IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO.

2, FEBRUARY 2010 709

Spectrum Sensing in Wideband

OFDM Cognitive Radios
Chien-Hwa Hwang, Member, IEEE, Guan-Long Lai, and Shih-Chang Chen

Abstract—In this paper, spectrum sensing of an orthogonal fre- To achieve the goal of CR, it is a fundamental requirement
quency division multiplexing (OFDM) based cognitive radio (CR) that the cognitive user performs spectrum sensing to identify
is addressed. The goal is to identify the portions of the spectrum the portions of spectrum that are unused at a specific time. Thus,
that are unused by primary user systems and other CR systems,
called existing user (EU) systems altogether, with the emphasis on over various frequency bands, sensing of primary user systems
conquering the challenge imposed by multipath fading channel. and other CR systems in operation should be performed reg-
The sensing of EU systems consists of two steps. In the first step, ularly. Each of the systems that have already been operating
the maximum likelihood (ML) estimates of the frequency bands in a particular band of interest is called an existing user (EU)
of EU systems are calculated; in the second step, detection is per- system. Digital signal processing techniques can be employed
formed at each suspected band to decide whether an EU system is
truly in operation. The idea is that an EU system appears at a seg- to promote the sensitivity of the EU signal sensing. Three com-
ment of continuous subcarriers. This fact can be exploited by em- monly adopted methods are matched filtering, energy detection
ploying measurements at a continual subcarriers and executing the [3]–[9], and signal feature detection with the cyclostationary
sensing along the frequency domain. An autoregressive (AR) model feature most widely adopted [10]–[13]. Moreover, cooperation
is adopted to track the variation of the received EU signal strength among cognitive users in spectrum sensing can not only reduce
along frequencies. It is shown by simulations that the proposed
spectrum sensing algorithm is robust in a severe frequency-selec- the detection time and thus increase the agility, but also alle-
tive fading channel. viate the problem that a cognitive user fails to detect EU signal
because it is located at a weak-signal region [7]–[9], [14]–[18].
Index Terms—Autoregressive model, cognitive radio (CR), dy- For an overview of these approaches and their properties, see
namic programming, orthogonal frequency division multiplexing
(OFDM), spectrum sensing. [19]–[21].
It is concluded in [22] that orthogonal-frequency-division
multiplexing (OFDM) is the best physical layer candidate for
I. INTRODUCTION a CR system since it allows easy generation of spectral signal
waveforms that can fit into discontinuous and arbitrary-sized
spectrum segments. Besides, OFDM is optimal from the view-
ADIO spectrum is the medium for all types of wireless point of capacity as it allows achieving the Shannon channel
R communications, such as cellular phones, satellite-based
services, wireless low-powered consumer devices, and so on.
capacity in a fragmented spectrum. Owing to these reasons, in
this paper, we consider the problem of spectrum sensing in an
Since most of the usable spectrum has been allocated to existing OFDM based CR system.
services, the radio spectrum has become a precious and scarce The main concern of this paper is to develop spectrum sensing
resource, and there is an urgent concern about the availability of algorithms that are robust to multipath fading channels. One
spectrum for future needs. The solution to the spectrum scarcity of the main challenges in performing reliable spectrum sensing
problem is dynamically looking for the spectrum “white spaces” arises from the fading channel of a wireless link. There are two
and using them opportunistically. Cognitive radio (CR) tech- types of fading, i.e., shadowing fading and multipath fading.
nology, defined first by Mitola [1], [2], is thus advocated as a The former does not cause large fluctuations in signal strength
candidate for implementing opportunistic spectrum sharing. over small changes of receiver’s location, while the latter gen-
erally results in significant variations of signal strength with a
small change of location [23]. To alleviate the difficulties re-
Manuscript received December 22, 2008; accepted August 19, 2009. First sulting from fading channels, it is advantageous to make use of
published September 18, 2009; current version published January 13, 2010. The
associate editor coordinating the review of this manuscript and approving it for diversity gains in various domains. Concerned with shadowing
publication was Prof. Hongbin Li. This research was supported in part by the fading, cooperative spectrum sensing among cognitive users sit-
National Science Council, Taiwan, under Contract NSC-95-2219-E-007-005,
and in part by Chung-Shan Institute of Science and Technology, Taiwan, under
uated at different locations can be employed, which is the user
Contract BU95W028. cooperation diversity. Regarding multipath fading (the concern
C.-H. Hwang and are with the Institute of Communications Engineering, Na- of this work), since channels may be quite different even for two
tional Tsing Hua University, Hsinchu 30013, Taiwan (e-mail: chhwang@ee.
nthu.edu.tw).
closely located receivers, a spectrum sensing device equipped
G.-L. Lai was with the Institute of Communications Engineering, National with multiple antennas is able to exploit the spatial diversity.
Tsing Hua University, Hsinchu 30013, Taiwan (e-mail: tim910282@hotmail. Moreover, particularly convenient for OFDM based CR, spec-
com). trum sensing using observations along the frequency domain
S.-C. Chen is with the Realtek Semiconductor Corporation, Hsinchu 300,
Taiwan (e-mail: iamcsc@realtek.com.tw). can take advantage of the fact that, once an EU system appears,
Digital Object Identifier 10.1109/TSP.2009.2032453 several subcarriers in a row are interfered simultaneously. Thus,
1053-587X/$26.00 © 2010 IEEE

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.
710 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO. 2, FEBRUARY 2010

spectrum sensing can be transformed into the model change de- where and are the complex-valued contri-
tection [24], which aims at detecting changes in the characteris- butions resulting from EU signal and additive white Gaussian
tics of physical systems and is an important practical problem. noise (AWGN), respectively.
However, due to frequency selectivity of multipath fading, the Suppose there are EU systems operating in the frequency
received EU signal strength may vary with frequencies. Part of band of interest, with the th system occupying
the band of the EU system may be deeply faded, making it dif- the band extending from the th to the th subcarriers of
ficult to detect its presence. A remedy is to endow the spec- the CR OFDM system. It is assumed that the frequency bands
trum sensing device with the capability of tracking EU signal of EU systems, i.e., with , are disjoint
strength variation in the frequency domain. In so doing, even and the knowledge of the number of EU systems and their fre-
though some portion of the band is severely faded, weak signals quency bands are unknown to the sensing device. The sensing
at this region can still be identified as a part of the entire signal. algorithm addressed in this paper decides how many EU sys-
In this paper, it is assumed that an unknown number of EU tems and where they are based on the observation
systems are operating in the spectrum segment under sensing. with denoting the observation
The detection of EU systems is composed of two stages: i) es- length at each subcarrier.
timating the most likely bands of EU systems, and ii) for each
suspected band, testing whether or not an EU system is really III. SENSING OF EU SYSTEMS IN A COGNITIVE OFDM SYSTEM
in operation. For the first stage, an autoregressive (AR) model
The observation is processed in two steps. We
is adopted to track the changes of signal magnitude in the fre-
first perform estimation for the most likely values of and
quency domain, and the maximum likelihood (ML) estimates of
, and we next execute a detection for each suspected
the parameters of concern are calculated. It turns out that a huge
band to decide whether or not an EU system is truly in op-
number of search is required for estimating the frequency bands
eration. The estimation and detection procedures are to be
of EU systems. A dynamic programming (DP) technique is em-
discussed in Sections III-A and B, respectively.
ployed to reduce the complexity in searching. For the second
stage, due to insufficient knowledge of statistics of EU signals,
A. Estimating Frequency Bands of EU Systems
an energy detector is employed. In each of stages i) and ii), both
single antenna and multiple antennas are considered. For each , we build an observation from
The organization of this paper is summarized as follows. In such that the estimation of the frequency bands of
Section II, the problem statement as well as the signal model of EU signals is based on the measurements along the frequency
a CR OFDM system interfered by an EU signal are described. domain. Assume it is known that the variances of in the
Two stages of spectrum sensing, i.e., estimation and detection, real and imaginary parts are both . We choose
are given in Section III when the sensing device has a single
antenna. In Section IV, results of Section III are extended to the
case that multiple antennas are equipped. Simulation results of
the proposed algorithm are demonstrated in Section V. Finally,
we conclude this paper in Section VI.
(2)
II. PROBLEM STATEMENT
with . It is seen
Consider a wideband cognitive OFDM system with sub- is the periodogram of the received signal at the th subcarrier
carriers. It is required that the CR system identifies the portions averaged over OFDM symbols, and the periodogram is an
of the spectrum that are unused at a specific time. Thus, over estimate of the true spectrum of a signal. The observation for EU
various frequency bands, sensing of primary user systems and system frequency bands estimation is , where we define
other CR systems in operation is performed regularly. Each of .
the systems that have already been functioning within a band We first consider the case that there is one EU system, i.e.,
of interest is called an EU system. The CR system performing , operating in the target band, and then we extend the
sensing is wideband in the sense that its bandwidth can accom- result to the more practical situation that there are unknown
modate more than one EU systems. number of EU systems in the band.
When a sensing of EU systems is performed, the CR OFDM Suppose that an EU system is present at subcarriers .
system ceases transmission. The received signal is amplified and We assume that correlation exists among with neigh-
frequency down-converted from the radio frequency (RF) under boring subcarrier indices. A first-order autoregressive (AR)
sensing to the baseband. After analog-to-digital (A/D) conver- model is adopted to fit the signal .
sion, cyclic prefix removal and some necessary processing, the That is,
output signal is passed through a -point discrete Fourier trans-
form (DFT). If an EU signal is present at the frequency of sub- (3)
carrier , the th DFT output corresponding to the th OFDM
symbol is given by where and are model parameters, and is a white
Gaussian random process with variance . All of , , and
(1) are unknown. If there is no EU signal in subcarriers ,

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.
HWANG et al.: SPECTRUM SENSING IN WIDEBAND OFDM COGNITIVE RADIOS 711

the model in (3) is revised by letting the model parameters Since we have set , the first term at the right-hand side
, i.e., of (10) is a constant, and the ML estimates of and are
obtained by solving
(4)
(11)
where is a white Gaussian process with unknown vari-
ance . whose solution can be found efficiently by the technique of DP
Since we have assumed , for notational simplicity, [25], [26] as follows. To resort to DP, some modifications to the
we use and to denote the starting and ending subcarrier derivation presented above are required. We can decompose the
indexes of the EU system. According to models (3) and (4), we second line of (5), i.e., the band without EU signal, into two
can perform ML estimation for and . Given the observation terms corresponding to segments and ,
, the likelihood function parameterized by unknowns , and the variance of in (4) is estimated individually
, , , and is [see (5), shown at the bottom of the page]. at the two segments. By doing so, the optimization in (11) can
Note that, in the third line of (5), is not considered in be rewritten as
the sum because the EU signal is present in
but absent in . The ML estimates of unknowns , ,
and are given by

(6)

(7)
(8)

where , , and
(12)
.. .. (9)
. .
where , , and and
have their expressions given in (7) but here they are affixed with
By plugging the above estimates into the likelihood function (5), subscripts to denote the interval in which the estimates are per-
its logarithm is given as formed. For , we define equations (13)–(14),
shown at the bottom of the page, and we let

(10) (15)

(5)

(13)

(14)

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.
712 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO. 2, FEBRUARY 2010

The optimization in (12) is equivalent to where , , and are ML esti-

mates obtained in (7). Note that in (21) is due
to the first term at the right-hand side of (10).
In the following, we first discuss the estimation of
(16) for a given , and we then consider how the value of can
be specified. Let the sum of the first two terms in (21) be repre-
sented by . For a fixed value of
, we can estimate by solving

(17)

(18)

To solve (18), for each , we compute in (15)

and the associated argument yielding . This argument is (22)
denoted as . Then, by (18), we obtain
where and are defined in (13) and (14), respec-
tively. Similarly to the case of , the optimization can be
solved by the DP technique. With and ,
due to the inner minimization of (17), we have we define as the minimal accumulated mod-
eling errors in segments

Note that the optimizations of (15) and hence of (18) are con-
cerned with solving in sequentially. Since
is the least square (LS) estimate, they can be ob- assuming EU systems are
operating at subcarriers
tained by the formulas of sequential LS estimates, e.g., [27, pp. , and all other segments
242–251]. contain only AWGN. Similarly, ( ) is defined
Now, we consider the situation that there may be more than as the minimal accumulated modeling errors in segments
one EU systems operating in the band of interest. Assume there
are EU systems with the th system operating at subcarriers
, and all of are disjoint. With
this setting, subcarriers can be divided into
segments, where each of where segments of have EU signals,
and all others contain only AWGN. Specifically, we have

(19)

contains samples of white noises, and every segment

(20)

contains an EU signal. Models of (4) and (3) are used to repre-

sent the signals in segments of (19) and (20), respectively. That
is, AR models and AWGN are adopted interlacedly to fit the ob-
servation . All of and are
unknown and need to be estimated. The target function for esti-
mation is given by

(21)

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.
HWANG et al.: SPECTRUM SENSING IN WIDEBAND OFDM COGNITIVE RADIOS 713

Fig. 1. DP algorithm for solving (22).

and so on. With the initial condition , it can be Suppose that the number of EU systems existing in the band
shown that and are is upper-bounded by . We solve (22) with . When
related recursively by the true number of EU systems is less than and a deep fade
occurs, two situations may happen in the solution of (22). First,
when the deep fade takes place at the middle region of the band,
the EU system may be judged as two, and the portion of the
band under deep fade is not found. Secondly, if the fade is at
(23)
the border, only part of the band is sensed, and a piece of band
containing only AWGN is falsely alarmed as the band of an
(24) EU system. Fig. 2 depicts the above two circumstances when
only one EU system is present at but is set as 2. Note
For each in (23), the value of argument that yields that, with the fading in Fig. 2(b), it is highly probable that the
is denoted by . Similarly, for each in (24), the value of second detected band starts at subcarrier . This is explained
argument that yields is denoted by . The pro- as follows. The signal strength at the left side of the EU band
cedure to computing the solution of (22) is stated in Fig. 1. The is strong, so it is natural that an AR model is chosen to fit it.
estimates can be determined back- Suppose that is modeled by an AR. Due to the pattern
ward by , , of models of AWGN-AR-AWGN-AR-AWGN when ,
, and so on. To summarize, we have the band should be modeled either by AWGN
(in this case, is modeled by AWGN-AR-AWGN) or
by AWGN-AR-AWGN. However, it is unlikely that the former
happens because the signal at results in an inaccurate
estimate of noise variance and hence a large value of
. Thus, the right side of the EU band is modeled by
AWGN, and is by AR-AWGN. The reason that
The benefit of DP is obvious. If we were to enumerate all pos- the second detected band starts exactly at subcarrier is
sible combinations of , it would be about for because it is the true transition point of the band of EU system
EU systems. However, from the algorithm in Fig. 1, it is seen and the band of AWGN.
the complexity of DP increases linearly with .
Regarding the estimate of , it is conventionally carried out B. Detecting EU Signals
by incorporating a penalty term into (21) using criteria such as The unfound bands due to deep channel fade can still be
Akaike information criterion (AIC) [28], minimum description hopefully recovered and detected if detection is performed at
length (MDL) [29], Bayesian information criterion (BIC) [30], each of bands ( estimated bands plus re-
and so on, and searching for the minimum with respect to . maining bands), and the final result of spectrum sensing is the
However, the specification of the penalty term requires the ex- union of the bands that are judged to have EU systems in op-
plicit model of the modeling error, which is generally unavail- eration.2 For example, when , and the bands
able when an EU system is present; moreover, the criteria are are labeled as . Suppose that the detections per-
usually based on asymptotic approximations.1 Since the accu- formed at , , and return positive result. Then, the two
racy of the estimate of is crucial in the performance of spec- bands , with ( and are adja-
trum sensing, we do not tackle the problem of estimating di- cent) and , are the final result of spectrum sensing.
rectly. Instead, we make a detour by specifying an upper bound Let us call the task performed in solving (22) as the estima-
for ; by means of some extra processing (to be described in tion stage and the detection mentioned above as the detection
Section III-B), it is still possible to find the estimates of EU stage. We will demonstrate later by simulations that, at a mod-
bands. erate signal-to-noise-ratio (SNR), the detection stage is able to
1A precise estimate of is difficult even for a very simple signal model. For m compensate for the errors made at the estimation stage. That is,
example, simulation results for estimation of number of change points based on the band in Fig. 2(a) can be detected, and the falsely
BIC are given in [31], in which the th segment is distributed as ( 1 I) with i N ;
1 an all-one vector and a constant having = for = . We have tried  6  i6 j 2For the M + 1 remaining bands, we can also set some rules to exclude those
several criteria to estimate . However, simulation results showed that these m very unlikely to be EU bands. As our objective is to re-include the bands that are
criteria do not lead to an accurate estimate of . m unfound due to deep fade, the rule of exclusion may be based on the bandwidth.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.
714 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO. 2, FEBRUARY 2010

Fig. 2. An EU system operates at subcarriers [a; b], and M is set as 2. The received EU signal strength is plotted as the shaded region. (a) A deep fade occurs at
the middle region of the band, and the result of band estimation is [a; c] and [d; b], and (b) deep fade occurs at the border, and the estimated bands are [a; c] and
[b + 1; d].

alarmed band in Fig. 2(b) can be correctly identified where by the central limit theorem, the distributions of the
as a band of AWGN. second and third terms at the right-hand side can be approx-
The goal is to test whether an EU system is truly present in imated as and
each of the suspected bands found in the estimation stage. For , respectively. Thus, given a threshold
notational simplicity, we address the detection of EU signal in , the probability of detection is
. Commonly adopted detection algorithms in CR contain
matched filtering, energy detection, and cyclostationary signal
feature detection [19]. However, due to the lack of knowledge
(26)
about EU signals, energy detection becomes the only choice
among the three. Actually, this is a very natural selection as the
receiver structure of an OFDM based CR is convenient for the
implementation of an energy detector. where is the right-tail probability of evaluated at
The detector is designated as . To relate the probability of miss and the probability
of false alarm, we approximate the test statistic under as
Gaussian, and the threshold in (25) has

where represents EU signal is absent, and otherwise. (27)

Under , the scaled test statistic is a central chi- Plugging (27) back to (26) and by some manipulations, we ob-
squared distribution with degree of freedom , tain
i.e., . Using the Neyman–Pearson philosophy, the
detection threshold is set as

(25) (28)
where is the average
SNR of the EU signal in the band.
with the target false alarm probability , where is
the inverse function of the right-tail probability evaluated at IV. MULTIPLE ANTENNAS
for distribution . To improve the performance of spectrum sensing, we may
Suppose that an EU system is present, and its average power equip the sensing device with ( ) receive antennas.
at the th subcarrier is denoted by , given as We suppose that, under , each antenna observes the same
EU transmission, and the channels from an EU to all antennas
are independent. The output of each antenna is processed
individually by a DFT. In the presence of an EU signal, let
stand for the th symbol DFT
The test statistic can be expressed as output of the th receive antenna at subcarrier , where
and are contributions from EU signal and AWGN,
respectively. In the following, the superscript is affixed to
a notation of single-antenna to represent the counterpart of the
notation in the th antenna for the multiantenna case. If no
ambiguity occurs, new notations augmented with superscript
will not be defined again.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.
HWANG et al.: SPECTRUM SENSING IN WIDEBAND OFDM COGNITIVE RADIOS 715

In the estimation stage, the observation from the th antenna

is

Consider that there is one EU system operating in the

band. Let the likelihood function at the th antenna be
. We assume the noise
variances and are common at all antennas, and
and are distinct for different ’s. Since the received sig-
nals at antennas are mutually independent, the joint likelihood
is given by

(29)

Fig. 3. Performance of an energy detector with the processing gain

We estimate the unknowns individually at each an- D = KN (q 0 q + 1) .
tenna, and and are estimated from the combined obser-
vations of all antennas. By following the same steps in the case
of single antenna, it can be shown the optimization presented in under and as
(16)–(18) still applies except that and should
be revised as

(30)

under . It can be shown the rates of miss and false alarm are
and [see (31), shown at the bottom of the page], respectively. On related by
the other hand, in the case that there may be arbitrary number
of EU systems in the band, the algorithm of Fig. 1 still holds
except that and in (23) and (24) are replaced with
(30) and (31), respectively, and the initial condition of is (32)
revised to be .
In the detection stage, to perform detection at the band where
, the test statistic is

Compare (28) and (32), we can see the effective observation

where is somewhat different in the form from its counterpart length is times that of the single antenna case; thus, the
in single antenna to take into account the distinct noise variances adoption of multiple antennas decreases the probability of miss
at antennas. With a target false alarm rate , the which often results from deep channel fade. The performance
of energy detection is plotted in Fig. 3 with processing gain
threshold is given as
.

V. SIMULATION RESULTS
By the central limit theorem, the test statistic is distributed as Throughout the simulations, an OFDM CR with the number
of subcarriers is adopted, and the observation length
is set as . The fading channels used in simulations are

(31)

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.
716 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO. 2, FEBRUARY 2010

and the estimates of EU bands . A hit

Fig. 4. Performance of the estimation stage with respect to the average SNR
of EU signal, frequency selectivity of fading channels, and the number of EU
systems. The length of observation is . The value of
chosen, and the estimator has single antenna.
Fig. 5. Performance of the estimation stage with respect to the average SNR of
EU signal, the number of antennas, and the number of EU systems. The length
of observation is N = 100
. The value of is correctly chosen, and the number
of channel paths is equal to 16.
multipath Rayleigh with uniform power delay profiles. Each of
the figures and tables is generated by Monte Carlo simulations
with 1000 runs. The channels change from one Monte Carlo run
to another. When multiple antennas are used, channels from an
EU to different antennas are independent. Thus, diversity gain
is obtained by the combining across antennas. We first show
the performance at the estimation stage in Figs. 4 and 5 and
Table I and then the overall performance of spectrum sensing,
i.e., composite of estimation and detection, in Tables II and III.
Before presenting the performance of the estimation stage, we
give the definition of a “hit”. We use and to denote the true
number of EU systems and the chosen number in running (22),
respectively. As is the upper bound of , we always have
. Given the true EU bands
happens if each EU band is equal to some estimated
EU band with one-subcarrier estimation error toler-
ance at both sides. That is, for each , one can find
some such that and .
In Figs. 4 and 5, we depict the performance of the estimation
stage when the value of (number of EU systems operating in
the band) is correctly chosen in running (22). Fig. 4 compares
the performance of the EU band estimator for fading channels
with different frequency selectivity when the number of antenna
. The numbers of multipaths in comparison are 4 and 16,
and we consider number of EU systems 1, 2, and 3. EU
systems #1, #2, and #3 appear at subcarriers ,
and , respectively. They all have bandwidths equal to 15.
When , #1 EU system is in the band; when , sys-
tems #1 and #2 exist, and so on. Every EU system has the same
average power. The horizontal axis of Fig. 4 is the ratio of the
average power of an EU signal to AWGN; the vertical axis is the
probability of hit. It is seen that the performance of estimation
is correctly N = 100 m degrades as the number of EU systems increases, and the fre-
quency selectivity of the channel deteriorates the performance
as well. Fig. 5 compares the performance of the EU band esti-
mator for the number of antennas and number of
EU systems . The number of multipaths is equal
to 16. The frequency bands of EU systems are the same as those
set in Fig. 4. It is shown that the performance of estimation im-
proves with the increase of .
Below, we investigate the performance at the estimation stage
when the value of in (22) is incorrectly specified, i.e.,
. Table I shows the performance of the estimation stage when
and . The EU system is located
at subcarriers , and the number of multipaths is equal to
16. Since , the estimator returns the result
and . For each , there contain four columns. The
first column denotes “hit”. Note that, since , we have
one falsely alarmed band when a hit occurs. However, simu-
lation results indicate that the bandwidth of a falsely alarmed
band is generally small (two or three subcarriers). The second
(miss of type-I) and third (miss of type-II) columns denote the
probabilities that the situations illustrated in Fig. 2(a) and (b),
respectively, happen. The fourth column is the sum of the values
m of the first three columns. It is seen that the sums are in general
close to 1, meaning that hit and miss of types I and II encompass
most situations. We observe that, as the average SNR of the EU
signal increases, the probability of hit decreases; regarding the
probability of miss, type-I increases with the average SNR, and
type-II tends to increase as well but not so obvious as type-I.
This is explained as follows. When a deep channel fade occurs,
the increase of average SNR magnifies the channel frequency
selectivity; the portion of the band not experiencing deep fade
tends to be fitted by an AR model, and the portion under deep
fade is modeled by AWGN. The EU system band is thus torn
into pictures of Fig. 2(a) and (b), and the probability of hit is
decreased with the increase of SNR. We also find that the ten-
dencies of increase/decrease of probabilities of hit and miss be-
come more and more evident when the number of antennas is
raised from to . This is because, with the increase
of the number of antennas, it is easier to identity the band of EU

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.
HWANG et al.: SPECTRUM SENSING IN WIDEBAND OFDM COGNITIVE RADIOS 717

TABLE I
PERFORMANCE AT THE ESTIMATION STAGE WHEN ONE EU SYSTEM IS PRESENT AT SUBCARRIERS [35; 49] (m = 1) BUT M = 2. THE NUMBER OF MULTIPATHS
IS EQUAL TO 16, AND THE OBSERVATION LENGTH IS N = 100. THE DATA ARE GENERATED BY THE MONTE-CARLO SIMULATION WITH 1000 RUNS

TABLE II
PROBABILITY OF DETECTION OF SPECTRUM SENSING FOR (m; M ) 2 f(1; 1); (2; 2); (3; 3)g AND K 2 f1; 2; 3g. THE NUMBER OF MULTIPATHS IS EQUAL TO 16,
AND THE OBSERVATION LENGTH IS N = 100. THE RESULTS ARE GENERATED BY THE MONTE-CARLO SIMULATION WITH 1000 RUNS

TABLE III
PROBABILITY OF DETECTION OF SPECTRUM SENSING FOR (m; M ) 2 f(1; 2); (1; 2); (2; 3)g AND K 2 f1; 2; 3g. THE NUMBER OF MULTIPATHS IS EQUAL TO 16,
AND THE OBSERVATION LENGTH IS N = 100. THE RESULTS ARE GENERATED BY THE MONTE-CARLO SIMULATION WITH 1000 RUNS

system, or equivalently, the band without EU system. However, EU system is located at ; when , EU systems are
due to the incorrect value of , the band of EU is to be at and ; when , a system at is
detected as two, leading to type-I miss. added. The number of channel paths used in simulations is 16.
In the following, the overall performance of spectrum For each pair of , each column shows the detection prob-
sensing, i.e., concatenation of estimation and detection stages, ability of EU system(s) at a specific value of . We observe that,
is demonstrated. Let denote the for each , the probability of detection is proportional to
final result of spectrum sensing. We have described in the first the number of antennas and the average SNR, and the detec-
paragraph of Section III-B how the set is ob- tion probability is inversely proportional to the number of EU
tained. Detection occurs if each EU band is covered systems. Table III shows the performance of the proposed spec-
by some with one-subcarrier tolerance at both ends. trum sensing algorithm when with ,
That is, we say the EU bands are detected if, for each EU band and , and . The environments of sim-
, , one can find some such that ulations are the same as those of Table II. Compare the results
. Note that the definitions of at and with that of in
“hit” and “detection” are somewhat different. In the former, it is Table II, it is observed that the detection probability gets worse
required that the estimated band is equal to the true band (with with the increase of discrepancy between and . The same
one-subcarrier tolerance); while in the latter, we require only inference is reached when comparing the result of
the estimated band to be a subset of a true band. Thus, in case in this table and the result of in Table II.
of detection, there may be some falselyalarmed subcarriers. Moreover, the second column of Table III gives the overall de-
However, simulation results indicate that the number of such tection performance for , while Table I shows
falselyalarmed subcarriers is generally small. the probabilities of hit and misses of types I and II at the esti-
Table II gives the performance when with mation stage with the same . Compare Table I and the
, and and . When , the second column of Table III, we can see the probability of de-

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.
718 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 58, NO. 2, FEBRUARY 2010

tection shown in the latter is generally higher than the column [8] D. Cabric, A. Tkachenko, and R. W. Brodersen, “Experimental study of
“Sum” in the former with the same . This is mainly because spectrum sensing based on energy detection and network cooperation,”
presented at the ACM Int. Workshop Technology Policy for Accessing
some unfound EU bands at the estimation stage belong to nei- Spectrum, Boston, MA, Aug. 2006.
ther type I nor type II misses, but they are recovered at the de- [9] S. M. Mishra, A. Sahai, and R. W. Brodersen, “Cooperative sensing
tection stage. among cognitive radios,” in Proc. IEEE Int. Conf. Commun. (ICC), Jun.
2006, pp. 1658–1663.
[10] M.Öner and F. Jondral, “On the extraction of the channel allocation in-
VI. CONCLUSION formation in spectrum pooling systems,” IEEE J. Sel. Areas Commun.,
vol. 25, pp. 558–565, Apr. 2007.
In this paper, the problem of EU signal sensing in an OFDM [11] M. Ghozzi, M. Dohler, F. Marx, and J. Palicot, “Cognitive radio:
Methods for the detection of free bands,” Comptes Rendus Physique,
based CR system is addressed, where the number of EU sys- Elsevier, vol. 7, pp. 794–804, 2006.
tems in operation and their frequency bands are all unknown. [12] N. Han, S. Shon, J. H. Chung, and J. M. Kim, “Spectral correlation
Our main idea is that, once an EU system appears, several sub- based signal detection method for spectrum sensing in IEEE 802.22
WRAN systems,” in Proc. Int. Conf. Adv. Commun. Technol., Feb.
carriers in a row are interfered. To exploit this fact, observations 2006, vol. 3, pp. 1765–1770.
along the frequency domain are employed. The emphasis of this [13] J. Lunden, V. Koivunen, A. Huttunen, and H. V. Poor, “Spectrum
work is on combating the challenge resulting from a severe fre- sensing in cognitive radios based on multiple cyclic frequencies,”
presented at the 2nd Int. Conf. Cognitive Radio Oriented Wireless
quency selective fading channel. We use an AR model to track Networks Communications, Jul. 3, 2007.
the variation of the received EU signal strength along the fre- [14] E. Visotsky, S. Kuffner, and R. Peterson, “On collaborative detection
quency axis. When a deep fade occurs at part of the band of an of TV transmissions in support of dynamic spectrum sensing,” in Proc.
IEEE Int. Symp. New Frontiers in Dynamic Spectrum Access Networks,
EU system, tracking enables the sensing device to identify the Nov. 2005, pp. 338–345.
weak signal under deep fade as a part of the entire signal. We [15] T. Weiss, J. Hillenbrand, and F. Jondral, “A diversity approach for the
also investigate the scenario of multiple antennas to enhance the detection of idle spectral resources in spectrum pooling systems,” in
Proc. 48th Int. Sci. Colloquium, Sep. 2003.
performance of spectrum sensing. [16] G. Ganesan and Y. G. Li, “Cooperative spectrum sensing in cognitive
The task of spectrum sensing is composed of two stages: es- radio—Part I: Two user networks,” IEEE Trans. Wireless Commun.,
timation and detection. In the estimation stage, due to the lack vol. 6, no. 6, pp. 2204–2213, Jun. 2007.
[17] G. Ganesan and Y. G. Li, “Cooperative spectrum sensing in cognitive
of precise knowledge of EU signals statistics, it is hard to ob- radio: Part II: Multiuser networks,” IEEE Trans. Wireless Commun.,
tain an accurate estimate for the number of EU systems in op- vol. 6, no. 6, pp. 2214–2222, Jun. 2007.
eration. We assume the knowledge of an upper bound of the [18] M. Gandetto and C. Regazzoni, “Spectrum sensing: A distributed ap-
proach for cognitive terminals,” IEEE J. Selt. Areas Commun., vol. 25,
number of EU systems, denoted by , is known, and we calcu- pp. 546–557, Apr. 2007.
late the ML estimates of the bands of EU systems. In the de- [19] D. Cabric, S. M. Mishra, and R. W. Brodersen, “Implementation
tection stage, energy detection is performed individually at each issues in spectrum sensing for cognitive radios,” in Proc. T38th
Asilomar Conf. Signals, Systems, Computers, Nov. 2004, vol. 1,
of bands, including estimated bands and re- pp. 772–776.
maining bands. Detection is executed at the remaining [20] S. M. Mishra, S. t. Brink, R. Mahadevappa, and R. W. Brodersen, “De-
bands because some of them are unfound bands due to deep tect and avoid: An ultra-wideband/WiMAX coexistence mechanism,”
IEEE Commun. Mag., vol. 45, no. 6, pp. 68–75, Jun. 2007.
channel fade. The final result of spectrum sensing is the union of [21] I. F. Akyildiz, W.-Y. Lee, M. C. Vuran, and S. Mohanty, “Next gen-
the bands that are judged to be “EU system present” by energy eration/dynamic spectrum access/cognitive radio wireless networks: A
detection. Simulations result demonstrate that the composite of survey,” Comput. Netw.: Int. J. Comput. Telecommun. Netw., vol. 50,
no. 13, pp. 2127–2159, Sep. 2006.
the estimation and detection is robust under the ill condition that [22] H. Tang, “Some physical layer issues of wideband cognitive radio sys-
the correct number of EU systems cannot be found. tems,” in Proc. IEEE Int. Symp. New Frontiers Dynamic Spectrum Ac-
cess Networks, Nov. 2005, pp. 151–159.
[23] A. Ghasemi and E. S. Sousa, “Spectrum sensing in cognitive radio
REFERENCES networks: Requirements, challenges and design trade-offs,” IEEE
[1] J. Mitola, “Cognitive radio: Making software radio more personal,” Commun. Mag., vol. 46, no. 4, pp. 32–39, Apr. 2008.
IEEE Pers. Commun., vol. 6, no. 4, pp. 48–52, Aug. 1999. [24] M. Basseville and I. V. Nikiforov, Detection of Abrupt Changes:
[2] J. Mitola, “Cognitive radio: An integrated agent architecture for soft- Theory and Application. Englewood Cliffs, NJ: Prentice-Hall,
1993.
ware defined radio,” Ph.D. dissertation, Royal Institute of Technology
[25] C. Myers and L. Rabiner, “A level building dynamic time warping al-
(KTH), Stockholm, Sweden, 2000.
gorithm for connected word recognition,” IEEE Trans. Signal Process.,
[3] H. Urkowitz, “Energy detection of unknown deterministic signals,”
vol. 29, no. 2, pp. 284–297, Apr. 1981.
Proc. IEEE, vol. 55, no. 4, pp. 523–531, Apr. 1967.
[26] F. K. Svendsen and F. K. Soong, “On the automatic segmentation of
[4] V. I. Kostylev, “Energy detection of a signal with random speech signals,” in Proc. 1987 Int. Conf. Acoust., Speech, Signal Pro-
amplitude,” in Proc. IEEE Int. Conf. Commun. (ICC), Apr. 2002, cessing, Dallas, TX, pp. 77–80.
pp. 1606–1610. [27] S. M. Kay, Fundamentals of Statistical Signal Procssing: Estimation
[5] F. F. Digham, M.-S. Alouini, and M. K. Simon, “On the energy de- Theory. Englewood Cliffs, NJ: Prentice-Hall PTR, 1993, vol. 1.
tection of unknown signals over fading channels,” in Proc. IEEE Int. [28] H. Akaike, B. N. Petrov and F. Csaki, Eds., “Information theory as
Conf. Commun. (ICC), May 2003, pp. 3575–3579. an extension of the maximum likelihood principle,” in Proc. 2nd Int.
[6] Marilynn and P. Wylie-Green, “Dynamic spectrum sensing by multi- Symp. Information Theory, Budapest, Akademiai Kiado, 1973, pp.
band OFDM radio for interference mitigation,” in Proc. IEEE Int. 267–281.
Symp. New Frontiers Dynamic Spectrum Access Networks, Nov. 2005, [29] J. Rissanen, “Modeling by the shortest data description,” Automatica,
pp. 619–625. vol. 14, pp. 465–471, 1978.
[7] A. Ghasemi and E. S. Sousa, “Collaborative spectrum sensing for op- [30] G. Schwarz, “Estimating the dimension of a model,” Ann. Statist., vol.
portunistic access in fading environments,” in Proc. IEEE Int. Symp. 6, no. 2, pp. 461–464, Mar. 1978.
New Frontiers in Dynamic Spectrum Access Networks, Nov. 2005, pp. [31] C.-B. Lee, “Bayesian estimation of the number of change points,” Sta-
131–136. tistica Sinica, vol. 8, pp. 923–939, 1998.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.
HWANG et al.: SPECTRUM SENSING IN WIDEBAND OFDM COGNITIVE RADIOS 719

Chien-Hwa Hwang (M’00) received the B.S. and
M.S. degrees from the National Taiwan University,
Taiwan, in 1993 and 1995, respectively, and the Ph.D.
degree from the University of Southern California,
Los Angeles, in 2003, all in electrical engineering.
In August 2003, he joined the Institute of Com-
munications Engineering of the National Tsing Hua
University, Taiwan, as an Assistant Professor. His
research interests include random matrix theory and
statistical signal processing.
Shih-Chang Chen received the B.S. degree from the
Department of Electrical Engineering and the M.S.
degree from the Institute of Communications Engi-
neering, both in the National Tsing Hua University,
Hsinchu, Taiwan, in 2005 and 2007, respectively.
He is currently an Engineer at the Digital IC De-
sign Department of Realtek Semiconductor Corpora-
tion, Hsinchu, Taiwan.

Guan-Long Lai received the B.S. degree from the

Department of Electrical Engineering at the National
Chung Hsing University, Taichung, Taiwan, in 2007
and the M.S. degree from the Institute of Communi-
cations Engineering, National Tsing Hua University,
Hsinchu, Taiwan, in 2009.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:452 UTC from IE Xplore. Restricon aply.