Sie sind auf Seite 1von 15

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
1

On the Spectral and Energy Efficiency of


Full-Duplex Small Cell Wireless Systems with
Massive MIMO
Yi Li, Pingzhi Fan, Fellow, IEEE, Anatolii Leukhin, and Lingjia Liu, Senior Member, IEEE

AbstractThe achievable rate of full-duplex (FD) small cell


systems with massive MIMO, in which a low-power base station
(BS) equipped with large antenna arrays sends/receives data
to/from multiple half-duplex (HD) users at the same time on the
same frequency is investigated. The BS uses imperfect channel
state information (CSI) obtained from received pilots, non-ideal
hardware and linear transmitter and receiver, i.e., zero-forcing
(ZF) or maximum-ratio transmission/maximum-ratio combining
(MRT/MRC) to process the signals. The approximate closedform expressions of the achievable rate for both the ZF and
MRT/MRC processing are derived and used to analyze the
effect of the number of antennas and the hardware imperfection
on the self-interference (SI), which is a bottleneck of the FD
systems. To maximize the spectral efficiency (SE) and the energy
efficiency (EE) of this system, two non-convex power allocation
optimization problems are formulated and solved by utilizing
the sequential convex approximation technique and fractional
programming technique. Two iterative algorithms are proposed
with proved local convergence. Numerical results illustrate that
the analytical approximation of achievable rate matches well
with the Monte-Carlo simulation. It is also shown that the ZF
processing has greater ability to suppress SI, compared with
MRT/MRT processing. The proposed power allocation algorithms
are shown to increase the SE and EE significantly, compared to
the uniform power allocation scheme when the BS is equipped
with moderately large antenna arrays.
Index TermsEnergy efficiency, full duplex, hardware imperfection, massive MIMO, power allocation, small cell, selfinterference, spectral efficiency.

I. I NTRODUCTION
Wireless systems have been facing continuously increasing
demands for higher data rates over limited available radio
Copyright (c) 2015 IEEE. Personal use of this material is permitted.
However, permission to use this material for any other purposes must be
obtained from the IEEE by sending a request to pubs-permissions@ieee.org.
The work of Y. Li and P. Fan is supported by the National Science Foundation of China (NSFC, No.61471302), the National
863 project (No.2014AA01A707), National Sci & Tech Major Project
(No.2016ZX03001018-002), the Sino-Russia joint project (MoST No.CR1915), the CSC Scholarship Programme and the 2016 Doctoral Innovation
Founds of Southwest Jiaotong University. The work of L. Liu is supported
by the U.S. National Science Foundation under grants ECCS-1228071, CCF1422241, and ECCS-1509514. The work of A. Leukhin is supported by the
grant of Russian Foundation of Basic Research (No.15-07-99514) and project
of Russian Ministry of Education and Science (No.1856).
Y. Li and P. Fan are with the Key Laboratory of Information Coding
and Transmission, Southwest Jiaotong University, Chengdu 611756, China
(e-mail: gobelieve1314@gmail.com, pzfan@swjtu.edu.cn).
A. Leukhin is with the Department of Science and Innovation Activities of
Mari State University, Yoshkar-Ola, Russia (e-mail:leukhinan@list.ru).
L. Liu is with the Department of Electrical Engineering and Computer Science, University of Kansas, Lawrence, KS 66045-7608 USA (email:lingjialiu@ittc.ku.edu).

spectrum. This motivates researchers in the field to start


working towards the fifth generation (5G) wireless systems
that are expected to be deployed after year 2020. Among
various technologies that exploit radio resources more efficiently, massive multiple input and multiple output (MIMO),
or large-scale MIMO, has been regarded as one key potential
technology for 5G [1], [2]. The key idea of massive MIMO is
to equip base stations (BS) with orders of magnitude more
antennas, e.g., 100 or more. The benefits brought by the
surplus degrees of freedom (DoF) include higher spectral and
energy efficiency [3], the possible reduction of transmit power
[4], the potential for asymptotically eliminating uncorrelated
interference and noise using simple linear signal processing
[5] and inexpensive hardware [6].
Another potential technology for 5G communications is
full-duplex (FD) transmission where each transceiver transmits
and receives simultaneously on the same frequency [7]. Obviously, the FD transmission has the potential to further improve
or even double the capacity of the conventional half-duplex
(HD) systems [8]. However, the self-interference (SI) from the
transmit antennas to the receive antennas may severely affect
the performance of the FD systems. The existing SI cancellation techniques based on advanced hardware processing are
summarized in [7], [9], including wireless-propagation domain
techniques, analog-circuit-domain techniques, digital-domain
techniques or the combination of them. The FD technique
is studied in point to point systems [8], [10], [11] , relay
systems [12][14] and cellular systems [15] on both theory and
experiment and show its superiority over the HD techniques
provided that the SI is properly handled. Particularly, the
transmit power allocation technique is developed in [14] to
effectively reduce residual SI and guarantees superiority of the
FD mode in an extended region of channel states. In addition,
the spatial processing which utilizes DoF offered by MIMO
to conduct transmit and receive beamforming can dramatically
improve the SI cancellation capability [11][13], [15].
To achieve better spectral and energy efficiency, recent
works sought to incorporate FD and massive MIMO together
to achieve the benefits from both of them in heterogeneous
networks [16], multi-pair MIMO FD relay networks [17][19]
and a single-cell scenario [20]. In [16], the macro BS equipped
with massive MIMO provides wireless backhaul to many FD
small cells. In [17], [18], multiple sources transmit simultaneously their signals to multiple destinations with the help of a
FD relay equipped with massive arrays. It is shown that with
simple linear zero-forcing (ZF) and maximum ratio combining

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
2

(MRC)/maximal ratio transmission (MRT) processing, the SI


effect can be reduced significantly by either using a massive
receive antenna array or using a massive transmit antenna array
together with very low transmit power at the relay. However,
it is still not clear to what extents the SI can be reduced when
the relays transmit and receive arrays are both massive. In
addition, the transmit power of the relay is uniformly allocated
for the destinations , whether a power allocation scheme can
be designed to alleviate or even eliminate the performance
ceilings on the achievable rate of the systems is not clear. In
[19], the effect of distortion noises caused by hardware impairments is studied and a hardware impairments aware transceiver
scheme for FD massive MIMO relay is proposed to mitigate
the distortion noises. However, it assumes perfect channel state
information (CSI) and the SI signal and multi-user interference
can be perfectly canceled by some sophisticated interference
cancellation methods. Reference [20] focuses on the uplink
(UL) spectral efficiency and studies how to utilize the spatial
resources to strike a tradeoff between receive diversity gain
and the SI suppression.
Recall that the current cellular systems with conventional
MIMO are designed to operate in HD mode, e.g., time division
duplex or frequency division duplex. Hence, such systems do
not achieve the maximal spectral efficiency yet. Reference [21]
shows that the small cell with low transmit power and short
transmission distance is considered to be especially suitable
for deployment of FD technology. In addition, small cell is
considered to be the main FD deployment scenario in the
DUPLO project [22]. Therefore, in this paper, we investigate
the achievable rate of FD small cell systems with massive
MIMO, where a low-power BS equipped with large antenna
arrays sends/receives data to/from multiple half-duplex user
equipments (UEs) at the same time on the same frequency. The
combination of two techniques brings the following benefits.
Firstly, the transmit power of the users in the UL can be
reduced significantly while maintaining a given quality of
service (QoS), thus decreasing the co-channel interference
(CCI) from the users in the UL channel to those in the DL
channel. Secondly, the transmit power of the BS can be also
reduced while maintaining a given QoS for the downlink (DL)
users, leading to smaller SI.
The low-complexsity linear ZF and MRT/MRC processing
are utilized at the FD BS. The UL channel and DL channel between the BS and UEs are assumed to be estimated using minimum mean-square-error (MMSE) estimation.
In addition, since the massive MIMO must be built with
low cost components to reduce the deploy cost and energy
consumption of circuits, hardware impairments are inevitable
due to nonlinear power amplifier, digital-to-analogue converter
impairments and so on [23]. The distortion noise is used to
model the effect of imperfect radio frequency chain by adding
an independent zero-mean complex Gaussian random variable
to each antenna whose power is proportional to the power of
the transmitted/received signal at that antenna. This additive
stochastic impairment model is used to study the influence of
hardware impairments in [19], [23], and its validity has been
confirmed in [24] by simulations. In addition, the impact of
hardware impairments on the performance of distributed DL

massive MIMO systems is analyzed in [25]. Therefore, this


paper also considers the effect of hardware impairments on
the achievable rate of the FD small cell system.
The main contributions of this paper are summarized as
follows:
To the best of the authors knowledge, we are the first
to study the ergodic achievable rate of the FD small cell systems with massive MIMO and linear processing, considering
both imperfect channel estimation and hardware impairments
due to the low-cost antennas. The approximate closed-form
expressions for the achievable rate with ZF and MRT/MRC
processing are derived.
It is shown that the massive MIMO technique can benefit
FD small cell systems. With ZF processing, it is shown that
the SI power can be significantly reduced as the number of
transmit antennas of FD BS, Nt , increases. More precisely,
the SI variance is proportional to 1/Nt when the BS with
massive MIMO transmits with full power to get as high SE as
possible, and proportional to 1/Nt2 when the BS transmits with
low power to maintain a given QoS for the DL UEs. However,
with MRT/MRC processing, the SI variance is proportional to
the number of receive antennas of FD BS, Nr , when the BS
transmits with full power, and converges to a constant when
the BS transmits with low power to simply maintain a given
QoS. Therefore, the ZF processing performs better when the
BS is equipped with large-scale antennas at both transmitter
and receiver. It is also shown that the hardware imperfection at
the FD BS equipped with finite number of antennas increases
the power of SI and decreases the achievable rate for both ZF
and MRT/MRC processing.
Two power allocation algorithms for the UL UEs and
DL of the FD BS are proposed to maximize the SE and
the EE. The nonconvex optimization problems are solved by
the sequential convex approximation technique in [26] and
fractional programming technique in [27]. The convergence
of the two proposed iterative algorithms are proved and their
effectiveness are verified by simulation.
It is important to note that the main goal of this paper is to
investigate potential gains of introducing the massive MIMO
to FD small cell along with the power allocation algorithms to
maximize the SE and EE as opposed to proposing a hardware
technique to cancel the SI or suppressing distortion noises
caused by hardware impairments. Although other issues may
still exist, it is expected that the FD small cell system equipped
with massive MIMO is particularly applicable in the scenarios
of hotspots, e.g., conference hall, shopping center and so on.
Notation: Bold lower and upper case letters represent vectors and matrices, respectively. HT and HH are the standard
transpose and Hermitian of H, respectively. The Euclidean
norm, the trace, and the expectation are denoted by ,
tr() and E{}, respectively. The notation and means
almost sure convergence and direct proportion, respectively.
x CN (0, ) denotes a complex Gaussian vector x with
zero mean and covariance matrix .
II. S YSTEM MODEL
Fig.1 shows the system model for FD small cell equipped
with massive MIMO where the FD capable BS equipped with

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
3

Nt transmit antennas and Nr receive antennas communicates


with KD single-antenna users in the DL channel and KU
single-antenna users in the UL channel at the same time on
the same frequency.

b d ), CN (0, Dd D
b d ), CN (0, D
b u ) and
CN (0, D
b
b
b u are
CN (0, Du Du ), respectively, where Dd and D
2
diagonal matrices whose kth diagonal elements are d,k
=
2
2
2
pp d,k /( pp d,k + 1) and u,k = pp u,k /( pp u,k + 1),
respectively, where is the length of the pilot sequences,
satisfying (KD + KU ), and pp is the transmit power
of each pilot symbol.

Nt

B. Data Transmission
The receive signal of KD users in the DL channel can be
rewritten as

=)057SUHFRGHU
SRZHUDOORFDWLRQ

SI

CCI

Nr
=)05&UHFHLYHU

Fig. 1.

System model for the full-duplex small cell with massive MIMO.

KD Nt
KD KU
Assume GH
, GH
and Gu
d C
u,d C
C
are the channel matrices from the BSs transmit
antennas to the KD users in the DL channel, from the KU
users in the UL channel to the KD users in the DL channel
and from the KU users in the UL channel to the BSs receive
antennas, respectively. More precisely, Gd , Gu,d and Gu can
1/2
1/2
be expressed as Gd = Hd Dd , Gu,d = Hu,d Du,d and
1/2
Gu = Hu Du where the small-scale fading matrices Hd ,
Hu,d and Hu have independent and identically distributed
(i.i.d) CN (0, 1) elements, while Dd , Du,d and Du are the
large-scale diagonal matrices whose kth diagonal elements are
denoted by d , u,d and u , respectively.
The SI channel Gs represents the residual interference due
to the imperfect SI cancellation techniques, and it is assumed
to be Rayleigh fading since the line-of-sight component can
be efficiently reduced by propagation-domain suppression.
Therefore, the entries of Gs are modeled as i.i.d CN
2
2
(0, SI
), where SI
is determined by the capability of the
SI cancellation technique. Note that the same assumption can
be found in [18], [20]. In fact, the conclusion of [10] is
that the Rician probability distribution with a small Rician
factor should be used to characterize the residual SI channel
after SI cancellation techniques. The performance analysis that
takes the Rician fading SI channel into consideration is an
interesting problem for future work.

H
yd = GH
(3)
d (sd + td ) + Gu,d su + n,

where sd = V Pd xd is the downlink transmit vector of the


BS, xd{ CKD}1 is the DL transmit symbols and it is assumed
= I. Pd CKD KD is the diagonal power
that E xd xH
d
matrices whose kth entry, Pd,k , represents the power allocated
to the kth DL UE. V CNt KD is the precoder at the BS. n is
additive Gaussian noise vector and its element are assumed to
be i.i.d CN (0, 1). As reference [23], the transmit distortion
noise td follows

td CN (0, vd diag(E{sd sH
d })),

Nr KU

where vd characterizes the level of transmit imperfection.


The receive signal at the BS from KU users in the UL
channel can be rewritten as
yu = ru + tu = Gu su + Gs (sd + td ) + n + tu , (5)

where su = Pu xu is the uplink transmit vector form the


Ku users, xu CK{U 1 is }the UL transmit symbols, and
= I. Pu CKU KU is the
it is assumed that E xu xH
u
diagonal power matrices whose kth entry, Pu,k , represents the
power allocated to the kth UL UE. n is additive Gaussian noise
vector and its element are assumed to be i.i.d CN (0, 1).
As reference [23], the receive distortion noise
tu CN (0, vu diag(E{ru rH
u })).

Assume the BS uses MMSE estimation to estimate Gd and


Gu . The MMSE channel estimates of Gd and Gu are given
by
b d = Gd d ,
G
(1)
b u = Gu u ,
G

KD

H
H
bd,k
bd,k
yd,k = Pd,k g
vk xd,k +
Pd,j g
vj xd,j
j=k
KD
KU

H
H
+
Pd,j d,k vj xd,j + gd,k td +
Pu,i gk,i xu,i + n,
i=1

(7)
where gd,k and vk are the kth columns of Gd and V,
respectively.
Similarly, the signal of kth UL UE after linear processing
WH at the BS can be expressed as
yu,k =

bu,k xu,k +
Pu,k wkH g

(2)

where d and u are the estimation error matrices of Gd


b d,
and Gu , respectively. According to [18], each column of G
b u and u are mutually independent and distributed as
d , G

(6)

where vu characterizes the level of receive imperfection.


Consider the imperfect channel model given by (1) and (2),
and transmit signal at the BS (3), the received signal at the
kth DL UE can be expressed as

j=1

A. Channel Estimation

(4)

KU

bu,j xu,j + wkH tu


Pu,j wkH g
j=k

KU

Pu,j wkH u,j xu,j + wkH Gs (V

Pd xd + td ) + wkH n,

j=1

(8)

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
4

where gu,k and wk are the kth columns of Gu and W,


respectively.
C. ZF and MRT/MRC Processing
In this work, we consider two conventional linear processing
techniques, namely ZF and MRC/MRT, which has been shown
to become near optimal with low complexity in massive
MIMO scenario [4].
1) ZF Processing
The ZF precoding matrix V can be written as V = A,
where
b d (G
b HG
b d )1 ,
A=G
(9)
d
and the diagonal matrix = diag(1 , 2 , ..., KD ) guarantees
that each column of the ZF precoder V satisfies
E{vkH vk } = k2 E{aH
k ak } = 1, for k = 1, ..., KD .

(10)

The normalization constant k can be computed as


1
1
1
=
=
H A] }
HG
1 ] }
b
b
E{[A
E{aH
a
}
kk
k
E{[(
G
k
d)
kk
d
2
KD d,k
2
,
=
= (Nt KD )d,k
E{tr(X1 )}
(11)
where X = HH
H
is
a
K
K
central
Wishart
matrix
with
d
D
D
d
Nt DoF and covariance matrix IKD , and the last equality is
obtained by using [28, Lemma 2.10]. The ZF receiver WH
can be written as
k2 =

b HG
b u )1 G
b H.
W H = (G
u
u

(12)

2) MRC/MRT Processing
The MRT precoding matrix V can be written as V = A,
where
b d,
A=G
(13)
and the diagonal matrix guarantees that each column of the
MRT precoder V satisfies (10). Therefore, the normalization
constant k can be computed as
k2 =

1
1
1
=
=
2 ,
Hg
N

b
E{aH
a
}
E{b
g
}
t d,k
k k
k k

(14)

and the MRC receiver WH can be written as


b H.
WH = G
u

(15)

III. ACHIEVABLE RATE ANALYSIS


In this section, we derive the achievable rate of the FD small
cell with massive MIMO for ZF and MRT/MRC processing.
According to the received signal expressions of the DL UE and
UL channels of the BS in (7) and (8), the ergodic achievable
rates of transmission links between the BS and the kth user
in the DL channel, between the BS and the kth user in the
UL channel is given by (16) and (17), shown at the top of the
next page, respectively.
Before the achievable rates of UL and DL channel are
derived, the covariance matrices of the transmit and receive
distortion noise due to the imperfect hardware with ZF and
MRC/MRT processing are provided by the following lemmas.

Lemma 1 With ZF processing, the covariance matrices of


the transmit and receive distortion noise td and tu can be
approximated as
KD
vd ( j=1
Pd,j )(Nt KD )
ZF
I,
d

Nt2
KU
KD
(

(18)
Pu,j u,j + (1 + vd )(
Pd,j )
u ZF vu
j=1

)
KD 2
)SI + 1 I.
(1
Nt

j=1

Proof: See Appendix A.


Lemma 2 With MRT/MRC processing, the covariance matrices of the transmit and receive distortion noise td and tu
can be calculated as
KD
vd j=1
Pj
M RT
d
=
I,
Nt
KU
KU
(
)

M RC
2
u
= vu
pu,j u,j + (1 + vd )
Pd,j SI
+ 1 I.
j=1

j=1

(19)
Proof: See Appendix B.
Then the approximate closed-form expressions of the
achievable rate for ZF and MRT/MRC processing at the BS
are given by the following theorems.
Theorem 1: With ZF precoder and ZF receiver at the BS,
the achievable rate of the FD system, for a finite number of
transmit and receive antennas at the BS, can be approximated
as (20), shown at the top of the next page.
Proof: See Appendix C.
Theorem 2: With MRT precoder and MRC receiver at the
BS, the ergodic achievable rate of the full-duplex small cell
system, for a finite number of transmit and receive antennas
at the BS, can be approximatively calculated as (21), shown
at the top of the next page.
Proof: See Appendix D.
Based on the above Theorem 1 and 2, we can gain some
insights which are summarized as the following propositions.
Proposition 1: Assume the transmit power for the users in
E
the DL channel at the BS is Pd,k = Nd,k
, and the transmit
t
E
power of the users in the uplink channel is Pu,k = Nu,k
, where
r
Ed,k and Eu,k are fixed regardless of Nt and Nr . As Nt
and Nr , the achievable rate of ZF and MRT/MRC
processing for the kth user in the DL channel and the kth user
in the UL channel are given respectively by
(
)
2
Rd,k log2 1 + Ed,k d,k
,
)
(
2
Eu,k u,k
(22)
.
Ru,k log2 1 +
1 + vu
2
2
Assume the estimation of large-scale fading d,k
, u,k
and
the parameter of receive hardware imperfection vu are fixed,
the achievable rate for each user in the FD system converges
to a given QoS which is only determined by the fixed power
constant Ed,k , Eu,k . That is to say, with Nt and Nr go to
infinity, the effect of SI, channel estimation error, CCI and
transmit hardware impairments all disappear with very low

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
5

bd,k
P
g
v


d,k
k

= E log2 1 + K
,


2 K

2

D
D

H

H

H 2 K
2

bd,k vj +
Pd,j g
Pd,j d,k vj + gd,k td +
Pu,i |gk,i | + 1

Rd,k

j=1

j=k

T
KU

(K
D

k=1

K
U

k=1

Pu,j ((1 + vu )u,j

j=1

MR

T
KU

k=1

(K
D

(
log2 1 +

k=1

(
log2 1 +

(1 + vu )

2
Pd,k (Nt KD )d,k

log2 1 + KD
( j=1 Pd,j )((1 + vd

(
log2 1 +

(17)

j=1

j=k

i=1

Hb


P
w
g
u,k

k u,k
= E log2 1 + K
.


2 K

2
2

2 H 2
U
U

H
H
H
H









bu,j +
Pu,j wk g
Pu,j wk u,j + wk tu + wk Gs V Pd + wk

Ru,k

ZF

(16)

KD vd
Nt )d,k

2 )+
d,k

KU
i=1

Pu,i k,i + 1

+
))
.

(20)

))

(21)

2
Pu,k (Nr KU )u,k
K
D

2 )+(
2
D
u,j
Pd,j )(1 K
Nt )SI (1 + vu + vd + vu vd ) + vu + 1
j=1

(1 + vd )(
KU
j=1

KD
j=1

Pu,j u,j

2
Pd,k d,k
(Nt + 1)
2 +
Pd,j )d,k Pd,k d,k

)
KU
i=1

Pu,i k,i + 1

2
Pu,k u,k
(Nr + 1)

K
D
2
2 +(
Pu,k u,k
j=1 Pd,j )SI (1 + vu + vd + vu vd ) + vu + 1

transmit power. Therefore, FD wireless system will be more


likely to fulfill its potential with the aid of massive MIMO
technique.
Proposition 2: The SI expression after ZF processing is given by (52). Particularly, if Nt = Nr and KD = KU , the SI exK
D

2
Pd,k )SI
k=1
2
Nt u,k

(1+vd )(

pression can be further simplified as SI =


.
Obviously, if the transmit power of the BS equipped with
massive MIMO scales proportionally to 1/Nt to just maintain
E
, the
a given QoS as in Proposition 1, i.e., Pd,k = Nd,k
t
2
SI power SI 1/Nt , and if the BS uses the maximum
transmit power Pd,max to get as high DL rate as possible,
K
D

i.e.,
Pd,k = Pd,max , SI 1/Nt . Therefore, the SI power
k=1

of ZF processing can be significantly reduced by increasing


the number of transmit antennas.
Proposition 3: The
after MRT/MRC proKDSI expression
2
2
cessing is (1 + vd )( j=1
Pd,j )SI
u,k
Nr , as shown in (61).
Particularly, if Nt = Nr and the transmit power of the
BS equipped with massive MIMO scales proportionally to
1/Nt to just maintain a given QoS as in Proposition 1,
E
, the SI power converges to a constant
i.e., Pd,k = Nd,k
KD t
2
2
(1+vd )( j=1 Ed,j )SI
u,k
, and if the BS uses the maximum
transmit power Pd,max to get as high DL rate as possible, the
SI power is proportional to Nr .

IV. POWER ALLOCATION


We can see from Theorem 1 and 2 that the achievable rate
is closely related to the transmit power of BS for each DL UE
Pd,k and the transmit power of each UL UE Pu,k . For example,
increasing Pd,k can increase the DL rate Rd,k on one hand,
on the other hand, it will cause more SI to the receiver at
BS. Therefore, designing a suitable power allocation scheme
is very important to the FD system. In this section, we propose
two power allocation schemes, aiming to maximize the SE and
the EE under the constrains of the maximum transmit power of
the BS and the UL UEs. In order to simplify the statements, the
ZF processing is taken as an example. However, the proposed
algorithms are also suitable for the MRT/MRC processing.
A. Power Allocation for Maximizing the SE
The optimization problem for maximizing the SE can be
formulated as (23), shown at the top of the next page, where
2
2
D vd
ak = (Nt KD )d,k
, bk = (1 + vd KN
)d,k d,k
,
t
(Nr KU ) 2

u,k
ck,i = k,i , dk =
, ej = u,j vuu,j
vu +1
+1 , f =
KD
2
(1 Nt )SI
(1 + vd ).
Because the objective function of Problem (23) is nonconvex, we first introduce the following auxiliary variables
to make it convex.
Pd,k ak
1 + KD
rd,k
(24)
KU
( j=1 Pd,j )bk + i=1
Pu,i ck,i + 1

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
6

max

{Pd,k },
{Pu,k }

subject to

)
Pd,k ak
log2 1 + KD
KU
( j=1 Pd,j )bk + i=1
Pu,i ck,i + 1
k=1
(
)
KU

Pu,k dk
log2 1 + KU
+
KD
j=1 Pu,j ej + (
j=1 Pd,j )f + 1
k=1
KD

Pd,k 0,
KD

(23)

k = 1, ..., KD ,

Pd,k Pd,max ,

k=1

0 Pu,k Pu,max , k = 1, ..., KU .

Pu,k dk
ru,k ,
KD
j=1 Pd,j )f + 1
j=1 Pu,j ej + (

1 + KU

(25)

Then the non-convex


objective function
of problem (23)
KD
KU
can be reformulated as k=1
log2 (rd,k ) + k=1
log2 (ru,k ).
Due to the monotonicity of the log function, it can be further
reformulated as
K
KU
D

rd,k
ru,k
(26)
k=1

k=1

The non-convex objective function of problem (23) can be


easily recognized equivalent to (24), (25) and (26) by noting
the fact that all constraints in (24) and (25) are met with
equality at the optimum.
Besides, due to the parameters in (24) and (25) are all
positive, two additional constraints as following are added
rd,k 1,

(27)

ru,k 1.

(28)

Note that maximizing a product of variables in (26) admits


an second order cone (SOC) representation [29]. Thus, we
only need to deal with the nonconvex constraints in (24) and
(25). Let us treat the constraint (24) first. It is without loss of
optimality to replace (24) by following two constraints
KD
KU

(
Pd,j )bk +
Pu,j ck,j + 1 d,k ,
j=1
KD

j=1

Pd,j )bk +

f (rd,k , d,k )

j=1

where d,k is a newly introduced variable. The equivalence


between (24) and the two inequalities in (29) and (30) follows
the same arguments as in [21]. The key idea is to prove that the
constraint (29) is met with equality at optimum. Assume that
the constraint in (29) holds with inequality at optimum, then

we form a new pair (d,k , rd,k ) as d,k = d,k


t and r d,k = trd,k
where t is a positive constant. Obviously, there exists a given
t > 1 such that (29) is still met when d,k is replaced by
d,k . Since d,k rd,k = d,k rd,k , (30) is also satisfied. However,
since rd,k > rd,k with t > 1, a strictly higher objective of
the optimization problem is obtained, which contradicts with
assumption of optimality.
We note that the constraint (29) is linear and the constraint
(30) is non-convex. Following the spirit of [21], [29], we will

(n)

r2
(n) d,k
2d,k

d,k
2

2
d,k
,

(n)

f (rd,k , d,k ) = g(rd,k , d,k , d,k ),

(32)

(n)

(33)

f (rd,k d,k ) = g(rd,k , d,k , d,k ),


(n)

when d,k =
reformulated as

KD

j=1

Pu,j ck,j + 1 + Pd,k ak rd,k d,k , (30)

(n)
g(rd,k , d,k , d,k )

(31)
(n)
which holds for every d,k > 0 and design an iterative
algorithm based on the sequential convex approximation. The
(n)
convex upper bound function g(rd,k , d,k , d,k ) has the following two important properties which can be used to prove
the convergence of the proposed algorithm. It can be easily
checked that

(29)

j=1
KU

replace the non-convex term f (rd,k , d,k ) = rd,k d,k in (30)


with its convex upper bound

Pd,j )bk +

rd,k
d,k .

KU

Therefore, the constraint (30) can be

Pu,j ck,j +1+Pd,k ak

j=1

(n)
d,k 2
2
r +
.
(n) d,k
2 d,k
2d,k

(34)
The constraint (25) can be converted to the following convex
constraint using a similar method
KU

Pu,j ej + (

j=1

KU

j=1

KD

Pd,j )f + 1 u,k ,

KD

Pu,j ej +(
Pd,j )f +1+Pu,k dk
j=1

(35)

j=1

(n)

u,k

2u,k

r2 +
(n) u,k

2
u,k
,

(36)
where u,k is a newly introduced variable.
Therefore, the problem (23) can be reformulated as a series
of convex approximate problem. The nth iteration will solve

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
7

functions R = h(Pd,k , Pu,k ) g(Pd,k , Pu,k ) where

the following convex problem


max

{Pd,k },{Pu,k },
{rd,k },{ru,k },
{d,k },{u,k }

subject to

K
D

rd,k

k=1

KU

h(Pd,k , Pu,k )
ru,k
=

k=1

KD

log2 ((

(29), (34), k = 1, ..., KD ,


(35), (36), k = 1, ..., KU ,
rd,k 1, k = 1, ..., KD ,
ru,k 1, k = 1, ..., KU ,

+
(37)

log2 (

KU

=
Pd,k Pd,max ,

KD

Pu,j ej + (

KU

KD

Based on (37), the proposed iterative algorithm is summarized in Algorithm 1, and its convergence is established in the
following Theorem 3.
Algorithm 1 Power Allocation for Maximizing the SE
(0)

STEP 1: Start with arbitrary d,k for k = 1, ..., KD and


(0)
u,k for k = 1, ..., KU . Let n = 0.
STEP 2: Solve (23) with the convex optimization solver

CVX [30] to find the optimal solutions Pd,k


, rd,k
, d,k
for

k = 1, ..., KD and Pu,k , ru,k , u,k for k = 1, ..., KU .


STEP 3: n n + 1.
(n)

STEP 4: Update u,k = ru,k


/u,k
for k = 1, ..., KD and
(n)

d,k = rd,k /d,k for k = 1, ..., KU .


STEP 5: Repeat step 2, 3, 4 until convergence.
Theorem 3: The proposed algorithm 1 is convergent to the
Karush-Kuhn-Tucker (KKT) point of the non-convex Problem
(23).
Proof: See Appendix E.

log2 (1 +

k=1

Pu,k + Pcir .

(38)

k=1

The optimization problem for maximizing the energy efficiency can be formulated as (39), shown at the top of the next
page.
This algorithm is based on the observation that the rate
function R can be expressed as a difference of two concave

(41)

KU

Pu,j ej + (

KD

Pd,j )f + 1),

j=1

ge(n+1) (Pd,k , Pu,k ) = log2 (1 + xd ) + (1 + xd )1


(n)

(n)

(n) 1
(xu x(n)
(xd xd ) + log2 (1 + x(n)
u ),
u ) + (1 + xu )
(42)
KU
KD
(n)
=
Pd,j )bk +
where xd = ( j=1
i=1 Pu,i ck,i , xd
KU
KU (n)
KD (n)
P
e
+
P
c
,
x
=
Pd,j )bk +
( j=1
u,j j
u
i=1 u,i k,i
KU (n)
Kj=1
KD
(n)
(n)
D
Pu,j ej + ( j=1
Pd,j )f .
( j=1
Pd,j )f and xu = j=1
Problem (39) will be solved iteratively and the n + 1st
iteration will solve the following fractional problem
(n)

subject to

EE is another important performance measure for wireless


communication systems. Although the transmit power consumption for each antenna is extremely low in the massive
MIMO systems, the effect of the hardware circuit power
consumption would be gradually increased by the factor
of the number of antennas. Hence, the power consumption
model which includes both the radio frequency circuit power
consumption Pcir and the transmit power consumption is
considered

Pu,i ck,i + 1)

i=1

j=1

{Pd,k },{Pu,k }

B. Power Allocation for Maximizing the EE

KU

which is in general non-convex.


We employ a linear approximation of the concave function
g(Pd,k , Pu,k ) to make R convex. Due to the concavity of the
logdet function, we have the inequality log(1 + x) log(1 +
x0 )+(1+x0 )1 (xx0 ) for x 0 by using the fact x log(1+
x) = (1 + x)1 . Suppose the value of {Pd,k } and {Pu,k } at
(n)
(n)
iteration n is denoted by {Pd,k } and {Pu,k }, respectively. The
(n)
upper bound of g(Pd,k , Pu,k ) at a neighborhood of {Pd,k } and
(n)
{Pu,k } at iterations n + 1 is

maximize

KU

(40)

Pd,j )f + 1 + Pu,k dk ),

Pd,j )bk +

j=1

k=1

Pd,k +

KD

j=1

log2 ((

k=1

k=1

KD

Pu,i ck,i + 1 + Pd,k ak )

g(Pd,k , Pu,k )

0 Pu,k Pu,max , k = 1, ..., KU .

Ptotal =

KU

i=1

j=1

k=1

Pd,k 0, k = 1, ..., KD ,
KD

Pd,j )bk +

j=1

k=1
KU

KD

h(Pd,k , Pu,k ) ge(n+1) (Pd,k , Pu,k )


Ptotal
Pd,k 0, k = 1, ..., KD ,
KD

(43)
Pd,k Pd,max ,

k=1

0 Pu,k Pu,max , k = 1, ..., KU


The fractional subproblem (43) belongs to the class of
concave-convex fractional programs, and can be globally
solved using a Dinkelbach-type algorithm [27], which is
based on solving a series of the following parametric convex
optimization problems for opt
maximize

{Pd,k },{Pu,k }

subject to

h(Pd,k , Pu,k ) ge(n+1) (Pd,k , Pu,k ) Ptotal


Pd,k 0, k = 1, ..., KD ,
KD

Pd,k Pd,max ,

k=1

0 Pu,k Pu,max , k = 1, ..., KU .


(44)

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
8

(
log
2 1+
k=1

KD
max

)
Pd,k ak
KD
KU
( j=1
Pd,j )bk + i=1
Pu,i ci,k +1

(
log
2 1+
k=1

KU

)
Pu,k dk
KD
j=1 Pu,j ej +(
j=1 Pd,j )f +1

KU

Ptotal

{Pd,k },{Pu,k }

subject to

Pd,k 0, k = 1, ..., KD ,
KD

(39)

Pd,k Pd,max ,

k=1

0 Pu,k Pu,max , k = 1, ..., KU

Based on the subproblem (43) and (44), the proposed


iterative algorithm to solve Problem (39) is summarized in
Algorithm 2.

STEP 1: Start with arbitrary {Pd,k } and {Pd,k


} for k =
(0)

1, ..., KD and {Pu,k } and {Pu,k } for k = 1, ..., KU . Let


n = 0.
STEP 2: Compute the linear approximation g(Pd,k , Pu,k )
(n)
(n)
of at a neighborhood of {Pd,k } and {Pu,k } to get
(n)
ge (Pd,k , Pu,k ).
(0)

Ntx=Nrx=200
70
Average Sum Rate (bits/s/Hz)

Algorithm 2 Power Allocation for Maximizing the EE

80

30
Ntx=Nrx=50
20
ZF simulation
ZF analytical
MRT/MRC simulation
MRT/MRC analytical
15

10

0
SNR (dB)

10

15

20

Fig. 2. Sum rate versus SNR for ZF and MRT/MRT processing (K =


2 = 1, P
10, SI
d,k = Pu,k = Pp , d,k = u,k = u,d,k = 1, vd = vu =
0.1).

the channel estimation error, SI, CCI and non-ideal transceiver


hardware when N is finite.
80
70

Average Sum Rate (bits/s/Hz)

In all illustrative results below, we choose the length of the


coherence interval to be T = 200 (symbols), the number of
communication users KD = KU = K, the training length
= 2K, and equal transmit and receive antennas at BS Nt =
Nr = N . The maximum transmit power of the BS is set to
20 dBm.
Fig. 2 shows that the analytical expression of achievable
rate for MRT/MRT processing obtained by Theorem 2 matches
well with the Monte-Carlo simulation for different number of
transmit (receive) antennas N . The performance gap between
the simulation results and the approximate analytic expression
for ZF processing is small, especially when N is larger.
Therefore, the proposed approximate expressions of achievable
rate in Theorem 1 and 2 are tight. When the SNR is low,
the MRT/MRC has almost the same or even better sum rate
than ZF processing. However, the ZF processing can obtain
larger achievable rate than MRT/MRC processing in high SNR
region, because the SI can be suppressed more effectively. In
addition, finite achievable rate ceilings can be observed due to

40

0
20

Theorem 4: The proposed Algorithm 2 produces a sequence


of u(n) with local convergence.
Proof: See Appendix F.
V. N UMERICAL AND S IMULATION R ESULTS

Ntx=Nrx=100
50

10

h(P ,P )e
g (n) (P ,P )

d,k
u,k
d,k
u,k
STEP 3: Compute (n) =
,P )
Ptotal (Pd,k
u,k
STEP 4: Solve (44) with the convex optimization solver

CVX [30] to find the optimal solutions Pd,k


for k =

1, ..., KD and Pu,k for k = 1, ..., KU .


STEP 5: Repeat step 3, 4 until convergence.
(n)
(n)

STEP 6: n n + 1. Update {Pd,k } = {Pd,k


}, {Pu,k } =

{Pu,k
}.
STEP 7: Repeat step 2, 3, 4, 5, 6 until convergence.

60

60
50
40
ZF vd=0,vu=0
ZF vd=0,vu=0.5
ZF vd=0.5,vu=0
MRT/MRC vd=0,vu=0
MRT/MRC vd=0,vu=0.5
MRT/MRC vd=0.5,vu=0

26

30

24
20
15.5 15 14.5 14
10
0
20

15

10

0
SNR (dBm)

10

15

20

Fig. 3. Sum rate versus SNR for ZF and MRT/MRT processing (N =


2 = 1, P
200, K = 10, SI
d,k = Pu,k = Pp , d,k = u,k = u,d,k = 1).

Fig. 3 shows that the effect of hardware impairment on


the achievable rate for ZF and MRT/MRT processing when
the BS is equipped with moderate number of antennas. It is
observed that the achievable rate decreases when hardware
impairment exists. Specifically, the imperfect receive hardware
is more detrimental to the achievable rate than the imperfect

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
9

70
vd=0, vu=0
vd=0, vu=0.25
vd=0.25, vu=0

5
10
QoS=2 bit/s/Hz

15
20
25
30
35
50

2SI=1, ZF

150

2SI=1, MRT/MRC
2SI=10, MRT/MRC

50

40

30

20

10

QoS=1 bit/s/Hz

100

2SI=10, ZF

60
Average Sum Rate (bits/s/Hz)

The selfinterference (dBm)

200
250
300
350
400
Number of transmit (receive) antennas

450

500

20

40
60
Number of downlink (uplink) users

80

100

(a) ZF processing.
45
vd=0, vu=0
vd=0, vu=0.25
vd=0.25, vu=0

Fig. 5. The sum rate versus the number of DL (UL) users (Pd,k = Pu,k =
Pp = 10 dBm, N = 100, d,k = u,k = 1, u,d,k = 2, vd = vu = 0.1).

The selfinterference (dBm)

40

35
QoS=2 bit/s/Hz
30

QoS=1 bit/s/Hz
25

20
50

100

150

200
250
300
350
400
Number of transmit (receive) antennas

450

500

(b) MRT/MRC processing.


Fig. 4.
The power of self-interference versus the number of transmit
antennas at the BS to achieve a required QoS per user, (K = 10, Pp =
10 dBm, d,k = u,k = u,d,k = 1).

channels for both ZF and MRT/MRC processing. This is


because when the number of users is small, increasing the user
number means more data stream is transmitted, leading to a
higher sum rate. However, with more and more users served,
the more severe SI and the CCI and the higher proportion
of pilot symbol leads to a smaller sum rate. The optimal
number of users decrease as the SI power increase. It is also
shown that when the number of antennas is far more than
the number of users, the ZF processing is superior to the
MRT/MRC processing. however, when the number of users
increases, the performance of ZF processing degrade more
rapidly than MRT/MRC processing, due to the insufficient
spatial dimension for interference elimination.
110
HD
FD

transmit hardware for the MRT/MRC processing at all range


of SNR. However, for the ZF processing, the imperfect receive
hardware decreases the achievable rate more at low SNR,
while the transmit hardware imperfection has greater influence
at high SNR.
Fig. 4 shows that for a given QoS, the SI is reduced
significantly with the number of transmit antennas at the BS
N increasing, especially when N is low to medium. For the
ZF processing, the power of SI is further reduced by 6 dB,
i.e., SI 1/Nt2 , as predicted by the Proposition 2, when N
increase from 250 to 500. However, the SI power converges
to constant for the MRT/MRC processing, as predicted by
Proposition 3. It is shown that the ZF processing can suppress
SI more effectively than MRT/MRC with the same number of
transmit and receive antennas. It can be observed that when the
number of antennas is small, even if we use infinite transmit
power, we cannot achieve a required spectral efficiency. Instead, we can add more antennas to reduce the SI and achieve
the required QoS. It can also be observed that the additional
SI power caused by transmitter imperfection is much stronger.
Fig. 5 illustrates that the average sum rate is a quasiconcave function of the number of users in the DL (UL)

Average Sum Rate (bits/s/Hz)

100
90
80
ZF
70
60
50
MRT/MRC
40
30
20

15

10

0
5
The SI power (dB)

10

15

20

Fig. 6. The sum rate comparison between half-duplex and full-duplex system
under different self-interference level (Pd,k = Pu,k = Pp = 10 dBm, N =
200, K = 10, d,k = u,k = u,d,k = 1, vd = vu = 0.1).

Fig. 6 compares the sum rate of the HD and FD mode. For


HD mode, two orthogonal time slots are allocated to the UL
and DL transmission, therefore, the SI and CCI disappear with
the cost of sum rate divided by 2. For fair comparison, the total
energies spent in a coherence interval for both modes are set
to be the same, hence the transmit power of BS and UL UEs

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
10

in HD mode is double the power using in FD model. when the


SI variance is low, the sum rate of FD mode is always larger
than the HD mode for both ZF and MRT/MRC processing.
2
However, when SI
is high, the performance of the HD mode
is superior.
In the following, we assume a more practical scenario
with UL and DL UEs uniformly distributed on the edge
of small cell. In this scenario, the power allocation scheme
is very important, because the signal power is weak and
severe interference is suffered. The simulation parameters are
summarized in Table I and the value of parameters follows the
settings of [31], [32].

30
28

Average Sum Rate (bits/s/Hz)

24
22
20
18
16
14
12
10

TABLE I
TABLE I. SIMULATION ASSUMPTIONS AND PARAMETERS FOR
NETWORK SCENARIOS

Fig. 7 illustrates the convergence rate of the proposed Algorithms 1 and 2 for a given set of channel realizations generated
randomly with the large-scale fading generated by parameters
in Table I. It is observed that the proposed two algorithms
exhibit monotonic convergence within a few iterations. It can
be proved by simulation that this convergence rate behavior is
typical for other channel realizations.
Fig. 8 shows that when the SNR is very low, the maximizing
SE and EE power allocation schemes have the same SE as
the uniform power allocation scheme for both ZF and the
MRT/MRC processing. When the SNR is high, the maximizing SE power allocation scheme offers a large performance
gain compared with the uniform power allocation scheme,
verifying its effectiveness. When the maximum transmit power
equals to 20 dBm, the sum rate achieved by the maximizing
SE scheme increases the sum rate by about 35%, 30% using
ZF and MRT/MRC processing, respectively, compared to the
uniform power allocation scheme. The maximizing EE scheme
has almost the sum rate as the the maximizing SE scheme.
This is because the circuit power consumption dominates in
the massive-MIMO small cell systems, the EE is maximized
by transmit with full power to achieve as high sum rate as
possible.
Fig. 9 shows that when the SNR is very low, the maximizing
SE and EE schemes obtain almost the same EE as the uniform
power allocation scheme. When the SNR is getting higher,
the EE of the uniform power allocation scheme decreases
sharply while the maximizing SE and EE scheme maintains
a constant. The maximizing EE scheme achieves the highest
energy efficiency at high SNR, which proves its effectiveness.
When the maximum transmit power equals to 20 dBm, the
maximizing EE power allocation scheme increases the EE by
about 45%, 59% using ZF and MRT/MRC processing, respectively, compared to the uniform power allocation scheme.

10

20
30
The Number of Iterations

40

50

0.026
0.024
Average Energy Efficiency (bits/J)

Value
50 m
2 GHz
10 MHz
103.8 + 20.9log10 (R), R in km
98.45 + 20log10 (R), R in km
-116 dBm/Hz
13 dB
9 dB

(a) The proposed Algorithm 1.

ZF SNR=20dB
MRT/MRC SNR=20dB
ZF SNR=0dB
MRT/MRC SNR=0dB

0.022
0.02
0.018
0.016
0.014
0.012
0.01

4
5
6
7
The Number of Iterations

10

(b) The proposed Algorithm 2.


Fig. 7. The illustration of the convergence of the proposed two alogorithms,
2 = 1, v =
(K = 5, N = 50, Pp = Pd,k,max = Pu,k,max = SN R, SI
d
vu = 0.1, Pcir = 1W ).

30
ZF
25
Average Sum Rate (bits/s/Hz)

Parameter
Small cell radius
Carrier Frequency
Bandwidth
Small cell to UE pathloss
UE to UE pathloss
Noise spectral density
Small cell noise figure
UE noise figure

ZF SNR=20dB
MRT/MRC SNR=20dB
ZF SNR=0dB
MRT/MRC SNR=0dB

26

20

15

10

MRT/MRC

0
10

Uniform power allocation


Optimal power allocationmaximizing SE
Optimal power allocationmaximizing EE
5

5
10
15
20
The maximum transmit power (dBm)

25

30

Fig. 8. The sum rate versus SNR when the SI and CCI is strong (N =
2 = 10, P
50, K = 5, SI
cir = 1W ).

VI.

CONCLUSION

In this paper, the approximate achievable rates of the fullduplex small cell with massive MIMO for both ZF and
MRT/MRC processing were analyzed, which considers both

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
11
(1+vd ) KD
( j=1 Pd,j )(Nt KD )Gs GH
=
s + I})
N2
2
tKU
KD
(1+vd )SI
( j=1 pu,j u,j +
( j=1 Pd,j )(Nt KD ) +
Nt

Average Energy Efficiency (bits/J)

0.025

vu
1)I.

0.02

A PPENDIX B
P ROOF OF L EMMA 2

ZF
0.015

0.01
MRT/MRC

0.005
Uniform power allocation
Optimal power allocationmaximizing SE
Optimal power allocationmaximizing EE
0
10

5
10
15
20
The maximum transmit power (dBm)

25

30

2
Fig. 9. The energy efficiency versus SNR (N = 50, K = 5, SI
=
10, vd = vu = 0.1, Pcir = 1W ).

the imperfect channel estimation and transceiver hardware.


With ZF processing, the SI power is proportional to N1t
when the BS uses all the transmit power, and proportional
to N12 when the BS uses low transmit power to maintain a
t
required QoS, where Nt represents the number of transmit
antennas at the BS. However, with MRT/MRC processing, the
SI power will converge to a constant even when the BS uses
low transmit power to maintain a required QoS. Therefore,
the ZF processing is superior to the MRT/MRT processing in
terms of SE. With imperfect hardware at the BS, the SI power
will be increased. The SE and the EE performance of the
uniform power allocation scheme is poor, especially when the
interference is strong, so two power allocation algorithms with
proved convergence were proposed, aiming at maximizing SE
and EE, respectively. Simulation results show that when the
BS with massive MIMO transmits with low power to maintain
a given QoS, the SI power can be reduced by about 6 dB
by doubling the number of transmit and receive antennas
at the BS with ZF processing. In addition, the proposed
algorithms can increase the SE and EE by more than 30% and
40%, respectively, compared to the uniform power allocation
scheme.
A PPENDIX A
P ROOF OF L EMMA 1
Following (4), the covariance matrices of the transmit distortion noise td can be written as ZF
= vd diag(E{sd sH
d
d }) =
H
vd diag(E{VPd V }) = vd diag(E{APd 2 AH }) =
b d (G
b HG
b d )1 Pd 2 (G
b HG
b d )1 G
b H }). When
vd diag(E{G
d
d
d
Nt K, we can use the law of large numb HG
b d Nt D
b d.
ber to obtain the approximation G
d
vd
ZF
2 b 2 b H
b
N 2 diag(E{Gd Pd Dd Gd }) =
Therefore, d
t
2
vd
vd KD j Pd,j
2 b 1 H
diag(E{H
P

D
H
})
=
(
)I =
2
2
2
d
d
d
j=1
d
Nt
Nt
d,j
vd KD
( j=1 Pd,j )(Nt KD )I.
Nt2
Following (6), the covariance matrices of the transmit
distortion noise tu can be written as u ZF = vu
H
diag(E{ru rH
+ Gs (sd sH
u }) = vu diag(E{Gu Pu G
d +
uKU
H
H
td td )Gs + I}) = vu diag(E{ j=1 pu,j u,j I +

Following (4), the covariance matrices of the transmit distortion noise td can be derived as ZF
= vd diag(E{sd sH
d
d }) =
H
b d Pd G
b H }) =
vd diag(E{VPd V }) = vd diag(E{G
d
KD
KD
vd j=1
Pj
vd j=1
Pj j2 d,j I =
I.
Nt
Based on (6), the covariance matrices of the transmit
distortion noise tu can be given as u M RC = vu
H
H
diag(E{ru rH
u Pu Gu + Gs (sd sd +
u }) = vu diag(E{G

KU
H
H
td td )Gs + I}) = vu ( j=1 pu,j u,j + (1 +
KU
2
vd ) j=1
Pd,j SI
+ 1)I.
A PPENDIX C
P ROOF OF T HEOREM 1
By the convexity of log2 (1 + x) and using Jensens
inequality, we can obtain the upper bound of the
ZF
downlink achievable rate Rd,k
in (16), given by
(45), shown at the top of the next page. In order

H

ZF
bd,k vk },
to derive Rd,k
, we need to compute E{ g
{
}
{
}

2

2
K
K
D
D

H

H

bd,k vj ,
Pd,j g
E
E
Pd,j d,k vj ,
j=1
{ j=k
}
{
}

K
U
H 2
2
E gd,k
td
and E
Pu,i |gk,i | .
i=1
{
}



2
K
D

H

H

bd,k vk } and E
bd,k vj :
Compute E{ g
Pd,j g
j=k

H
b H A = I , i.e., g
bd,k
Due to (12), we have G
ak =
d
H
bd,k aj =} 0 for {j = k.} Therefore, we can get
1, {g

2

2
H

H

2
bd,k vk
bd,k ak
E g
= k2 E g
= k2 = (Nt KD )d,k
{
}

2
K
D

H

bd,k vj
and E
Pd,j g
= 0.
j=k
{
}

2
K
D

H

Compute E
Pd,j d,k vj : Since d,k and vj
j=1
{
}

2
K
D

H

are uncorrelated, we obtain E
Pd,j d,k vj
=
K
D

j=1

}
{
Pd,j E vjH bd,k bH
d,k vj

j=1

K
D

Pd,j (d,k

j=1

K
D
{
}

2
2
) due to (10).
)E vjH vj = (
Pd,j )(d,k d,k
d,k
j=1
{
}

H 2
H
Compute E gd,k
td : Since gd,k
and td are
{
}
2
{
}
H
H
uncorrelated, E gd,k
td
= E gd,k
ZF
=
d gd,k
v d K D
to Lemma1.
j=1 Pd,j )(Nt {KD )d,k according
Nt (
}
K
U

2
Compute E
Pu,i |gk,i | : We can obtain
i=1
{K
} K
U
U
2
E
Pu,i |gk,i | =
Pu,i k,i .
i=1

i=1

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
12

bd,k
P
g
v


d,k
k

log2 1 + E K
.

2 K

2
2 K
D
D
U

H
H
H

bd,k vj +
Pd,j g
Pd,j d,k vj + gd,k td +
Pu,i |gk,i | + 1

ZF
Rd,k

j=k

j=1

i=1

)
(
2
Pd,k (Nt KD )d,k
ZF
Rd,k
log2 1 + KD
.
KU
2 )+
D vd
P

+
1
( j=1 Pd,j )((1 + vd KN
)

u,i
k,i
d,k
i=1
d,k
t

Substituting the above results into (45), the approximate


ZF
closed-form expression for Rd,k
, is given by (46), shown at
the top of this page.
ZF
The approximate closed-form expression for Ru,k
can be
computed following the same methodology, we can obtain

2
bu,k } = 1,
E{ wkH g

(47)

KU

2

bu,j } = 0,
E{
Pu,j wkH g

(48)

j=k

b HG
b u )1 ]kk }
E{wkH wk } = E{[(G
u
E{tr(X1 )}
1
=
= 2
,
2
KU u,k
u,k (Nr KU )

(49)

where X = HH
u Hu is a KU KU central Wishart matrix
with Nr degrees of freedom and covariance matrix IKU , and
the last equality is obtained by using [28, Lemma 2.10].

(45)

(46)

H
E{Gs GH
s } and E{wk wk }. Therefore, we can obtain
H
ZF
H
E{wkH Gs VPd VH GH
s wk } + E{wk Gs d Gs wk }

KD
(1 + vd )
(
Pd,j )(Nt KD )E{wkH Gs GH
s wk }
Nt2
j=1

KD
2

(1 + vd )SI
(
Pd,j )(Nt KD )E{wkH wk }
=
Nt
j=1
KD
2
(1 + vd )( j=1 Pd,j )(Nt KD )SI
=
.
2 (N K )
Nt u,k
r
U

(52)

Substituting (47), (48), (49), (50), (51) and (52) into (17),
ZF
the closed-form expression for Ru,k
with ZF processing, is
given by (53), shown at the top of the next page.
ZF
ZF
Given the approximate expressions of Rd,k
and Ru,k
in
(46) and (53), respectively, and considering the overhead of
channel estimation, the approximate closed-form expression
for the achievable rate can be derived as (20) in Throrem 1.
A PPENDIX D
P ROOF OF T HEOREM 2

By the convexity of log2 (1+x) and using Jensens inequality, we can obtain the upper bound of the downlink achievable
KU
M RT
KU
rate Rd,k
in (16), given by (54), shown at the top of the
2

H
2
Pu,j (u,j u,j
)
j=1
next
page.
To
compute (54), we need to compute the following
E{
Pu,j wk u,j } =
.
(50)
2 (N K )
u,k
r
U
j=1
items.

2

2
H

H

bd,k vk }: We have E{ g
bd,k vk } =
Compute E{ g

2
H

2
4
bd,k g
bd,k } = k2 d,k
k2 E{ g
E{|X| }, where X = hH
d,k hd,k is
H 2
H ZF
a
central
Wishart
matrix
and
X

W
(N
,
I).
Based
on [28,


1
t
E{ wk tu } = E{wk u wk }
2
2 4
2 4
Lemma
2.9],
we
can
obtain

E{|X|
}
=

N
k d,k t (Nt +
KU
KD
(

2 k d,k



H
2
= vu
Pu,j u,j + (1 + vd )(
Pd,j )
bd,k
1). Therefore, E{ g
vk } = d,k
(Nt + 1).
{
}
j=1
j=1
2

K
D

)


H
KD 2
bd,k

Compute
E
vj :
We
have
Pd,j g
(1
)SI + 1 E{wkH wk }
j
=
k
Nt
{
}
{
2
2 }

KU
KD
K
K
D
D

2
D


H
H
2
vu ( j=1
Pu,j u,j + (1 + vd )( j=1
Pd,j )(1 K
b
bd,k vj
b
=
P

E
g
=
P
g
g
Nt )SI + 1)E



d,j
d,j
d,j
j
d,k
=
. j=k
2
j
=
k
u,k (Nr KU )
{
}
K
D
D

(51) K
2
H b
2
bd,k
Pd,j j2 d,j
E g
gd,k =
Pd,j d,k
, due to (13) and
j=k
j=k
More
importantly,
the
self-interference
term
H


2
2
{
}
E{ wk Gs V Pd + wkH Gs td } can be rewritten (14).

2
K
D

H
ZF
H H
H
H


H
as E{wk Gs VPd V Gs wk } + E{wk Gs d Gs wk },
Compute E
Pd,j d,k vj : Since d,k and vj are
respectively.
j=1
{
}
K
D

Because V, ZF
H
d , Gs and wk are uncorrelated, to comuncorrelated, we obtain
Pd,j E bH
d,k vj vj d,k
H
pute the power of SI, we need to compute E{VPd V },
j=1

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
13

ZF
Ru,k
log2 1 + K
U

Pu,j ((1 + vu )u,j

Pu,k (Nr KU )u,k

.
K
D

KD
2
2
u,j ) + (
Pd,j )(1 Nt )SI (1 + vu + vd + vu vd ) + vu + 1

(53)

j=1

j=1

bd,k vk
P
g

d,k

log2 1 + E K
.






K
K
2
2
2
D
D
U

H



H
2

bd,k vj +
Pd,j g
Pd,j H
v
+
t
+
P
|g
|
+
1

g

u,i k,i
d,k j
d,k d

M RT
Rd,k

j=1

j=k

{
}
K
D

2
2
Pd,j k2 d,j
E H
Pd,j )(d,k d,k
).
d,k d,k = (
j=1
j=1
{
}

H 2
H
Compute E gd,k
td : Since gd,k
and td are
}
{
2
{
}
H
H
RT
uncorrelated, E gd,k
td
= E gd,k
M
gd,k
=
d
{
}
KD

KD
vd
H
( j=1 Pd,j ) N
E gd,k
gd,k = ( j=1
Pd,j )vd d,k accordt
ing to Lemma2.
{K
}
U
2
Compute E
Pu,i |gk,i | : We can obtain
i=1
{K
} K
U
U
2
E
Pu,i |gk,i | =
Pu,i k,i .
=

K
D

i=1

i=1

Substituting the above results into (16), the approximate


M RT
closed-form expression for Rd,k
, is given by (55), shown at
the top of the next page.
The approximate achievable rate of UL UEs can be computed following the same methodology, we can obtain
2

4
bu,k } = Pu,k u,k
Nr (Nr + 1),
E{ wkH g
KU
KU


2
2
2
bu,j } = (
E{
Pu,j wkH g
Pu,j u,j
)u,k
Nr ,
j=k

(56)

i=1

The self-interference term can be computed as


H
M RT
E{wkH Gs VPd VH GH
GH
s wk } + E{wk Gs d
s wk }
H
H H
b
b
bu,k }
= E{b
g Gs Gd Pd G G g

u,k
d
s
H
M RT
H
bu,k }
+ E{b
gu,k Gs d
Gs g
KD

2
H
bu,k }
(1 + vd )
Pd,j d,j
j2 E{b
gu,k
Gs GH
s g
j=1

= (1 + vd )

KD

= (1 + vd )(

KD

(61)
Substituting (56), (57), (58), (59), (60) and (61) into (17),
M RC
the approximate closed-form expression for Ru,k
, is given
by (62), shown at the top of the next page.
M RT
M RC
Given the approximate expressions of Rd,k
and Ru,k
in (55) and (62), respectively, and considering the overhead of
channel estimation, the approximate closed-form expression
for the achievable rate can be derived as (21) in Theorem 2.
A PPENDIX E
P ROOF OF T HEOREM 3
(n)

KU
KU


2
2
2
E{
Pu,j wkH u,j } =
Pu,j (u,j u,j
)u,k
Nr . (59)
j=1

j=1

KU


2
RC
E{ wkH tu } = E{wkH M
w
}
=
v

(
pu,j u,j
k
u
u
j=1

+ (1 + vd )

KU

2
Pd,j SI
+ 1)E{wkH wk }

j=1
KU

= vu (

j=1

Pu,j u,j + (1 + vd )

KU

2
2
Pd,j SI
+ 1)u,k
Nr .

j=1

(60)

2
2
Pd,j )SI
u,k
Nr .

j=1

(57)

(58)

2
2
H
bu,k }
g
Pd,j d,j
j2 Nt SI
E{b
gu,k

j=1

j=k

H
2
bu,k } = u,k
g
Nr ,
E{wkH wk } = E{b
gu,k

(54)

(n)

(n)

(n)

Suppose {rd,k }, {d,k }, {ru,k }, {u,k } are optimal values


of Problem (37) at iteration n and are chosen as the linear
approximation points at iteration n+1 according to the up(n+1)
(n) (n)
(n+1)
(n)
(n)
dating rule d,k
= rd,k /d,k and u,k
= ru,k /u,k in
Algorithm 1. Consider the (n + 1)st iteration of Algorithm
(n)

(n)

(n+1)

(a)

(n)

(n)

(b)

1, we can obtain g(rd,k , d,k , d,k ) = f (rd,k , d,k )


(c)
KU (n)
(n) (n) (n)
(n)
KD
g(rd,k , d,k , d,k ) ( j=1
Pd,j )bk + j=1
Pu,j ck,j + 1 +
(n)
Pd,k ak . Step (a) is due to the updating rule in Algorithm 1
(n+1)
(n) (n)
(i.e., d,k
= rd,k /d,k ) and the property (32), step (b) is
(n)
due to the fact that g(rd,k , d,k , d,k ) is an upper bound of
f (rd,k , d,k ) according to (31) and step (c) is due to the second
constraint of Problem (37) at the nst iteration. Similarly, we
KU (n)
KD (n)
(n) (n) (n+1)
have g(ru,k , u,k , d,k ) j=1
Pu,j ej + ( j=1
Pd,j )f +
(n)
1 + Pu,k dk . This means that the optimal solution of the nth
iteration is a feasible point of the problem in the (n + 1)st
iteration and thus u(n+1) u(n) where u(n) is the optimum
of (37) at iteration n. Furthermore, the value of u(n) is

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
14

M RT
Rd,k

(
log2 1 +

(
M RC
Ru,k

log2 1 +

K
U

2
Pu,j u,j

j=k

(1 + vd )(

KD
j=1

2
Pd,k d,k
(Nt + 1)
2 +
Pd,j )d,k Pd,k d,k

A PPENDIX F
P ROOF OF T HEOREM 4
Because the subproblem (43) has continuous and quasiconvex objective function and its feasible set is compact, the
primal Dinkelbach-type algorithm can guarantee the convergence point of (44) converges to the optimal solution of (43)
according to [27, Theorem 8.7]. Then the following important
property of the linear approximation (42) is used to show that
the sequence produced by iteratively solving (43) is convergent
(n)

(n)

ge(n+1) (Pd,k (n) , Pu,k (n) ) = g(Pd,k , Pu,k ),

(63)

where the optimal value of (43) at nst iteration


{Pd,k (n) }, {Pu,k (n) } is chosen as the linear approximation
point at iteration n + 1 according to the updating rule in
Algorithm 2. Further letting u(n) be the optimal objective of
(43) at iteration n, we get
(n+1)

(a)

(n+1)

(n+1)

(n+1)

h(Pd,k , Pu,k ) ge(n+1) (Pd,k , Pu,k


KD (n+1) KU (n+1)
+ k=1 Pu,k + Pcir
k=1 Pd,k
(n)

(n)

(n)

(n)

h(Pd,k , Pu,k ) ge(n+1) (Pd,k , Pu,k )


KD (n) KU (n)
k=1 Pd,k +
k=1 Pu,k + Pcir
(n)

(n)

(n)

(n)

h(Pd,k , Pu,k ) g(Pd,k , Pu,k )


= K
KU (n)
(n)
D
k=1 Pd,k +
k=1 Pu,k + Pcir

(b)

(n)

(n)

(n)

(n)

h(Pd,k , Pu,k ) ge(n) (Pd,k , Pu,k )


K
= u(n)
KU (n)
(n)
D
P
+
P
+
P
cir
k=1 d,k
k=1 u,k

(c)

KU
i=1

Pu,i k,i + 1

(55)

)
2
Pu,k u,k
(Nr + 1)
. (62)
KD
K U
2 )+(
2 (1 + v + v v ) + v + 1
P
)
+ j=1
Pu,j ((1 + vu )u,j u,j
u
u d
u
SI
j=1 d,j

bounded above due to the limited transmit power, and thus


it is guaranteed to converge.
Because the objective function of (37) is a continuously
differentiable and admits an convex SOC representation, and
the constraints of (37) form a nonempty convex and compact
set, then the objective function is strongly convex based on
[26, Lemma 3.1]. With the update rule in Algorithm 1 and
properties (32) and (33), the limit point of the sequence
generated by Algorithm 1 are KKT points of problem (23)
based on [26, Proposition 3.2].

u(n+1) =

(64)
(n+1)
(n+1)
where step (a) is due to the optimality of {Pd,k }, {Pu,k },
step (b) follows from property (63) and step (c) is due to
the fact that ge(n) (Pd,k (n) , Pu,k (n) ) is an upper bound of
g(Pd,k (n) , Pu,k (n) ) at {Pd,k (n) }, {Pu,k (n) } according to (42).
Therefore, the sequence {u(n) } is nondecreasing. Moreover,
its bounded due to the limited transmit power, so the proposed
algorithm 2 is locally convergent.

R EFERENCES
[1] J. Andrews, S. Buzzi, W. Choi, S. Hanly, A. Lozano, A. Soong, and
J. Zhang, What will 5G be? IEEE J. Sel. Areas Commun., vol. 32,
no. 6, pp. 10651082, Jun. 2014.
[2] F. Boccardi, R. Heath, A. Lozano, T. Marzetta, and P. Popovski, Five
disruptive technology directions for 5G, IEEE Commun. Mag., vol. 52,
no. 2, pp. 7480, Feb. 2014.
[3] H. Q. Ngo, E. Larsson, and T. Marzetta, Energy and spectral efficiency
of very large multiuser MIMO systems, IEEE Trans. Commun., vol. 61,
no. 4, pp. 14361449, Apr. 2013.
[4] F. Rusek, D. Persson, B. K. Lau, E. Larsson, T. Marzetta, O. Edfors,
and F. Tufvesson, Scaling up MIMO: Opportunities and challenges with
very large arrays, IEEE Signal Process. Mag., vol. 30, no. 1, pp. 4060,
Jan. 2013.
[5] T. Marzetta, Noncooperative cellular wireless with unlimited numbers
of base station antennas, IEEE Trans. Wireless Commun., vol. 9, no. 11,
pp. 35903600, Nov. 2010.
[6] E. Larsson, O. Edfors, F. Tufvesson, and T. Marzetta, Massive MIMO
for next generation wireless systems, IEEE Commun. Mag., vol. 52,
no. 2, pp. 186195, Feb. 2014.
[7] Z. Zhang, X. Chai, K. Long, A. Vasilakos, and L. Hanzo, Full duplex
techniques for 5G networks: self-interference cancellation, protocol
design, and relay selection, IEEE Commun. Mag., vol. 53, no. 5, pp.
128137, May. 2015.
[8] M. Duarte and A. Sabharwal, Full-duplex wireless communications
using off-the-shelf radios: Feasibility and first results, in Proc. 2010
AsilomarConf. Signals, Syst., Comput., Nov 2010, pp. 15581562.
[9] A. Sabharwal, P. Schniter, D. Guo, D. Bliss, S. Rangarajan, and
R. Wichman, In-band FD wireless: Challenges and opportunities, IEEE
J. Sel. Areas Commun., vol. 32, no. 9, pp. 16371652, Oct. 2014.
[10] M. Duarte, C. Dick, and A. Sabharwal, Experiment-driven characterization of full-duplex wireless systems, IEEE Trans. Wireless Commun.,
vol. 11, no. 12, pp. 42964307, Dec. 2012.
[11] S. Huberman and T. Le-Ngoc, Self-interference threshold-based MIMO
full-duplex precoding, IEEE Trans. Veh. Technol., vol. 64, no. 8, pp.
38033807, 2014.
[12] T. Riihonen, S. Werner, and R. Wichman, Mitigation of loopback selfinterference in full-duplex MIMO relays, IEEE Trans. Signal Process.,
vol. 59, no. 12, pp. 59835993, Dec. 2011.
[13] H. Suraweera, I. Krikidis, G. Zheng, C. Yuen, and P. Smith, Lowcomplexity end-to-end performance optimization in MIMO full-duplex
relay systems, IEEE Trans. Wireless Commun., vol. 13, no. 2, pp. 913
927, Feb. 2014.
[14] T. Riihonen, S. Werner, and R. Wichman, Hybrid full-duplex/halfduplex relaying with transmit power adaptation, IEEE Trans. Wireless
Commun., vol. 10, no. 9, pp. 30743085, Sep. 2011.
[15] D. Nguyen, L.-N. Tran, P. Pirinen, and M. Latva-Aho, Precoding for
full duplex multiuser MIMO systems: Spectral and energy efficiency
maximization, IEEE Trans. Signal Process., vol. 61, no. 16, pp. 4038
4050, Aug. 2013.
[16] B. Li, D. Zhu, and P. Liang, Small cell in-band wireless backhaul in
massive MIMO systems: A cooperation of next-generation techniques,
IEEE Trans. Wireless Commun., vol. 14, no. 12, pp. 70577069, Dec.
2015.
[17] H. Q. Ngo, H. A. Suraweera, M. Matthaiou, and E. G. Larsson, Multipair massive MIMO full-duplex relaying with MRC/MRT processing,
in Proc. IEEE ICC, Jun. 2014, pp. 48184824.
[18] H. Q. Ngo, H. Suraweera, M. Matthaiou, and E. G. Larsson, Multipair
full-duplex relaying with massive arrays and linear processing, IEEE
J. Sel. Areas Commun., vol. 32, no. 9, pp. 17211737, Sep. 2014.

0018-9545 (c) 2016 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.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TVT.2016.2577636, IEEE
Transactions on Vehicular Technology
15

[19] X. Xia, D. Zhang, K. Xu, W. Ma, and Y. Xu, Hardware impairments


aware transceiver for full-duplex massive MIMO relaying, IEEE Trans.
Signal Process., vol. 63, no. 24, pp. 65656580, Dec. 2015.
[20] Y. Jang, K. Min, S. Park, and S. Choi, Spatial resource utilization to
maximize uplink spectral efficiency in full-duplex massive MIMO, in
Proc. IEEE Int. Conf. on Commun. (ICC), Jun. 2015, pp. 15831588.
[21] D. Nguyen, L. N. Tran, P. Pirinen, and M. Latva-aho, On the spectral
efficiency of full-duplex small cell wireless systems, IEEE Trans.
Wireless Commun., vol. 13, no. 9, pp. 48964910, Sep. 2014.
[22] System scenarios and technical requirements for full-duplex concept,
DUPLO project, deliverable d1.1. [online], Available: http://www.fp7duplo.eu/index.php/deliverables,.
[23] E. Bjornson, J. Hoydis, M. Kountouris, and M. Debbah, Massive
MIMO systems with non-ideal hardware: Energy efficiency, estimation,
and capacity limits, IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7112
7139, Nov. 2014.
[24] U. Gustavsson, C. Sanchz-Perez, T. Eriksson, F. Athley, G. Durisi,
P. Landin, K. Hausmair, C. Fager, and L. Svensson, On the impact of
hardware impairments on massive MIMO, in Proc. IEEE GLOBECOM,
Dec 2014, pp. 294300.
[25] E. Bjornson, M. Matthaiou, A. Pitarokoilis, and E. G. Larsson, Distributed massive MIMO in cellular networks: Impact of imperfect
hardware and number of oscillators, in Proc. EUSIPCO, Aug 2015,
pp. 24362440.
[26] A. B.-T. A. Beck and L. Tetruashvili, A sequential parametric convex
approximation method with applications to nonconvex truss topology
design problems, SJ. Global Optim., vol. 47, no. 12, pp. 2951, 2010.
[27] J. B. Frenk and S. Schaible, Handbook of Generalized Convexity and
Generalized Monotonicity. New York: SpringerVerlag, 2006.
[28] A. M. Tulino and S. Verd, Random matrix theory and wireless
communications, Found. Trends Commun. Inf. Theory, vol. 1, no. 1,
p. 1C182, Jun. 2004.
[29] L. Tran, M. Hanif, A. Tolli, and M. Juntti, Fast converging algorithm
for weighted sum rate maximization in multicell MISO downlink, IEEE
Commun. Lett., vol. 19, no. 12, pp. 872875, Dec. 2012.
[30] CVX: Matlab Software for Disciplined Convex Programming, version
2.0, http://cvxr.com/cvx, Aug. 2012.
[31] M. Lin, J. Ouyang, and W. Zhu, Joint beamforming and power control
for device-to-device communications underlaying cellular networks,
IEEE J. Sel. Areas Commun., vol. 34, no. 1, pp. 138150, Jan. 2016.
[32] 3GPP TR 36.828 v11.0.0, Technical Specification Group Radio Access
Network; Evolved Universal Terrestrial Radio Access (EUTRA); Further
Enhancements to LTE Time Division Duplex (TDD) for DownlinkUplink (DL-UL) Interference Management and Traffic Adaptation (Release 11), 3GPP 3rd Generation Partnership Project, Jun. 2012.

Pingzhi Fan (M93-SM99-F15) received his PhD


degree in Electronic Engineering from the Hull University, UK. He is currently a professor and director
of the Institute of Mobile Communications, Southwest Jiaotong University, China. He is a recipient of
the UK ORS Award, the NSFC Outstanding Young
Scientist Award, and chief scientist of a national 973
major project. He served as general chair or TPC
chair of a number of international conferences, as
well as guest editor-in-chief, guest editor or editorial
member of several international journals. He is the
founding chair of IEEE VTS BJ Chapter, IEEE ComSoc CD Chapter, and
IEEE Chengdu Section. He also served as a board member of IEEE Region
10, IET(IEE) Council, and IET Asia-Pacific Region. He has over 200 research
papers published in IEEE and other academic journals, and 8 books (incl.
edited) published by John Wiley & Sons Ltd, Springer, IEEE Press, etc.
His current research interests include high mobility wireless communications,
5G technologies, wireless networks for big data, etc. He is an IEEE VTS
Distinguished Lecturer (2015-2017), and a fellow of IEEE, IET, CIE and
CIC.

Anatolii Leukhin received his Ph.D. degree in


Physics and Mathematical Sciences from Samara
State Aerospace University, Samara, Russia. He is
currently a professor and vice-rector in the Institute
of Science and Innovations of Mari State University.
He is a recipient of the Medal of Russian Space Federation in 1999, State Honour For Science Research
and Technical Applications in 2001 and Award of
Russian Ministry of Science and Education in 2010.
His current research interests include sequence design, digital signal and image processing and grid
computing.

Lingjia Liu (M07-SM15) received his B.S. degree in Electronic Engineering from Shanghai Jiao
Tong University and Ph.D. degree in Electrical and
Computer Engineering from Texas A&M University.
Prior to joining the faculty in the EECS at KU, he
spent more than three years at Samsung Research
America where he received the Global Samsung Best
Paper Award in 2008 and 2010. He was leading
Samsungs efforts on downlink multiuser MIMO,
coordinated multipoint (CoMP), and heterogeneous
networks (HetNet) in LTE/LTE-Advanced standards.

Yi Li (S14) received the B.S. in Traffic Information


Engineering & Control in 2012 from Southwest
Jiaotong University, Chengdu, China. He is currently
working toward the Ph.D. degree with the Key
Laboratory of Information Coding and Transmission at Southwest Jiaotong University. His research
interests include the interference management in
multicell networks, the optimization of full duplex
and massive MIMO systems, and the non-orthogonal
multiple access techniques.

Dr. Liu is currently an Editor for the IEEE Trans. Wireless Commun., an
Editor for the IEEE Trans. Commun. and Associate Editor for the EURASIP
J. on Wireless Commun. and Netw. and Wileys Intl J. on Commun. Systems.
He is co-editor of several special issues. His general research interests lie in
emerging technologies for 5G cellular networks including massive MIMO and
mmW wireless communications. He has been serving as the Workshop Chair
for the past four IEEE GLOBECOM 5G Workshops. Dr. Liu was selected by
the National Engineers Week Foundation Diversity Council as New Faces of
Engineering 2011, Air Force Summer Faculty Fellow in 2013, 2014, 2015
and 2016, Miller Scholar at KU in 2014 and Miller Professional Development
Award for Distinguished Research in 2015.

0018-9545 (c) 2016 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.

Das könnte Ihnen auch gefallen