2690
6, JULY 2013
Energy-Efficient Multichannel Cooperative Sensing
Scheduling With Heterogeneous Channel Conditions
for Cognitive Radio Networks
Salim Eryigit, Student Member, IEEE, Suzan Bayhan, Member, IEEE, and Tuna Tugcu, Member, IEEE
Abstract—Spectrum sensing is an important aspect of cognitive
radio networks (CRNs). Secondary users (SUs) should periodi-
cally sense the channels to ensure primary-user (PU) protection.
Sensing with cooperation among several SUs is more robust and
less error prone. However, cooperation also increases the energy
spent for sensing. Considering the periodic nature of sensing,
even a small amount of savings in each sensing period leads
to considerable improvement in the long run. In this paper, we
consider the problem of energy-efficient (EE) spectrum sensing
scheduling with satisfactory PU protection. Our model exploits
the diversity of SUs in their received signal-to-noise ratio (SNR)
of the primary signal to determine the sensing duration for each
user/channel pair for higher energy efficiency. We model the given
problem as an optimization problem with two different objectives.
The first objective is to minimize the energy consumption, and the
second objective is to minimize the spectrum sensing duration to
maximize the remaining time for data transmission. We solve both
problems using the outer linearization method. In addition, we
present two suboptimal but efficient heuristic methods. We pro-
vide an extensive performance analysis of our proposed methods
under various numbers of SUs, average channel SNR, and channel
sampling frequency. Our analysis reveals that all proposals with
an energy minimization perspective provide significant energy
savings compared with a pure transmission-time maximization
(TXT) technique.
Index Terms—Cooperative sensing scheduling (CSS), energy-
efficient (EE) sensing, heterogeneous sensing, sensing task
assignment.
I. I NTRODUCTION
T HE increasing demand for wireless communications calls
for better spectrum utilization. One of the most promising
solutions is the dynamic spectrum access (DSA) paradigm,
which allows the opportunistic access of a spatiotemporal un-
used wireless spectrum by cognitive radio networks (CRNs). In
a CRN, a secondary user (SU) transmits through a frequency
channel that is licensed to primary users (PUs) but is currently
Manuscript received August 30, 2012; revised November 26, 2012 and
January 28, 2013; accepted February 4, 2013. Date of publication February 13,
2013; date of current version July 10, 2013. This work was supported in
part by the State Planning Organization of Turkey under Grant 07K120610,
in part by the Scientific and Technical Research Council of Turkey under
Grant 108E101, and in part by the COST Action IC0902. The review of this
paper was coordinated by Dr. E. K. S. Au.
S. Eryigit and T. Tugcu are with the Department of Computer Engineering,
Bogazici University, Istanbul 34342, Turkey (e-mail: eryigit@boun.edu.tr;
tugcu@boun.edu.tr).
S. Bayhan is with Helsinki Institute for Information Technology, Aalto
University, 00076 Espoo, Finland (e-mail: bayhan@hiit.fi).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TVT.2013.2247070
0018-9545/$31.00 © 2013 IEEE
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO.
unoccupied. However, DSA has the challenge to guarantee
that SUs vacate the spectrum whenever a PU simultaneously
transmits at the band of SU transmission. Hence, SUs must
perform spectrum sensing. The accuracy of spectrum sens-
ing is paramount for both finding the spectral voids and for
protecting the PU communications. Hence, a sensing period
is reserved at the beginning of each frame for the spectrum
sensing task.
Previous works showed the increase in sensing accuracy with
the increase in sensing time [1]. On the other hand, SUs that
are mostly mobile devices should be energy-efficient (EE)1
as they use their battery power. Therefore, from the energy
(throughput) efficiency perspective, the more time is spent on
sensing, the more energy is consumed for overhead, and less
time remains for transmission. On the other hand, the through-
put of the network is a function of the detection accuracy (i.e.,
probability of detection Pd and probability of false alarm Pf ).
Hence, there is a tradeoff between sensing and transmission
duration for both throughput and EE.
In addition, it is shown that cooperation among the SUs
increases the detection reliability of spectrum sensing at the
expense of additional communication overhead that increases
with the number of cooperating SUs [2]. Different from cooper-
ative sensing in a single channel, cooperative sensing schedul-
ing (CSS) has to balance the tradeoff between the detection
accuracy of a single channel and the number of channels being
sensed in a multichannel CRN. That is, if there are more SUs
assigned to sense a single channel, there will be a higher prob-
ability of detection for that channel at the expense of leaving
some channels being unexplored. While cooperative sensing
has been well investigated, CSS still remains unexplored. It
is shown in previous works that CSS is NP-hard [3]. Taking
the EE concerns into account makes this problem even more
complicated.
In this paper, we focus on the EE of cooperative spectrum
sensing in a multichannel CRN with heterogeneous PU chan-
nels in terms of received SNRs. Since scheduling the SUs to
sense a number of channels in a CRN is a difficult task [3], we
propose three schemes for EE CSS. The first scheme uses outer
linearization to find the optimal solution, whereas the latter two
are efficient heuristic methods. Apart from these three, we also
analyze the problem from the transmission-time point of view
as time spent for sensing is also the time lost for transmission.
1 We use EE for both energy efficiency and energy-efficient throughout the
paper. The exact meaning can be derived from the context.
ERYIGIT et al.: EE MULTICHANNEL CSS WITH HETEROGENEOUS CHANNEL CONDITIONS FOR CRNS
The rest of this paper is organized as follows. In Section II,
we revise the related work on EE in spectrum sensing and state
our contributions to the literature. In Section III, we define
the cooperative sensing system model and introduce the basic
theorems used in the formulated EE CSS scheme. Section IV
first formulates the problem and presents the methodology for
finding the optimal solution. Proposed heuristic schemes are
described in Section V, and their performances are evaluated
in Section VI. Finally, Section VI concludes this paper.
II. R ELATED W ORK AND C ONTRIBUTIONS
EE of CRNs has recently gained interest, and most of the
initial works focus on the EE of spectrum sensing [4]–[9] and
cooperative sensing [10]–[14]. Su and Zhang minimized the
sensing energy consumption while meeting the constraint on
undiscovered spectrum opportunities [5], and they adapted the
period of spectrum sensing to attain a balance between energy
consumption and missed spectrum opportunities for a random
access CRN [7]. Optimal sensing duration and transmission
duration for an SU under both high and low SU power capacity
are analytically derived in [8]. The effect of transmission,
idling, and sensing power consumption are analyzed in that
work. Pei et al. devised an optimal policy for a single SU to
decide on the order of channels to be sensed and when to stop
sensing and start transmission [9]. While all these previous
works have valuable contributions, they fall short of practi-
cality. In practical CRNs, there are multiple and most likely
heterogenous primary channels. In this setting, one of the major
concern of the operator is to explore as many PU channels as
possible, meeting the PU detection and false alarm constraints.
Therefore, in this paper, we enforce the SUs to sense all
PU channels collaboratively to maximize the discovered spec-
trum opportunities.
In [12] and [13], the communication cost for determining
the number of cooperating SUs is considered. Maleki et al.
found the minimum number of cooperating SUs that attains
the required detection and false alarm probability performance
[12]. If fewer SUs are engaged in sensing, less time will be
spent reporting the sensing outcomes; thereby, more time will
be spent for transmission. Moreover, sensing reports from unre-
liable SUs may decrease the sensing performance. SUs with un-
reliable sensing information are refrained from reporting their
sensing results to save energy in [13]. In addition, a cluster-
based decision collection instead of a high power-consuming
broadcasting scheme is also proposed in [13]. The cluster-based
scheme adapts the transmission power considering the most
distant node. The decision fusion rule (i.e., how the collected
sensing information is processed to give the final decision on
the existence of PU) at the fusion center also affects the EE.
Peh et al. [14] tuned the k parameter in k-out-of-N fusion rule
at each frame and the threshold for energy detection scheme
at the fusion center. Heterogeneity of PU channels and SU link
conditions were ignored in these works, making the investigated
scenarios less realistic. In addition, assigning the same sensing
duration for all SUs, regardless of their link conditions, may
result in wasted energy at SUs with good link conditions. In
contrast to these works, we incorporate the effect of SNRs of
2691
SUs into our scheme to calculate appropriate sensing duration
for each SU and frequency pair.
In [15]–[17], various solutions for improving the EE of
CSS are presented. Sensing scheduling is modeled as a utility
maximization problem subject to a certain cooperative detec-
tion probability in [15]. In addition, a constraint on minimum
discovered transmission time is imposed to ensure a certain
quality-of-service, together with heterogeneous detection prob-
ability requirements. Similarly, Zhang and Tsang determined
the number of SUs to sense each channel and the sensing
duration in a slot [16], whereas Hao et al. study the optimal
partition of the SUs into coalitions, such that the total EE of all
coalitions are maximized [17]. In [16], a partially observable
Markovian decision process framework is utilized, and the
punishment parameter for higher EE is tuned. A distributed
solution using coalition formation is proposed in [17].
Apart from these works, the main contributions of this paper
can be summarized as follows.
• We consider a scenario in which the number of SUs is
larger than the number of primary channels. Therefore, our
main concern is to select the SUs to sense all channels,
whereas in [15]–[17], a subset of primary channels to be
sensed by all SUs is selected. In addition, in the previous
works, an SU can sense at most one channel, whereas in
this paper, SUs can sense multiple channels, as long as
they finish sensing in the dedicated time.
• Unlike these works, we account for the heterogeneity of
the SU link conditions (i.e., received SNR of the PU
signal at the SU). Therefore, our CSS solution additionally
determines which SUs should sense a channel. This paper
diverges from the previous works, which only determine
the number of SUs to sense a specific primary channel.
• Moreover, sensing duration associated with an SU is
adjusted according to the link SNR as opposed to the
prior works, which consider identical sensing duration for
all SUs. Simply, our approach is based on the fact that
channels with high SNRs require less sensing time for a
required detection probability and false alarm probability.
Hence, an SU can save energy by sensing one of the
channels with a higher SNR, as opposed to the fixed
sensing duration scheme.
III. S YSTEM M ODEL
We assume an infrastructure-based CRN with N SUs,
M channels, and a cognitive-radio base station (CBS). Our
consideration is a specific case where the number of channels is
less than the number of SUs, i.e., N  M . We believe that, in
a cellular network, this assumption generally holds as there are
many users within the coverage area of the base station. If that
is not the case, the CBS may select a subset of the channels
based on their past data, such as availability, capacity, etc.,
such that there are enough SUs to sense all selected channels.
This selection procedure has the potential to reduce energy
consumption by eliminating the less-favorable channels. SUs
operate in a time-synchronized manner within a frame-based
communication protocol. Each frame starts with a fixed-length
quiet sensing period of duration T s during which SUs sense the
2692 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 6, JULY 2013

where fs is the sampling frequency, and Q is the complemen-

tary cumulative distribution of a standard Gaussian.
d
Theorem 1: Pm, n is an increasing function of τm, n . Further-
more, it is also concave if
√
1 γm, n fs  −1  f   
−√ + Q Pm, n − τm, n fs γm, n < 0.
τm,n (2γm, n +1)
(2)
This theorem is well known, and its proof is given in
Fig. 1. Frame starts with a sensing period followed by reporting and transmis-
sion periods. Appendix A.
d d
Lemma 1: Pm,n is a concave function of τm,n if Pm,n > 0.5.
Lemma 1 is a straightforward application of Theorem 1. The
channel(s) assigned to them. An SU may sense multiple chan-
proof can be found in Appendix B.
nels during the quiet period as long as the total time dedicated
Lemma 2: 1 − Pm, d
n is a nonnegative decreasing function of
to sensing by the SU does not exceed T s . Then, all SUs that
τm, n . It is also convex if the condition in (2) is satisfied.
sense at least one channel report their hard decisions about these
Let Sm and Qdm denote the set of SUs sensing channel
channels (0 or 1, indicating the absence or presence of PU) to
m, and the cooperative detection probability for channel m,
the CBS. We assume that the secondary network has a dedicated
respectively. Using OR rule for decision combining gives
common control channel that is used for this reporting task
 
and other control messages. The CBS combines the decisions Qdm =1− 1−Pm, d
n
using OR rule. The remaining time is used for transmission. We n∈Sm
also assume that the SUs and the channels are heterogeneous.    f   
That is, the SNR of each SU over each channel is different Q−1 Pm, n − τm, n fs γm, n
=1− 1−Q  .
due to different proximity from the PUs and different channel 2γm, n +1
n∈Sm
conditions (shadowing, fading, etc.). We assume the existence
of a receiver block at each SU to estimate the SNR level and to Theorem 2: Qdm is an increasing function of τm, n . More-
feed it back to the CBS through the error-free feedback channel over, it is also concave if the condition in (2) is satisfied
[18]. With the help of the receiver block, we assume that the ∀n ∈ Sm .
instantaneous SNRs are known. However, if that is not the case, The proof of this theorem is given in Appendix C.
long-term SNRs can also be used. This time, the techniques
discussed in this paper can also be applied. However, the main IV. O PTIMIZATION M ODEL AND
objective becomes the minimization of expected energy, instead S OLUTION M ETHODOLOGY
of the actual energy. The frame structure is shown in Fig. 1
where T , T rep , and τm, n are the total frame length, the time A. Energy Consumption Model
dedicated for reporting sensing results, and the time that SUn Let P s and Em,s
n be the power consumed during channel
senses channel m, respectively. sensing and energy dissipated by SUn for sensing channel m,
Our main goal is to sense all M channels with minimum s
respectively. Em, s
n is equal to P τm, n . Then, energy consump-
energy and sufficient accuracy, such that cooperative detection tion for channel sensing (denoted by E s ) can be written as
probability of each channel is greater than some predefined
threshold value (denoted by th Qd ), and cooperative false alarm

probability is smaller than another threshold (denoted by th Qf ). Es = P s τm, n .

Since channel sensing consumes energy, an SU may not utilize
all of the quiet period duration for sensing if not necessary.
On the other hand, it is desirable to sense a channel with
a couple of SUs instead of a single SU (although, it may
satisfy the thresholds) to increase robustness. Hence, there is
a tradeoff between energy consumption and sensing reliability.
The problem includes the assignment of SUs to channel(s) for
the sensing task, together with the decision of the sensing time
for the channel(s) to be sensed by each SU.
f d
Let Pm, n , Pm, n , and γm, n denote the probability of false
alarm, probability of detection, and SNR for SUn over channel
f
m, respectively. If we assume that Pm, n is fixed, then for a
complex-valued PSK channel with circularly symmetric com-
m=1 n=1
In addition to channel sensing, SUs also consume energy
by transmitting their local results to the CBS. We assume that
SU transmits its sensing report as a single packet, regardless
of the number of channels sensed, and the reporting period is
long enough such that all SUs can send their packets. Let Enrep
denote the energy consumed for reporting the sensing result to
CBS, which depends on the location of SUn relative to the CBS.
In addition, let S rep denote the set of SUs that perform sensing
in this frame and are required to report their local decisions to
the CBS. Then, the total energy consumption for reporting is
given by

plex Gaussian noise, Pm, d

n is given by [1] E rep = Enrep .
  f    n∈S rep
Q−1 Pm, n − τm, n fs γm, n
Pm, n = Q
d
 (1) This model assumes that all reporting packets are success-
2γm, n + 1 fully transmitted. If that is not the case, the model can be
ERYIGIT et al.: EE MULTICHANNEL CSS WITH HETEROGENEOUS CHANNEL CONDITIONS FOR CRNS 2693

modified as follows. Let p denote the probability of successful subject to

packet transmission that is geometrically distributed. Then, the
expected number of transmission attempts τm, n ≥ τm,
min
n xm, n ∀m ∈ M ; ∀n ∈ N (6)
for an SU is given
by 1/p, and E rep is given by E rep = 1/p n∈S rep Enrep .

M
τm, n ≤ T s yn ∀n ∈ N (7)
m=1
B. Optimization Model for EE Sensing

We first define the decision variables that are used in the xm, n ≥ δ min ∀m ∈ M (8)
optimization model. Let n=1

N
τm, n = time spent by SUn for sensing channel m xm, n ≤ δ max ∀m ∈ M (9)
 n=1
1, if channel m is sensed by SUn

M
xm, n =
0, otherwise xm, n ≤ M yn ∀n ∈ N (10)
 m=1
1, if SUn transmits sensing result to CBS d
th Q − Qdm ≤ 0 ∀m ∈ M (11)
yn =
0, otherwise.
xm, n , yn ∈ {0, 1} ∀m ∈ M ; ∀n ∈ N (12)
d
From (1), for a given Pm, n value, the required τm, n can be
τm, n ≥ 0 ∀m ∈ M ; ∀n ∈ N (13)
written as
where, this time, Qdm is defined as
  f   d  2
Q−1 Pm, −1     
n −Q Pm, n 2γm, n + 1 N
Q−1 (P f )− τm, n fs γm, n
τm, n = √ .
γm, n fs
d
Qm = 1− 1−Q  xm, n .
n=1
2γm, n +1
(3)
min
In addition, let τm, n denote the sensing time required for SUn
Hence, SUs with xm, n value of 0 contribute 1 to the given
d
to achieve a Pm, n value of 0.5. It can be calculated from (3) as multiplication, whereas those with xm, n value of 1 contribute
 (1 − Pm,d
n ).
 f  2
Q−1 Pm, The objective in (5) minimizes the total energy consumption
min
τm, n = √ n . associated with sensing for a frame. Constraint (6) specifies
γm, n fs
that, if SUn senses channel m, the sensing duration should
min
be at least τm, n . This way, we guarantee that the concavity
We assume that a channel should be sensed by at least δ min
condition always holds. Constraint (7) denotes that total time
SUs. δ min defines the minimum number of cooperating SUs for
spent by an SU for sensing should be less than or equal to
a channel. The selection of δ min value is a design criterion. To
the sensing duration of a frame. It also forces all τm, n values
encourage cooperation and improve robustness, a δ min value
associated with SUn to 0 if yn = 0. Constraint (8) requires
greater than 1 is preferred. On the other hand, regarding EE
that each channel should be sensed by at least δ min SUs.
concern, δ min should not be high as each additional SU used
Similarly, Constraint (9) limits the number of cooperating SUs
for sensing incurs sensing energy consumption and, maybe,
for a channel to satisfy the false alarm probability threshold.
reporting energy.
Constraint (10) forces yn value for an SU to 1 if that SU
n = P ∀m, n, then Qm is given by
f f f
If we assume that Pm,
senses any channels. The requirement for cooperative detection

Qfm = 1 − (1 − P f ). probability being greater than the threshold for each channel is
n∈Sm
expressed by constraint (11). Finally, constraints (12) and (13)
specify the types of variables.
Since Qfm ≤ th Qf , then the maximum number of cooperating The given problem is a mixed-integer nonlinear program-
SUs, which is denoted by δ max , can be calculated as ming problem because of constraint (11), although its objective
 is linear. We resort to the outer linearization algorithm to solve
log (1 − th Qf ) the given problem.
δ max = . (4)
log (1 − P f )

In other words, δ max is the maximum number of cooperating C. Outer Linearization

SUs that satisfy the cooperative false alarm constraint. The
As proven before, once the xm, n values are fixed, Qdm
solution methodology that we apply can also be used for the
f value is concave in terms of τm, n . Thus, constraint (11) is
case where Pm, n values differ. We discuss this case in detail convex, and the outer linearization procedure can be used to
at the end of Section IV-C. The optimization model can be
find the optimal solution [19]. Outer linearization works by first
written as
ignoring the mixed-integer nonlinear constraints to obtain an

N initial solution. If the solution satisfies all previously ignored

P1 : min w = P s τm, n + Enrep yn (5) constraints, then it is optimal. On the other hand, if it does not,
m=1 n=1 n=1 then the most violated constraint is linearized using the current
2694 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 6, JULY 2013

solution and is added to the current problem as a new constraint ability constraint for each channel) need to be checked for
to obtain another solution. The linearization process goes on feasibility. The other steps of the procedure are the same.
until all constraints are satisfied with an  tolerance. Since the
constraints are convex, the procedure is guaranteed to terminate D. TXT
in finite number of steps [20]. The steps of the procedure are
as follows. The model mentioned previously optimizes the total en-
ergy dedicated to the sensing task while achieving satisfactory
Step 1: Initialize the iteration counter, i.e., k = 1. Solve the
sensing performance in terms of detection and false alarm
initial mixed-integer linear programming problem (P2)
probabilities. However, in this approach, sensing duration of
formed by ignoring constraint (11), and obtain the initial
1 1 1 a frame (denoted by T s ) is constant. Hence, if we denote the
solution τm, n , xm, n , and yn . frame duration by T and reporting time of the sensing outcomes
Step 2: Identify the most violated constraint gm among the M
k k by T rep , which are also constant, then the transmission time for
constraints of (11) with the current solution (τm, n , xm, n , data packets is given by T − T s − T rep . Another approach is to
and ynk ). That is, gm is the cooperative detection probability maximize the data transmission duration of a frame. This time,
constraint corresponding to the channel that deviates from we treat T s as a decision variable.
the threshold value most. Let vm denote the corresponding  Assuming a quiet sensing
period, T s is given by maxn { M m=1 τm, n }. In other words,
deviation. T s is the maximum of total sensing times for all SUs as the
Step 3: If the maximum violation is smaller than , then stop; network should wait for the SU with the longest total sensing
the current solution is optimal with  feasibility tolerance. time before moving on the next phase of a frame. Then, the ob-
Otherwise, proceed with Step 4. jective becomes max z1 = T − T rep − maxn { M m=1 τm, n }.
Step 4: Linearize the most violated constraint by adding the Since T and T rep  are constants, this objective is equivalent to
following linear constraint to P2: min z2 = maxn { M m=1 τm, n } subject to constraints (6)–(13).
⎡ .. ⎤ To solve this problem, we resort to the outer linearization
. procedure again as the constraints are almost the same.
⎢ ⎥
⎢ xm, i − xkm, i ⎥
⎢ ⎥
T ⎢ .. ⎥
∇gm (. . . xkm, i , . . . τm,
k
, . . .) ⎢ . ⎥ + vm ≤ 0 V. H EURISTIC A PPROACHES
i
⎢ ⎥
⎢ τm, i − τ k ⎥
⎣ m, i ⎦ Here, we propose two suboptimal but fast heuristic ap-
.. proaches for the EE sensing problem. The first one focuses
.
on greedily minimizing sensing energy while disregarding the
where ∇gm (. . . xkm, i , . . . τm,
k reporting energy. On the other hand, the second heuristic ini-
i , . . .) is the gradient of gm
evaluated at the current solution. Its individual entries are tially considers the reporting energy, and then, it considers the
given by sensing energy.
  Unlike the previous two approaches that support different

∂gm Q−1 (P f ) − τm, i fs γm, i detection probabilities for different channel and user pairs,
= −Q  Bm, i these heuristics require a fixed detection probability P d for
∂xm, i 2γm, i + 1
all channels and users for the sake of simplicity and quick
√
∂gm xm, i γm, i fs execution time. This approach is frequently applied in the
=− √ √  Am, i Bm, i literature [4], [12], [14]. For both heuristics, we sense each
∂τm, i 2 τm, i 2π 2γm, i + 1
channel with δ min SUs. Thus, the required P d value can be
where Bm, i is given by calculated as
       1/δmin 

N
Q−1 (P f ) − τm, n fs γm, n P d = max 1 − 1 − th Qd d
, Pmin
1−Q  xm, n .
n=1, n =i
2γm, n + 1
d
which guarantees a minimum detection probability of Pmin . As
f
Set k = k + 1, and solve the current problem to obtain the Pm, n values are assumed to be the same for all SU–channel
k
τm, k k pairs as before, the goal of the heuristics is to find the best
n , xm, n , and yn values. Proceed with Step 2.
SU/channel assignment.
In the remainder of this paper, we refer to the application of
outer linearization to Problem P1 as EE.
f A. SEM Heuristic
For the case where Pm, n values differ, the false alarm
constraint assumes the following form: This heuristic minimizes the sensing energy by selecting SUs
with high SNRs for a channel while disregarding reporting

N
  energy. Initially, the remaining sensing time of all SUs are equal
1− 1 − Pm,
f
n xm, n − th Q ≤ 0.
f

n=1
The outer linearization procedure can still be applied in this
case, but this time, 2M constraints (cooperative false alarm
probability constraint in addition to cooperative detection prob-
to T s . The heuristic starts with the first channel, sorts the SUs
in descending order based on their γm, n values, and selects the
first SU in the list. Then, it calculates the required τm, n value
for the selected SU to obtain a detection probability of P d . If the
remaining sensing time of the selected SU is greater than τm, n ,
ERYIGIT et al.: EE MULTICHANNEL CSS WITH HETEROGENEOUS CHANNEL CONDITIONS FOR CRNS 2695

the selected SU is assigned to sense channel m. Otherwise, we 3: for m = 1 to M do

move on to the next SU. The algorithm runs until δ min SUs are 4: Sort SUs in S rep in descending order of γm, n , and let
assigned to all channels. The pseudocode for this heuristic is indexRep be the list of indices of the sorted entries.
given in Algorithm 1. 5: assignmentN o = 0, k = 1
6: while (assignmentN o < δ min ) && (k ≤ |S rep |) do
7: n = indexRep[k]
Algorithm 1 Sensing Energy Minimization (SEM) Heuristic 8: Select SUn ∈ S rep as a candidate, and calculate τm, n
value to achieve P d using (3).
Require:P d , δ min M , N , γm, n , T s 9: if τm, n ≤ remainingT ime[n] then
1: remainingT ime[n] = T s ∀n 10: remainingT ime[n] = remainingT ime[n] −
2: for m = 1 to M do τm, n
3: Sort SUs in descending order of γm, n and let index 11: assignmentN o = assignmentN o + 1
be the list of indices of the sorted entries such that 12: end if
index[1] corresponds to the index of SU with the 13: k =k+1
highest γm, n and index[N ] corresponds to the index 14: end while
of SU with the lowest γm, n . 15: if assignmentN o < δ min then
4: assignmentN o = 0, k = 1 16: Sort SUs in S nrep in descending order of γm, n , and
5: while assignmentN o < δ min do let indexN rep be the list of indices of the sorted
6: n = index[k] entries.
7: Select SUn as a candidate and calculate the τm, n value 17: k=1
to achieve P d using (3). 18: while assignmentN o < δ min do
8: if τm, n ≤ remainingT ime[n] then 19: n = indexN rep[k]
9: remainingT ime[n] = remainingT ime[n] − 20: Select SUn ∈ S nrep as a candidate and calculate
τm, n τm, n value to achieve P d using (3).
10: assignmentN o = assignmentN o + 1 21: if τm, n ≤ remainingT ime[n] then
11: end if 22: remainingT ime[n] = remainingT ime[n] −
12: k =k+1 τm, n
13: end while 23: assignmentN o = assignmentN o+1, S rep =
14: end for S rep ∪ {SUn }, S nrep = S nrep \ {SUn }
24: end if
25: k =k+1
Starting with the first channel, the heuristic selects δ min 26: end while
SUs with the best γm, n values and enough remaining sens- 27: end if
ing time for the sensing task. The outer loop takes O(M ) 28: end for
steps. Sorting SUs based on their γm, n values is O(N log N ),
whereas the inner loop is O(N ). Hence, the total running time
is O(M N log N ). This time, both inner WHILE loops (lines 6 and 18) take
O(N ), and the sorting operations are still O(N log N ). As in
B. REM Heuristic the previous case, the total running time is O(M N log N ).
The main difference between the reporting energy mini- VI. R ESULTS
mization (REM) heuristic and SEM heuristic is that REM first
considers SUs that are already assigned to sense a channel. Let We assume that the received SNR at SU (γm, n ) follows an
S rep be the set of SUs that are going to perform sensing and exponential distribution with mean μSNR . To be consistent, we
transmit their reports for this frame. Similarly, S nrep is the set use the same γm, n values for a given μSNR across different
of SUs that are not assigned to sense a channel yet. Initially, runs. For a given parameter set, we first run the transmission-
S rep = ∅, S nrep = {SU1 , SU2 , . . . , SUN }. The heuristic first time maximization (TXT) method to obtain the ideal sensing
s
looks for SUs among those in S rep to save reporting energy. time denoted by Topt . For the other methods, we scale this value
s
If enough SUs are not found, then it moves on to S nrep . As with an α value (α > 1), and use αTopt as the sensing time for
in the previous case, SUs in S rep and S nrep are processed in the other methods. The values for the other parameters are given
decreasing order of γm, n values for the considered channel. in Table I.
The pseudocode of REM heuristic is given in Algorithm 2. By using (4), we obtain δ max = 10 for the given P f and
f
th Q values. Note that the presented results are for a single
frame. Hence, the cumulative effect will be much higher if
Algorithm 2 Reporting Energy Minimization Heuristic multiple frames are considered. Furthermore, the processing
order of the channels is important for the given heuristics as
Require:P d , δ min , M , N , γm, n , T s they converge to local optimal solutions. Although the channels
1: remainingT ime[n] = T s ∀n are ordered naturally in the given pseudocode, we also run both
2: S rep = ∅, S nrep = {SU1 , SU2 , . . . , SUN } heuristics with randomly ordered channels 20 times. The results
2696
TABLE I
PARAMETERS VALUES
Fig. 2. Energy consumption profiles with N = 200, δ min = 3, and α = 2.
(a) Low SNR: μSNR = −5 dB. (b) High SNR: μSNR = 2 dB.
given in the following for the heuristics are the best of the
21 runs in terms of energy consumption.
We first observe the total energy consumption and its individ-
ual components in Fig. 2(a) and (b) for μSNR values of −5 and
2 dB, respectively. For low μSNR , the sensing component of the
energy consumption is more dominant. On the other hand, the
reporting energy consumption becomes the major component
when μSNR is higher. As we can see, the reporting energy
IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 6, JULY 2013
Fig. 3. Effect of μSNR on total sensing energy consumption with N = 200,
δ min = 3, and α = 2. (a) μSNR between −10 and 5 dB. (b) μSNR between
−2 and 3 dB.
consumption is similar in both cases. Hence, the difference
stems from the sensing energy consumption. With high μSNR ,
the time required to achieve a particular detection probability
decreases, which in turn decreases the required sensing time.
In both cases, TXT achieves the worst performance since its
objective does not consider the energy consumption at all.
On the other hand, the performance of EE is always superior
compared with other methods. Furthermore, SEM is slightly
superior compared with REM for low SNRs because it prior-
itizes the sensing energy. On the contrary, REM achieves lower
total energy for high SNR since it first considers the reporting
energy component.
The effect of changing μSNR on the total sensing energy
consumption is shown in Fig. 3. Fig. 3(a) shows a broader
range, whereas Fig. 3(b) shows the high-SNR regime. Initially,
increasing μSNR values have a significant impact on the total
energy consumption for all methods, whereas beyond a certain
point, the benefits are minimal. In this case, EE provides 7%
improvement over the next best method, namely SEM heuristic,
when μSNR is −10 dB. Moreover, the improvement over other
methods is much better when μSNR assumes higher values, as
shown in Fig. 3(b). As an example, using EE results in 22%
ERYIGIT et al.: EE MULTICHANNEL CSS WITH HETEROGENEOUS CHANNEL CONDITIONS FOR CRNS
Fig. 4. Effect of the number of SUs on the total sensing energy consumption
with μSNR = −5 dB, δ min = 3, and α = 2.
Fig. 5. Energy consumption profiles with N = 200, δ min = 3, α = 2, fs =
10 kHz, and μSNR = −5 dB.
reduction in total energy consumption compared with the next
best method, which is REM this time, when μSNR is 0 dB.
In addition, both figures support our previous claim that SEM
achieves better performance than REM for low SNRs, whereas
the reverse is true for high SNRs.
Fig. 4 shows the change in total energy consumption with
respect to the increase in the number of SUs. Apart from
TXT method, all schemes yield better results as N increases.
The main reason for this performance improvement is the
diversity brought by the added SUs. That is, with more SUs, the
probability of finding an SU with a high γm, n value increases
for a given channel m. On the other hand, the TXT method
shows slight variations since its goal is not related to energy
consumption.
The total energy consumption and its individual components
for fs = 10 kHz case are presented in Fig. 5. In comparison
with Fig. 2, increasing fs has a similar effect as increasing SNR.
However, the effect of SNR is more prominent. For instance,
with all other factors constant, increasing μSNR from −5 to
2 dB (almost a fivefold increase) results in nearly 83% reduction
in energy consumption for EE. On the other hand, increasing fs
2697
tenfold from 1 to 10 kHz gives a 76% decrease for EE. These
observations are in accordance with (3). In addition, similar to
the case in Fig. 3(b), with a higher sampling rate, the REM
heuristic provides lower energy consumption than the SEM
heuristic.
The energy consumption values for various values of α are
given in Fig. 6(a) and (b) for μSNR values of −5 and 2 dB,
respectively. As α is not a parameter for TXT, it is not affected
by the change in α. For low α values, the results for SEM and
REM are not shown because both heuristics fail to provide a
feasible solution. For the low-SNR regime, both EE and SEM
produce lower energy consumption with increasing α, but the
decrease is marginal. Unlike SEM and EE, the results for REM
first decrease and then start to increase. The rationale behind
this pattern can be explained as follows: Since REM prefers
SUs that are already assigned with a channel for sensing when
selecting SUs for channel m, long sensing duration causes SUs
with low γm, n to be assigned to channel m. We observe that
sensing energy component dominates in the low-SNR regime;
therefore, this causes an increase in total energy consumption
for REM. On the contrary, for high-SNR regime, REM pro-
duces lower energy consumption values as the reporting energy
component is the dominating factor. Both figures show that with
only a small amount of additional sensing time, great energy
savings are possible.
To sum up, all three energy minimization methods (EE, SEM,
and REM) provide significant energy savings compared with a
pure TXT technique. In all cases, EE achieves the best energy
values, whereas the performance of SEM and REM depends
on the parameter values. On one hand, a low μSNR or high α
favors SEM. On the other hand, a high μSNR or high fs supports
REM. As both heuristics have very low complexity, both can
be executed in a short amount of time, and one can select the
method with the better energy consumption.
VII. C ONCLUSION
In this paper, we have formulated the EE CSS problem for
a CRN and presented various approaches for this problem.
Each scheme ensures the minimum detection probability con-
straint as a PU protection criterion and the maximum false
alarm probability constraint as a CRN operability criterion in
each channel. EE, SEM, and REM aim to minimize energy
expenditure for sensing, whereas TXT minimizes time spent for
the sensing task to leave more time for data transmission. We
have investigated the performance of our proposals with various
parameters. To find the optimal solution, we have employed
the outer linearization method. Numerical evaluations have
shown that by sacrificing very little data transmission time, a
significant amount of energy can be saved. Furthermore, re-
porting energy is an important factor in energy consumption,
particularly when the SNR or sampling frequency is high.
As future work, we plan to incorporate different fusion rules,
e.g., AND, MAJORITY, etc., into our model. Moreover, we also
would like to analyze the impact of channel switching delay and
energy consumption of channel switching on sensing energy
consumption. Another point to pursue is to treat false alarm
probabilities as decision variables and jointly optimize them
together with sensing times.
2698 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 62, NO. 6, JULY 2013

Fig. 6. Effect of sensing duration (Ts ) under low and high SNRs with N = 200, δ min = 3. (a) Low SNR: μSNR = −5 dB. (b) High SNR: μSNR = 2 dB.

A PPENDIX A A PPENDIX B
P ROOF OF T HEOREM 1 P ROOF OF L EMMA 1
d
The first derivative of Pm, n with respect to τm, n is By combining (1) and (2), we get
√ √  d 
d
dPm, n γ f 1 γm, n fs Q−1 Pm, n
= √ √m, n s Am, n −√ +  < 0.
dτm, n 2 τm, n 2π 2γm, n + 1 τm, n 2γm, n + 1

where The first term is always negative, whereas the second term is
d d
⎛   f   2 ⎞ negative if Pm, n > 0.5. Since it is reasonable to assume a Pm, n
1 Q−1 Pm, − τm, n fs γm, n d
value greater than 0.5, we can safely say that Pm, n is a concave
Am, n = exp ⎝− n ⎠.
function of τm, n most of the time.
2 2γm, n + 1

The first derivative is always positive; hence, Pm, d A PPENDIX C

n is an
increasing function of τm, n . P ROOF OF T HEOREM 2
d
The second derivative of Pm, n with respect to τm, n is Let τn denote the τ vector with n entries that consists of τm, n
given as values for channel m. Moreover, let fm, k and hm, k denote (1 −
⎡ d
√ √ Pm, k ), and fm, 1 fm, 2 , . . . , fm, k , respectively. The proof is by
d2 Pm,
d
n γ m, n f A
s m, n ⎣ 1 γm, n fs induction on the number of elements in Sm denoted by |Sm |.
2
= √  − +
dτm, n 4 2π 2γm, n +1 3
τm, n
τm, n (2γm, n +1) • |Sm | = 2: Without loss of generality, assume that SUs 1
⎤ and 2 are in Sm . We can rewrite Qdm as 1 − hm, 2 .
  The gradient of hm, 2 is given by
 f   ⎥ 
× Q−1 Pm, n − τm, n fs γm, n ⎦. √
∂hm, 2 γm, 1 fs Am, 1
= − √ √  fm, 2 ,
∂τ2 2 τm, 1 2π 2γm, 1 + 1
The second derivative is negative if √ 
√ γm, 2 fs Am, 2
− √ √  fm, 1
−1 γm, n fs 2 τm, 2 2π 2γm, 2 + 1
 +
3
τm, τm, n (2γm, n + 1)
n
where
  f    ⎛   f   2 ⎞
× Q−1 Pm, n − τ f γ
m, n s m, n < 0. 1 Q−1 Pm, − τ f γ
= exp ⎝− ⎠.
m, n s m, n
Am, n n
2 2γm, n + 1
Reducing the τm, n term leads to
√
1 γm,n fs  −1  f   
Both terms are always negative; thus, hm, 2 is a decreasing
−√ + Q Pm,n − τm, n fs γm, n < 0.
τm, n (2γm,n + 1) function of τ2 . Therefore, Qdm is an increasing function
of τ2 since (∂Qdm /∂τ2 ) = −(∂hm, 2 /∂τ2 ). In addition,
d
Thus, Pm, n is a concave function of τm, n if the condition in as shown in Lemma 2, both fm, 1 and fm, 2 are nonneg-
(2) is satisfied. ative, decreasing, and convex functions; therefore, their
ERYIGIT et al.: EE MULTICHANNEL CSS WITH HETEROGENEOUS CHANNEL CONDITIONS FOR CRNS
multiplication hm, 2 is also convex [21], which leads to the
concavity of Qdm .
• Let us assume that the theorem holds for |Sm | = k and
show that it also holds for |Sm | = k + 1. This time, Qdm
can be written as
Qdm = 1 − hm, k+1 = 1 − hm, k fm, k+1 .
The gradient of hm, k+1 is given by
 
∂hm, k+1 ∂hm, k ∂fm, k+1
= fm, k+1 , hm, k . 2011, pp. 2723–2731.
∂τk+1 ∂τk+1 ∂τk+1
Let us focus on the first term. Since (∂hm, k /∂τk+1 ) is
negative by induction and fm, k+1 is nonnegative, then
their multiplication is negative. For the second term, hm, k
is a nonnegative function, and (∂fm(k+1) /∂τk+1 ) is neg-
ative by Lemma 2. Thus, their multiplication is also nega-
tive. Since both terms are negative, hm, k+1 is a decreasing
function of τk+1 .
We apply the same logic as in the previous step to
prove the convexity of hm, k+1 . Both hm, k and fm, k+1
are decreasing convex functions (convexity of hm, k comes
from induction); then, their multiplication hm, k+1 is also
convex. Thus, Qdm is a concave and increasing function
of τk+1 .
Proving both the base step and the induction step leads to the
conclusion that Qdm is an increasing concave function of τm, n
if (2) is satisfied.
R EFERENCES
[1] Y. Liang, Y. Zeng, E. Peh, and A. Hoang, “Sensing-throughput tradeoff for
cognitive radio networks,” IEEE Trans. Wireless Commun., vol. 7, no. 4,
pp. 1326–1337, Apr. 2008.
[2] G. Ganesan and Y. Li, “Cooperative spectrum sensing in cognitive radio
networks,” in Proc. 1st IEEE Int. Symp. New Frontiers Dyn. Spectr. Access
Netw., 2005, pp. 137–143.
[3] C. Song and Q. Zhang, “Cooperative spectrum sensing with multi-channel
coordination in cognitive radio networks,” in Proc. IEEE ICC, 2010,
pp. 1–5.
[4] Y. Liu, S. Xie, Y. Zhang, R. Yu, and V. C. M. Leung, “Energy-efficient
spectrum discovery for cognitive radio green networks,” Mobile Netw.
Appl., vol. 17, no. 1, pp. 64–74, Feb. 2011.
[5] H. Su and X. Zhang, “Opportunistic energy-aware channel sens-
ing schemes for dynamic spectrum access networks,” in Proc. IEEE
GLOBECOM, 2010, pp. 1–5.
[6] H. Su and X. Zhang, “Power-efficient periodic spectrum sensing for
cognitive MAC in dynamic spectrum access networks,” in Proc. IEEE
WCNC, 2010, pp. 1–6.
[7] H. Su and X. Zhang, “Energy-efficient spectrum sensing for cognitive
radio networks,” in Proc. IEEE ICC, 2010, pp. 1–5.
[8] Y. Wu and D. Tsang, “Energy-efficient spectrum sensing and transmission
for cognitive radio system,” IEEE Commun. Lett., vol. 15, no. 5, pp. 545–
547, May 2011.
[9] Y. Pei, Y. Liang, K. Teh, and K. Li, “Energy-efficient design of sequen-
tial channel sensing in cognitive radio networks: Optimal sensing strat-
egy, power allocation, and sensing order,” IEEE J. Sel. Areas Commun.,
vol. 29, no. 8, pp. 1648–1659, Sep. 2011.
[10] S. Huang, H. Chen, Y. Zhang, and F. Zhao, “Energy-efficient cooperative
spectrum sensing with amplify-and-forward relaying,” IEEE Commun.
Lett., vol. 16, no. 4, pp. 450–453, Apr. 2012.
[11] R. Deng, J. Chen, C. Yuen, P. Cheng, and Y. Sun, “Energy-efficient coop-
erative spectrum sensing by optimal scheduling in sensor-aided cognitive
radio networks,” IEEE Trans. Veh. Technol., vol. 61, no. 2, pp. 716–725,
Feb. 2012.
2699
[12] S. Maleki, S. Chepuri, and G. Leus, “Energy and throughput efficient
strategies for cooperative spectrum sensing in cognitive radios,” in Proc.
12th IEEE Int. Workshop SPAWC, 2011, pp. 71–75.
[13] C. Lee and W. Wolf, “Energy efficient techniques for cooperative spec-
trum sensing in cognitive radios,” in Proc. 5th IEEE CCNC, 2008,
pp. 968–972.
[14] E. C. Peh, Y.-C. Liang, Y. L. Guan, and Y. Pei, “Energy-efficient co-
operative spectrum sensing in cognitive radio networks,” in Proc. IEEE
GLOBECOM, 2011, pp. 1–5.
[15] X. Sun, T. Zhang, and D. Tsang, “Optimal energy-efficient cooperative
sensing scheduling for cognitive radio networks with QoS guarantee,” in
Proc. 7th IEEE IWCMC, 2011, pp. 1825–1830.
[16] T. Zhang and D. Tsang, “Optimal cooperative sensing scheduling for
energy-efficient cognitive radio networks,” in Proc. IEEE INFOCOM,
[17] X. Hao, M. Cheung, V. Wong, and V. Leung, “A coalition formation
game for energy-efficient cooperative spectrum sensing in cognitive radio
networks with multiple channels,” in Proc. IEEE GLOBECOM, 2011,
pp. 1–6.
[18] M. Alouini and A. Goldsmith, “Capacity of Rayleigh fading channels un-
der different adaptive transmission and diversity-combining techniques,”
IEEE Trans. Veh. Technol., vol. 48, no. 4, pp. 1165–1181, Jul. 1999.
[19] A. Conejo, E. Castillo, R. Minguez, and R. Garcia-Bertrand, Decomposi-
tion Techniques in Mathematical Programming. New York, NY, USA:
Springer-Verlag, 2006.
[20] C. Floudas, Nonlinear and Mixed Integer Optimization. Fundamentals
and Applications. New York, NY, USA: Oxford Univ. Press, 1995.
[21] M. M. Kantrowitz and R. Neumann, “Optimization for products of con-
cave functions,” Rend. Circ. Mat. Palermo, vol. 54, no. 2, pp. 291–302,
Jun. 2005.
Salim Eryigit (S’12) received the B.S. degree
in industrial engineering and the M.S. degree in
computer engineering from Bogazici University,
Istanbul, Turkey, in 2003 and 2008, respectively.
He is currently working toward the Ph.D. degree in
computer engineering with Bogazici University.
His research interests include network optimiza-
tion, cognitive radio networks, green communica-
tions, and scheduling.
Suzan Bayhan (M’13) received the B.S., M.S.,
and Ph.D. degrees in computer engineering from
Bogazici University, Istanbul, Turkey, in 2003, 2006,
and 2012, respectively.
She is currently a Postdoctoral Researcher with
the Helsinki Institute for Information Technology,
Aalto University, Espoo, Finland. Her current re-
search interests include cognitive radio networks,
small cells, green communications, and mobile op-
portunistic networks.
Tuna Tugcu (M’00) received the B.S. and Ph.D.
degrees in computer engineering from Bogazici
University, Istanbul, Turkey, in 1993 and 2001,
respectively, and the M.S. degree in computer and
information science from the New Jersey Institute of
Technology, Newark, NJ, USA, in 1994.
He worked as a Postdoctoral Fellow and as a Vis-
iting Professor with the Georgia Institute of Technol-
ogy, Atlanta, GA, USA. He is currently an Associate
Professor with the Department of Computer Engi-
neering, Bogazici University. His research interests
include cognitive radio networks, WiMAX, and molecular communications.