5306 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO.

12, DECEMBER 2008

Resource Allocation for Spectrum Underlay in

Cognitive Radio Networks
Long Bao Le, Member, IEEE, and Ekram Hossain, Senior Member, IEEE

Abstract—A resource allocation framework is presented for of the primary network from excessive interference due to
spectrum underlay in cognitive wireless networks. We consider spectrum access from secondary networks. As described in
both interference constraints for primary users and quality of [3], users from a secondary network (who are referred to as
service (QoS) constraints for secondary users. Specifically, inter-
ference from secondary users to primary users is constrained to secondary users) can access the spectrum owned by the pri-
be below a tolerable limit. Also, signal to interference plus noise mary network provider (whose users are referred to as primary
ratio (SINR) of each secondary user is maintained higher than users) using spectrum underlay or spectrum overlay. Spectrum
a desired level for QoS insurance. We propose admission control underlay and overlay techniques are the basis for designing
algorithms to be used during high network load conditions the emerging cognitive radio networks. Also, in contrast to
which are performed jointly with power control so that QoS
requirements of all admitted secondary users are satisfied while resource allocation problems in traditional wireless networks
keeping the interference to primary users below the tolerable [4]-[6], in cognitive radio networks primary users should be
limit. If all secondary users can be supported at minimum protected while secondary users access the spectrum.
rates, we allow them to increase their transmission rates and
In the spectrum overlay paradigm, secondary users are only
share the spectrum in a fair manner. We formulate the joint
power/rate allocation with proportional and max-min fairness allowed to access spectrum resources (i.e., channels) owned
criteria as optimization problems. We show how to transform by the primary network provider if these channels are not
these optimization problems into a convex form so that their being used by primary users. Here, spectrum sensing needs to
globally optimal solutions can be obtained. Numerical results be performed to avoid possible collision with primary users
show that the proposed admission control algorithms achieve
[7], [8]. Given that the spectrum opportunities are detected
performance very close to that of the optimal solution. Also,
impacts of different system and QoS parameters on the network by a spectrum sensing technique, it is important to coordinate
performance are investigated for the admission control, and the spectrum sharing among secondary users in such a way
rate/power allocation algorithms under different fairness criteria. that different design objectives can be achieved. Centralized
architectures for spectrum sensing and sharing in cognitive
Index Terms—Cognitive radio, spectrum sharing, spectrum wireless networks were proposed in [9] and [10]. In general,
underlay, spectrum overlay, interference temperature limit, ad- a medium access control (MAC) layer protocol with sensing
mission control, rate and power allocation. capability is important to perform fair resource allocation
among secondary users while avoiding collision with primary
I. I NTRODUCTION users.

T HE emerging high-speed wireless access technologies For the spectrum underlay paradigm, it is required that an
and the requirements of different wireless applications interference limit corresponding to an interference temperature
are expected to create huge demand for spectral resources in level be maintained at receiving points of the primary network.
the next generation wireless systems. Achieving high spectrum A graph-theoretic model for spectrum sharing/access among
utilization is, therefore, one of the most critical research secondary users was proposed in [11] where different objective
objectives in designing wireless communication systems today. functions were investigated. In [12], the channel allocation
In fact, it has been reported that current utilization of some problem was formulated using game theory. Here, the pro-
allocated spectrum can be as low as 15 % [1]. Therefore, posed utility functions capture the interference perceived by
there is an increasing interest in developing efficient methods one user on each channel and/or the interference this user
for spectrum management and sharing which is encouraged by creates to its neighboring users. In these papers, primary
both industry and FCC [2]. This motivates to exploit spectrum users were not explicitly protected from interference due to
opportunities in space, time, frequency while protecting users spectrum access by secondary users. In [13], a heuristic-
based channel and power allocation algorithm was proposed
Manuscript received August 10, 2007; revised October 15, 2007; accepted where interference constraints for primary users were con-
October 26, 2007. The associate editor coordinating the review of this paper sidered. However, how good the performance of the pro-
and approving it for publication was M. Guizani. This work was supported
in part by the University of Manitoba Graduate Fellowship (UMGF) and in posed algorithm compared to the optimal solution was not
part by a grant from the Natural Sciences and Engineering Research Council known. In [14], a user removal algorithm based on the tree-
(NSERC) of Canada. pruning algorithm was proposed so that QoS constraints for
L. B. Le is with the Massachusetts Institute of Technology, USA (e-mail:
longble@mit.edu). secondary users and interference temperature constraints for
E. Hossain is with the Department of Electrical and Computer Engi- primary users are satisfied. The proposed removal algorithm
neering, University of Manitoba, Winnipeg, MB, Canada R3T 5V6 (e-mail: is, however, computationally extensive.
ekram@ee.umanitoba.ca).
Digital Object Identifier 10.1109/T-WC.2008.070890 In this paper, we propose a resource allocation frame-
1536-1276/08$25.00 
c 2008 IEEE
LE and HOSSAIN: RESOURCE ALLOCATION FOR SPECTRUM UNDERLAY IN COGNITIVE RADIO NETWORKS 5307

work for spectrum underlay in cognitive wireless networks. network or simply a mobile node in an ad hoc network. The
Specifically, we consider design problems of cognitive wire- CDMA technology will be assumed although the model may
less networks using code-division multiple access (CDMA) be extended for other technologies as well.
technology with explicit interference protection for primary Simultaneous communications among users (i.e., both pri-
users and QoS constraints for secondary users. We assume mary and secondary users) will interfere with each other. We
that secondary users have the requirement of a minimum will consider the interference constraints at the receiving ends
transmission rate in terms of SINR (or equivalent bit error rate of the primary network which will be referred to as primary
(BER)) and maximum power constraints. When the network receiving points in the sequel. We assume that each primary
load is high, we show how to perform admission control to receiving point can tolerate a maximum interference level.
guarantee QoS constraints for secondary users and interference Also, secondary links have desired QoS performance in terms
constraints for primary users. If all of the secondary users can of BER. Fig. 1 illustrates the transmission setting considered
be supported, we solve the joint rate and power allocation in this paper.
problems under two popular fairness criteria, namely, the max-
min and the proportional fairness (PF) criteria [15], [16].
Similar design problems exist in conventional wireless A. Modeling of QoS and Interference Constraint
networks. In [17], different user removal algorithms for joint Assume that there are M primary receiving points and N
admission control and power control were proposed which secondary communication links in the considered geographical
employed the power control technique developed in [18]. area. Let us denote the channel gain from the transmitting
These admission control algorithms were proposed for the node of secondary link i to receiving node of secondary link
traditional cellular wireless networks and did not take into (s)
j by gj,i while the channel gain from the transmitting node
consideration the interference constraints which are present (p)
in a cognitive radio network based on spectrum underlay. of secondary link i to primary receiving point j as gj,i . If Ni
Moreover, joint rate and power allocation problems in CDMA denotes the total noise and interference due to primary users’
wireless network were solved in several works in the literature transmissions at the receiving side of secondary link i, for
[19]-[21]. Most of the existing works in the literature, however, wireless access based on CDMA, the corresponding effective
aim to maximize the network throughput. However, it is im- bit-energy-to-noise power spectral density ratio can be written
portant to maintain throughput fairness among users to avoid as [19]
(s)
severe QoS degradation for users with unfavorable channel B gi,i Pi
conditions. In addition, for the spectrum sharing problem, both μi =  (s)
(1)
Ri N gi,j Pj + Ni
j=1,j=i
QoS and interference constraints need to be considered. These
design aspects will be considered in the resource allocation where B is the spectrum bandwidth, Ri is the transmission
framework proposed in this paper. rate of secondary link i. Here, B/Ri is the processing gain
The rest of this paper is organized as follows. System model which is usually required to be larger than a particular value.
and problem definition are presented in Section II. Section III The processing gain is simply equal to one for other multiple
presents the solutions for the admission control problem. The access technologies such as FDMA and μi denotes the SINR.
solutions of the power/rate allocation problem for the spectrum In the sequel, we will refer to μi as SINR in all cases. Now,
underlay model are described in Section IV. Section V presents if a particular modulation scheme is employed, there will be
the numerical results. Conclusions are stated in Section VI. an explicit relationship between BER and SINR. Thus, for
a specific required level of BER for secondary link i, μi is
II. S YSTEM M ODEL AND P ROBLEM D EFINITION required to be larger than a corresponding value γi . Hence,
the QoS requirement for secondary link i can be expressed as
We consider the hierarchical spectrum sharing problem
in a cognitive wireless network where several unlicensed μi ≥ γi , i = 1, 2, · · · , N. (2)
users (secondary users) and licensed users (primary users)
transmit data in the same spectral region. Specifically, the Now, let Ij be the maximum interference level tolerable
spectrum underlay model is investigated in this paper where by primary receiving point j. The interference constraint for
secondary users and primary users transmit simultaneously primary receiving point j can be written as
in a common frequency band (e.g., as in a CDMA-based
N

wireless network). Interested readers are referred to [3] for (p)
gj,i Pi ≤ Ij , j = 1, 2, · · · , M (3)
an excellent survey and for more details about the taxonomy
i=1
of these spectrum access models. The entities we work with
are the communication links each of which corresponds to a where total interference at the primary receiving point j
pair of users who wish to communicate with each other. We should be smaller than the tolerable limit. We assume that
will refer to the communication links belonging to secondary transmission rate of secondary link i can be adjusted in an
networks as secondary links. allowable range where the minimum and the maximum values
We assume that there are a number of primary and sec- are Rimin and Rimax , respectively. Also, power of secondary
ondary users communicating with their partners simultane- link i is constrained to be smaller than the maximum limit
ously. Here, the term “user” will be used broadly where it can Pimax . We assume that all channel gains can be made available
be a mobile node or base station/access point in a centralized at a control point.
5308 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008

Remark: If the channel is frequency-selective and an L-finger III. A DMISSION C ONTROL A LGORITHMS
RAKE receiver is used at each receiver, the effective bit- As has been mentioned in Section II, we will consider the
energy-to-noise spectral density ratio can be written as [21] admission control problem when the network load is high and
L (s) all of the secondary links transmit with their minimum rate (if
B Pi l=1 gi,i,l
μi =  L (s) (4) admitted). Now, using equation (1), we can rewrite the QoS
Ri 2 N
3 j=1,j=i Pj l=1 gi,j,l + Ni constraint in (2) as follows:
N (s)
(s)  γi Rimin gi,j γi Rimin Ni
where gi,j,l is the channel gain for the l-th path from the Pi ≥ Pj + , i = 1, 2, · · · , N.
transmitter of link j to the receiver of link i. Let us denote B g (s) B g (s)
(s) L (s) (s) L (s) j=1,j=i i,i i,i
gi,i = l=1 gi,i,l and gi,j = 2/3 l=1 gi,j,l (i = j).
(6)
Substituting these quantities into (4), we can obtain equation These constraints for all secondary links can be written in
(s)
(1) with the corresponding “equivalent” channel gains gi,j . the matrix form as follows:
As a consequence, the same analysis can be applied to both (I − F )P  u (7)
cases.
where I is an identity matrix of order N × N , u is a column
vector which can be written as
B. Admission Control Problem  
γ1 R1min N1 γ2 R2min N2 min
γN RN NN
The admission control problem considered in this paper u= , ,··· ,
is similar to that investigated in [17]. Specifically, secondary B g (s) B g (s) B g
(s)
1,1 2,2 N,N
links requesting access to the spectrum licensed to the primary 
where (.) denotes the matrix/vector transpose. And F is an
network have QoS requirements expressed in (2). The major N × N matrix whose (i, j)-th element is
challenge for the spectrum sharing problem comes from the ⎧
interference protection for primary users expressed in (3) ⎨ γi Rmin (s)
gi,j
i
(s) , if i = j
which is usually not considered in the conventional fixed Fi,j = B gi,i
⎩ 0, if i = j.
spectrum allocation paradigm.
In particular, we are interested in the scenario where a
number of secondary links wish to access the spectrum with A. Constrained Power Control
minimum transmission rate (i.e., Ri = Rimin ) and both Recall that we are interested in a scenario where not all N
the QoS requirements (in (2)) as well as the interference secondary links can be admitted into the network while satis-
constraints (in (3)) are satisfied. The problem is how to choose fying the QoS and the interference constraints stated in (2) and
the subset of requesting links with maximum size such that (3), respectively. We will first focus on the power allocation
the constraints in (2) and (3) are both satisfied. problem under maximum power constraint (i.e., P  P max )
and QoS constraints while ignoring the interference constraints
for the time being.
C. Joint Rate and Power Allocation Problem In [18], the authors proposed an efficient iterative power
When the network load is low, all requesting secondary control algorithm which can be implemented synchronously or
links with minimum transmission rates can be supported while asynchronously. Specifically, let Pi (t) and Pi (t + Δt) denote
satisfying the QoS and the interference constraints in (2) the power levels of secondary link i after two consecutive
and (3), respectively. If this is the case, secondary links power updates at time instants t and t + Δt, respectively. The
would increase their transmission rates above the minimum power of secondary link i at time t+Δt is updated as follows:
values and share the spectrum in a fair manner. For notational γi
convenience, we will arrange power, rate and other quantities Pi (t + Δt) = min Pimax , Pi (t) (8)
μi (t)
of all secondary links into the corresponding vectors. For
example, P will denote a column vector whose element Pi where μi (t) is the instantaneous SINR at the receiving side of
is the transmission power of secondary link i. The joint rate secondary link i at time instant t with power vector P (t). It
and power allocation problem can be stated as was shown in [18] that this power control algorithm converges
to the fixed point solution of
maximize f (R)
subject to Rmin  R  Rmax , P  P max (5) P = min {P max , F P + u} (9)
and constraints (2), (3) which will be referred to as the stationary power vector. Let Ω
be the set of secondary links and P Ω be the stationary power
where the objective function f (R) strikes a balance between
vector when the power algorithm with the rule as in (8) is run
maximizing throughput and maintaining fairness for different
with the secondary link set Ω. From the results obtained in
secondary links. Specifically, we will investigate two fairness
[17], we have the following facts:
criteria, i.e., f (R) =  max {mini Ri } to achieve max-min
N
fairness and f (R) = i=1 ln(Ri ) to achieve proportional Fact 1: If all secondary links in Ω can be supported (i.e., the
fairness [15], [16]. We will show how to solve the admission power control algorithm in (8) results in a stationary power
control as well as the joint rate and power allocation problems vector P Ω satisfying the QoS constraints in (2)), the QoS
in Sections III and IV, respectively. constraints will be satisfied with equality.
LE and HOSSAIN: RESOURCE ALLOCATION FOR SPECTRUM UNDERLAY IN COGNITIVE RADIO NETWORKS 5309


Fact 2: If a subset Ω ⊆ Ω is the set of secondary links Proof: The proof is provided in Appendix A.
which are not supported with stationary power vector P Ω , Because of the complexity of the problem, we propose a

then PiΩ = Pimax for i ∈ Ω . low-complexity admission control algorithm in this subsection.
Now, let us define the following “interference measures”: The proposed algorithm also removes the “worst” link one-
⎡ ⎤ by-one. In each step, we perform the power control algorithm
N (s)
 (s) gi,i B Ω as in (8) and remove one secondary link from the network
Ω
αi (P ) = ⎣Pi Ω ⎦
gj,i + Ni − P (10)
γi Ri i until the remaining set of links satisfies both the QoS and the
j=1,j=i interference constraints. Here, the key issue is to construct a
⎡ ⎤
 N (s)
gi,i B Ω removal criterion which achieves good overall performance.
(s)
βi (P Ω ) = ⎣ gi,j PjΩ + Ni ⎦ − P (11) Because there are two different kinds of constraints, we
γi Ri i
j=1,j=i consider the following cases in each removal step.

DΩ0 (P Ω ) = βi (P Ω ), for Ω0 ⊆ Ω. (12) Case 1: Interference constraints for all primary receiving
i∈Ω0 points stated in (3) are satisfied but QoS constraints in (2)
are violated.
We can easily see that
  In this case, we employ the SMIRA algorithm as presented
DΩ (P Ω ) = βi (P Ω ) = αi (P Ω ). (13) in Section III.A.
i∈Ω i∈Ω
Case 2: Interference constraints for primary receiving points
We can also see that if the QoS constraint for secondary stated in (3) are violated.
link i is satisfied with equality, then βi (P Ω ) = 0. Also, Note that this case covers both the scenarios where QoS
DΩ (P Ω ) = 0 if and only if all secondary links in Ω are constraints in (2) are violated or not. In this case, we would
supported. In general, we have βi (P Ω ) ≥ 0 and the value of remove the link which violates both QoS and interference
βi (P Ω ) reflects the degree by which the QoS constraint for constraints the most in each step. Now, we define the measure
secondary link i is violated. Also, it is intuitive that αi (P Ω ) which quantifies degree of violation at primary receiving point
quantifies the aggregate relative interference that secondary j as follows:
link i creates for other links in Ω. We will present single and
N

multiple link removal algorithms for admission control in the (p)
ηj (Ω) = Ij − gj,i PiΩ . (15)
following subsection.
i=1

We propose the following removal criterion in which the

B. Admission Control with QoS and Interference Constraints
removal metric quantifies the aggregate interference effect that
In [17], the authors proposed several removal algorithms each link creates to other links:
which aim at maximizing the number of secondary links which 
can be admitted into the network while satisfying the QoS ∗ DΩ (P Ω )
i = argmax M ×
requirements. Among these proposed algorithms, SMIRA and i∈Ω DΩ (P Ω ) + k=1 ηk (Ω)
SMART(R) are the two most efficient ones. Note that the ⎡ ⎤
 (s)  (s)
interference measures defined in (10)-(12) are not the same max ⎣ gj,i PiΩ , gi,j PjΩ ⎦
as those in [17]. However, the spirit of the SMIRA and j∈Ω,j=i j∈Ω,j=i
SMART(R) algorithms remain the same in this paper. ⎫
1) Single Removal Algorithm: SMIRA runs the power
M
 ηj (Ω) ⎬
(p) Ω
+  g j,i Pi . (16)
control algorithm in (8) until reaching the stationary power DΩ (P Ω ) + M ⎭
j=1 k=1 ηk (Ω)
vector. Based on this stationary power vector, the algorithm
removes one link from the network and runs the power control  (s) Ω
In fact, j∈Ω,j=i gj,i Pi denotes the total interference
algorithm again. This procedure is repeated until the remaining that secondary link i creates to other secondary links while
set of links can be supported. The removal criterion of SMIRA  (s) Ω
j∈Ω,j=i gi,j Pj is the total interference received at the
is as follows: (p)
   receiving end of link i. Moreover, gj,i PiΩ denotes the interfer-
i∗ = argmaxi∈Ω max αi (P Ω ), βi (P Ω ) . (14) ence that secondary link i creates for primary receiving point
j. Recall that DΩ (P Ω ) quantifies the degree of violation for
Intuitively, SMIRA algorithm removes the link which vio-
QoS constraints and ηj (Ω) quantifies the degree of violation
lates the QoS constraints the most and/or creates the largest
for the interference constraint of primary receiving point j.
amount of interference to other links in each removal step.
Therefore, the proposed criterion removes in each step the
Thus, it can potentially remove the least number of links from
secondary link which creates the largest amount of interfer-
the network. In our spectrum sharing problem, besides QoS
ence for primary receiving points and receives or creates the
constraint, admission and power control should be done such
largest amount of interference from/for other secondary links
that interference constraints for primary links stated in (3) are
in the weighted average sense. As a result, it would potentially
also satisfied. We have the following result on the complexity
remove the least number of secondary links from the network.
of this admission control problem.
We will refer to this algorithm as interference constraint-
Proposition 1: The admission control problem with QoS and aware SMIRA (I-SMIRA) in the sequel. The computational
interference constraints is NP-hard. complexity of I-SMIRA is only of O(N 2 ).
5310 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008

2) Multiple Removal Algorithm: In practice, it may be R in each removal step which satisfies
desired to have a fast admission control algorithm, especially 
when many links are to be removed in a highly congested DΩ (P Ω )
Ω1 = argmax M
network. In this case, multiple links may be removed from the Ω1 DΩ (P Ω ) + k=1 ηk (Ω)
network in each removal step. In [17], the authors proposed ⎧ ⎫
 ⎨   (s) ⎬
the SMART(R) algorithm which removes R links from the (s) Ω
× max gj,i Pi , gi,j PjΩ
network in each step. Before presenting the removal metric, ⎩ ⎭
i∈Ω1 j∈Ω,j=i j∈Ω,j=i
let us define ⎫
 M
ηj (Ω) ⎬
(p)
+  M
gj,i PiΩ . (20)
Ω/Ω PiΩ , if i ∈ Ω0 DΩ (P Ω ) + ηk (Ω) ⎭
Pi 0 = (17) i∈Ω1 j=1 k=1
0, otherwise.
In fact, the removal metric in (20) is similar to that for single
removal in (16). However, the total interference effects of the
Now, the SMART(R) algorithm removes links in the set Ω\Ω0
potential removal set is accounted for in this metric. Note that
which attains
Ω0 is the remaining set of links after R links in the set Ω1 =
  Ω \ Ω0 have been removed from the network. As before, the
max DΩ (P Ω ) − DΩ0 (P Ω/Ω0 ) (18) power control algorithm in (8) is run until the stationary power
Ω0 ⊂Ω
vector is attained before the removal algorithm is activated.
The computational complexity of I-SMART(R) is of O(N 2 )
where
 |Ω \ Ω0 | = R and recall that DΩ0 (P Ω/Ω0 ) = which is the same as that of I-SMIRA algorithm.
Ω/Ω0
i∈Ω0 βi (P ). In essence, the SMART(R) algorithm re-
moves a set of links of size R which contributes the most to
DΩ (P Ω ). Note that DΩ (P Ω ) reflects the degree of violation C. Some Implementation Issues
of QoS constraints. Thus, the SMART(R) algorithm tends to In order to guarantee both the QoS requirements for sec-
remove the set of worst links in each removal step. To simplify ondary users and interference constraints for primary users,
the calculation of (18), we can use the following result. (s) (p)
all channel gains (i.e., gi,j and gi,j ) must be continuously
estimated for online admission control operation. Also, per-
Proposition 2: The removal metric of the SMART(R) algo-
forming distributed power and admission control for spectrum
rithm can be rewritten as
underlay may be difficult because of the following reasons.
 First, interference constraints are required to be satisfied at
DΩ (P Ω ) − DΩ0 (P Ω/Ω0 ) = βi (P Ω ) + αi (P Ω ) all times in general. This requirement may only be fulfilled
i∈Ω
/ 0 by centralized implementation because secondary links are
  ⎡ ⎤ not allowed to cause even a short-term violation of QoS
(s)
gi,i
B Ω  
+ P − Ni −⎣
(s)
gi,j PjΩ ⎦ . (19) and interference constraints as in the conventional wireless
γi Ri i networks [17], [18]. Second, primary users usually do not
i∈Ω
/ 0 j ∈Ω
/ 0 ,j=i (p)
assist secondary users in estimating the channel gains gi,j .
Thus, secondary users are responsible for making these chan-
Proof: The proof is provided in Appendix B. nel gains available at the central controller for power and
admission control purposes.
The formula to calculate the removal metric in this propo- (s)
sition is simpler to compute because it only calculates all the The channel gains between secondary links (i.e., gi,j ) can
terms over the removal set of size R which is usually very be estimated by using pilot-aided or other approaches [22],
small. Note that if R = 1, the SMART(R) algorithm removes [23]. Channel gains from transmitters of secondary links
(p)
a single link in each removal step. Now, for our admission to primary receiving points (i.e., gi,j ) can be estimated by
control problem with both QoS and interference constraints, employing sensors near all receiving points which are then
we propose an interference constraint-aware SMART(R) re- forwarded to the central controller [8].
moval algorithm which will be referred to as I-SMART(R) in
the sequel. Similar to the I-SMIRA algorithm, we consider the IV. O PTIMIZATION OF J OINT R ATE AND P OWER
following cases in each removal step. A LLOCATION
Case 1: Interference constraints for all primary receiving When the network load is low, all secondary links can be
points stated in (3) are satisfied but QoS constraints in (2) admitted into the network. In this case, they can increase
are violated. their transmission rates above the minimum values. In this
In this case, we employ the SMART(R) algorithm as section, we show how to solve the optimization problem stated
presented above. in (5). The decision variables are transmission rates R and
powers P . As in [19], we will assume that continuous rate
Case 2: Interference constraints for primary receiving points and rate variation is achieved by varying the processing gain.
stated in (3) are violated The results obtained in this paper, therefore, can be considered
In this case, we remove the set of links Ω1 = Ω \ Ω0 of size as upper bounds for the discrete case.
LE and HOSSAIN: RESOURCE ALLOCATION FOR SPECTRUM UNDERLAY IN COGNITIVE RADIO NETWORKS 5311

A. Rate and Power Allocation with Max-Min Fairness Proposition 3: The optimal solution of the joint rate and power
It is observed that the objective function in (5) is equivalent allocation problem satisfies Ri = Rj , ∀i, j.
to min {maxi 1/Ri }. By introducing a new variable t and Proof: The proof is provided in Appendix C.
writing down all the constraints explicitly, the optimization Hence, the rate and power allocation problem under max-
problem (5) becomes equivalent to min fairness achieves perfectly fair rates for all secondary links
in the sense that optimal transmission rates for all links are
minimize t the same.
subject to
1/Ri ≤ t, i = 1, 2, · · · , N B. Rate and Power Allocation with Proportional Fairness
(s)
B gi,i Pi For proportional fairness criterion, the objective function
≥ γi , i = 1, 2, · · · , N N
Ri N (s)
+ Ni f (R) = i=1 ln(Ri ) can be rewritten as
j=1,j=i gi,j Pj
N N 
 (p) 
gj,i Pi ≤ Ij , j = 1, 2, · · · , M maximize ln Ri . (23)
i=1 i=1
Rimin ≤ Ri ≤ Rimax , Pi ≤ Pimax , i = 1, 2, · · · , N. N
This objective function is equivalent to min (1/ i=1 Ri ).
(21) After changing variables and using the procedure as in Section
This optimization problem is not convex. However, we can IV. A, the optimization in (5) with proportional fairness
transform it into a geometric program which can be solved criterion can be written as
efficiently (Chapter 4, [25]). Specifically, the optimization 
N

problem in (21) is equivalent to minimize − yi

i=1

minimize t subject to
⎛ ⎞
subject to γi 
N
γi Ni
ln ⎝ e(yi −xi ) ⎠ ≤ 0,
(s)
gi,j e(xj −xi +yi ) +
t−1 Ri−1 ≤ 1, i = 1, 2, · · · , N (s)
Bgi,i j=1,j=i Bgi,i
(s)

N i = 1, 2, · · · , N
γi −1 (s) γi Ni  
(s)
R i Pi gi,j Pj + R P −1 ≤ 1,
(s) i i N (p)
gj,i xi
Bgi,i j=1,j=i Bg i,i ln e ≤ 0, j = 1, 2, · · · , M
Ij
i = 1, 2, · · · , N i=1

N (p) ln(Rimin e−yi ) ≤ 0 ln((Rimax )−1 eyi ) ≤ 0,

gj,i
Pi ≤ 1, j = 1, 2, · · · , M ln((Pimax )−1 exi ) ≤ 0, i = 1, 2, · · · , N.
i=1
Ij
This equivalent optimization problem is convex. Therefore,
Rimin Ri−1 ≤ 1, (Rimax )−1 Ri ≤ 1, (Pimax )−1 Pi ≤ 1,
the globally optimal solution can be obtained using standard
i = 1, 2, · · · , N. (22) convex optimization techniques.
Now, let us define Pi = exi , Ri = eyi and t = es , substitute
these new variables into (22), and take ln in both the objective V. N UMERICAL R ESULTS
and the constraint functions. Then, we have the following We consider a simple network setting as shown in Fig. 1.
optimization problem: Assume that primary users communicate with its BS in the
minimize s uplink direction (i.e., a single cell is considered). Transmitting
subject to nodes corresponding to the secondary links are randomly
ln(e(−s−yi ) ) ≤ 0, i = 1, 2, · · · , N located in a rectangular area and the BS of primary network
⎛ ⎞ is located at the center of the rectangular area. The size of
γ 
N
γ N
ln ⎝
i (s)
gi,j e (x j −x i +y i )
+
i i
e (y i −x i ) ⎠ ≤ 0, the rectangular area is 2000m × 2000m. Also, receiving node
(s) (s)
Bgi,i j=1,j=i Bgi,i of each secondary link is generated randomly in a 1000m ×
i = 1, 2, · · · , N 1000m rectangle with its transmitting node being at the center.
 (p)
 The channel gains are modeled as gi,j
(s)
=
N
gj,i xi
ln e ≤ 0, j = 1, 2, · · · , M (s)
(s) (p) (p)
(p)
Ij K0 .10βi,j /10 .(di,j )−4 , gi,j = K0 .10βi,j /10 .(di,j )−4 , where
i=1
(s) (p) (s) (p)
ln(Rimin e−yi ) ≤ 0, ln((Rimax )−1 eyi ) ≤ 0, di,j and di,j are the corresponding distances, βi,j and βi,j
are random Gaussian variables with zero mean and standard
ln((Pimax )−1 exi ) ≤ 0, i = 1, 2, · · · , N.
deviation equal 6 dB, K0 = 103 which captures system and
transmission effects such as antenna gain, carrier frequency,
Since log-sum-exp functions are convex [25], this optimiza- etc. The total noise and interference at the receiving node
tion problem is a convex one. Therefore, the globally optimal of all secondary links is chosen to be Ni = N0 = 10−10
solution can be found by interior point methods [25]. We have W. The maximum transmission power on secondary links
the following property on the solution of the joint rate and is Pimax = 0.1 W. The spectrum bandwidth is B = 5.12
power allocation problem. MHz. We will denote the tolerable interference limit at
5312 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008

Optimal, I = 3N0
0.55
I−SMIRA, I = 3N0
0.5
I−SMART(2), I = 3N0
0.45 Optimal, I = 10N0

Outage probability
0.4 I−SMIRA, I = 10N0
0.35 I−SMART(2), I = 10N0

0.3
0.25
0.2
0.15
0.1
Secondary users Communication link
0.05
Primary users Interference to primary user 5 10 15 20
Target SINR (dB)
Fig. 1. Spectrum sharing among primary and secondary users.
Fig. 3. Outage probability versus target SINR for different interference
constraint requirements (for N = 7).
0.5
Optimal removal
0.45 SMIRA 10
I−SMIRA I−SMIRA
0.4 SMART(1) I−SMART(1)
I−SMART(1) 9
I−SMART(2)
Outage probability

0.35 I−SMART(2) I−SMART(3)

Number of admitted links
8
0.3

0.25 7

0.2
6
0.15
5
0.1

0.05 4
5 10 15 20
Target SINR (dB)
3
5 10 15
Fig. 2. Outage probability versus target SINR of different removal algorithms Number of requesting links
(for I = 5N0 , N = 7).
Fig. 4. Number of admitted links versus number of requesting links from
secondary users (for I = 5N0 , SINR = 15 dB).
the primary receiving point (i.e., BS) as I. The minimum
transmission rate on secondary links is Rimin = 64 kbps and
the maximum transmission rate is Rimax = B/P G, where are obtained by exhaustive search. As is evident, both of
P G is the minimum processing gain. For each simulation the proposed single removal algorithms (i.e., I-SMIRA, I-
run, the locations of secondary links (i.e., transmitting and SMART(1)) achieve outage probability very close to that of
receiving nodes) are generated randomly. The measure of optimal removal, especially for low and medium values of
interest is obtained by averaging over 103 simulation runs. target SINR (i.e., target SINR < 10 dB). Fig. 2 shows that
We will investigate performance of four removal algorithms, SMIRA and SMART(1) algorithms, which do not account
namely, SMIRA, SMART(R), I-SMIRA, and I-SMART(R) al- for interference constraints in the removal metrics, consis-
gorithms. For SMIRA and SMART(R) algorithms, we remove tently achieve higher outage probability than I-SMIRA and
links in each removal step using the metrics in (14), (18), I-SMART(1) for all target SINR values. Outage performance
respectively, until the remaining links satisfy both the QoS for I-SMART(2), which removes two links in each removal
and the interference constraints. In contrast, we consider two step, is also shown in these two figures. It can be observed
different cases in each removal step as presented in Section that I-SMART(2) has higher outage probability than all other
III.B for the I-SMIRA, I-SMART(R) removal algorithms. Two removal algorithms.
performance measures will be investigated, namely, outage We illustrate the number of admitted links and outage
probability and number of admitted links. Outage probability probability versus the number of requesting links for different
is calculated as the ratio of average number of removed links removal algorithms in Figs. 4-5, respectively. These two fig-
and average number of requesting links. ures show that I-SMIRA and I-SMART(1) algorithms achieve
The outage probability versus target SINR under different almost the same outage performance. In addition, outage per-
set of constraints for different removal algorithms and optimal formance of I-SMART(R) algorithm degrades rapidly with the
removal is shown in Figs. 2-3. Results for optimal removal number of removed links R. In order to quantify the tradeoff
LE and HOSSAIN: RESOURCE ALLOCATION FOR SPECTRUM UNDERLAY IN COGNITIVE RADIO NETWORKS 5313

0.7 2200
I−SMIRA PF, SINR = 10dB, I = 20N0
I−SMART(1) 2000 PF, SINR = 15dB, I = 5N
0.6 0
I−SMART(2)
I−SMART(3) 1800 Max−min, SINR = 10dB, I = 20N
0

Total throughput (kbps)

Max−min, SINR = 15dB, I = 5N0
1600
Outage probability

0.5

1400
0.4
1200

0.3 1000

800
0.2
600

400
0.1 10 15 20 25 30 35 40
5 10 15 Minimum processing gain
Number of requesting links

Fig. 7. Total throughput under different fairness criteria (for N = 5).

Fig. 5. Outage probability versus number of requesting links from secondary
users (for I = 5N0 , SINR = 15 dB).
1200
0.5
I−SMIRA 1100
0.45 I−SMART(1)
I−SMART(2) 1000

Total throughput (kbps)

0.4
I−SMART(3)
0.35 900

0.3 800
Probability

0.25
700
0.2
600
0.15
500 PF
0.1
Max−min
0.05 400
10 12 14 16 18 20
0 Target SINR (dB)
0 2 4 6 8 10 12
Number of removal steps Fig. 8. Total throughput versus target SINR under different fairness criteria
(for Rmin
i = 256 Kbps, N = 5, I = 5N0 ).
Fig. 6. Probability distribution for the number of removal steps under
different removal algorithms (for N = 15, I = 5N0 , SINR = 15 dB).

higher throughput than that with max-min fairness criterion.

between outage performance and the speed of removal, the In addition, when the minimum processing gain increases, the
probability distribution of the number of removal steps is differences among total throughput under different fairness
shown in Fig. 6 for different removal algorithms. It is evident criteria and different sets of QoS and interference constraints
that the number of removal steps tends to decrease rapidly become smaller. This is because, when the minimum pro-
with R (i.e., the number of removed links in each removal cessing gain increases, the maximum rate decreases which
step) for the I-SMART(R) algorithm. Therefore, while it may essentially reduces the feasible region. Therefore, the gaps
be desirable to adopt single removal algorithms in slow fading among different throughput curves become smaller.
environments to achieve good outage performance, multiple Throughput performance under max-min and PF crite-
removal algorithms may be employed in scenarios where the ria versus the target SINR values is illustrated in Fig. 8.
channel experiences fast fading. Again, it is observed that rate and power allocation under
PF criterion consistently achieves higher throughput than
When the network load is low and all requesting links
that under max-min fairness criterion. Recall that rate/power
can be admitted while satisfying both QoS and interference
constraints, we perform the joint rate and power allocation allocation under max-min fairness achieves perfectly fair rates
for all links. To show the fairness behavior for the PF
as presented in Section IV. In Fig. 7, we show the total
case, we plot its fairness index versus minimum processing
throughput versus the minimum processing gain under dif-
ferent sets of constraints for both max-min and proportional gain in !Fig. 9. Here,
"2 !thefairness "index is calculated as
N N
fairness criteria. As expected, the more stringent the QoS and FI = i=1 Ri / N i=1 Ri2 [24] which becomes
the interference constraints are, the lower is the throughput closer to one as the rate allocation becomes fairer. Fig. 9
that can be achieved. This figure also shows that joint rate shows that rate/power allocation under PF criterion is not very
and power allocation with PF criterion achieves significantly bad in maintaining fairness for different links. Also, fairness
5314 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 7, NO. 12, DECEMBER 2008

0.98
DΩ (P Ω ) − DΩ0 (P Ω/Ω0 )
0.96  
= βi (P Ω ) − βi (P Ω/Ω0 ) (24)
0.94 i∈Ω i∈Ω0
 #
  
Fairness index

0.92
= βi (P Ω ) + βi (P Ω ) − βi (P Ω/Ω0 ) (25)
0.9 i∈Ω
/ 0 i∈Ω0 i∈Ω0
⎧ ⎡ ⎤
0.88  ⎨  (s)
gi,i B Ω
= Ω
βi (P ) + ⎣ (s) Ω
gi,j Pj + Ni − P ⎦
0.86 ⎩
i∈Ω0 j∈Ω,j=i
γi Ri i
i∈Ω
/ 0
⎡ ⎤⎫
0.84   g
(s)
B ⎬
⎣ i,i
PiΩ ⎦
(s)
PF, SINR = 10dB, I = 20N
0 − gi,j PjΩ + Ni − (26)
0.82
PF, SINR = 15dB, I = 5N γi Ri ⎭
0 i∈Ω0 j∈Ω0 ,j=i
0.8 ⎡ ⎤
10 15 20 25 30 35 40    
Minimum processing gain = βi (P Ω ) + ⎣ (s)
gi,j PjΩ − gi,j PjΩ ⎦
(s)

i∈Ω
/ 0 i∈Ω0 j∈Ω,j=i j∈Ω0 ,j=i
Fig. 9. Fairness index under proportional fairness (for N = 5). (27)
 Ω
  (s)
= βi (P ) + gi,j PjΩ (28)
i∈Ω0 j ∈Ω
improves when the minimum processing gain increases (i.e., i∈Ω
/ 0
⎡
/ 0 ,j=i

smaller feasible rate region) and/or QoS and interference   

βi (P ) + ⎣
Ω (s)
constraints are less stringent. Results from Fig. 7-9, therefore, = gi,j PjΩ
i∈Ω
/ 0 i∈Ω j ∈Ω
/ 0 ,j=i
suggest that the PF criterion may be chosen for rate/power ⎤
allocation to achieve better throughput if some degree of  
gi,j PjΩ ⎦
(s)
unfairness can be tolerated by different secondary links. − (29)
i∈Ω
/ 0 j ∈Ω
/ 0 ,j=i
$  
VI. C ONCLUSIONS  (s)
gi,i B Ω
= βi (P Ω ) + αi (P Ω ) + Pi − Ni
We have presented a framework to solve the spectrum γi Ri
i∈Ω
/ 0
sharing problem in cognitive wireless networks with spectrum ⎡ ⎤
underlay. In particular, admission control algorithms for spec-  
−⎣ gi,j Pj ⎦ . (30)
(s) Ω
trum underlay in CDMA networks have been proposed which i∈Ω
/ 0 j ∈Ω
/ 0 ,j=i
aim to remove the least number of secondary links such that
both the QoS constraints in terms of target SINR for accepted Here, equation (24) results immediately from the definition in
links and the interference constraints for primary links are (12). Equation (25) is obtained from equation (24) simply by
satisfied. Then, we have formulated the joint rate and power splitting the first term in (24) into the first two terms in (25).
allocation problems for the secondary links as optimization Equation (26) is again due to the definition of βi (P Ω ) in (11).
problems with both QoS and interference constraints under Finally, equation (30) is obtained by applying the definition
low network load conditions. We have shown how to transform of αi (P Ω ) to the second term in equation (29). Thus, the
these optimization problems into geometric programs in the proposition is proved.
convex form so that the globally optimal solutions can be
obtained. Numerical results have revealed the impacts of A PPENDIX C
different system, QoS and interference constraint parameters P ROOF OF P ROPOSITION 3
on the network performance. The proof can be done using the procedure similar to
that for proposition 3 in [19] although the objective func-
A PPENDIX A
tion and the set of constraints of the optimization problem
P ROOF OF P ROPOSITION 1
here is different from that in [19]. Specifically, we will
It was shown in [17] that the admission control with prove the proposition by contradiction using the objective
only QoS constraint as presented in Section III.A is NP- function min {maxi 1/Ri }. Now, let P ∗ and R∗ be the
hard. The admission control with both QoS and interference optimal solutions of the optimization problem. Then, we
constraints includes that with only QoS constraint as a special should have maxi 1/Ri∗ ≤ maxi 1/Ri for any feasible R.
case. Specifically, the admission control with both QoS and Suppose that maxi 1/Ri∗ > mini 1/Ri∗ (i.e., optimal rates are
interference constraints degenerates into that with only QoS not equal). Assume that link m achieves the minimum, i.e.,
constraint in [17] if the interference limits in (3) becomes m = argmini 1/Ri∗ . We will rewrite the SINR constraints in
infinity (i.e., Ij → ∞ for j = 1, 2, · · · , M ). Therefore, it is (21) as follows:
also NP-hard.
(s)
gi,i Pi Ri γi
A PPENDIX B N (s)
≥ , i = 1, 2, · · · , N. (31)
P ROOF OF P ROPOSITION 2 j=1,j=i gi,j Pj + Ni B

The removal metric of the SMART(R) algorithm can be We observe that in the left hand side of (31) the numerator
written as is a strictly increasing function of Pi and the denominator is a
LE and HOSSAIN: RESOURCE ALLOCATION FOR SPECTRUM UNDERLAY IN COGNITIVE RADIO NETWORKS 5315

strictly decreasing function of Pj , j = i. Therefore, there must [11] H. Zheng and C. Peng, “Collaborative and fairness in opportunistic
be some power decrement ΔP and rate decrement ΔR for link spectrum access,” in Proc. IEEE ICC’05.
m and a small rate increment ΔR1 for all other links such [12] N. Nie and C. Comaniciu, “Adaptive channel allocation spectrum
etiquette for cognitive radio networks,” in Proc. IEEE DySPAN’05.
that the constraints in (31) still hold. Also, with this small [13] A. T. Hoang and Y. -C. Liang, “A two-phase channel and power allo-
power decrement for link m, the interference constraints in cation scheme for cognitive radio networks,” in Proc. IEEE PIMRC’06.
(21) continue to hold. Specifically, the following constraints [14] Y. Xing, C. N. Mathur, M. A. Haleem, R. Chandramouli, and K. P. Sub-
balakshmi, “Dynamic spectrum access with QoS and interference temper-
hold ature constraints,” IEEE Trans. Mobile Comp., vol. 6, no. 4, pp. 423-433,
(s)∗ ∗ Apr. 2007.
gm,m (Pm − ΔP ) (Rm − ΔR)γm [15] F. P. Kelly, A. Maulloo, and D. Tan, “Rate control for communication
N (s)
≥
∗ networks: Shadowing prices, proportional fairness, and stability,” J. Oper.
j=1,j=m gm,j Pj + Nm
B
Res. Soc. , vol. 49, no. 3, pp. 237-252, Mar. 1998.
(s)
gi,i Pi∗ [16] J. Mo and J. Walrand, “Fair end-to-end window-based congestion
N control,” IEEE/ACM Trans. Networking, vol. 8, no. 5, pp. 556-567, Oct.
(s) (s)
j=1,j=i,j=m gi,j Pj∗ + gi,m (Pm
∗ − ΔP ) + N
i 2000.
[17] M. Andersin, Z. Rosberg, and J. Zander, “Gradual removals in cellular
(Ri∗ + ΔR1 )γi PCS with constrained power control and noise,” ACM/Baltzer Wireless
≥ , i = m Networks J., vol. 2, no. 1, pp. 27-43, 1996.
B
N [18] S. A. Grandhi and J. Zander, “Constrained power control,” Wireless
 (p) (p) Personal Commun., vol. 1, no. 4, 1995.
gj,i Pi∗ + gj,m (Pm
∗
− ΔP ) ≤ Ij , j = 1, 2, · · · , M. [19] A. Muqattash, M. Krunz, and T. Shu, “Performance enhancement of
i=1,i=m adaptive orthogonal modulation in wireless CDMA systems, ” IEEE J.
Select. Areas Commun., vol. 24, no. 3, pp. 565-578, Mar. 2006.
Suppose we choose the rate and power variation small [20] D. I. Kim, E. Hossain, and V. K. Bhargava, “Downlink joint rate and
enough such that power allocation in cellular multirate WCDMA systems,” IEEE Trans.
Wireless Commun., vol. 2, no. 1, pp. 69-80, Jan. 2003.
∗ [21] S. A. Jafar and A. Goldsmith, “Adaptive multirate CDMA for uplink
1/(Rm − ΔR) ≤ max 1/(Ri∗ + ΔR1 ). throughput maximization,” IEEE Trans. Wireless Commun., vol. 2, no. 2,
i=m
pp. 218-228, Mar. 2003.
Under such small rate and power variations, we have [22] A. J. Weiss and B. Friedlander, “Channel estimation for DS-CDMA
downlink with aperiodic spreading codes,” IEEE Trans. Commun., vol.
1 1 1 47, no. 10, pp. 1561-1569, Oct. 1999.
max = max > max [23] G. Chen, X. -H. Yu, and J. Wang, “Adaptive channel estimation and
i Ri∗ i=m Ri∗ i=m Ri∗ + ΔR dedicated pilot power adjustment based on the fading-rate measurement
1 1 for a pilot-aided CDMA system,” IEEE J. Select. Areas Commun., vol.
= max max , ∗ . 19, no. 1, Jan. 2001.
i=m Ri∗ + ΔR Rm − ΔR [24] R. Jain, A. Durresi, and G. Babic, “Throughput fairness index: An
explanation,” ATM Forum Document Number: ATM Forum/990045, Feb.
This contradicts with the condition maxi 1/Ri∗ ≤ 1999.
maxi 1/Ri for any feasible R. Therefore, we must have [25] S. Boyd and L. Vandenberge, Convex Optimization. Cambridge Univer-
maxi 1/Ri∗ = mini 1/Ri∗ . sity Press, 2004.

Long Bao Le (S’04-M’07) received the B.Eng.

R EFERENCES degree with highest distinction from Ho Chi Minh
City University of Technology, Vietnam, in 1999, the
[1] FCC, Spectrum policy task force report, FCC 02-155, Nov. 2002. M.Eng. degree from Asian Institute of Technology
[2] FCC, Facilitating opportunities for flexible, efficient, and reliable spec- (AIT), Thailand, in 2002 and the Ph.D. degree
trum use employing cognitive radio technologies, notice of proposed rule from University of Manitoba, Canada, in 2007. He
making and order, FCC 03-322, Dec. 2003. is currently a Postdoctoral Research Associate at
[3] Q. Zhao and B. M. Sadler, “A survey of dynamic spectrum access: Signal Massachusetts Institute of Technology, USA. His
processing, networking, and regulatory policy,” IEEE Signal Processing current research interests include cognitive radio,
Mag., pp. 79-89, May 2007. link and transport layer protocol issues, cooperative
[4] L. B. Le, E. Hossain, and A. S. Alfa, “Delay statistics and throughput diversity and relay networks, stochastic control and
performance for multi-rate wireless networks under multiuser diversity,” cross-layer design for communication networks.
IEEE Trans. Wireless Commun., vol. 5, no. 11, pp. 3234-3243, Nov. 2006.
[5] L. Le and E. Hossain, “Multihop cellular networks: Potential gains, re- Ekram Hossain (S’98-M’01-SM’06) is currently
search challenges, and a resource allocation framework,” IEEE Commun. an Associate Professor in the Department of Elec-
Mag., vol. 45, no. 9, pp. 66-73, Sept. 2007. trical and Computer Engineering at University of
[6] L. Le and E. Hossain, “Cross-layer optimization frameworks for multihop Manitoba, Winnipeg, Canada. Dr. Hossain’s current
wireless networks using cooperative diversity,” IEEE Trans. Wireless research interests include design, analysis, and op-
Commun., vol. 7, no. 7, pp. 2592-2602, July 2008. timization of wireless communication networks and
[7] M. Oner and F. Jondral, “On the extraction of the channel allocation in- cognitive radio systems. He is a co-author/editor/co-
formation in spectrum pooling systems,” IEEE J. Select. Areas Commun., editor for the books Dynamic Spectrum Access and
vol. 25, no. 3, pp. 558-565, Apr. 2007. Management in Cognitive Radio Networks (Cam-
[8] T. Banerjee, C. Ghosh, and D. P. Agrawal, “Wireless sensor based bridge University Press, 2009), Heterogeneous Wire-
dynamic channel selection in cellular communication by cognitive radio less Access Networks (Springer, 2008), Introduction
approach,” in Proc. 1st International Conference on Cognitive Radio to Network Simulator NS2 (Springer, 2008), Cognitive Wireless Communica-
Oriented Wireless Networks and Communications, June 2006. tion Networks (Springer, 2007), and Wireless Mesh Networks: Architectures
[9] C. T. Chou, S. Shankar. N, H. Kim, and K. G. Shin, “What and how and Protocols (Springer, 2007). Dr. Hossain serves as an Editor for the IEEE
much to gain by spectrum agile?” IEEE J. Select. Areas Commun., vol. T RANSACTIONS ON M OBILE C OMPUTING, the IEEE T RANSACTIONS ON
25, no. 3, pp. 576-588, Apr. 2007. W IRELESS C OMMUNICATIONS, the IEEE T RANSACTIONS ON V EHICULAR
[10] A. R. Fattahi, F. Fu, M. v. d. Schaar, and F. Paganini, “Mechanism- T ECHNOLOGY, IEEE W IRELESS C OMMUNICATIONS, and several other
based resource allocation for multimedia transmission over spectrum agile international journals. He is a registered Professional Engineer in the Province
wireless networks,” IEEE J. Select. Areas Commun., vol. 25, no. 3, pp. of Manitoba, Canada.
601-612, Apr. 2007.