieves a P
MD
of 0.1 at an SNR of
27.5dB and P
FA
of 0.1 with a 19.03ms sensing time, which
translates to a 2.5dB improvement over existing proposals.
We also compared our proposal against two of the popu-
lar detection techniques, the energy detection and matched
lter detection, and exposed the shortcomings of the latter
schemes at low SNR environments with noise uncertainty and
timing synchronization error, respectively. Thus, the proposed
cyclostationarity detectors are strong candidates for spectrum
sensing when the signal of interest has a nonzero spectral
correlation at some known cyclic frequency.
REFERENCES
[1] J. Mitola-III Jr, Cognitive radio: An integrated agent architecture for software
dened radio, Ph.D. dissertation, Royal Institute of Technology, Sweden, 2000.
[2] S. Shellhammer, Spectrum sensing in IEEE 802.22, Proc. of IAPR Workshop on
Cognitive Information Processing, Jun. 2008.
[3] I. F. Akyildiz et al. NeXt generation/dynamic spectrum access/cognitive radio
wireless networks: A survey, Comp. Networks, vol. 50, Sep. 2006.
[4] D. Cabric, S. M. Mishra, and R. W. Brodersen, Physical layer design issues
unique to cognitive radio systems, IEEE PIMRC, 2005.
[5] H.-S. Chen, W. Gao, and D. G. Daut, Spectrum sensing using cyclostationary
properties and application to IEEE 802.22 WRAN, Globecom, 2007.
[6] K. Kim et al. Cyclostationary approaches to signal detection and classication
in cognitive radio, IEEE DySPAN, 2007.
[7] J. Lund en, V. Koivunen, A. Huttunen, and H. V. Poor , Spectrum sensing in
cognitive radios based on multiple cyclic frequencies, CrownCom, 2007.
[8] X. Chen, W. Xu, Z. He, and X. Tao, Spectral correlation based multi-antenna
spectrum sensing technique, IEEE WCNC, 2008.
[9] H. Sadeghi and P. Azmi, Cyclostationarity-based cooperative spectrum sensing
for cognitive radio networks, IST, Aug. 2008.
[10] J. Lund en et al. , Collaborative cyclostationary spectrum sensing for cognitive
radio systems, IEEE Trans. Sig. Proc., Nov. 2009.
[11] R. Tandra and A. Sahai, SNR walls for signal detection, IEEE J. of Sel. Topics
in Sig. Proc., vol. 2, no. 1, pp. 417, Feb. 2008.
[12] D. Bhargavi and C. Murthy, Performance comparison of energy, matched-lter
and cyclostationarity-based spectrum sensing, Proc. IEEE SPAWC, 2010.
[13] W. A. Gardner, Exploitaion of spectral redundancy in cyclostationary signals,
IEEE Sig. Proc. Magazine, pp. 1435, Apr. 1991.
[14] R. S. Roberts, W. A. Brown, and H. H. Loomis, Computaionally efcient
algorithms for cyclic spectral analysis, IEEE Sig. Proc. Magazine, Apr. 1991.
[15] B. Deepa, Design and Analysis of Cyclostationarity Based Spectrum Sensing for
Cognitive Radio, M. E. Thesis, Indian Institute of Science, India, May 2009.
[16] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series and Products, 1994.
[17] G. C. Carter, C. H. Knapp, and A. H. Nutall, Estimation of the magnitude-squared
coherence function via overlapped fast fourier transform processing, IEEE Trans.
on Audio and Electroacoustics, vol. AU-21, no. 4, pp. 337344, Aug. 1973.
[18] H. Gish and D. Cochran, Invariance of the magnitude-squared coherence estimate
with respect to second-channel statistics, IEEE Trans. on Acoustics, Speech and
Sig. Proc., vol. ASSP-35, no. 12, pp. 17741776, Dec. 1987.
[19] R. Lugannani, Distribution of the sample magnitude-squared coherence obtained
using overlapped fourier transforms, IEEE ICASSP, Apr. 1981.
[20] H. Urkowitz, Energy detection of unknown deterministic signals, Proc. of the
IEEE, vol. 55, no. 4, pp. 523531, Apr. 1967.
[21] C. Hwang, G. Lai, and S. Chen, Spectrum sensing in wideband OFDM cognitive
radios,, IEEE Trans. Sig. Proc., Feb. 2010.
978-1-4244-5638-3/10/$26.00 2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.