IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO.

2, FEBRUARY 2015

921

Group Partition and Dynamic Rate Adaptation

for Scalable Capacity-Region-Aware
Device-to-Device Communications
Yi-Shing Liou, Rung-Hung Gau, Senior Member, IEEE, and Chung-Ju Chang, Fellow, IEEE

AbstractIn this paper, we propose using group partition and

dynamic rate adaptation for scalable throughput optimization
of capacity-region-aware device-to-device communications. We
adopt network information theory that allows a receiving device
to simultaneously decode multiple packets from multiple transmitting devices, as long as the vector of transmitting rates is
inside the capacity region. Based on graph theory, devices are
first partitioned into subgroups. To optimize the throughput of a
subgroup, instead of directly solving an integer-linear programming problem, we propose using a fast iterative algorithm to select
active devices and using aggression levels for rate adaptation based
on channel state information. Simulation results show that the
proposed algorithm is scalable and could significantly outperform
the greedy algorithm by more than 50%.
Index TermsDevice-to-device communications, capacity region, graph coloring, rate adaptation, channel state information.

I. I NTRODUCTION

O deal with the exponential growth of mobile traffic,

device-to-device (D2D) communications have been considered one of the key techniques in the Third Generation
Partnership Project (3GPP) Long Term Evolution Advanced
(LTE-A) [1], [2]. With D2D communications capability, two
adjacent devices could directly send data to each other without using base stations as relays. D2D communications could
increase the throughput of cellular networks and decrease the
energy consumption of mobile devices [1][3]. Unlike mobile
ad-hoc networks based on the IEEE 802.11 standards, D2D
communications in cellular networks could use licensed bands
to avoid interference from devices that use unlicensed bands.
An overview of D2D standardization activities in 3GPP can be
found in [1].
In this paper, we study the case in which the network consists
of cellular user equipments (UEs) and devices. A cellular user
equipment sends data through the base station, while a source
device directly sends data to the corresponding destination
Manuscript received April 15, 2014; revised July 28, 2014; accepted
October 2, 2014. Date of publication October 9, 2014; date of current version
February 6, 2015. This work was supported in part by the Ministry of Science and Technology, Taiwan under Grant NSC 101-2221-E-009-005-MY3.
This paper was presented in part at the IEEE Wireless Communications and
Networking Conference (WCNC), Istanbul, Turkey, April 2014. The associate
editor coordinating the review of this paper and approving it for publication was
S. Cui.
The authors are with the Department of Electrical Engineering, National
Chiao Tung University, Hsinchu 300, Taiwan (e-mail: kilik820@hotmail.com;
runghung@mail.nctu.edu.tw; cjchang@mail.nctu.edu.tw).
Digital Object Identifier 10.1109/TWC.2014.2362523

device. We take into consideration the case in which a number

of source devices could have the same destination device.
We adopt network information theory [4] for multiple access
channels. In particular, a receiving device could concurrently
receive/decode distinct packets from multiple transmitting devices as long as the transmission rate vector is inside the
capacity region. When the proposed approach is used, devices
never severely interfere with cellular user equipments.
For D2D communications, Liou et al. [5] proposed using a
graph coloring algorithm to partition devices into subgroups
and solving an integer-linear programming problem to maximize the throughput of a subgroup. In this paper, we focus on
scalable algorithms for large-scale networks in which solving
integer-linear programming problems is infeasible. Our major
technical contributions of the paper include the following. First,
for each subgroup, we propose using a fast approximation
algorithm to choose active devices and dynamically adjusting the data transmission rates based on aggression levels.
Unlike most of the previous work on D2D communications,
the approximation algorithm exploits the information-theoretic
capacity region of a multiple access channel to improve the
network throughput. In addition, the approximation algorithm
remains applicable when only partial channel state information
is available. Furthermore, we show how to extend the proposed
approach to the case in which there are both short-lived D2D
communication sessions and long-lived D2D communication
sessions. Our simulation results show that the proposed approach could significantly outperform the greedy algorithm and
could be near-optimal.
The rest of the paper is organized as follows. In Section II, we
briefly introduce related work. In Section III, we present system
models for D2D communications. In Section IV, for scalable
D2D communications, we show how to create a conflict graph
in order to partition devices into subgroups. In Section V, we
first formulate an integer-linear programming problem for maximizing the throughput of a subgroup. In addition, we propose
a fast approximation algorithm to decide the set of active devices and adjust data transmission rates based on channel state
information. In Section VI, we show mathematical properties
of the proposed algorithm. In Section VII, we show simulation
results that justify the usage of the proposed approach. In
Section VIII, we outline an implementation and an extension
of the proposed approach in cellular networks. Our conclusions
are included in Section IX. Mathematical proofs are included in
the Appendix.

1536-1276 2014 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.

922

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 2, FEBRUARY 2015

II. R ELATED W ORK

Resource allocation had been extensively studied for enhancing the network performance. Yu et al. [6] studied the problem
of optimal resource allocation with power control between
cellular connections and D2D connections. Le [7] proposed a
maxmin fair resource allocation algorithm for cellular users
and a resource allocation algorithm with the rate protection for
cellular users and D2D users. Feng et al. [8] studied a resource
allocation problem to maximize the overall network throughput
while guaranteeing the quality of service for both D2D users
and regular cellular users. In [9], stochastic Petri Nets were used
for performance analysis of D2D communications with random
packet arrivals. Chai et al. [10] proposed a resource assignment
algorithm which allows D2D users to utilize the resource of
multiple cellular users to reduce the interference to cellular
users. Lee et al. [11] proposed a semi-distributed resource
allocation algorithm based on the multiple set covering problem
to maximize the spatial reuse of radio resources. Zhang et al.
proposed [12] a graph-theoretic resource allocation scheme for
D2D communications in OFDM-based cellular networks. In
their scheme, to reflect that channel responses for distinct resource blocks could be different, an edge weight in the conflict
graph is a matrix rather than a real number.
Power control and interference management had been widely
used to guarantee the quality of service for cellular users and
improve the network throughput. Kaufman et al. [13] proposed
a distributed spectrum sharing mechanism which includes
power control and route determination for the D2D users.
Wang et al. [14] proposed resource sharing through power
control for D2D users. In particular, their design goal was to
maximize the throughput of the D2D network while guaranteeing the quality of service of cellular users. Fodor et al.
[15] investigated the performance of D2D communications
with various LTE-based power control methods. Min et al.
[16] proposed an interference management scheme to improve
the reliability of D2D communications without decreasing the
transmission power of cellular user equipments. Min et al. [17]
proposed an interference limited area control scheme to manage
the interference from cellular connections to D2D connections.
Bao and Yu [18] used the partial location information to develop
an interference limited area in which the resources allocated to
cellular users cannot be reused by D2D users. In [19], under the
assumption that power control is ideal, the transmitting power
and SINR distributions of D2D networks in which cellular users
and D2D users use orthogonal resources were derived. In this
paper, we focus on binary power control and all active source
devices in a time slot have the same transmission power.
Recently, game theory has been applied to networks with
cellular users and D2D users. Xu et al. [20] proposed an
auction-game-based resource allocation algorithm for D2D
users to mitigate the interference between cellular users and
D2D users. Zhang et al. [21] proposed a distributed coalitiongame-based resource assignment for both cellular users and
D2D users assuming that both types of users have the same
priority. Wang et al. [22] modeled the power control on the D2D
users as an auction game to extend the battery life. Wang et al.
[23] developed a Stackelberg-game-based resource allocation

scheme to adjust the transmitting power of D2D users and

guarantee the fairness. In this paper, we study the case in which
all devices are cooperative rather than selfish. Game theory is
beyond the scope of this paper.
Various medium access control algorithms were proposed to
increase the network performance. Hakola et al. [24] proposed
means for selecting the optimal communication mode, when
both cellular mode and D2D mode are available. Seppala et al.
[25] introduced a concept of reliable multicast for D2D communications underlying cellular networks. Han et al. [26] designed an optimal channel reusing algorithm for the single
cell scenario based on the Hungarian algorithm. Zhou et al.
[27] proposed an intra-cluster D2D retransmission scheme with
optimal resource utilization. Pei and Liang [28] proposed a
novel spectrum sharing protocol, which allows D2D users to
assist the two-way communications between the base station
and cellular users.
Many previous works on D2D communications do not take
successive interference cancellation [29] into consideration. In
this paper, we study the case in which successive interference
cancellation is used. Recently, cooperative interference cancellation in the context of 3GPP D2D communications has been
proposed and discussed [30]. To the best of our knowledge,
our previous work [5] is the first in the literature that exploits
network information theory for D2D communications. While
it is possible to extend other previous works on D2D communications to benefit from network information theory, such
extensions are beyond the scope of the paper.
III. S YSTEM M ODELS AND P ROBLEM F ORMULATION
Consider a cell in a cellular network where there is a base
station (BS) in the cell. Let C be the total number of orthogonal channels that could be used for D2D communications
and W be the bandwidth of a channel. The C channels are
indexed by 1, 2, 3, . . . , C, respectively. Let N 2 be the total
number of devices in the cell. The devices are indexed by
1, 2, 3, . . . , N , respectively. A device is either a source device
or a destination device. Let Ns be the total number of source
devices and Nd = N Ns be the total number of destination
devices. Without loss of essential generality, it is assumed that
the source devices are indexed by 1, 2, 3, . . . , Ns , respectively.
Let S = {1, 2, . . . , Ns } be the set composed of the indexes
of source devices and D = {Ns + 1, Ns + 2, . . . , N } be the
set composed of the indexes of destination devices. A source
device always has data to send and is equipped with an omnidirectional antenna. We focus on unicast applications in which a
source device has only one destination. Let Di be the index of
destination device associated with source device i. According
to network information theory [4], two or more source devices
could successfully send different data to the same destination
device at the same time through the same frequency band as
long as the data rate vector is inside the capacity region. Let
Sj be the set composed of the indexes of source devices associated with destination device j. Define Sj = {i|Di = j, 1
i Ns }. Namely, Sj is composed of the indexes of source
devices that want to transmit data to destination device j. Note
that Ns Nd . Regarding the total number of cellular user

LIOU et al.: GROUP PARTITION AND DYNAMIC RATE ADAPTATION FOR D2D COMMUNICATIONS

equipments (UEs) in the network, we study two cases. In the

first case, there is no UE. It corresponds to the case when an
UE and a source device always use orthogonal channels. In the
second case, there are C UEs. In addition, it is assumed that UE
k uses channel k, 1 k C. While a cellular UE has to use
a BS to relay its data, a source device directly sends data to the
corresponding destination device.
The time domain is partitioned into time slots of equal length.
The length of a time slot is smaller than the channel coherence
time of the wireless channel. Let P k be the maximum transmission power of UE k. Let P k (t) be the transmission power
of UE k in time slot t. Let Rk (t) be the data transmission
rate of UE k in time slot t. Let hk0 (t) be the gain of channel
k from UE k to the BS in time slot t. Let hkj (t) be the gain
of channel k from UE k to destination device j in time slot t.
white
Let 2 be the power spectral densityof the background


P k hk (t)
.
Gaussian noise. Define Rk = 12 E W log2 1 + 20


k k
P h (t)
Rk , P k (t) = P k and Rk (t) =
If W log2 1 + 20
k

R . Otherwise, P (t) = 0 and R (t) = 0.

k
(t) be the gain of channel k from device i to device j
Let gi,j
k
in time slot t. Let gi,0
(t) be the gain of channel k from device i
to the BS in time slot t. Let Pi (t) be the transmission power of
device i in time slot t. Let Ri (t) be the data transmission rate of
device i in time slot t. Let Fi (t) be the index of the channel used
by source device i to transmit data in time slot t. Define S(t) =
{i|i S, Pi (t) > 0, Ri (t) > 0}. Then, S(t) corresponds to the
set of source devices that are selected by the base station to
transmit data in time slot t. Define S k (t) = {i|i S, Pi (t) >
0, Ri (t) > 0, Fi (t) = k}, k. Namely, S k (t) corresponds to
source devices that use channel k to transmit data in time
slot t. Define Sjk (t) = S k (t) Sj . Then, Sjk (t) corresponds to
source devices that use channel k to transmit data to destination
device j in time slot t. Define Ijk (t) = S k (t) Sjk (t). Ijk (t)
corresponds to source devices that use channel k in time slot
t but do not want to transmit data to destination device j in
time slot t. Define D(t) = iS(t) Di and Dk (t) = iS k (t) Di ,
k. If |Sjk (t)| 2, in time slot t, through a multiple access
channel [4], source devices with indexes in Sjk (t) use channel k
to transmit data to destination device j with interference from
source devices with indexes in Ijk (t) and UE k. According to
network information theory [4], for each source device with
index in S(t) to successfully transmit data to the corresponding
destination device in time slot t, the following constraints have
to be satisfied for each (k, j), where 1 k C and j Dk (t).
Ri (t)

W log2 1+

k
(t)
Pi (t)gi,j


,
k (t)
2 +P k (t)hkj (t)+ u:uI k (t) Pu (t)gu,j
j

i:iT

i Sjk (t),
Ri (t)


W log2 1+

k
i:iT Pi (t)gi,j (t)


,
k (t)
2 +P k (t)hkj (t)+ u:uI k (t) Pu (t)gu,j
j

T Sjk (t).

(1)

923

In addition, the BS should be able to receive/decode data

transmitted from UE k in time slot t. Therefore,


P k (t) hk0 (t)
k
. (2)
R (t) W log2 1+ 2 
k (t)
+ u:uS k (t) Pu (t)gu,0
Let P be the maximum transmission power of a device in
a time slot. In this paper, we study the case in which Pi (t)
{P, 0}, i, t. Transmission power control is beyond the scope
of this paper. Then, if i S(t), Pi (t) = P . Otherwise, Pi (t) =
0. In addition, it is assumed that at the beginning of time slot
t, with cooperation from UEs and devices, the base station
could know the values of all channel gains, which are obtained
through channel estimation.
At time slot t, given the values of channel gains, to optimize
the network throughput, the base station solves the following
integer-linear programming problem.
max

C
k=1

i:iS k (t)

Ri (t)

subject to
S k (t) {1, 2, . . . , Ns }, 1 k C,
S (t) S (t) = , 1 < C,


Dk (t) = S k (t) , 1 k C,


P k (t) hk0 (t)
k
R (t) W log2 1 + 2 
,
k (t)
+ u:uS k (t) Pu (t) gu,0
1 k C,
Ri (t)


k
(t)
Pi (t)gi,j

,
1+ 2
k (t)
+P k (t)hkj (t)+ u:uI k (t) Pu (t)gu,j

W log2

j Dk (t), i Sjk (t),

i:iT

Ri (t)


W log2 1+

k
i:iT Pi (t)gi,j (t)

2+P k (t)hkj (t)+

k
u:uIjk (t) Pu (t)gu,j (t)

j Dk (t), T Sjk (t),

Ri (t) 0, 1 i Ns ,
k
S(t) = C
k=1 S (t),

Ri (t) = 0, i
S(t).

(3)

Note that we have to find the optimal values of S k (t), 1

k C, and Ri (t), i S(t). The second constraint reflects
that a source device uses at most one channel in time slot t.
The third constraint means that Dk (t) is a function of S k (t).
The last constraint means that the data rate of source device i
must be zero if it does not use any channel.

924

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 2, FEBRUARY 2015

We now show that the optimization problem in (3) is an

integer-linear programming problem [31]. Define binary variables Bi,k,t s as follows. If source device i uses channel k to
transmit data in time slot t, Bi,k,t = 1. Otherwise, Bi,k,t = 0.
Then, Bi,k,t = 1 if and only if i S k (t). In addition, the first
two constraints in (3) are equivalent to the two
 constraints,
Bi,k,t {0, 1}, 1 i Ns , 1 k C and C
k=1 Bi,k,t
1, 1 i Ns . Furthermore, for each fixed collection of
Bi,k,t s, the optimization problem in (3) becomes a linear
programming problem [31]. Therefore, the above optimization
problem is an integer-linear programming problem.
Since an integer-linear programming problem is NP-hard in
general, we propose using the following three-stage approach
for scalable D2D communications. First, we create a conflict
graph to partition devices into subgroups. In particular, source
devices in subgroup k are candidate devices that could use
channel k to transmit data, k {1, 2, . . . , C}. We elaborate
on the first stage in Section IV. Second, for each channel,
we use a fast iterative algorithm to select active devices that
actually use the channel. Last, we use aggression levels for
fast rate adaptation based on channel state information. We
elaborate on the last two stages in Section V.

two source devices transmit data to the same destination device

through a multiple access channel. Otherwise, source device
i interferes with the reception at destination device Dj , while
source device j interferes with the reception at destination
device Di . Recall that log2 (x) + log2 (y) = log2 (xy). Then,
based on network information theory [4],




P gj,Dj (t)
P g
(t)
i

W log2 1 + 2 +Pi,D
1
+
,

2 +P g
g
(t)

(t)

j,Di
i,Dj

Ts (t) =

 if Di
= Dj ,

P (gi,Di (t)+gj,Dj (t))

, if Di = Dj .
W log 1 +
2
2

(5)
Since log2 (1 + x) is a concave function of x, based on
Jensens inequality [4], for an arbitrary random variable X,
E[log2 (1 + X)] log2 (1 + E[X]). Therefore,
 


1
P gi,Di (t)
E
W log2 1 +
2
2


P gj,Dj (t)
+ W log2 1 +
2





P mj,Dj
1
P mi,Di
W log2 1+
+W
log
1+
.
2
2
2
2

IV. PARTITIONING D EVICES BASED ON

C OLORING A C ONFLICT G RAPH
We [5] proposed an algorithm that partitions source devices
into subgroups based on graph theory. When the algorithm is
used, source devices in subgroup k become candidate devices
that could use channel k, 1 k C. For the completeness
of the paper and the convenience of the readers, we briefly
introduce the algorithm in this section.
The proposed algorithm is composed of two phases. In
the first phase, the algorithm creates a conflict graph and
determines the number of required subgroups based on graph
coloring. In the second phase, the algorithm partitions source
devices into subgroups.
In the first phase, to determine the total number of required
subgroups, a conflict graph G = (V, E), where V is the vertex
set and E is the edge set, is created. In particular, vertex i V
corresponds to source device i in the network. In addition,
(i, j) E if and only if source device i and source device j
should not use the same channel to transmit data at the same
time. Let mi,j be the mean channel gain from device i to
device j. Recall that Di is the index of the destination device
associated with source device i. Consider source device i and
source device j. Let To (t) be the throughput per channel in time
slot t when the two source devices use two orthogonal channels.
Then, based on information theory [4],



1
P gi,Di (t)

To (t) =
W log2 1 +
2
2


P gj,Dj (t)
+ W log2 1 +
. (4)
2
Let Ts (t) be the throughput per channel in time slot t when
the two source devices share the same channel. If Di = Dj , the

(6)
Define To the approximated average throughput per channel
in a time slot when the two source devices use two orthogonal
channels as follows.





P mj,Dj
1
P mi,Di
+W log2 1+
. (7)
To = W log2 1+
2
2
2
Similarly, define Ts the approximated average throughput per
channel in a time slot when the two source devices share the
same channel as follows.




P mj,Dj
P mi,Di

W
log
1
+
1
+

2
2 +P mj,Di
2 +P mi,Dj ,

Ts =

 if Di
= Dj , (8)

P (mi,Di +mj,Dj )

, if Di = Dj .
W log2 1 +
2

To reduce the computational complexity, we propose partitioning devices at time zero based on the value of (To , Ts ) rather
than partitioning devices every time slot based on the value of
(To (t), Ts (t)).
If To > Ts , it is better for source device i and source device
j to transmit data via orthogonal channels. In this case, source
device i and source device j are seen as two conflicting devices.
Therefore, i, j V and (i, j) E. Otherwise, (i, j)
E (and
i, j
V ). In [5], it has been proved that if Di = Dj , Ts > To .
Thus, it is more efficient for two source devices with the same
destination to share a channel rather than use two orthogonal
channels. In this paper, all sources to the same destination use
the same channel.
Let Nc be the total number of colors required for coloring
the vertices of G so that adjacent vertices are colored with
different colors. Given the conflict graph, we could use the
sequential coloring algorithm [32] to find out Nc . Fig. 1 shows

LIOU et al.: GROUP PARTITION AND DYNAMIC RATE ADAPTATION FOR D2D COMMUNICATIONS

Fig. 1.

high-priority vertices to low-priority vertices. Let (i) be the

index of the color to which source device i is assigned. Let
V0 be the set composed of the indexes of source devices that
cannot not transmit data. Let Q be the set composed of the
indexes of source devices that transmit data. If (i) > C, i
V0 . Otherwise, i Q. We propose using the above iterative
algorithm with Equation (9) to determine the index of the
subgroup to which source device i belongs, i Q.
We now take into consideration the source devices that do
not correspond to vertices in the conflict graph. Define V c =
{i|1 i Ns , i
V }. The source devices with indexes in V c
are first sorted in descending order according to their priorities.
Next, an iterative algorithm is used as follows. Consider source
device i. Let ei,k be the interference level from source device i
to destination devices in subgroup k. In particular, if |Vk | 1,
ei,k is set to maxj:jVk mi,Dj . Otherwise, ei,k is set to 0. To
minimize the maximum interference from source device i to
destination devices, source device i is assigned to subgroup
k0 (i), where

A conflict graph [5].

a conflict graph in which a circle represents a source device and

a cross represents a destination device. In addition, a dashed
line represents the association between a source device and
the corresponding destination device. The number adjacent
to a circle is the index of the corresponding source device.
A solid line represents an edge of the conflict graph. For
example, source device 6 and source device 7 cannot use the
same channel, since (6, 7) E. In the conflict graph, there are
7 vertices and 6 edges.
We now elaborate on the second phase. Pseudo codes for the
second phase can be found in [5] and are omitted due to limit of
space. The second phase is used to assign each source device a
subgroup. Recall that C is the total number of channels. Source
devices are assigned priorities based on the average throughput up to date. In particular, to improve fairness, the source
devices with smaller throughput up to date are assigned higher
priorities. Define (H) = iH Di , H {1, 2, 3, . . . , Ns }.
If Nc C, we use an iterative algorithm to sequentially assign
source devices to subgroups. Let Vk be the set composed of
the indexes of the source devices that currently belong to
subgroup k, 1 k C. Define Vki = Vk {i}. According to
the iterative algorithm, source device i is assigned to subgroup
k1 (i), where

k:1kC

j:j(Vk {i})


log2 1+

P
2 +P

j:j(Vk )

k0 (i) = arg min ei,k

k:1kC

= arg min

max mi,Dj .

k:1kC j:jVk

(10)

When the partition algorithm terminates, Vk consists of the

indexes of the candidate source devices that might use channel
k to transmit data. A source device with index i belongs to
subgroup k if and only if i Vk .
V. T HROUGHPUT O PTIMIZATION FOR A S UBGROUP
In this section, we focus on the problem of throughput
optimization for each subgroup. Recall that a source device
is associated with a single destination device. Consider a subgroup of source devices in a time slot. In order to optimize the
throughput for the subgroup in the time slot, it is necessary
to find the optimal set of active source devices and the corresponding optimal data rates. We first formulate the optimization
problem as an integer-linear programming problem and then
propose a fast approximation algorithm for solving it.
A. An Integer-Linear Programming Approach

k1 (i) = arg max W



925

log2 1+




P
2 +P

l:lVki Sj

l:lVki(Vki Sj )




ml,j

l:lVk Sj

ml,j


ml,j

l:lVk(Vk Sj )

ml,j

.
(9)

Note that in each step, the iterative algorithm assigns a source

device to the optimal subgroup that maximizes the throughput
increment.
If Nc > C, the following procedure is used. First, the sequential coloring algorithm is used to color vertices in V from

Consider subgroup k in time slot t. Recall that Vk is the set of

candidate source devices that might use channel k to transmit
data. Recall that S k (t) is the set of source devices that use
channel k to transmit data in time slot t. Then, S k (t) Vk . In
order to maximize the throughput of subgroup k in time slot t,
we have to solve the following optimization problem.

max
Ri (t)
k
i:iS (t)

subject to
S k (t) Vk ,


Dk (t) = S k (t) ,

R (t) W log2
k


P k (t) hk0 (t)
,
1+ 2 
k (t)
+ u:uS k (t) Pu (t) gu,0

926

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 2, FEBRUARY 2015

Ri (t)


k
(t)
Pi (t)gi,j

,
1+ 2
k (t)
+P k (t)hkj (t)+ u:uI k (t) Pu (t)gu,j

W log2

Algorithm 2: Selecting the active source devices that use

channel k to transmit data in time slot t

i:iT

j Dk (t), i Sjk (t),

Input: W , 2 , P , P k , Rk , hk0 (t), gi,j (t).

Output: P k (t), Rk (t), S k (t).
1: k (t) 


Ri (t)


W log2 1+

k
i:iT Pi (t)gi,j (t)

2 +P k (t)hkj (t)+

k
u:uIjk (t) Pu (t)gu,j (t)

P k hk (t)

j Dk (t), T Sjk (t),

Ri (t) 0, i S k (t).

(11)

The third constraint is used to ensure that UE k could successfully use channel k to transmit data to the BS in time slot t.
The above problem is an integer-linear programming problem,
which is NP-hard in general. In the worst case, there are 2|Vk |
feasible solutions for S k (t). For each fixed value of S k (t),
the above optimization problem becomes a linear programming
problem. In principle, a linear programming problem can be
solved by the simplex algorithm or the interior-point algorithms
[31]. However, for each fixed value of S k (t), the corresponding

k
linear programming problem has j:jDk (t) 2|Sj (t)| 1 constraints. Therefore, when the total number of source devices is
large, it is impractical to find an optimal solution for the integerlinear programming problem in (11).

Rk then
2: if W log2 1 + 20
3: // UE k transmits data in time slot t.
4: Set P k (t) = P k and Rk (t) = Rk .
5: else
6: Set P k (t) = 0 and Rk (t) = 0.
7: end if
8: repeat
9: Obtain (v , j ) by Equation (13) or Equation (14).
10: if v > k,t (k (t)) then
11:
Add j to k (t).
12: end if
13: until v k,t (k (t))
14: Set S k (t) = k (t).
15: if Rk (t) == Rk then



16:

Rk (t) W log2 1 +

2 +

P k hk
0 (t)
k (t)
P g,0
k

:S (t)

17: end if

B. A Fast Approximation Algorithm

Instead of finding an optimal solution for the above integerlinear programming problem, to reduce the computational
complexity, we propose a fast approximation algorithm. The
proposed approximation algorithm is composed of two phases.
In the first phase, the algorithm selects active source devices
that transmit data in the current time slot. There are two
variants for the first phase. The first variant is source-based,
while the second variant is destination-based. When the sourcebased approximation algorithm is used, in each round, the
algorithm selects an additional active source device. When
the destination-based approximation algorithm is used, in each
round, the algorithm chooses a destination device and all associated source devices become active source devices. In the
second phase, the algorithm determines the data rates of active
source devices. Pseudo codes for the main function of the
approximation algorithm are shown in Algorithm 1. Pseudo
codes for the first phase of the approximation algorithm are
shown in Algorithm 2 and pseudo codes for the second phase
of the approximation algorithm are shown in Algorithm 3.
Algorithm 1: The approximation algorithm (for subgroup k)
Input: W , 2 , P , P k , Rk , hk0 (t), hkj (t), gi,j (t).
Output: P k (t), Rk (t), S k (t), Ri (t), i S k (t).
1: Determine P k (t), Rk (t), S k (t) by Algorithm 2.
2: Calculate Ri (t), i S k (t) by Algorithm 3.

Algorithm 3: Rate adaptation based on channel state information for subgroup k

Input: W , 2 , P , hkj (t), gi,j (t), P k (t), S k (t).
Output: Ri (t), i S k (t).
1: // Calculate date transmission rate for source device i,
i S k (t).
2: for j (S k (t)) do
3: Calculate jk (t) by Equation (15).
4: for i Sjk (t) do
5:
Calculate Ri (t) by Equation (16).
6: end for
7: end for

Consider channel k in time slot t. We now elaborate on

the first phase, which is an iterative algorithm. We first set
the temporary
values of P k (t) and Rk (t) as follows. If

W log2 1 +

P k hk
0 (t)
2
k

Rk , P k (t) = P k and Rk (t) = Rk .

Otherwise, P (t) = 0 and Rk (t) = 0. Recall that S k (t) is the

set composed of the indexes of the source devices that actually
use channel k to transmit data in time slot t. Let k (t) be a set
that is used to determine the value of S k (t). Initially, k (t) =
. If the source-based approximation algorithm is used, define
j = {j}, j S. On the other hand, if the destination-based
approximation algorithm is used, define j = Sj , j D. Let
k,t (X) be the throughput of channel k in time slot t, when the

LIOU et al.: GROUP PARTITION AND DYNAMIC RATE ADAPTATION FOR D2D COMMUNICATIONS

set of active source devices that use channel k to transmit data

in time slot t is equal to X. In particular,
k,t (X)

=W

log2


1+

i:i(X)

:XSi

2 +

k
P g,i
(t)

:XSi

k (t)
P g,i

(12)
When the source-based approximation algorithm is used,
the iterative algorithm first solves the following optimization
problem.


max k,t k (t) {j}

927

data to destination device j in time slot t. The proposed

algorithm obtains the value of Ri (t), i Sjk (t), as follows.
Recall that Ijk (t) constitutes of the indexes of the source devices
that use channel k to transmit data to destination devices other
than destination device j in time slot t. Define jk (t) = P k (t)

k
hkj (t) + :I k (t) P g,j
(t). Then, jk (t) equals the power
j
of the interference at destination device j through channel k in
time slot t.
Let jk (t) be the aggression level of the source devices that
use channel k to transmit data to destination device j in time
slot t. In particular,



k
jk (t) = 

subject to

iSjk (t)

j Vk (t),

P gi,j (t)
iS k (t)
j
2 +jk (t)

1+

log2


log2 1 +

k (t)
P gi,j

.

(15)

2 +jk (t)

Rk (t) W log2

1+

P (t)
k

hk0 (t)

:k (t){j}


.

k (t)
P g,0

(13)
Note that the last constraint is used to assure that UE k could
successfully transmit data to the base station with rate Rk (t).
Let v be the optimal value and j be an optimal solution for
the above optimization problem. When the above optimization
problem is infeasible, v = . If v k,t (k (t)), the first
phase terminates. Otherwise, the integer j is added into the set
k (t) and the iterative algorithm continues.
Similarly, when the destination-based approximation algorithm is used, the iterative algorithm first solves the following
optimization problem.


max k,t k (t) j
subject to


j Vk k (t) ,

Rk (t) W log2

P k (t) hk0 (t)

1+ 2 
k (t)
+ :k (t)j P g,0


.

Let v be the optimal value and j be an optimal solution for

the above optimization problem. If v k,t (k (t)), the first
phase terminates. Otherwise, all elements of the set j are
added into the set k (t) and the iterative algorithm continues.
When the first phase terminates, S k (t) is set to be
the current value of k (t). In addition, if the temporary
k
value of Rk (t) is Rk , the final value
 of R (t) is set to
2 +

P k hk
0 (t)
k (t)
P g,0
k

Then,

Ri (t) = jk (t)W

i:iSjk (t)

log2

1+

iSjk (t)


= W log2

1+

iSjk (t)

k
(t)
P gi,j

2 + jk (t)

k
P gi,j
(t)

2 + jk (t)


.

(17)

Namely, the sum rate for source devices that transmit data to
destination device j through channel k in time slot t reaches the
upper bound of the sum rate for a multiple access channel [4].
However, (17) does not imply (1) in general.
VI. P ROPERTIES OF THE P ROPOSED A LGORITHM

(14)

W log2 1 +

In time slot t, for each i Sjk (t), the proposed algorithm

chooses the value of Ri (t) as follows.


k
(t)
P

g
i,j
Ri (t) = jk (t)W log2 1 + 2
.
(16)
+ jk (t)

in order to increase the

:S (t)

throughput of UE k. Note that the same algorithm is used for

each channel in each time slot.
We now introduce the second phase. Consider destination
device j, where j (S k (t)). Recall that Sjk (t) is composed
of the indexes of source devices that use channel k to transmit

We show analytical results in this section. We first prove that

when two source devices simultaneously use the same channel
to transmit data to the same destination device according to (15)
and (16), both source devices succeed for sure.
Theorem 1: If |Sjk (t)| = 2 and the values of Ri (t)s, i
k
Sj (t), are determined by (15) and (16), both source devices
with indexes in Sjk (t) successfully use channel k to transmit
data to destination device j in time slot t.
Proof: See Appendix.
The above theorem means that if a destination device has two
sources in a time slot, the proposed aggression-level-based approach maximizes the sum rate with a very low computational
complexity.
We now study the case in which not all channel gains are
known to the base station. Consider source devices that transmit
data to destination device j through channel k in time slot t.
Let i be an integer in Sjk (t). Let mki,j be the mean channel
gain of channel k from source device i to destination device j.

928

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 2, FEBRUARY 2015

k
Suppose that the base station knows the values of g,j
(t)s,
k
k
Sj (t) {i} but does not know the value of gi,j . In this
k
case, based on (15), jk (t) is approximated by
i,j
(t), which is
defined as follows.
k

i,j
(t)


= log2

1+

log2

P mki,j +

:Sjk (t){i}

k
P g,j
(t)

2 + jk (t)


1+

P mki,j
2 + jk (t)


log2

1+

Sjk (t){i}

k
(t)
P g,j
2 + jk (t)

 1

(18)

Note that the above equation is very similar to (15) except that
k
gi,j
(t) is replaced by mki,j .
k
k
Define f (gi,j
(t)) = jk (t)
i,j
(t). Note that jk (t) dek
k
k
pends on gi,j (t) but
i,j (t) is independent of gi,j
(t). If
k
f (gi,j (t)) < 0, destination device j fails to receive any data
k
from source devices with indexes in Sjk (t). If f (gi,j
(t)) 0
k
and |Sj (t)| = 2, based on Theorem 1, two source devices
with indexes in Sjk (t) successfully transmit data to destination
device j through channel k in time slot t.
Consider the case in which Sjk (t) = {i+ , i }, the value of
k
gi+ ,j (t) is known, and the value of gik ,j (t) is unknown. To
analyze the function f in this case, for each fixed positive real
number a, define the function a (b) as follows.
a (b) =
Define + =
s(gik ,j (t)) =
gik ,j (t).

ln(1 + a + b)
, b [0, ).
ln(1 + a + b + ab)

P gik

+ ,j

(t)

2 +jk (t)

P gik ,j (t)

2 +jk (t)

(19)

. Define a function s such that

. Then, s is an increasing function of

Based on (15),

 

jk (t) = + s gik ,j (t) .

(20)

In addition, based on (18),

 


ik ,j (t) = + s mki ,j .

(21)

Thus,


f gik ,j (t) = jk (t)
ik ,j (t)

 
 

= + s gik ,j (t) + s mki ,j .
(22)
It was conjectured that if gik ,j (t) > mki ,j , f (gik ,j (t)) > 0,
since the unknown channel gain is underestimated. It is clear
that f (mki ,j ) = 0. Then, it is conjectured that the function f

has only one positive root. However, the conjectures are incorrect. It is possible that the function f have two positive roots and
f (gik ,j (t)) < 0 even when gik ,j (t) > mki ,j . The following
theorem contains a sufficient condition for the function f to
have two positive roots.
Theorem 2: If Sjk (t) = {i+ , i }, the value of gik+ ,j (t) is
known, the value of gik ,j (t) is unknown, and + (mki ,j ) >
minx:x0 + (x), then there exists a positive real number b
(0, ) such that (b) = 0 and the function f has two or more
positive roots.
Proof: See Appendix.
When the condition of Theorem 2 is satisfied, f (x) is not a
monotonic function of x.
We now analyze the overall computational complexity of the
proposed approach. Consider a graph G = (V, E). Let deg(v)
be the degree of vertex v, v V . For each vertex v, it takes
deg(v) steps to collect the colors that have been used by its
neighbors and at most |V | more stepsto find an appropriate
color. According to graph theory [32], vV deg(v) = 2|E|
|V |(|V |1)
< |V |2 . Recall that N is the total number of devices
2
in the network. Thus, there are at most N vertices in the
conflict graph and the computational complexity of the group
partition based on the sequential coloring algorithm is at most
O(N 2 + N N ) = O(N 2 ). The computational complexity of
the first phase of the fast approximation algorithm is upper
bounded by O(N 3 ). The computational complexity of the second phase of the fast approximation algorithm is upper bounded
by O(N 2 ). Therefore, the overall computational complexity is
upper bounded by O(N 3 ).
VII. S IMULATION S ETUP AND R ESULTS
In this section, we show simulation setup and results. To
obtain discrete event simulation [33] results, we wrote C++
programs that could call MATLAB for solving linear programming problems. The service region is a 20 by 20 square. We
randomly create 10 networks. In a network, the BS is deployed
at the center of the service region. Regarding the total number
of UEs in the network, we study two cases. In the first case,
there are no UEs in the network. The case corresponds to the
case in which an UE and a device use orthogonal channels. In
the second case, there are C UEs in the network and different
UEs use distinct channels. Each UE is randomly deployed 3
units away from the BS. Source devices are randomly deployed
in the service region. A source device is associated with a
destination device. There are two classes of destination devices.
A destination device in the first class is associated with a
source device, while a destination device in the second class
is associated with two or more source devices. The distance
between a source device and the corresponding destination
device is a random variable that is uniformly distributed in
[0, 3]. In this section, P = 1, 2 = 0.01, and W = 10 (MHz).
Unless explicitly stated, a communication link between a source
device and a destination device is a Rayleigh fading channel.
In addition, for each link, channel gains at distinct time slots
are independent and identically distributed random variables.
The mean channel gain of each link is the reciprocal of the

LIOU et al.: GROUP PARTITION AND DYNAMIC RATE ADAPTATION FOR D2D COMMUNICATIONS

929

Fig. 2. The performance of the global optimal algorithm and the proposed
approximation algorithm in small networks.

Fig. 3. The performance of the local optimal algorithm and the proposed
approximation algorithm in small networks.

squared distance between the source device and the destination

device. In addition, C = 5. To obtain the network throughput,
a simulation instance contains 10 000 time slots. In this section,
we show normalized throughput in units of bps/Hz.
We compare the performance of five algorithms: the global
optimal algorithm, the local optimal algorithm [5], the greedy
algorithm, the proposed source-based approximation algorithm,
and the proposed destination-based approximation algorithm.
The global optimal algorithm solves the integer-linear programming problem in (3) to maximize the overall throughput of
the network. The local optimal algorithm uses the proposed
graph coloring approach to partition devices into subgroups and
solves the integer-linear programming problem in (11) to maximize the throughput of each subgroup. Both the global optimal
algorithm and the local optimal algorithm are feasible only in
very small networks, since a general integer programming problem is NP-hard. When the greedy algorithm is used, in a time
slot, only the D2D source devices with the largest C channel
gains might transmit data. In a time slot, the active set contains
all active source devices in the time slot. When the sourcebased approximation algorithm is used, in each round, a source
device is added into the active set. When the destination-based
approximation algorithm is used, in each round, a destination
device is selected and the corresponding source device is added
into the active set. None of the five algorithms are aware of the
locations of devices.
In Fig. 2, we show the performance of the proposed approximation algorithm and the global optimal algorithm, when there
are two orthogonal channels, 4 source devices and 4 destination
devices. In this case, the source-based approximation algorithm
is identical to the destination-based approximation algorithm.
When there is no user equipment in the network, the network
throughput of the proposed approximation algorithm is at least
93% of that of the global optimal algorithm. When there are
user equipments in the network, the network throughput of the
proposed approximation algorithm is at least 82% of that of the
global optimal algorithm.

In Fig. 3, we show the network throughput of the local optimal algorithm, the greedy algorithm, and the sourcebased approximation algorithm. The network throughput of the
destination-based approximation algorithm is almost identical
to that of the source-based approximation algorithm and is
not shown due to limit of space. We study the case when the
network contains 13 source devices and 10 destination devices.
Seven destination devices belong to the first class, while three
destination devices belong to the second class. A destination
device in the first class is associated with one source device,
while a destination device in the second class is associated
with two source devices. The network throughput of the sourcebased approximation algorithm is very close to that of the local
optimal algorithm. When there are UEs in the network, on average, the network throughput of the approximation algorithm
could be 20% larger than that of the greedy algorithm.
We also study the case when the total number of source devices in the network is either 25 or 50. Five destination devices
belong to the second class. In addition, a destination device in
second class is associated with three source devices. In Fig. 4,
we show the network throughput for the two approximation
algorithms and the greedy algorithm, when there is no UE.
Regardless of the total number of source devices in the network, the approximation algorithms are superior to the greedy
algorithm. The throughput of the source-based approximation
algorithm is almost identical to that of the destination-based approximation algorithm. Furthermore, when there are 25 source
devices, the throughput of the source-based approximation
algorithm could be 1.77 times larger than that of the greedy
algorithm. When there are 50 source devices, the throughput of
the source-based approximation algorithm could be 2.42 times
larger than that of the greedy algorithm.
In Fig. 5, we show the network throughput for the approximation algorithms and the greedy algorithm, when there are five
UEs and the total number of source devices is either 25 or 50.
The throughput difference between the two approximation algorithms is negligible. When there are 25 source devices, the

930

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 2, FEBRUARY 2015

Fig. 4. Throughput comparison, when there is no UE.

Fig. 5. Throughput comparison, when there are UEs.

throughput of the approximation algorithms is about 1.53 times

larger than that of the greedy algorithm. When there are 50
source devices, the throughput of the approximation algorithms
is about 2.03 times larger than that of the greedy algorithm.
When there are UEs, a source device should not interfere with
the UE that uses the same channel. Thus, a source device may
not be able to transmit data even if its instantaneous channel
gain is quite large. Therefore, the throughput when there are
UEs is smaller than that when there is no UE.
We also evaluate the performance of the approximation
algorithm when the base station is unaware of some channel
gains. Recall that mki,j is the mean channel gain of channel k
from source device i to destination device j. If the base station
k
(t), where i Sjk (t), it uses
does not know the value of gi,j
k
mi,j instead to determine the aggression level and the value of
Ri (t). We study the case in which there are 15 source devices
and 1 destination device. In Fig. 6, the variable in the x-axis
is the total number of known channel gains at the base station.

Fig. 6. The network throughput, when some of the channel gains are
unknown.

Fig. 7. Performance comparison of the graph coloring approach and the

modulo approach.

Note that in the simulation, the expected value of a Rayleigh

random variable is the same as that of a Rician random variable.
When all communication links are Rayleigh fading channels,
as the total number of known channel gains increases from
2 to 8, the network throughput increases as expected. Similarly,
when all communication links are Rician fading channels, as
the number of known channel gains increases from 2 to 8, the
network throughput increases. On the other hand, as the number
of known channel gains increases from 10 to 14, the network
throughput decreases irrespective of the fading channel model.
In Fig. 7, we compare the performance of two group partition approaches: the proposed graph coloring approach and
the modulo approach. According to the modulo approach, the
source device with identification number x can only use channel
y = x mod C. We study the case in which C = 5, Ns = Nd =
25, and devices form 5 clusters. In particular, in the service
region, there are 5 circles with radius 3 centered at (0, 0), (7, 7),

LIOU et al.: GROUP PARTITION AND DYNAMIC RATE ADAPTATION FOR D2D COMMUNICATIONS

(7, 7), (7, 7), and (7, 7), respectively. Destination device
i is randomly deployed in circle i mod C. In addition, each
source device associated with destination device i is randomly
deployed in circle i mod C. When each destination device is
associated with a unique source device, on average, the network
throughput of the proposed graph coloring approach is 58.79%
larger than that of the modulo approach. When each of the first
5 destination devices has 3 sources and each of the remaining
destination devices has only one source, on average, the network throughput of the proposed graph coloring approach is
10.87% larger than that of the modulo approach. We also study
the performance of the two group partition when source devices
are randomly deployed according to the first paragraph of the
section. When there are 50 source devices in the network, on
average, the network throughput of the graph coloring approach
is 7.9% larger than that of the modulo approach. Due to the limit
of space, the corresponding figure is not included.
In terms of network throughput, partitioning based on both
network topology and instantaneous channel gains is expected
to outperform the graph coloring approach in which the locations of devices are not known/used. On the other hand,
the proposed graph coloring approach does not require any
localization techniques and does not have to collect information
on the locations of devices.
VIII. D ISCUSSIONS
We now outline an implementation of the proposed approach
in cellular networks. When Orthogonal Frequency Division
Multiplexing (OFDM) is used at the physical layer, a channel
is composed of adjacent subcarriers. In addition, a subcarrier
belongs to only one channel. A time slot is composed of two
parts: the control sub-slot and the data sub-slot. The control
sub-slot consists of four phases. Consider time slot t. In the
first phase of the control sub-slot, there are Ns time intervals
for channel estimation. At the ith channel estimation time
interval, source device i broadcasts pilot signals for channel
estimation through C orthogonal channels and each destination
1
2
C
(t), gi,j
(t), . . . , gi,j
(t)). It is
device j estimates the value of (gi,j
assumed that channel estimation is perfect in the paper. The impacts of channel estimation errors on the network performance
are beyond the scope of this paper. More details on channel
estimation in MIMO-OFDM wireless communication systems
can be found in [34]. In the second phase of the control sub-slot,
there are Nd time intervals for feedback. At the jth feedback
k
(t)s
time interval, destination device j sends the values of gi,j
to the base station. Although it is possible to use an appropriate
random multiple access scheme instead in the second phase,
random multiple access is beyond the scope of this paper. The
base station could collect the values of hkj (t)s through a similar
procedure. In the third phase of the control sub-slot, to optimize
the network throughput (at time slot t), the base station uses
the proposed approach to approximately solve the optimization
problem in (3). In the fourth phase of the control sub-slot,
the base station broadcasts its decision. In particular, the base
station sends the value of (Pi (t), Ri (t)) to source device i. In
the data sub-slot, each source device i transmits data with power
Pi (t) and rate Ri (t).

931

Fig. 8. Network throughput when there are long-lived and short-lived D2D
sessions.

We now discuss how to extend the proposed approach to

the case in which there are short-lived D2D communication
sessions in addition to Ns long-lived D2D communication
sessions. A long-lived D2D communication session always has
data to send from the source device to the destination device. The arrivals of short-lived D2D communication sessions
are modeled as a Poisson process with rate . The lifetimes
of short-lived D2D communication sessions are modeled as
independent and identically distributed exponential random
variables with mean 1 . Among the C channels, Cs channels
are reserved for short-lived D2D sessions, while C Cs channels are used by long-lived D2D sessions. When a short-lived
session arrives/starts, the associated source device randomly
chooses one of the Cs channels to use till the end of the session.
According to M/M/ queueing theory [35], the expected
value of the total number of short-lived sessions is equal to .
x
Let x be a real number such that /
Ns = Cx . We propose using
a loading-aware algorithm in which Cs = x. In Fig. 8, we
show the performance of the proposed loading-aware algorithm
and the baseline algorithm in which Cs =  C2 , when C = 6.
When x = 1, the proposed loading-aware algorithm always
outperforms the baseline algorithm. The average performance
improvement is 13%. When x = 2, on average, the network
throughput of the proposed loading-aware algorithm is 3.6%
larger than that of the baseline algorithm. When x = 3, the
network throughput of the loading-aware algorithm is identical
to that of the baseline algorithm.

IX. C ONCLUSION
In this paper, we have proposed using group partition and
dynamic rate adaptation for scalable throughput optimization of
capacity-region-aware device-to-device communications. We
have adopted network information theory that allows a receiving device to simultaneously decode distinct packets from multiple transmitting devices, as long as the vector of transmitting

932

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 2, FEBRUARY 2015

rates is inside the capacity region. We have proposed creating

a conflict graph to partition devices into subgroups based on
graph theory. To optimize the throughput of a subgroup, instead
of directly solving an integer-linear programming problem,
we have proposed an iterative algorithm to quickly select
the set of active devices and using aggression levels for fast
rate adaptation based on channel state information. Simulation
results show that the proposed approach is scalable and could
significantly outperform the greedy algorithm when it is infeasible to obtain an optimal solution. Future work includes
further improving the network throughput when the arrivals of
D2D communication sessions form a renewal process and the
lifetime of a D2D communication session is a general random
variable based on queueing theory.

Since all variables in the right-hand side of the above

equation are positive, jk (t) > 0.
2) Note that

k
P gm,j
(t)
2 +jk (t)

> 0 and

k
P gn,j
(t)
2 +jk (t)

> 0. In addition,

1 + a + b < (1 + a)(1 + b), a, b > 0. Furthermore,

since log2 (x) is an increasing function of
x,
log2 (1 + a +b) < log2 ((1 + a)(1 +b)),
a, b > 0,

log2 1 +

P g k

(t)+P g k (t)

m,j
n,j
<
log2 1 +
2 +jk (t)



k
k
P gm,j (t)
P gn,j (t)
1 + 2 +
and therefore
k (t)
2 + k (t)

Thus,
j

jk (t) < 1.
3) Since
0 < jk (t) < 1,
i Sjk (t),
Ri (t) =




k
k
P gi,j (t)
P gi,j
(t)
k
j (t) log2 1 + 2 +k (t) < log2 1 + 2 +k (t) .
j

4) Based on the definition of jk (t),

A PPENDIX
For the convenience of readers, we list key variables as
follows.
hk0 (t)
the channel gain from cellular UE k to the base
station in time slot t.
k
(t)
the gain of channel k from source device i to the
gi,0
base station in time slot t.
the channel gain from cellular UE k to destinahkj (t)
tion device j in time slot t.
k
(t)
the gain of channel k from source device i to
gi,j
destination device j in time slot t.
the transmission power of cellular user equipP k (t)
ment k in time slot t.
the transmission power of source device i in
Pi (t)
time slot t.
the data transmission rate of cellular user equipRk (t)
ment k in time slot t.
the data transmission rate of source device i in
Ri (t)
time slot t.
S k (t)/Sjk (t) the set composed of the indexes of source devices that use channel k to transmit data (to
destination device j) in time slot t.
the set composed of the indexes of destination
Dk (t)
devices that use channel k to receive data in time
slot t.
the set composed of the indexes of source deIjk (t)
vices that use channel k but do not want to
transmit data to destination device j in time
slot t.
Theorem 1: If |Sjk (t)| = 2 and the values of Ri (t)s, i
k
Sj (t), are determined by (15) and (16), both source devices
with indexes in Sjk (t) successfully use channel k to transmit
data to destination device j in time slot t.
Proof:
1) Since |Sjk (t)| = 2, without loss of essential generality, it
is assumed that Sjk (t) = {m, n}. Then, according to (15),

jk (t) =

k
k
P gm,j
(t)+P gn,j
(t)
k
2
+j (t)
k
k (t)
P gm,j
(t)
P gn,j
2 +jk (t)
2 +jk (t)

log2 1 +

log2 1 +



1+

 .

Ri (t)

= jk (t)

i:iSjk (t)


log2

1+

i:iSjk (t)


= log2

1+

i:iSjk (t)

k
(t)
P gi,j

2 + jk (t)

k
P gi,j
(t)

2 + jk (t)


.

5) Based on 3 and 4, for channel k in time slot t,

(Rm (t), Rn (t)) is inside the capacity region of the multiple access channel from source device m and source
device n to destination device j. Therefore, according
to network information theory, all source devices with
indexes in Sjk (t) successfully use channel k to transmit
data to destination device j in time slot t.
Theorem 2: If Sjk (t) = {i+ , i }, the value of gik+ ,j (t) is
known, the value of gik ,j (t) is unknown, and + (mki ,j ) >
minx:x0 + (x), then there exists a positive real number b
(0, ) such that (b) = 0 and the function f has two or more
positive roots.
Proof:
1) Based on the proof of Theorem 1, a (b) (0, 1], a,
b > 0. In addition, a (0) = 1, a > 0.
2) Consider a fixed value of a.
lim a (b)

ln(1 + a + b)
b ln(1 + a + b + ab)

= lim

= lim

ln(1 + a + b)
ln(1 + a) + ln(1 + b)

ln(1 + a)
ln(1 + b)
= lim
+ lim
b ln(1 + a + b)
b ln(1 + a + b)

1

= 1.
3) a (b) = {(1+b) ln(1 + a)[(1 + a + b) ln(1 + a + b)
(1 + b) ln(1 + b)]} {(1 + b)(1 + a + b) ln2 ((1 +
a)(1 + b))}1 .

LIOU et al.: GROUP PARTITION AND DYNAMIC RATE ADAPTATION FOR D2D COMMUNICATIONS

1
4) Define c = (1+b)(1+a+b) ln
2 ((1+a)(1+b)) . Then, c > 0.
Define 1 (b) = (1 + b) ln(1 + a). In addition, define
2 (b) = (1 + a + b) ln(1 + a + b) (1 + b) ln(1 + b).
Furthermore, define (b) = 1 (b) 2 (b). Then, a (b) =
c [1 (b)2 (b)] = c(b). In addition, 1 (0) = ln(1 +
a) and 2 (0) = (1 + a) ln(1 + a). Thus, 1 (0)2 (0) <
0 and a (0) <0.
5) Define = ln(1 + a) 1. Then, > 0. Since
limb 1+a+b
1+b = 1, there exists b1 (0, ) such that
1+a+b
1+b

ln(1+a+b)
ln(1+b) = 1,
ln(1+a+b)
ln(1+b) < 1 + ,

< 1 + , b b1 . Since limb

there exists b2 (0, ) such that

b b2 . Define b+ = max(b1 , b2 ). Then, 0 < b+ < .
In addition,
1 (b+ ) 2 (b+ )
= (1 + b+ ) ln(1 + a) (1 + a + b+ ) ln(1 + a + b+ )
+ (1 + b+ ) ln(1 + b+ )
= (1 + b+ ) ln(1 + a + b+ + ab+ )
(1 + a + b+ ) ln(1 + a + b+ ).
Furthermore,
(1 + b+ ) ln(1 + a + b+ + ab+ )
(1 + a + b+ ) ln(1 + a + b+ )
=

1 + b+
ln(1 + b+ )
ln(1 + a)

1 + a + b+
ln(1 + a + b+ )

> (1 + )1 (1 + )1 (1 + )2
= 1.
Thus, (b+ ) = 1 (b+ ) 2 (b+ ) > 0. Since (0) <
0, (b+ ) > 0, and (b) is a continuous function of b,
based on the intermediate value theorem for continuous
functions, there exists a positive real number b (0, b+ )
such that (b) = 0.
6) Recall
that
f (gik ,j (t)) = + (s(gik ,j (t)))
+ (s(mki ,j )). Since + (b) is a continuous function
of b, + (b) (0, 1], b [0, ), + (0) = 1 =
limb + (b) and there exists b (0, ) such
that + (b ) < 1, minx:x0 + (x) exists. Define
b = arg minx:x0 + (x). Note that b (0, ).
since
+ (b ) < + (mki ,j ) <
If
mki ,j < b ,
limb + (b), based on the intermediate value theorem
for continuous functions, there exists a positive real
number r (b , ) such that + (r) = + (mki ,j )
and therefore f (r) = f (mki ,j ) = 0. On the other hand,
if mki ,j > b , since + (b ) < + (mki ,j ) < + (0),
there exists a positive real number r (0, b ) such that
+ (r) = + (mki ,j ) and therefore f (r) = 0. Thus,
both mki ,j and r are roots of the function f . Namely, the
function f has two or more positive roots.
7) Based on 5 and 6, we have completed the proof.

933

R EFERENCES
[1] X. Lin, J. G. Andrews, A. Ghosh, and R. Ratasuk, An overview of 3GPP
device-to-device proximity services, IEEE Commun. Mag., vol. 52, no. 4,
pp. 4048, May 2014.
[2] D. Feng et al., Device-to-device communications in cellular networks,
IEEE Commun. Mag., vol. 52, no. 4, pp. 4955, May 2014.
[3] M. N. Tehrani, M. Uysal, and H. Yanikomeroglu, Device-to-device communication in 5G cellular networks: Challenges, solutions, future directions, IEEE Commun. Mag., vol. 52, no. 5, pp. 8692, May 2014.
[4] T. M. Cover and J. A. Thomas, Elements of Information Theory, 2nd ed.
Hoboken, NJ, USA: Wiley, 2006.
[5] Y.-S. Liou, R.-H. Gau, and C.-J. Chang, Group partition for capacityregion-aware device-to-device communications, in Proc. IEEE WCNC,
Apr. 2014, pp. 17081713.
[6] C.-H. Yu, K. Doppler, C. B. Ribeiro, and O. Tirkkonen, Resource sharing optimization for device-to-device communication underlaying cellular
networks, IEEE Trans. Wireless Commun., vol. 10, no. 8, pp. 27522763,
Aug. 2011.
[7] L. B. Le, Fair resource allocation for device-to-device communications
in wireless cellular networks, in Proc. IEEE GLOBECOM, Dec. 2012,
pp. 54515456.
[8] D. Feng et al., Device-to-device communications underlaying cellular networks, IEEE Trans. Commun., vol. 61, no. 8, pp. 35413551,
Aug. 2013.
[9] L. Lei, Y. Zhang, X. Shen, C. Lin, and Z. Zhong, Performance analysis of device-to-device communications with dynamic interference using
stochastic Petri nets, IEEE Trans. Wireless Commun., vol. 12, no. 12,
pp. 61216141, Dec. 2013.
[10] Y. Chai, Q. Du, and P. Ren, Partial timefrequency resource allocation
for device-to-device communications underlaying, in Proc. IEEE ICC,
Jun. 2013, pp. 60556059.
[11] D. H. Lee, K. W. Choi, W. S. Jeon, and D. G. Jeong, Resource allocation scheme for device-to-device communication for maximizing spatial
reuse, in Proc. IEEE WCNC, Apr. 2013, pp. 112117.
[12] R. Zhang, X. Cheng, L. Yang, and B. Jiao, Interference-aware graph
based resource sharing for device-to-device communications underlaying
cellular networks, in Proc. IEEE WCNC, Apr. 2013, pp. 140145.
[13] B. Kaufman, J. Lilleberg, and B. Aazhang, Spectrum sharing scheme
between cellular users and ad-hoc device-to-device users, IEEE Trans.
Wireless Commun., vol. 12, no. 3, pp. 10381049, Mar. 2013.
[14] J. Wang, D. Zhu, C. Zhao, J. C. F. Li, and M. Lei, Resource sharing of
underlaying device-to-device and uplink Cellular communications, IEEE
Commun. Lett., vol. 17, no. 6, pp. 11481151, Jun. 2013.
[15] G. Fodor, D. D. Penda, M. Belleschi, M. Johansson, and A. Abrardo,
A comparative study of power control approaches for device-to-device
communications, in Proc. IEEE ICC, Jun. 2013, pp. 60086013.
[16] H. Min, W. Seo, J. Lee, S. Park, and D. Hong, Reliability improvement
using receive mode selection in the device-to-device uplink period underlaying cellular networks, IEEE Trans. Wireless Commun., vol. 10, no. 2,
pp. 413418, Feb. 2011.
[17] H. Min, J. Lee, S. Park, and D. Hong, Capacity enhancement using an
interference limited area for device-to-device uplink underlaying cellular
networks, IEEE Trans. Wireless Commun., vol. 10, no. 12, pp. 3995
4000, Dec. 2011.
[18] P. Bao and G. Yu, An interference management strategy for device-todevice underlaying cellular networks with partial location information,
in Proc. IEEE PIMRC, Sep. 2012, pp. 465470.
[19] M. C. Erturk, S. Mukherjee, H. Ishii, and H. Arslan, Distributions of
transmit power and SINR in device-to-device networks, IEEE Commun.
Lett., vol. 17, no. 2, pp. 273276, Feb. 2013.
[20] C. Xu et al., Efficiency resource allocation for device-to-device underlay
communication systems: A reverse iterative combinatorial auction based
approach, IEEE J. Sel. Areas Commun., vol. 31, no. 9, pp. 348358,
Sep. 2013.
[21] R. Zhang, L. Song, Z. Han, X. Cheng, and B. Jiao, Distributed resource
allocation for device-to-device communications underlaying cellular networks, in Proc. IEEE ICC, Jun. 2013, pp. 18891893.
[22] F. Wang et al., Energy-aware resource allocation for device-to-device
underlay communication, in Proc. IEEE ICC, Jun. 2013, pp. 60766080.
[23] F. Wang, L. Song, Z. Han, Q. Zhao, and X. Wang, Joint scheduling
and resource allocation for device-to-device underlay communication,
in Proc. IEEE WCNC, Apr. 2013, pp. 134139.
[24] S. Hakola, T. Chen, J. Lehtomaki, and T. Koskela, Device-to-device
(D2D) communication in cellular networkPerformance analysis of optimum and practical communication mode selection, in Proc. IEEE
WCNC, Apr. 2010, pp. 16.

934

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 14, NO. 2, FEBRUARY 2015

[25] J. Seppala, T. Koskela, T. Chen, and S. Hakola, Network controlled

device-to-device (D2D) and cluster multicast concept for LTE and LTE-A
networks, in Proc. IEEE WCNC, Mar. 2011, pp. 986991.
[26] T. Han, R. Yin, Y. Xu, and G. Yu, Uplink channel reusing selection
optimization for device-to-device communication underlaying cellular
networks, in Proc. IEEE PIMRC, Sep. 2012, pp. 559564.
[27] B. Zhou, H. Hu, S.-Q. Huang, and H.-H. Chen, Intracluster device-todevice relay algorithm with optimal resource utilization, IEEE Trans.
Veh. Technol., vol. 62, no. 5, pp. 23152326, Jun. 2013.
[28] Y. Pei and Y.-C. Liang, Resource allocation for device-to-device communications overlaying two-way cellular networks, IEEE Trans. Wireless
Commun., vol. 12, no. 7, pp. 36113621, Jul. 2013.
[29] D. Tse and P. Viswanath, Fundamentals of Wireless Communications.
Cambridge, U.K.: Cambridge Univ. Press, 2005.
[30] R. Tanbourgi, H. Jakel, and F. K. Jondral, Cooperative interference cancellation using device-to-device communications, IEEE Commun. Mag.,
vol. 52, no. 6, pp. 118124, Jun. 2014.
[31] D. G. Luenberger and Y. Ye, Linear and Nonlinear Programming, 3rd ed.
New York, NY, USA: Springer-Verlag, 2008.
[32] G. Chartrand and O. R. Oellermann, Applied and Algorithmic Graph
Theory. New York, NY, USA: McGraw-Hill, 1993.
[33] S. M. Ross, Simulation, 5th ed. New York, NY, USA: Academic Press,
2012.
[34] Y. S. Cho, J. Kim, W. Y. Yang, and C. G. Kang, MIMO-OFDM Wireless Communications With MATLAB. Hoboken, NJ, USA: Wiley-IEEE
Press, 2010.
[35] D. Gross, J. F. Shortle, J. M. Thompson, and C. M. Harris, Fundamentals
of Queueing Theory, 4th ed. Hoboken, NJ, USA: Wiley, 2008.

Yi-Shing Liou received the B.S. degree in industrial technology education from National Kaohsiung
Normal University, Kaohsiung, Taiwan, in 2006, the
M.S. degree in applied electronics technology from
National Taiwan Normal University, Taipei, Taiwan,
in 2009, and the Ph.D. degree in communications
engineering from National Chiao Tung University,
Hsinchu, Taiwan, in 2014. He is currently with
the Department of Electrical Engineering, National
Chiao Tung University. His research interests include cross-layer medium access control in wireless
networks, device-to-device communications in cellular networks, and access
network selection in heterogeneous networks.

Rung-Hung Gau received the B.S. degree in electrical engineering from National Taiwan University,
Taipei, Taiwan, in 1994, the M.S. degree in electrical engineering from the University of California at
Los Angeles, Los Angeles, CA, USA, in 1997, and
the Ph.D. degree in electrical and computer engineering from Cornell University, Ithaca, NY, USA,
in 2001. He is currently a Professor with the Department of Electrical Engineering, National Chiao Tung
University, Hsinchu, Taiwan. His research interests
include cross-layer design for medium access control
in wireless networks, device-to-device communications in cellular networks,
machine-to-machine communications/Internet of Things, mobility management, multicast flow control, and stochastic processes and queueing theory with
applications to communications networks.

Chung-Ju Chang (F06) was born in Taiwan in

1950. He received the B.E. and M.E. degrees in
electronics engineering from National Chiao Tung
University, Hsinchu, Taiwan, in 1972 and 1976, respectively, and the Ph.D. degree in electrical engineering from National Taiwan University, Taipei,
Taiwan, in 1985. From 1976 to 1988, he was with the
Telecommunication Laboratories, Directorate General of Telecommunications, Ministry of Communications, Taiwan, as a Design Engineer, Supervisor,
Project Manager, and then Managing Director. He
also acted as a Science and Technology Advisor for the Minister of the
Ministry of Communications during 1987 and 1989. In 1988, he joined the
Faculty of the Department of Electrical and Computer Engineering, College of
Electrical Engineering and Computer Science, National Chiao Tung University,
as an Associate Professor. He has been a Professor since 1993 and a Chair
Professor since 2009. He was the Director of the Institute of Communication
Engineering from August 1993 to July 1995, Chairman of the Department
of Communication Engineering from August 1999 to July 2001, Dean of the
Research and Development Office from August 2002 to July 2004, and Director
of the Center for Information and Communications Research, Aim for Top
University Plan, sponsored by the Ministry of Education from 2006 to 2010.
Also, he was an Advisor for the Ministry of Education to promote the education
of communication science and technologies for colleges and universities in
Taiwan during 1995 and 1999. He was acting as a Committee Member of
the Telecommunication Deliberate Body, Taiwan. His research interests include
performance evaluation, radio resources management for cellular mobile communication systems, and traffic control for broadband networks. He is a member
of the Chinese Institute of Engineers and the Chinese Institute of Electrical
Engineers. He serves as the Editor of the IEEE Communications Magazine
and an Associate Editor of the IEEE T RANSACTIONS ON V EHICULAR T ECH NOLOGY . He was a recipient of the Outstanding Research Award in 2003 and
2009 as well as the Outstanding Scholar Research Project in 2008 from the
National Science Council, Taiwan; the TECO Award from TECO Technology
Foundation in 2006; and the Science and Technology Chair from Far Eastern
Y.Z. Hsu Science and Technology Memorial Foundation in 2013.