This article has been accepted for inclusion in a future issue of this journal.
Content is final as presented, with the exception of pagination.
IEEE/ACM TRANSACTIONS ON NETWORKING
Multicast in Multihop CRNs Under Uncertain
Spectrum Availability: A Network
Coding Approach
Yuben Qu, Chao Dong, Member, IEEE, Haipeng Dai, Fan Wu, Member, IEEE, Shaojie Tang, Member, IEEE,
Hai Wang, Senior Member, IEEE, and Chang Tian, Member, IEEE
Abstract— The benefits of network coding on multicast in
traditional multihop wireless networks have already been
extensively demonstrated in previous works. However, most
existing approaches cannot be directly applied to multihop
cognitive radio networks (CRNs), given the unpredictable
primary user occupancy on licensed channels. Specifically, due to
the unpredictable occupancy, the channel’s available bandwidth
is time-varying and uncertain. Accordingly, the capacity of the
link using that channel is also uncertain, which can significantly
affect the network coding subgraph optimization and may
result in severe throughput loss if not properly handled. In this
paper, we study the problem of network coding-based multicast
in multihop CRNs while considering the uncertain spectrum
availability. To capture the uncertainty of spectrum availability,
we first formulate our problem as a chance-constrained program.
Given the computational intractability of the above-mentioned
program, we then transform the original problem into a
tractable convex optimization problem, through appropriate
Bernstein approximation with relaxation on link scheduling.
We further leverage Lagrangian relaxation-based optimization
techniques to propose an efficient distributed algorithm for the
original problem. Extensive simulation results show that the
proposed algorithm achieves higher multicast rates, compared
with a state-of-the-art non-network coding algorithm in multihop
CRNs, and a conservative robust network coding algorithm that
treats the link capacity as a constant value in the optimization.
Manuscript received June 25, 2015; revised March 11, 2016,
October 22, 2016, and February 4, 2017; accepted February 7, 2017; approved
by IEEE/ACM T RANSACTIONS ON N ETWORKING Editor D. Goeckel. This
work was supported in part by the State Key Development Program for
Basic Research of China (973 Project) under Grant 2014CB340303, in part
by the National NSF of China under Grant 61103224, Grant 61371124,
Grant 61472445, Grant 61422208, and Grant 61272443, in part by the NSF
of Jiangsu Province under Grant BK2011118 and Grant BK20140076, in
part by the CCF-Intel Young Faculty Researcher Program, and in part by the
CCF-Tencent Open Fund. (Corresponding author: Chao Dong.)
Y. Qu is with the College of Communications Engineering, PLA Uni-
versity of Science and Technology, Nanjing 210007, China, and also with
the Rocket Force University of Engineering, Xi’an 910025, China (e-mail:
quyuben@gmail.com).
C. Dong, H. Wang, and C. Tian are with the College of Communications
Engineering, PLA University of Science and Technology, Nanjing
210007, China (e-mail: dch999@gmail.com; haiwang@ieee.org; tianchang
126@126.com).
H. Dai is with the Department of Computer Science and Technology, Nan-
jing University, Nanjing 210023, China (e-mail: dhpphd2003@gmail.com).
F. Wu is with the Department of Computer Science and Engineer-
ing, Shanghai Jiao Tong University, Shanghai 200240, China (e-mail:
fwu@cs.sjtu.edu.cn).
S. Tang is with the Naveen Jindal School of Management, University
of Texas at Dallas, Richardson, TX 75080 USA (e-mail: tangshaojie@
gmail.com).
This paper has supplementary downloadable material available at
http://ieeexplore.ieee.org, provided by the authors. This consists of detailed
proofs of Theorems 1 and 2.
Digital Object Identifier 10.1109/TNET.2017.2666788
1063-6692 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1
Index Terms— Cognitive radio networks, multicast, network
coding, chance-constrained optimization, convex optimization.
I. I NTRODUCTION
D UE to the pressing demand of efficient frequency
spectrum usage, cognitive radio networks (CRNs) [1]
have spurred a wide range of research interests and emerged
as a promising technology to improve the spectrum utilization.
In CRNs, with the proliferation of computationally powerful
wireless devices and compelling demand on diversified ser-
vices, providing multicast services for secondary users (SUs),
especially in a multi-hop fashion, is urgently needed [2]–[5].
Therefore, it is crucial to improve the performance of multicast
in multi-hop CRNs.
During the past decade, network coding [6] has been
proposed to improve the network resource utilization. By
combining multiple input packets into one packet algebraically
before forwarding, network coding is able to increase the
multicast capacity [6]. The throughput benefits of network
coding on multicast in multi-hop wireless networks have been
well recognized in previous works, e.g., [7]–[14]. However,
almost all these works focus on traditional multi-hop wireless
networks rather than multi-hop CRNs. Then, a question nat-
urally arises: whether network coding can still enhance the
multicast throughput in multi-hop CRNs or not?
Indeed, most existing works on network coding-based multi-
cast cannot handle new challenges arising from CRNs, e.g., the
uncertainty of spectrum availability. Particularly, in the opti-
mization of traditional wireless networks, the link capacity is
usually assumed as a fixed value [7]–[9], [13], [14]. However,
in CRNs, due to unpredictable primary occupancy on licensed
channels, the availability of the channel (i.e., on/off) follows
some probability distribution in the time domain, while the
unoccupied bandwidth of an available channel is uncertain
in the frequency domain, in the sense that it is a random
variable satisfying some certain distribution [15]–[18]. That
is to say, the capacity of a link using that channel is also
a random variable. The randomness of capacity invalidates
all of the previous approaches on the optimization of coding
subgraph, which is the core of network coding in multi-hop
networks [6]–[11]. Besides, simply replacing the uncertain
link capacity by its expected constant value is not preferable,
since the fruitful information of spectrum availability is not
exploited, which can incur great performance degradation.
In this paper, we study how to maximize the data rate of
a multicast session by intra-session network coding [9] in
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2 IEEE/ACM TRANSACTIONS ON NETWORKING
multi-hop CRNs. To the best of our knowledge, this is the first
work to study the problem of network coding-based multicast
in multi-hop CRNs, by taking into account the uncertainty
of spectrum availability. Different from existing works, we
model the link capacity as a random variable instead of a fixed
value, which captures the uncertainty of spectrum availability
in CRNs. To deal with the random variables in the constraints,
we cast these constraints as chance constraints [19]–[21],
i.e., candidate solutions are required to satisfy the random
constraints with a given probability.
However, chance constraints are usually more difficult to
tackle because these constraints are typically nonconvex and
sometimes difficult to be expressed in a closed form [19].
Another challenge in our problem is that since the available
channels at different nodes could be different in multi-hop
CRNs, how to efficiently schedule the links under the channel
availability is also a critical concern. Note that the scheduling
for SUs’ data transmission in multi-hop CRNs can be done
either in frequency domain or in time domain. In this work,
we focus on the scheduling in the frequency domain.
To overcome the above challenges, we first employ
Bernstein approximation techniques [19] to translate the
aforementioned chance constraints into convex computable
constraints. Although Bernstein approximation has been
applied in previous studies, the main novelty as well as
difficulty here is how to properly choose the approximation
bounds according to our problem, to ensure both feasibility
and tractability in the transformation. Further, we encapsulate
the constraints of hyperarc-channel scheduling in a two-
dimensional conflict graph (TDCG), and eventually obtain an
optimization problem that can be solved by common convex
optimization approaches. Specifically, we decompose the
overall problem into two relatively independent subproblems,
i.e., a multiple-shortest-paths problem with channel
availabilities constraints and a maximum-weighted-stable-set
problem on the TDCG, respectively. Then, we propose an
efficient distributed algorithm for network coding-based
multicast in multi-hop CRNs, to simultaneously optimize the
flow rate and coding subgraph with channel selection.
Our main contributions are summarized as follows:
1) We for the first time study the problem of network
coding-based multicast in multi-hop CRNs under uncer-
tain spectrum availability, and formulate the problem as
a chance-constrained program. Specifically, we capture
the uncertainty of link capacity in chance constraints,
and encapsulate both hyperarc scheduling and channel
selection into multiple hyperarc-channel tuples repre-
sented by the TDCG.
2) We transform the formulated chance-constrained pro-
gram into a computationally tractable convex problem,
through appropriate Bernstein approximation and relax-
ation of finding stable sets on the TDCG. We decom-
pose the convex problem into two separate subproblems
by using Lagrangian relaxation, and then propose an
efficient distributed algorithm for network coding-based
multicast in multi-hop CRNs.
3) Simulation results show that, the proposed decentralized
algorithm achieves higher multicast rate than the
state-of-the-art non-network coding algorithm in
multi-hop CRNs, which achieves the promising benefit
of network coding in such scenarios. Moreover, our
decentralized algorithm outperforms the conservative
robust network coding algorithm without regard for
uncertain spectrum availability, in terms of multicast
rate.
The rest of this paper is organized as follows. Section II
and Section III describe the network model and preliminaries,
and problem formulation, respectively. In Section IV, we use
Bernstein approximation to translate the chance constraints
into tractable convex constraints. An efficient distributed algo-
rithm is developed in Section V. VI extends our solution to the
multiple multicast flows case. We conduct performance evalu-
ation via extensive simulations in Section VII. In Section VIII,
we review the related work. Finally, Section IX concludes the
paper.
II. N ETWORK M ODEL AND P RELIMINARIES
In this section, we will first introduce the basic network
model used in our work, then interpret the modeling of
uncertain spectrum availability in CRNs, and lastly, explain
the coding subgraph optimization in multi-hop CRNs.
A. System Model
We consider a multi-hop CRN, where N secondary
users (SUs)1 share C orthogonal channels that can be accessed
by SUs when they are not occupied by primary users (PUs).
The set of nodes is denoted by N = {1, 2, . . . , N }, and the
set of channels is denoted by C = {1, 2, . . . , C}. Each node
is equipped with two radios: one is a half-duplex cognitive
radio (CR) that can switch among the channels for data
transmissions, and the other is a half-duplex normal radio
working on a common control channel (CCC) for exchanging
control messages [5]. In this work, we assume that the location
of all SU nodes is fixed and known.2 By exchanging the
geographical location information through CCC, the distance
between any two neighbor nodes can be calculated. With CR,
each node can identify the set of current available channels
through existing spectrum sensing techniques [22], [23].
Following most previous works [3], [5], we suppose that
the available channels at a SU may be different from those
at another one in the network, considering the geographical
location differences of the SUs. Let Ci ⊆ C represent the set
of available channels at SU i ∈ N , and CJ ⊆ C denote the
set of 
common available channels at a set of nodes J, i.e.,
CJ = j∈J Cj . Following [3], [5], [18], we assume that the
channel availability at SUs is assumed quasistatic, i.e., the
available channel set Ci varies with time but does not change
in a short period of time. In this work, we mainly consider
the scheduling in frequency domain other than time domain,
in terms of assigning frequency bands (channels). We note
that our assumption of an available channel c at SU i means
that SU i can transmit and receive data on channel c without
causing interference to the licensed PU of channel c.
1 SU and node will be used interchangeably in the rest of this paper.
2 When the locations is not fixed but randomly moving, we can pursue the
approaches like stochastic geometry and random graph to model the network
as in many previous works [44], [45].
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
QU et al.: MULTICAST IN MULTIHOP CRNs UNDER UNCERTAIN SPECTRUM AVAILABILITY
TABLE I
A L IST OF N OTATIONS
We leverage the broadcast nature for efficient multicast
in wireless multi-hop CRNs, i.e., a node can transmit data
simultaneously to multiple nodes in its transmission range
by switching to the same available channel [5]. To illustrate
the broadcast relationship, following [11], [14], we model the
network by a hypergraph H = (N , A), where N is the set of
nodes and A is the set of hyperarcs. A hyperarc is a pair (i, J)
(at least one common channel available in node i and set J),
where i ∈ N is the start node, and J ⊂ N is the set of end
nodes. Every hyperarc (i, J) represents a broadcast link from
node i to the nodes in the nonempty set J.
For flow service, we assume that there is a multicast
connection of rate R with a source s ∈ N and sinks
D = {1, 2, . . . , D} ⊂ N . In this study, we only consider the
scenario where all multicast sinks request the same informa-
tion, e.g., audio/video file downloads and software updates,
etc. This kind of application can be attributed to reliable
multicast. Since one of the key performance metrics of reliable
multicast is throughput, in this work, we will focus on how
to maximize the multicast rate under multiple constraints.
How to exploit network coding to enhance the performance
of multicast in CRNs in terms of more than rate is left for
future work. Thus, R can be defined as the total amount of
data that are successfully delivered to all sinks per unit time.
Table I provides a list for the major notations used in this
paper.
B. Modeling of Uncertain Spectrum Availability
The uncertain spectrum availability of CRNs distinguishes
it from traditional (multichannel) networks. That is, in CRNs,
the available channels may be known at SUs, while the
unoccupied bandwidth of an available channel is uncertain
in the frequency domain, due to the unpredictable bandwidth
occupancy of PUs [18]. To model this unique feature of CRNs,
following [18], let B c denote the unoccupied bandwidth of
an available channel c ∈ C, where B c is a random variable
satisfying some certain probability distribution. Notice that
based on [15]–[17], the availability of a channel follows an
3
exponential-like distribution that is not independent over time.
However, this will not change the optimization framework
in our problem, since we study the multicast problem with
link scheduling in the frequency domain, given the channel
availabilities at SUs. According to the well known Shannon-
Hartley theorem, the capacity of a link has strong positive
relationship with the bandwidth of the channel it uses. Since
the channel bandwidth is random in CRNs, the link capacity
is also a random variable, which may pose great challenges
to the optimization of network performance.
C. Coding Subgraph Optimization in Multi-Hop CRNs
In this study, we consider intra-session network coding [9]
to improve the multicast performance of SUs. The core of
intra-session network coding is, to allow network nodes mix
(code) received data packets before data transmission, where
the coding is restricted in the same session only. Further,
we employ the probabilistic version of network coding in
our work, i.e., random linear network coding approach [7].
Specifically, each coded packet sent is a linear combination
of the stored packets in a node with coefficients drawn
randomly from a sufficiently large finite field GF (2q ). And
any destination node can recover several data packets when
receiving enough coded packets.
However, how to optimize the packet injection rate at
intermediate nodes including the source, which is so called
the coding subgraph optimization, is the key issue in the
problem of network coding-based multicast. Let ziJ (c) denote
the average packet injection rate into hyperarc (i, J) through
channel c. The rate vector z = (ziJ (c))(i,J)∈A,c∈Ci Ì CJ is
termed as the coding subgraph [8]. Following [8], we denote
ziJK (c) the rate at which packets are received by precisely
the set of nodes K ⊆ J through hyperarc (i, J) on channel c.
Worth noting that K can be J, if
there are no transmission
errors. Hence, we have ziJ (c) = K⊆J ziJK (c).
For two different hyperarcs, they may utilize different
available channels due to interference constraints, and thus the
coding subgraph is different at different channels. The choice
of ziJ (c) generally depends on a number of factors, one of
which the aggregated link rate should be no more than the link
capacity. As stated previously, the link capacity is a random
variable in CRNs. This may change the pattern of the whole
optimization framework on coding subgraph, which will be
discussed in detail in next section.
D. Preliminaries on Bernstein Approximation
Bernstein approximation is originated from [43] and further
significantly improved in [19]. It aims to transform a proba-
bilistic (chance) constraint into a tractable convex constraint,
which guarantees that any feasible solution of the transformed
constraint is also feasible for the original constraint. Consider
the following chance constraint:
 
 k
Pr f0 (x) + ξi fi (x) ≥ 1 − ε,
i=1
where x is a decision parameter vector, ε ∈ (0, 1) is a
given tolerance, and {ξi }i=1,...,k are independent variables
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4 IEEE/ACM TRANSACTIONS ON NETWORKING
with known probability distribution {Pi }i=1,...,k . Bernstein
approximation can be applied if the two following assumptions
hold:
1) {fi (x)} are affine in x for i = 1, . . . , k;
2) The distributions from Pi are commonly supported on
[−1, 1], i.e., for all i = 1, . . . , k, −1 ≤ ξi ≤ 1.
Note that the final transformed constraint is determined by
both the probability distribution information of {ξi }i=1,...,k
and the form of {fi (x)}.
III. P ROBLEM F ORMULATION
In this section, we formulate the problem of network coding-
based multicast in multi-hop CRNs. The objective is to find
the flow rates and coding subgraphs with channel selection,
with the consideration of uncertain spectrum availability,
such that the multicast rate R is maximized. In general, we
should consider the following constraints: flow rate constraints,
coding subgraph constraints, and link capacity constraints.
Note that with network coding, the routing issue has been
taken into account by the coding subgraph optimization as
well as the flow balance and link capacity. We now specify
these constraints separately in detail.
A. Flow Rate Constraints
A flow rate constraint requires that the amount of incoming
traffic rate equals to the amount of outgoing traffic at every
node in the network. We denote by fiJj d
(c) the data rate
for sink d from node i to node j ∈ J, when packets are
actually transmitted on virtual wireless broadcast link (i, J),
using channel c. To explicitly demonstrate the dependency of
a broadcast link on a specific channel, we define
 One may just follow the similar approach in [8] and [11]
Tic = {j|dij ≤ rT , j = i, c ∈ Ci Cj }, (1)
where dij is the distance between node i and j, and rT is the
transmission range of node i. Tic is thus the set of nodes that
node i can simultaneously transmit data to through channel
c. Note that the difference between Tic and J in a hyperarc
(i, J) is that the former clearly specifies the common available
channel, while the latter does not.
According to the max-flow min-cut theorem in [6], there
exists a flow vector fiJjd
(c) for each sink d, which satisfies
⎧ of simultaneously active links, which results in throughput
   ⎪
⎨R i=s

fiJj (c) −
d
fjIi (c ) = −R i ∈ D
d In this work, inspired by [24], we develop a two-

⎪
⎩ dimensional conflict graph (TDCG) to describe the conflicts
J⊆Tic j∈J {j|I⊆Tjc ,i∈I} 0 otherwise,
(2)
for all d ∈ D. The first and second parts in the left of (2)
correspond to the total outgoing rate and total incoming rate,
respectively. If node i is the source node, its outgoing rate is
R and incoming rate is 0. If i is one of the sinks, its outgoing
rate is 0 and incoming rate is R. Otherwise, node i is an
intermediate node, and its outgoing rate should be equal to its
incoming rate.
B. Coding Subgraph Constraints
Similar to [7]–[9], and [11], the constraints of coding
subgraph include two types. One is that the link rate for any
sink cannot exceed the corresponding injection rate restricted
by the coding subgraph. The other is about the scheduling of
hyperarcs, i.e., two hyperarcs cannot simultaneously transmit
if they conflict.
1) Link Capacity Constraint on the Coding Subgraph:
We define biJK (c) as the ratio of the sum of any successful
reception rate that relates to subset K ⊆ J, to the injection
rate on hyperarc (i, J) through channel c, i.e.,

Ì
{S|S⊂J,S K=∅}
ziJS (c)
biJK (c) = . (3)
ziJ (c)
If channel c is lossless, we have biJK (c) = 1 for all nonempty
K ⊆ J and biJ∅ (c) = 0. Then, for the link rate requirement on
the coding subgraph [8], [11], i.e., the flow rate for any sink
cannot exceed the rate determined by both coding subgraph
and channel quality, we have

d
fiJj (c) ≤ ziJ (c)biJK (c). (4)
j∈K
2) Hyperarc Scheduling Constraint on the Coding
Subgraph: Next, we deal with more complex scheduling
constraints on hyperarcs. According to [8], z should be
restricted with a convex subset. The core is to decide
which sets of hyperarcs can transmit simultaneously without
conflict. However, this depends on the interference model
of the network. In this study, we consider a more general
secondary interference model [11]. That is, since there is
only one radio for data transmission at each node, at a certain
time, each node is constrained to receive from at most one
other sending node, and can only successfully receive if all
other neighbors on the same channel are silent.
using a conflict graph on the hypergraph only to capture the
conflict of hyperarcs. This kind of hyperarc conflict graph,
however, cannot be directly applied mainly because the hyper-
arc conflict graph is insufficient in describing the extra conflict
in competing for channels. Take a simple example. According
to the hyperarc conflict graph, hyperarc (i1 , J1 ) interferes
with hyperarc (i2 , J2 ). Actually, the two hyperarcs may not
interfere, if they employ different channels. Using such a
hyperarc conflict graph will definitely decrease the number
degradation in the end.
over hyperarcs and channels jointly. Each vertex in the TDCG
represents a hyperarc-channel tuple, a basic resource point in
the multi-hop CRN. The TDCG is thus a graph that illustrates
the conflict relationships among the resource points in the
multi-hop CRN. A tuple w is defined in the format as follows:
hyperarc-channel tuple: [(i, J), c],

where (i, J)∈ A and c ∈ Ci CJ . The tuple [(i, J), c] indicates
that the hyperarc (i, J) operates on channel c. Consider two
tuples w1 = [(i1 , J1 ), c1 ] and w2 = [(i2 , J2 ), c2 ]. We say that
w1 and w2 do not conflict if the following conditions hold:

i1 = i2 ; 2) i1 ∈ J2 ; 3) i2 ∈ J1 ; 4) J1 J2 = ∅; 5)
1) 
J1 Tic22 = ∅ and J2 Tic11 = ∅ when c1 = c2 .
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
QU et al.: MULTICAST IN MULTIHOP CRNs UNDER UNCERTAIN SPECTRUM AVAILABILITY
Fig. 1. Example of a hypergraph.
Fig. 2. TDCG corresponding to the hypergraph in Fig. 1.
Definition 1: The TDCG of a hypergraph H with the channel
availability (Ci )i∈N is an undirected graph G = (V, E), with
V corresponding to the set of all hyperarc-channel tuples, and
an edge (w1 , w2 ) belongs to the edge set E if w1 and w2
interfere, for all w1 , w2 ∈ V.
Note that the above TDCG well captures different kinds of
interference in multi-hop CRNs, including the self-interference
at the physical layer, the interference on the same channel
from transmission and reception, and the interference in the
transmission range of a node, etc.
Given the TDCG, a stable (independent) set I is a set of
5
The above constraint not only captures the classical coding
subgraph requirements including the scheduling of hyperarcs,
but also restricts the channel selection of different hyperarcs.
Although link scheduling in CRNs has been studied in pre-
vious works, few study how to employ network coding to
improve the performance of multicast in multi-hop CRNs, by
scheduling the links with channel selection and determining
the packet forwarding rate. Furthermore, they seldom consider
the uncertainty of spectrum availability in CRNs, which will
be introduced in the next subsection.
C. Link Capacity Constraints
Since we consider wireless broadcasting nature, we should
take the capacity of a broadcast link rather than a point-
to-point link into consideration. We define the capacity of
a broadcast link as the minimum of the capacities of its
component point-to-point links [14].
Because the link capacity is significantly affected by the
channel selection, we define uij (c) as the 0-1 indicator repre-
senting whether node i and j operate on channel c

1 if i transmits data to j on channel c
uij (c) = (7)
0 otherwise,

where c ∈ Ci Cj . We note that the value of uij (c) is
determined by the aforementioned coding subgraph constraint.
Based on the channel selection, following [18], we can define
the capacity of a link (i, j) on channel c, i.e., hij (c), which is
given by
vertices belonging to V, any two of which are nonadjacent.
I includes the tuples that can be active simultaneously in the
network. If adding any one more vertex into a stable set I
results in a non-stable set, I is a maximal stable set (MLSS).
A maximum stable set (MMSS) is one that is not contained
in any other stable set. Similar to [11], we define πIv as the
incidence vector of I, which is given by

1 if node v ∈ I,
πI =
v
(5)
0 otherwise.
Then, the stable set polytope of G, i.e., PSTAB (G), is thus the
convex hull of the incidence vectors of all MMSSs of G.
We use a toy example to illustrate the notation. Consider the
hypergraph in Fig. 1 (the symbol next to each node denotes

gij Q
hij (c) = uij (c)B log2 1 +
c
, (8)
η
where gij is the power propagation gain between node i and
node j, Q is the power spectral density from the transmitter,3
and η is the ambient Gaussian noise density. Note that we
further consider channel selection in the link capacity based
upon [18]. And this definition takes into account many prac-
tical issues in the physical layer, including channel fading,
interference, transmission range, channel selection and one of
the most typical features of CRNs, the uncertainty of channel
available bandwidth. Accordingly, the capacity of a broadcast
link (i, J) on channel c can be given by
hiJ (c)
the current available channels), and its corresponding TDCG  

in Fig. 2. Consider the channel availability, as shown in Fig. 2, gij Q
= min hij (c) = min uij (c)B log2 1 +
c
. (9)
V = {[(1, 2), 1], [(1, 3), 1], [(1, {2, 3}), 1], [(2, 5), 1], [(3, 5), 1], j∈J j∈J η
[(3, 4), 2], [(3, 5), 2], [(3, {4, 5}), 2]}, which describes all 
By defining uiJ (c)  j∈J uij (c), we further have
resource points in Fig. 1. From Fig. 2, for example, we


easily obtain that tuples [(1, 2), 1] and [(3, 4), 2] can be gij Q
scheduled at a slot, since the two corresponding vertices hiJ (c) = uiJ (c)B min log2 1 +
c
, J ⊆ Tic . (10)
j∈J η
are not adjacent. In this example, the stable set polytope is
thus the convex hull of the incidence vectors of the three When uiJ (c) = 0, we have hiJ (c) = 0. It should be noted
stable sets {[(1, 2), 1], [(3, 4), 2]}, {[(1, 2), 1], [(3, 5), 2]} and that (10) can handle the possible interference incurred by
{[(1, 2), 1], [(3, {4, 5}), 2]}. multiple parallel transmissions, since uiJ (c) in (10) has indi-
Finally, the constraints of scheduling on the coding subgraph cated the link-channel selection, which is determined by the
can be obtained as follows aforementioned TDCG. In other words, if we set uiJ (c) as 1,

z = (ziJ (c))(i,J)∈A,c∈C ∈ PSTAB (G). (6) 3 We here assume that all nodes use the same power for transmission.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE/ACM TRANSACTIONS ON NETWORKING

it means that there exists no interference among the broadcast to verify. Specifically, the feasible set of (12) can be either
link (i, J). convex or non-convex, which depends on the distribution
Intuitively, one may easily have the following constraints of hiJ (c) [19]. Typically, if [aT bT ]T has a symmetric
for the flow rate requirement, i.e., the aggregate data rates on logarithmically concave density, Pr{aT x < b} ≥ 1 −  is
each broadcast link (i, J) cannot exceed the link’s capacity on convex for  < 1/2 [19]. However, it is still unclear how
channel c [11], [14] to compute the closed form of the probability even if it is
 convex [19], making the problem intractable.
d
fiJj (c) ≤ hiJ (c), c ∈ Ci , , K ⊆ J ⊆ Tic . (11)
d∈D j∈K
IV. B ERNSTEIN A PPROXIMATION TO THE
However, the link capacity hiJ (c) in (11) is actually a C HANCE -C ONSTRAINED P ROGRAM
random variable and thus uncertain — not knowing what the
exact value is. The standard way to treat such uncertainty is In this section, we propose a conservative convex approx-
to pass from the uncertain constraint to its chance-constrained imation approach. Bernstein approximations are known as
version [19]–[21] an useful class of approximation techniques for chance con-
⎧ ⎫ straints [19]–[21]. However, it should be careful to choose the
⎨  ⎬ appropriate approximation form according to the characteristic
Pr d
fiJj (c) − hiJ (c) ≤ 0 ≥ 1 − εiJ (c), (12) of the link capacity constraints. And in the transformation,
⎩ ⎭
d∈D j∈K both the feasibility and tractability should be ensured. Specif-
where εiJ (c)∈ (0, 1)
is thecorresponding violation tolerance. ically, we should replace the chance constraints with a series
Define ε  i∈N c∈Ci J⊆T c εiJ (c). By setting ε 1, of constraints Υ, such that (i) any solution satisfies (12) when
i
we can guarantee that for all c ∈ Ci , J ⊆ Tic , K ⊆ J, it is feasible for Υ (safe approximation), and (ii) the con-
constraints like (11) hold with a high probability 1-ε.4 The straints in Υ are convex and efficiently computable (tractable
chance constraint based approach to modeling data uncertainty approximation). In the rest of this section, we describe the
in optimization theory was initially proposed in [33] and approximation approach in two steps, presented by Theorem 1
followed by a huge literature (please refer to [21] for the and Theorem 2, respectively. 
references therein). For simplicity, we define fiJj (c)  d∈D fiJj d
(c). Then,
(12) turns to be
⎧ ⎫
D. Network Coding-Based Multicast in Multi-Hop CRNs ⎨ ⎬
Putting the three types of constraints together, we have the Pr fiJj (c) − hiJ (c) ≤ 0 ≥ 1 − εiJ (c). (13)
⎩ ⎭
following chance-constrained program: j∈K

maximize R Enlightened by the Bernstein approximation [19], we con-

struct the following function
subject to (2), (4), (6), (12). ⎧ ⎫
⎨ ⎬
In the program, fiJjd
(c) ≥ 0 and ziJ (c) ≥ 0 are optimization Φ(f)  inf fiJj (c) + ρΩ(−ρ−1 ) − ρ log(εiJ (c)) ,
variables, corresponding to flow rate and coding subgraph, ρ>0 ⎩ ⎭
j∈K
respectively. Worth noting that, spectrum sensing and
(14)
dynamic spectrum access have been jointly incorporated with
network coding and routing in the program. Specifically, where f = (fiJj (c))j∈K ≥ 0 and Ω(y)  log E[exp(yhiJ (c))].
in the program, Ci , the set of available channels at node i, Here E[·] is excuted on the distribution information of hiJ (c).
is identified in advance through existing spectrum sensing For a given εiJ (c) > 0, consider the following inequality
techniques [22], [23]. uij (c) is the 0-1 indicator representing
whether node i and j operate on channel c, which contains Φ(f) ≤ 0. (15)
the dynamic spectrum access. The coding subgraph variable
We can prove that any solution f that is feasible for (15) is also
ziJ (c) and flow rate variable fiJjd
(c) jointly determine the
feasible for the chance constraint (13) through the following
routing issue in the network. The optimization of fiJjd
(c) and theorem.
ziJ (c) (which also determines uij (c)) is based on the channel Theorem 1: Define Φ(f) as in (14) and hiJ (c) is a ran-
availabilities {Ci }i∈N and other information. dom variable following a specific probability distribution.
The above optimization problem is intractable, owing to For a given εiJ (c) > 0, suppose that there exists f =
two main reasons. One is that the constraint in (6) requires (fiJj (c))j∈K ≥ 0, such that (15) holds. Then, f satisfies (13).
to find all MMSSs in a given TDCG. However, searching Proof:  Inspired by the Bernstein approximation [19], let
all MMSSs is NP-hard [11], [34], due to the combinatorial F (f, h) = j∈K fiJj (c) − hiJ (c), then (15) is equivalent to
difficulty encapsulated in PSTAB (G). The other, which is more
 
challenging, is the probabilistic constraint in (12), rendering inf ρE[exp(ρ−1 F (f, h))] − ρεiJ (c) ≤ 0.
ρ>0
the convexity of the feasible set defined by (12) difficult
4 Note
If Pr{F (f, h) > 0} ≤ E[exp(ρ−1 F (f, h))] holds and
that there exists a small risk that no solution is feasible for (12). If
this happens, we can enlarge the parameter ε and corresponding εiJ (c) to E[exp(ρ−1 F (f, h))] ≤ εiJ (c), then we can have (13). In fact,
obtain a feasible solution. the former equality can be obtained by the property of exp(·)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

QU et al.: MULTICAST IN MULTIHOP CRNs UNDER UNCERTAIN SPECTRUM AVAILABILITY 7

and the latter one can be derived from (15). The detailed proof techniques. However, the new problem is still NP-hard in
is provided in Appendix A. general, owing to the aforementioned combinatorial difficulty
In light of Theorem 1, we can use a series of constraints in (6). Fortunately, enlightened by [11], after relaxing (6) of
 finding all MLSSs rather than MMSSs in the TDCG, we can
Υ : inf { d
fiJj (c) + ρ log E[exp(−ρ−1 hiJ (c))]
ρ>0 employ the dual decomposition method to obtain an efficient
d∈D j∈K
and decentralized solution.d
−ρ log(εiJ (c))} ≤ 0, i ∈ N , J ⊆ Tic , c ∈ Ci (16)
Let fijd (c)  J⊆T c fiJj (c), where ∀i ∈ N and j ∈ Tic ,
i
to replace (12) in the chance-constrained program, so that an with the clear understanding that fiJj d
(c) = 0 if j ∈ Tic .
optimal solution to the new problem is a feasible suboptimal Therefore, fijd (c) represents the flow rate
 of innovative packets
solution to the original problem. However, the closed form of on link (i, j) using channel c (c ∈ Ci Cj ) for sink d.
Υ is still hard to compute, owing to the uncertain expectation The optimization problem is converted into
value and the inf(·) function.
Next, we exploit some useful distribution features including maximize R
the support, unimodality, symmetry of the distribution, to subject to (2), (4), (6), and
 
obtain a computationally tractable approximation, which is fijd (c) ≤ ĥiJ (c), i ∈ N , K ⊆ Tic ,
presented by the following theorem. d∈D j∈K J⊆Tic
Theorem 2: Suppose that for a given εiJ (c) > 0, there (18)
exists f = (fiJj (c))j∈K ≥ 0 such that (15) holds, where Φ(f) 
is defined by (14). If the distribution of hiJ (c) has bounded where ĥiJ (c)  βiJ (c) − αiJ (c)ν − + αiJ (c)σ 2 log( εiJ1(c) ).
support [aiJ (c), biJ (c)], we then have

d
fiJj (c) ≤ βiJ (c) − αiJ (c)ν − A. Dual Decomposition and Subgradient Method
d∈D j∈K We form the dual problem by introducing the Lagrange
 multipliers for constraints in (4) and (18). By introducing dual
1
+ αiJ (c)σ 2 log( ), (17) variables λ  (λdiK (c)) ≥ 0, μ  (μiK (c)) ≥ 0 to relax
εiJ (c) (4) and (18) respectively, and moving them to the objective
which implies (13), where αiJ (c) = 12 (biJ (c) − aiJ (c)) = 0, function, we have
βiJ (c) = 12 (biJ (c) + aiJ (c)), and ν − , σ are the constants θ(λ, μ) = max L(R, f, z, λ, μ)
related to the probability distribution of hiJ (c). R,f,z

Proof: The proof mainly utilizes the following inequality subject to (R, f) ∈ QF (19)
2
Ω(y) ≤ max{ν − y, ν + y} + σ2 y 2 , where ν − , ν + ∈ [−1, 1] z ∈ PSTAB (G), (20)
and σ ≥ 0 are constants that depend on the given fami-
lies of probability distributions of y with bounded support where f = (fijd (c)) ≥ 0, z = (ziJ (c)) ≥ 0, and the flow
[−1, 1]. Since hiJ (c) may not be bounded in [-1, 1], e.g., polytope QF is defined by (2). The corresponding Lagrangian
hiJ (c) ∈ [aiJ (c), biJ (c)], we need to normalize hiJ (c) as is given by
follows: ζiJ (c)  hiJ (c)−β iJ (c)
, where αiJ (c)  12 (biJ (c) − λ, μ)
L(R, f, z,
αiJ (c)  
aiJ (c)) = 0 and βiJ (c)  12 (biJ (c)+aiJ (c)). And substituting = R+ λdiK (c)
ζiJ (c) in the inequality can derive the final approximation. The c
⎛ d∈D i∈N c∈Ci K⊆Ti ⎞
detailed proof is provided in Appendix B.  
The above approximation can be implemented in ×⎝ ziJ (c)biJK (c) − fijd (c)⎠
reality as follows. First, in practice, SUs learn the exact J⊆Tic
⎛ j∈K ⎞
probability distribution of hiJ (c) to determine the values    
of αiJ (c), βiJ (c), ν − , σ. In this study, we assume that SUs + μiK (c)⎝ ĥiJ (c)− fijd (c)⎠
can accurately obtain the probability distribution of B c by c c

the methods proposed in existing measurement data studies ⎛ i∈N c∈Ci K⊆Ti J⊆Ti d∈D j∈K

  
[15]–[17], [42] in advance. The estimation of the probability = ⎝R − λdiK (c) fijd (c)
distribution of B c can be seen as a warmup for the evaluation d∈D i∈N c∈Ci K⊆Tic j∈K
of the proposed algorithm in real world. Since we focus on ⎞
  
the long-duration application, the cost of this warmup will be − μiK (c) fijd (c)⎠
small, compared with the performance gain. The effect of the i∈N c∈Ci K⊆Ti d∈D j∈K c

estimation inaccuracy in the proposed algorithm will be stud- ⎛ ⎞

ied in the future. Second, after obtaining αiJ (c), βiJ (c), ν − , σ,   
+⎝ λdiK (c) ziJ (c)biJK (c)⎠
which can be exchanged between SUs by CCC, SUs substitute
d∈D i∈N c∈C K⊆Tic J⊆Tic
(12) by (17), and then solve the overall optimization problem.  i 
+ μiK (c) ĥiJ (c). (21)
V. D ECENTRALIZED A LGORITHM i∈N c∈Ci K⊆Tic J⊆Tic

We note that (17) is linear and thus convex, which may The first term in the above equation depends only on the
enable the whole problem to be solved by convex optimization flow rate variables fijd (c), while the second term relies solely
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE/ACM TRANSACTIONS ON NETWORKING


on the coding subgraph variables ziJ (c). Therefore, the dual where qijd
(c) = {K|K⊆Tic ,j∈K} (λiK (c) + μiK (c)). This
d

function can be computed by decomposing the optimization subproblem is a multiple-shortest-paths problem owing to the
problem into following subproblems: conservation constraints in the flow polytope QF . Addition-
θ(λ, μ) = max L(R, f, z, λ, μ) ally, the subproblem is similar to the subgraph optimization
{R,f,z} problem in [11], except that we should consider the channel
= θ1 (λ, μ) + θ2 (λ) + θ0 , (22) selection for every link under a given set of available channels.
Solving (28) is equivalent to searching for each sink d ∈ D,
where d
the shortest path with respect to the cost qij (c) from the source
θ1 (λ, μ) s to sink d, ensuring that there is at least a common channel
= max available for each link in the path. Moreover, to maximize (28),
(R,f)∈QF ,0≤R≤1
⎛ we need to increase R and make the right expression small.
   Nevertheless, R and f are coupled by QF . Fixing R, we should
× ⎝R − λdiK (c) fijd (c) minimize the cost:
d∈D i∈N c∈Ci K⊆Tic j∈K
⎞   
   min d
qij (c)fijd (c), (29)
fijd (c)⎠,
f
− μiK (c) (23) d∈D i∈N c∈Ci j∈Tic
i∈N c∈Ci K⊆Tic d∈D j∈K
for s to each d supporting end-to-end throughput R. (29) can
θ2 (λ) ⎛ be seen as a minimum-cost network problem without capacity
  constraints [11]. Consequently, it is sufficient for each of the
= max ⎝ λdiK (c)
z∈PSTAB (G) D parallel optimization problems to find a single minimum-
d∈D i∈N c∈Ci K⊆Tic
⎞ cost path from s to d with an available channel for each link.
 Since each of the flows in the minimum-cost path carries the
× ziJ (c)biJK (c)⎠, (24) total throughput R as there is only one active path from s to
J⊆Tic
   d and one channel for each link, (28) can be reformulated as
θ0 = μiK (c) ĥiJ (c). (25) ⎛ ⎞
i∈N c∈Ci K⊆Tic J⊆Tic  
max ⎝R(1 − d
qij (c))⎠, (30)
In the first subproblem, similar as in [11], we add a R,0≤R≤1
d∈D (i,j,c)∈Psd
redundant constraint on R (0 ≤ R ≤ 1), since by the
definitions of ziJ (c) and biJK (c), and constraint (4), the rate R where Psd denotes the collection of links (i, j) with channel c
cannot be larger than one. This constraint is to avoid handling (c ∈ Ci Cj ) in the minimum-cost path from s to d. If the sum
possibly unbounded solutions. Then, we have the following of costs in all paths from source s to all sinks is less than 1,
dual problem the maximum is 1 for R̂ and for those link flows fˆijd (c) with
(i, j, c) ∈ Psd . Otherwise, the maximum is zero by setting
minimize θ(λ, μ) both R and f to zero.
subject to λ ≥ 0, μ ≥ 0. To solve (30), we have to determine D shortest paths, which
Since θ(λ, μ) may not be differentiable, we apply the well- can be distributively solved by the asynchronous Bellman-Ford
known subgradient projection method to solve the above dual algorithm [35]. Since all qijd
(c) are greater than zero, there
problem. And the update rules are given by exist no cycles with negative costs and thus the shortest path
algorithms converge [35] in O(M N ) time, where M and N
λ[t + 1] = {λ[t] − η1 [t]Δ1 [t]}+ (26) are the number of arcs and nodes, respectively. In fact, in
μ[t + 1] = {μ[t] − η2 [t]Δ2 [t]}+ , (27) our problem, the time complexity is O(M N C), where C is
where {x}+  max(x, 0), η1 [t] and η2 [t] are the maximal number of available channels.
We now consider the second subproblem, which can be
corresponding
 appropriate stepsizes,
 and Δ1 [t] =
( J⊆T c ẑiJ (c)[t]b̂iJK (c) − fˆd (c)[t]), Δ2 [t] =
rewritten as
 i   j∈K ij ⎛ ⎞
( J⊆T c ĥiJ (c) − d∈D j∈K fˆijd (c)[t]). f̂[t] = (fˆijd (c)[t])  
i
and ẑ[t] = (ẑiJ (c)[t]) are the solutions of subproblems (23) ẑ = arg max ⎝ wiJ (c)ziJ (c)⎠, (31)
z∈PSTAB (G) i∈N c∈Ci J⊆Tic
and (24) respectively, at step t. And b̂iJK (c) can be locally
 
calculated by node i. We will solve these subproblems in a where weights wiJ (c) = d∈D K⊆Tic (λiK (c)biJK (c)).
d
decentralized way in Section V-B. The weight wiJ (c) of a node i is just a function of local
B. Flow Optimization and Scheduling Subproblem Solution variables λdiK (c) and biJK (c)), which enables us to design a
Rearranging (23) yields distributed scheduling algorithm. This is a standard maximum-
weight-stable-set (MWSS) problem in the scheduling litera-
(R̂, f̂) ture [36], which is NP-hard in general. A worth emphasizing
⎛ ⎛ ⎞⎞
   point is that the weight is not only related to the hyperarc
= arg max ⎝R− ⎝ d
qij (c)fijd (c)⎠⎠, but also the channel they choose. We have encapsulated both
(R,f)∈QF ,0≤R≤1 i∈N c∈Ci j∈Tic
d∈D the hyperarc and channel in a tuple in the proposed TDCG.
(28) To obtain an efficient solution, we relax the maximum weight
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

QU et al.: MULTICAST IN MULTIHOP CRNs UNDER UNCERTAIN SPECTRUM AVAILABILITY 9

stable set constraint and instead search a maximal stable link (i, J), using channel c. Similar to (2), there exists a flow
set. The relaxed problem can be solved using the distributed vector xdfiJj (c) for each sink d ∈ Df , which satisfies
algorithm in [36]. The input of the algorithm is a weighted   df 

indirected graph and the output is a stable set with high weight. xiJj (c) − xdf
jIi (c )

Further, the output stable set is guaranteed to be maximal too. J⊆Tic j∈J {j|I⊆Tjc ,i∈I}
And the algorithm converges in just 2δ(G) steps, where δ(G) ⎧
⎨ Rf i = sf
is the stability number of the conflict graph G [36]. = −Rf i ∈ Df (33)
The sequence of f̂[t] and ẑ[t] iterated by subgradient ⎩
0 otherwise,
optimization methods might be infeasible [37]. To recover
the primal solution, we employ a classical primal recovery for all f ∈ F and d ∈ Df .
method [37], which averages the intermediate optimal values, 2) Coding Subgraph Constraints: The coding subgraph
i.e. constraints contain the restriction of coding subgraph on the
link rate and the conflict on the hyperarc scheduling. Note that
1 1  1
T T T
R∗ = R̂[t], f∗ = f̂[t], z∗ = ẑ[t]. (32) in the multiple flows case, the definitions of biJK (c), ziJ (c)
T t=1 T t=1 T t=1 and ziJS (c) are the same as in the single flow case. For the
This approach can converge to the primal optimal solution first constraint, similar to (4), the flow rate for any sink of a
a    flow should be no more than the rate restricted by the coding
for a stepsize b +c  t with a > 0, b ≥ 0 and c > 0 [37].

Since both the flow optimization subproblem and scheduling
subproblem can be distributively solved, our proposed solu-
tion is also distributed. In the implementation, each node i
firstly initializes its variables, i.e., dual variables λ[0], μ[0],
flow rate variables fijd (c)[0] and coding subgraph variables
ziJ (c)[0]. Then, given λ[t] and μ[t], each node solves the two
subproblems to obtain the locally optimal flow rate fˆijd (c)[t]
and injection rate ẑiJ (c)[t], and exchanges this information to
its corresponding neighbors by the CCC. Each node updates
the dual variables using (26) and (27), and repeats the afore-
mentioned process until the algorithm converges.
VI. E XTENSION TO M ULTIPLE M ULTICAST F LOWS
So far, we only focus on the single multicast flow case.
In this section, we introduce how to extend our theoretic
framework to the multiple multicast flows case. Generally, in
this case, the optimization problem turns into how to maximize
the sum rates of all multicast flows subject to the three types
of constraint similar as in the single multicast flow case. In the
following, we will first present the problem formulation, and
then briefly discuss how to extend the solution to the multiple
multicast flows case.
Suppose there are F multicast flows in the network and
denote the flow set as F = {1, . . . , F }. The f th multicast
flow is associated with a source sf ∈ N and sinks Df =
{1, . . . , Df } ⊂ N , for f ∈ F. Let Rf be the multicast rate
of flow f ∈ F. Note that since we consider the intra-flow
network coding policy in this paper, the coding operation at
the intermediate nodes is executed in the same flow only.
First of all, the objective function becomes f ∈F U (Rf ),
where the utility function U (Rf ) is non-decreasing and
strictly concave. With the properly chosen utility function,
maximizing the objective function can achieve the fairness
among the flows [38]. Two classical examples of U (Rf )
R1−γ
are wf log(1 + Rf ) and wf 1−γ f
, where wf is the weight
associated with flow f , and 0 ≤ γ < 1. Next, we interpret
the three types of constraint one by one in details.
1) Flow Rate Constraints: Let xdfiJj (c) be the data rate for
sink d ∈ Df (f ∈ F) from node i to node j ∈ J, when
packets are actually transmitted on virtual wireless broadcast
subgraph combined with channel quality. That is,
 df
xiJj (c) ≤ ziJ (c)biJK (c), ∀d ∈ Df , f ∈ F. (34)
j∈K
For the second constraint of scheduling on the coding sub-
graph, it deals with the scheduling of hyperarcs with channels
only, which has nothing to do with the flows and their
rates. When there are multiple multicast flows, the con-
structed TDCG keeps unchanged. Accordingly, this constraint
is exactly the same as (6).
3) Link Capacity Constraints: In the multiple flows case,
the aggregate data rate on broadcast link (i, J) is the
sum
 offlow  rate over all sinks in all data flows, i.e.,
df
f ∈F d∈D j∈K xiJj (c). The definition of link capacity
hiJ (c) is same as (10). In a similar way, to deal with the
uncertainty of link capacity, we employ chance constraints to
model link capacity constraints in the following:
⎧ ⎫
⎨   ⎬
xdf
Pr iJj (c) − hiJ (c) ≤ ≥ 1 − εiJ (c), (35)
⎩ ⎭
f ∈F d∈D j∈K
where εiJ (c) ∈ (0, 1) is the corresponding violation tolerance.
Therefore, the problem of network coding-based multicast
in multi-hop CRNs with multiple data flows can be mathemat-
ically formulated as

maximize U (Rf )
f ∈F
subject to (33), (34), (6), (35),
where xdfiJj (c) ≥ 0 and ziJ (c) ≥ 0 are optimization variables,
corresponding to flow rate and coding subgraph, respectively.
It is not difficult to see that the above problem shares a
very similar form with the chance-constrained program for a
single multicast flow defined in Section III-D, which can be
solved by a similar approach, since the objective function is
convex. Specifically, via Bernstein approximation introduced
in Section IV, we can first transform the chance constraint (35)
into a safe and tractable constraint. Then, the transformed
problem can be efficiently solved using a dual decomposition
and subgradient method similar as in Section V.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
10
VII. E VALUATION
In this section, we evaluate the performance of our proposed
solution by extensive simulations executed on Matlab. The
main goal is to demonstrate the superior performance of
our algorithm over existing approaches. We first introduce
the objects of comparison in our evaluation, then explain the
explicit simulation parameters setting, and lastly present the
evaluation results for a single multicast flow and multiple
multicast flows.
A. Objects of Comparison
Our decentralized algorithm offers a feasible solution
to the chance-constrained program. Firstly, to evaluate the
gap between our algorithm and the optimal solution, we
can remove constraint (12) in the program, whose solution
forms an upper bound of the performance of our proposed
algorithm. Since the modified program is apparently a normal
convex problem, we then solve it using the similar dual
decomposition and subgradient method and label the new
solution as “Upper bound” in the figures. As is known to all,
there is no available network coding-based multicast algorithm
for multi-hop CRNs in existing literatures. Therefore, we
employ the following algorithm to validate the use of the
chance-constrained program.
Specifically, we incorporate an approach called as Robust
network coding that simply substitutes the bandwidth B c to the
expectation value of B c in the link capacity hiJ (c). It should
be noted that this algorithm employs network coding in the
multicast transmission but does not consider the uncertainty
of spectrum availability. Note that the robust network coding
algorithm also employs a dual decomposition and subgradient
method similar as in our proposed decentralized algorithm.
Besides, to show the benefits of network coding on multicast in
multi-hop CRNs, we implement the approach (labeled “Non-
network coding” in the figures) in [3] that studies multicast in
multi-hop CRNs without network coding.
B. Setup
For ease of scalability, the distance, rate, bandwidth and
power are normalized with appropriate dimensions in our
simulation. We conduct simulations with a CRN over random
network topologies, where nodes are randomly placed on
a square region. We evaluate all algorithms in a network
topology with 20 nodes on a 80 × 80 square. Specifically,
the location of each node is randomly chosen from the square
region with a uniform probability L1 , where L is the number
of locations. We consider the leftmost node to always be
the source s and two sets of sinks for the single multicast
flow case, i.e., the two rightmost nodes and the five rightmost
nodes, respectively. And for the multiple multicast flows case,
we suppose that there are three multicast flows (each with
three sinks). Under our interference model, the link-channel
interference and self-interference at the physical layer are
developed in our simulations after the above settings.
Following previous works [3], [18], the explicit settings
of each node are as follows. The power propagation gain
between node i and node j is gij = γd−n ij , where dij denotes
IEEE/ACM TRANSACTIONS ON NETWORKING
the distance of node i and node j, γ is an antenna related
constant, and n is the path loss factor. We set γ and n to
be 1 and 4, respectively. The power spectral density from
the transmitter is Q = 1.6 × 105 η, where η is the ambient
Gaussian noise density. The transmission range for a node
1 1
is thus rT = (γQ/η) n = (1.6 × 105 ) 4 = 20. To evaluate
the performance of our proposed algorithm under different
channel qualities, we add a uniform artificial channel erasure
rate p in the erasure model, i.e., any successful reception is
dropped with probability p, where we change p from 0.05
to 0.25 in the simulations. We also suppose that the erasures
across different receivers are independent. For the tolerance
level, we set εiJ (c) = 0.01 for all i ∈ N , c ∈ Ci , J ⊂ Tic .
For the channel setting, we assume that there are C = 5
licensed channels that can be opportunistically used by SUs.
The number of available channels at each node is randomly
selected from [1, C], and the exact set of available channels is
randomly constructed from the channel pool {1, 2, . . . , C}.
As to the probability distribution of the channel band-
width B c , its exact expression is determined by the cor-
responding occupancy of PUs. Based on the measurement
study on spectrum utilization in [15], the distribution can be
exponential. Accordingly, we first consider the exponential
case, i.e., B c ∼ E(υ c ), where υ c ∈ (0, 3]. Note that different
channel c may have different distribution parameter υ c . In
our simulations, we randomly choose υ c from (0, 3] for each
channel c to simulate heterogeneous channels. Moreover, to
validate the effectiveness of our proposed algorithm, we also
conduct a similar simulation study on a normal bandwidth
distribution, i.e., B c follows a normal distribution, B c ∼
2
N (μ , σ  ). In the simulations, μ and σ  are set as 2 and 1,
respectively. Worth noting that the analysis and results can be
easily extended to any other distribution.
As to the initialization of Lagrange multipliers in our
proposed algorithm and three other algorithms, we set the
initial values of λdiK (c) and μiK (c) to be different random
values chosen from (0, 0.1). Also, we run these algorithms
with the stepsizes chosen to be η1 [t] = 1t and η2 [t] = 1t . Due to
the randomly deployment of both nodes and channels as well
as the inherent spectrum scarce and limited transmission range
of nodes, there may exist disconnections between the source
and some destinations, which results in no feasible solution
for some specific data set. Therefore, in this study, we focus
only on the data sets with feasible solutions, and present the
corresponding results in the following subsection.
C. Single Multicast Flow
1) Exponential Distribution Case: We first present
the simulation results of a single multicast flow with
two and five sinks, when the bandwidth B c follows
exponential distribution, as illustrated in Fig. 3 and Fig. 4,
respectively. Fig. 3(a) shows the throughput performance of
our decentralized algorithm over a multicast flow with two
sinks under Exponential distribution, compared to the upper
bound, robust network coding algorithm and non-network
coding algorithm, when the channel erasure rate p varies.
According to the figure, firstly, our algorithm outperforms the
non-network coding algorithm, no matter how the channel
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

QU et al.: MULTICAST IN MULTIHOP CRNs UNDER UNCERTAIN SPECTRUM AVAILABILITY 11

Fig. 3. Maximum multicast rate as a function of channel erasure rate for Fig. 5. Maximum multicast rate as a function of channel erasure rate for
a single multicast flow with two sinks under Exponential distribution, i.e., a single multicast flow with two sinks under Normal distribution,
B c ∼ E(υc ), υc ∈ (0, 3]. (a) Performance comparison of four algorithms. i.e., B c ∼ N (2, 1). (a) Performance comparison of four algorithms.
(b) Performance ratio of our decentralized algorithm. (b) Performance ratio of our decentralized algorithm.

Fig. 6. Maximum multicast rate as a function of channel erasure rate for

Fig. 4. Maximum multicast rate as a function of channel erasure rate for
a single multicast flow with five sinks under Normal distribution,
a single multicast flow with five sinks under Exponential distribution, i.e.,
i.e., B c ∼ N (2, 1). (a) Performance comparison of four algorithms.
B c ∼ E(υc ), υc ∈ (0, 3]. (a) Performance comparison of four algorithms.
(b) Performance ratio of our decentralized algorithm.
(b) Performance ratio of our decentralized algorithm.

quality deteriorates. This is because with network coding, both

wireless broadcast nature and multipath benefits are effectively
exploited in multi-hop CRNs, while without network coding,
packets are transmitted for each sink on one single path only.
Further, we find that the performance of the robust network
coding algorithm is a little worse than the non-network coding
algorithm, due to the overly conservative constraints on link
capacity. And this shows the importance of considering Fig. 7. Sum of multicast rate as a function of channel erasure rate for
the uncertainty of spectrum availability. Moreover, from multiple multicast flows under Exponential distribution, i.e., B c ∼ E(υc ),
Fig. 3(b), our decentralized algorithm achieves a competitive υc ∈ (0, 3]. (a) Performance comparison of four algorithms. (b) Performance
performance compared with the upper bound. And the gain ratio of our decentralized algorithm.
of our decentralized algorithm over robust network coding
and 1.8182× maximum multicast rate over robust network
and non-network coding is up to 1.9412× and 1.6837×,
coding and non-network coding, respectively. Moreover, from
respectively. The above phenomena are similar when there are
Fig. 6, compared to the two sinks case, the performance gain
five sinks, as shown in Fig. 4. And we find that the gain of our
of our decentralized algorithm is higher, i.e., up to 3.368× over
proposed algorithm over both robust network coding and non-
non-network coding and 2.844× over robust network coding,
network coding is relatively stable, up to ∼1.91 and ∼2.11,
respectively. And as illustrated in Fig. 6(b), the “Over robust
respectively. The ratio of our proposed algorithm to the upper
network coding” curve and “Over non-network coding“ curve
bound in the five sinks case is steady at [0.73, 0.79]. This illus-
has a cross point when the channel erasure rate varies. This
trates that our algorithm is effective in a bigger set of sinks.
is because the multicast rates of the two algorithms are small
2) Normal Distribution Case: Fig. 5 and Fig. 6 show the
and unstable as shown in Fig. 6(a).
evaluation results of a single multicast flow with two and five
sinks under Normal distribution, respectively. Specifically, in
Fig. 5, we evaluate the performance of all algorithms under D. Multiple Multicast Flows
Normal distribution with two sinks, when the channel erasure Fig. 7 and Fig. 8 illustrate the performance of the proposed
rate p changes from 0.05 to 0.25. As shown in the figure, decentralized algorithm when there are three multiple flows
similar to Fig. 3, compared to the robust network coding in the network, under Exponential and Normal distribution,
algorithm and the non-network coding algorithm, our proposed respectively. From these figures, we can see that, compared
algorithm always achieves higher maximum multicast rates to both robust network coding and non-network coding, our
when the channel erasure rate varies. Similarly, the perfor- decentralized algorithm still achieves higher multicast rate
mance of the robust network coding algorithm is as low in both distributions. Moreover, according to Fig. 7(b) and
as that of non-network coding algorithm. And according to Fig. 8(b), we find that the performance gain of the proposed
Fig. 5(b), our decentralized algorithm achieves at most 2.5× algorithm in multiple multicast flows case is higher than that
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
12
Fig. 8. Sum of multicast rate as a function of channel erasure rate for
multiple multicast flows under Normal distribution, i.e., B c ∼ N (2, 1).
(a) Performance comparison of four algorithms. (b) Performance ratio of our
decentralized algorithm.
Fig. 9. Convergence of the distributed algorithm. (a) Single multicast flow.
(b) Multiple multicast flows.
in single multicast flow case. This is because under multiple
multicast flows, with network coding and careful consider-
ation into the uncertain spectrum availability, our proposed
algorithm not only better uses the benefits of network coding
over multi-path routing, but also fully utilizes the available
spectrums in CRNs.
E. Convergence
In Fig. 9, we illustrate the convergence of the decentralized
algorithm with a single multicast flow and multiple multicast
flows. Since the optimal value of the maximum multicast rate
(sum of multicast rate in the multiple multicast flows case),
i.e., R∗ , cannot be efficiently obtained, we use instead the
improvement in the maximum multicast rate in each iteration,
IEEE/ACM TRANSACTIONS ON NETWORKING
schemes in multi-hop multichannel wireless ad hoc networks.
With opportunistic overhearing, Chieochan and Hossain [14]
proposed a suboptimal, auction-based solution for the overall
network throughput optimization with network coding in
MR-MC networks. Lin and Yang [29] studied the problem of
multirate multichannel multicast with intra-session network
coding. However, these studies for MR-MC networks are
not applicable to our problem. This is because in MR-MC
networks, the set of common available channels for every
node in the network is assumed fixed and corresponding link
capacity is definite, which is impractical for CRNs.
B. Multicast in Multi-Hop CRNs
Since there are multiple channels in CRNs, the problem
of multicast in multi-hop CRNs always incorporates the
routing and channel assignment jointly [4], [25], [26].
Kim et al. [25] improved the scalability of the classical
routing protocol ODMRP [27] to support multicast
services in a cognitive radio ad hoc wireless network.
Almasaeid et al. [26] proposed an on-demand multicast
routing algorithm for cognitive radio mesh networks. In [4],
a reactive joint channel allocation and multicast scheme is
proposed for a multi-hop CRN. Gao et al. [3] formulated
the problem of multicast in multi-hop CRNs as mixed
linear program to minimize the required resources. Besides,
Almasoud and Kamal [28] studied the robust multicasting in
the face of failures in CRNs, and proposed two algorithms
that provide robust multicasting in such scenarios. However,
none of the aforementioned works utilizes network coding to
pursue higher throughput performance.
C. The Utilization of Network Coding in CRNs
A large body of work has been carried out on exploiting the
various possible benefits of network coding in CRNs from dif-
ferent angles recently. Using network coding as an assistance
strategy, Almasaeid and Kamal [5] proposed an algorithm
i.e., R∗ = R∗ [t + 1] − R∗ [t]. From the figures, we note that to reduce the effect of the channel heterogeneity property
our algorithm achieves fast convergence in both cases, where on the multicast throughput in cognitive radio wireless mesh
R∗ quickly converges close to zero within ∼60 iterations.
VIII. R ELATED W ORK
In this section, we review the related work in the following
three aspects: network coding-based multicast in MR-MC
networks, multicast in multi-hop CRNs, and the utilization of
network coding in CRNs.
A. Network Coding-Based Multicast in MR-MC Networks
Although how to utilize network coding to enhance
multicast throughput in singleradio-singlechannel networks
has been extensively studied, the benefits of network
coding over multicast in multiradio-multichannel (MR-MC)
networks are also validated in recent works [12]–[14], [29].
Zhang and Li [12] showed that network coding can further
increase the capacity of multichannel mesh networks.
Su and Zhang [13] modeled the throughput gains of
conventional network coding and analog network coding
networks. Qu et al. [30] develops an efficient network coding
strategy for SUs to improve the broadcast throughput while
considering the uncertain idle durations in CRNs. Note that
in [5] and [30], network coding is restricted in single hop,
which restricts its application in our problem. Network coding
is viewed as a spectrum shaper to increase the spectrum avail-
ability and predictability for SUs, through exploiting network
coding on PUs only [31]. Shu et al. [32] proposed a network
coding-aware channel allocation and routing scheme for multi-
hop CRNs. However, network coding is actually not utilized
in SUs’ data transmissions in [32]. Besides, Jin et al. [2]
studied multicast scheduling with cooperation and network
coding in CRNs, where network coding is mainly used to
reduce overhead and perform error control.
D. Multi-Path Routing and Multiple
Unicast in Multi-Hop CRNs
In essence, our work is a kind of multi-path routing in multi-
hop CRNs. In related studies, Beltagy et al. [39] proposed
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
QU et al.: MULTICAST IN MULTIHOP CRNs UNDER UNCERTAIN SPECTRUM AVAILABILITY
a novel routes selection metric for multi-path routing in
multi-hop CRNs, and validated the effectiveness of the new
metric in enhancing end-to-end throughput and increasing
connection reliability by extensive simulations. Shu et al. [41]
studied the problem of minimizing the end-to-end delay
with multi-path routing in multi-hop CRNs, and proposed
a channel allocation and rate allocation scheme with low
complexity based on graph coloring. Shi and Hou [40]
investigated how to maximize the data rate for several user
communication sessions with multi-path routing in multi-hop
CRNs, and presented a distributed optimization algorithm that
increases the data rate iteratively for the sessions. Besides,
considering the uncertain spectrum supply, Pan et al. [18]
studied the joint multi-path routing and frequency scheduling
problem in multi-hop CRNs. However, compared to the above
studies, there are at least two differences from our work.
First, we intend to employ network coding to pursue further
performance improvement, while they do not use network
coding in the multi-path routing. When network coding is
introduced, these studies cannot be applied since they do
not consider the coding subgraph optimization, which is the
key issue of network coding-based multicast problems [8].
Second, all the aforementioned works focus on the service of
multiple unicast flows in multi-hop CRNs, which only fits for
the data transmission of point-to-point (unicast) other than
point-to-multipoint (multicast).
IX. C ONCLUSION
In this paper, we for the first time studied the network
coding-based multicast problem in multi-hop CRNs, while
considering the uncertain spectrum availability. We formulated
the overall problem as a chance-constrained program. We
constructed a TDCG to encapsulate both hyperarc scheduling
and channel selection into hyperarc-channel tuples. Though
appropriate Bernstein approximation and relaxation of finding
stable sets in the TDCG, we proposed an efficient distributed
algorithm for the original problem. Simulation results showed
that our algorithm obtains higher multicast rates than the
non-network coding algorithm and the robust network coding
algorithm in CRNs. We also extend our theoretic framework
to multiple multicast flows.
R EFERENCES
[1] J. Mitola, III, “Cognitive radio: An integrated agent architecture for soft-
ware defined radio,” Ph.D. dissertation, Royal Inst. Technol., Stockholm,
Sweden, 2000.
[2] J. Jin, H. Xu, and B. Li, “Multicast scheduling with cooperation and
network coding in cognitive radio networks,” in Proc. IEEE INFOCOM,
Mar. 2010, pp. 1–9.
[3] C. Gao, Y. Shi, Y. T. Hou, H. D. Sherali, and H. Zhou, “Multicast
communications in multi-hop cognitive radio networks,” IEEE J. Sel.
Areas Commun., vol. 29, no. 4, pp. 784–793, Apr. 2011.
[4] Z. Shu, Y. Qian, Y. Yang, and H. Sharif, “Channel allocation and
multicast routing in cognitive radio networks,” in Proc. IEEE WCNC,
Apr. 2013, pp. 1703–1708.
[5] H. M. Almasaeid and A. E. Kamal, “Exploiting multichannel diversity
for cooperative multicast in cognitive radio mesh networks,” IEEE/ACM
Trans. Netw., vol. 22, no. 3, pp. 770–783, Jun. 2014.
[6] R. Ahlswede, N. Cai, S.-Y. R. Li, and R. W. Yeung, “Network infor-
mation flow,” IEEE Trans. Inf. Theory, vol. 46, no. 4, pp. 1204–1216,
Jul. 2000.
[7] T. Ho et al., “A random linear network coding approach to multicast,”
IEEE Trans. Inf. Theory, vol. 52, no. 10, pp. 4413–4430, Oct. 2006.
13
[8] D. S. Lun et al., “Minimum-cost multicast over coded packet networks,”
IEEE Trans. Inf. Theory, vol. 52, no. 6, pp. 2608–2623, Jun. 2006.
[9] T. Ho and H. Viswanathan, “Dynamic algorithms for multicast with
intra-session network coding,” IEEE Trans. Inf. Theory, vol. 55, no. 2,
pp. 797–815, Feb. 2009.
[10] A. Khreishah, I. M. Khalil, and J. Wu, “Distributed network coding-
based opportunistic routing for multicast,” in Proc. ACM MobiHoc,
2012, pp. 115–124.
[11] D. Traskov, M. Heindlmaier, M. Médard, and R. Koetter, “Scheduling
for network-coded multicast,” IEEE/ACM Trans. Netw., vol. 20, no. 5,
pp. 1479–1488, Oct. 2012.
[12] X. Zhang and B. Li, “On the benefits of network coding in multi-channel
wireless networks,” in Proc. IEEE SECON, Jun. 2008, pp. 341–349.
[13] H. Su and X. Zhang, “Modeling throughput gain of network coding in
multi-channel multi-radio wireless ad hoc networks,” IEEE J. Sel. Areas
Commun., vol. 27, no. 5, pp. 593–605, Jun. 2009.
[14] S. Chieochan and E. Hossain, “Channel assignment for through-
put optimization in multichannel multiradio wireless mesh networks
using network coding,” IEEE Trans. Mobile Comput., vol. 12, no. 1,
pp. 118–135, Jan. 2013.
[15] D. Chen, S. Yin, Q. Zhang, M. Liu, and S. Li, “Mining spectrum
usage data: A large-scale spectrum measurement study,” in Proc. ACM
MobiCom, 2009, pp. 13–24.
[16] M. Mehenry. (Jun. 2003). Spectrum White Space Measurements.
[Online]. Available: http://www.newamerica.net/Download_Docs/pdfs/
Doc_File_185_1.pdf
[17] M. Wellens, J. Riihijärvi, and P. Mähönen, “Evaluation of adaptive mac-
layer sensing in realistic spectrum occupancy scenarios,” in Proc. IEEE
DySPAN, Apr. 2010, pp. 1–12.
[18] M. Pan, C. Zhang, P. Li, and Y. Fang, “Spectrum harvesting and sharing
in multi-hop CRNs under uncertain spectrum supply,” IEEE J. Sel. Areas
Commun., vol. 30, no. 2, pp. 369–378, Feb. 2012.
[19] A. Nemirovski and A. Shapiro, “Convex approximations of chance
constrained programs,” SIAM J. Optim., vol. 17, no. 4, pp. 969–996,
2006.
[20] A. Ben-Tal and A. Nemirovski, “Selected topics in robust convex
optimization,” Math. Program., vol. 112, no. 1, pp. 125–158, 2008.
[21] A. Nemirovski, “On safe tractable approximations of chance con-
straints,” Eur. J. Oper. Res., vol. 219, no. 3, pp. 707–718, 2012.
[22] G. Ganesan and Y. G. Li, “Cooperative spectrum sensing in cognitive
radio networks,” in Proc. IEEE DySpAN, Nov. 2005, pp. 137–143.
[23] T. Yucek and H. Arslan, “A survey of spectrum sensing algorithms for
cognitive radio applications,” IEEE Commun. Surveys Tuts., vol. 11,
no. 1, pp. 116–130, 1st Quart., 2009.
[24] Y. Cheng, H. Li, and P.-J. Wan, “A theoretical framework for optimal
cooperative networking in multiradio multichannel wireless networks,”
IEEE Wireless Commun., vol. 19, no. 2, pp. 66–73, Apr. 2012.
[25] W. Kim, S. Y. Oh, M. Gerla, and J.-S. Park, “COCAST: Multicast
mobile ad hoc networks using cognitive radio,” in Proc. IEEE MILCOM,
Oct. 2009, pp. 1–7.
[26] H. M. Almasaeid, T. H. Jawadwala, and A. E. Kamal, “On-demand
multicast routing in cognitive radio mesh networks,” in Proc. IEEE
GLOBECOM, Dec. 2010, pp. 1–5.
[27] S.-J. Lee, W. Su, and M. Gerla, “On-demand multicast routing protocol
in multihop wireless mobile networks,” Mobile Netw. Appl., vol. 7, no. 6,
pp. 441–453, 2002.
[28] A. M. Almasoud and A. E. Kamal, “Robust provisioning of multi-
cast sessions in cognitive radio networks,” in Proc. IEEE IWCMC,
Aug. 2014, pp. 417–422.
[29] K. C.-J. Lin and D.-N. Yang, “Multicast with intraflow network coding
in multirate multichannel wireless mesh networks,” IEEE Trans. Veh.
Technol., vol. 62, no. 8, pp. 3913–3927, Oct. 2013.
[30] Y. Qu et al., “Towards near optimal network coding for secondary users
in cognitive radio networks,” in Proc. IEEE SECON, Jun./Jul. 2014,
pp. 522–530.
[31] A. Fanous, Y. E. Sagduyu, and A. Ephremides, “Reliable spectrum
sensing and opportunistic access in network-coded communications,”
IEEE J. Sel. Areas Commun., vol. 32, no. 3, pp. 400–410, Mar. 2014.
[32] Z. Shu, J. Zhou, Y. Yang, H. Sharif, and Y. Qian, “Network coding-
aware channel allocation and routing in cognitive radio networks,” in
Proc. IEEE GLOBECOM, Dec. 2012, pp. 5590–5595.
[33] A. Charnes, W. W. Cooper, and G. H. Symonds, “Cost horizons and
certainty equivalents: An approach to stochastic programming of heating
oil,” Manage. Sci., vol. 4, no. 3, pp. 235–263, 1958.
[34] M. Garey and D. Johnson, Computers and Intractability: A Guide to the
Theory of NP-Completeness. San Francisco, CA, USA: Freeman, 1979.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
14
[35] D. P. Bertsekas and R. G. Gallager, Data Networks, 2nd ed.
Upper Saddle River, NJ, USA: Prentice-Hall, 1992.
[36] S. Basagni, “Finding a maximal weighted independent set in wireless
networks,” Telecommun. Syst., vol. 18, nos. 1–3, pp. 155–168, 2001.
[37] T. Larsson, M. Patriksson, and A.-B. Strömberg, “Ergodic, primal
convergence in dual subgradient schemes for convex programming,”
Math. Program., vol. 86, no. 2, pp. 283–312, 1999.
[38] T. Bonald and L. Massoulié, “Impact of fairness on Internet perfor-
mance,” in Proc. ACM SIGMETRICS, 2001, pp. 82–91.
[39] I. Beltagy, M. Youssef, and M. El-Derini, “A new routing metric and
protocol for multipath routing in cognitive networks,” in Proc. IEEE
WCNC, Mar. 2011, pp. 974–979.
[40] Y. Shi and Y. T. Hou, “A distributed optimization algorithm for multi-
hop cognitive radio networks,” in Proc. IEEE INFOCOM, Apr. 2008,
pp. 1966–1974.
[41] Z. Shu, Y. Qian, and R. Q. Hu, “Delay based channel allocations in
multi-hop cognitive radio networks,” in Proc. IEEE IWCMC, Aug. 2015,
pp. 1475–1480.
[42] S. Geirhofer, L. Tong, and B. M. Sadler, “Dynamic spectrum access in
WLAN channels: Empirical model and its stochastic analysis,” in Proc.
ACM TAPAS, 2006, pp. 271–275.
[43] J. Pintér, “Deterministic approximations of probability inequalities,”
Zeitschrift Oper. Res., vol. 33, no. 4, pp. 219–239, 1989.
[44] P. Wang, G. Mao, Z. Lin, X. Ge, and B. D. O. Anderson, “Network
coding based wireless broadcast with performance guarantee,” IEEE
Trans. Wireless Commun., vol. 14, no. 1, pp. 532–544, Jan. 2015.
[45] H. Fukui, P. Popovski, and H. Yomo, “Physical layer network coding:
An outage analysis in cellular network,” in Proc. 22nd Eur. Signal
Process. Conf. (EUSIPCO), 2014, pp. 1133–1137.
Yuben Qu received the B.S. degree in mathematics
and applied mathematics from Nanjing University
in 2009, and the M.S. degree in communication
and information systems and the Ph.D. degree in
computer science and technology from the PLA Uni-
versity of Science and Technology in 2012 and 2016,
respectively. He is currently a Research Assistant
with the Rocket Force University of Engineering,
China. His research interests include D2D commu-
nication, crowdsensing, and network coding.
Chao Dong (M’06) received the Ph.D. degree in
communication engineering from the PLA Univer-
sity of Science and Technology, China, in 2007.
From 2008 to 2011, he held a post-doctoral position
with the Department of Computer Science and Tech-
nology, Nanjing University, China. He is currently an
Associate Professor with the College of Communi-
cation Engineering, PLA University of Science and
Technology. His research interests include wireless
cognitive networks, software defined networks, and
network coding. He is a member of the IEICE.
Haipeng Dai received the B.S. degree from the
Department of Electronic Engineering, Shanghai
Jiao Tong University, Shanghai, China, in 2010, and
the Ph.D. degree from the Department of Com-
puter Science and Technology, Nanjing University,
Nanjing, China, in 2014. He is currently a Research
Assistant with the Department of Computer Science
and Technology, Nanjing University, Nanjing, China.
His research interests are mainly in the areas of
wireless sensor networks, wireless power transfer,
and smartphone.
IEEE/ACM TRANSACTIONS ON NETWORKING
Fan Wu received the B.S. degree in computer
science from Nanjing University in 2004, and the
Ph.D. degree in computer science and engineering
from the State University of New York at Buffalo
in 2009. He has visited the University of Illinois at
Urbana–Champaign as a Post Doc Research Asso-
ciate. He is currently an Associate Professor with the
Department of Computer Science and Engineering,
Shanghai Jiao Tong University. His research interests
include wireless networking and mobile computing,
algorithmic network economics, and privacy preser-
vation. He has authored over 70 peer-reviewed papers in leading technical
journals and conference proceedings. He is a receipt of the China National
Natural Science Fund for Outstanding Young Scientists, the CCF-Intel Young
Faculty Researcher Program Award, the CCF-Tencent Rhinoceros bird Open
Fund, and the Pujiang Scholarship. He has served as the Chair of the
CCF YOCSEF Shanghai, on the Editorial Board of Computer Communi-
cations, and as a member of technical program committees of more than
40 academic conferences.
Shaojie Tang (S’09–M’14) received the B.S. degree
in radio engineering from Southeast University,
China, in 2006, and the Ph.D. degree from the
Department of Computer Science, Illinois Institute
of Technology, in 2012. He is currently an Assis-
tant Professor with the Naveen Jindal School of
Management, University of Texas at Dallas. His
main research interests focus on wireless networks
(including sensor networks and cognitive radio net-
works), social networks, security and privacy, and
game theory. He has served on the Editorial Board
of the Journal of Distributed Sensor Networks. He also served as a TPC
Member of a number of conferences, such as the ACM MobiHoc, the IEEE
ICNP, and the IEEE SECON.
Hai Wang (SM’14) received the Ph.D. degree in
communications engineering from the PLA Univer-
sity of Science and Technology, China, in 1999.
He is currently a Professor with the College of
Communications Engineering, PLA University of
Science and Technology. His current research inter-
ests include wireless network, cognitive network,
and network coding. He was the Publication Chair
of Wireless Communications and Signal Processing
2011.
Chang Tian received the B.S. degree in commu-
nication and information system and the M.S. and
Ph.D. degrees in computer science and engineering
from the Institute of Communications Engineering,
Nanjing, China, in 1984, 1989, and 2001, respec-
tively. From 1995 to 1996, he was a Visiting Scholar
with the Università di Pisa of Italy. He is currently
a Professor with the PLA University of Science
and Technology, China. He has published widely
in the areas of distributed system, wireless ad hoc
network, signal processing for communications, and
information theory. His current research interests include wireless sensor
network and network coding.