Multi-Pair Two Way AF Full-Duplex Massive
MIMO Relaying with ZFR/ZFT Processing
Ekant Sharma, Rohit Budhiraja and K Vasudevan
Department of Electrical Engineering, IIT Kanpur, India
email: {ekant, rohitbr, vasu}@iitk.ac.in
Abstract—We consider two-way amplify and forward
relaying, where multiple full-duplex user pairs exchange
information via a shared full-duplex massive multiple-
input multiple-output (MIMO) relay. We derive closed-
form lower bound for the spectral efficiency with zero-
forcing processing at the relay, by using minimum mean
squared error channel estimation. The zero-forcing lower
bound for the system model considered herein, which is
valid for arbitrary number of antennas, is not yet derived
in the massive MIMO relaying literature. We numerically
demonstrate the accuracy of the derived lower bound and
the performance improvement achieved using zero-forcing
processing. We also numerically demonstrate the spectral
gains achieved by a full-duplex system over a half-duplex
one for various antenna regimes.
Index Terms—Full-duplex, relay, spectral efficiency.
I. I NTRODUCTION
Relay based communication is being extensively in-
vestigated to expand the coverage, improve the diversity,
increase the data rate and reduce the power consumption
of wireless communication systems [1], [2]. The current
generation relays are mostly half-duplex due to their
implementation simplicity. Full-duplex technology is be-
coming popular after recent studies, e.g., [3], [4], demon-
strated a significant reduction in the loop interference,
caused due to transmission and reception on the same
channel. A full-duplex relay [5], commonly known as
full-duplex one-way relay, transmits and receives on the
same channel, and can theoretically double the spectral
efficiency, when compared with a half-duplex one-way
relay [1], [2].
Full-duplex two-way relaying [6], wherein two users
exchange two data units in one channel use via a relay,
further improves the spectral efficiency. Two-way full-
duplex relaying is recently extended to multi-pair two
-way full-duplex relaying [7]–[9] wherein multiple user
pairs exchange data via a shared relay in a single chan-
nel use. A multi-pair two-way full-duplex relay system
has following interference sources: i) co-channel (inter-
pair) interference due to multiple users simultaneously
accessing the channel; ii) loop interference at the relay
and at the users; and iii) inter-user interference caused
due to simultaneous transmission and reception by full-
duplex nodes.
978-1-5386-3531-5/17/$31.00 ⃝
c 2017 IEEE
Massive multiple-input multiple-output (MIMO) sys-
tems have become popular as they cancel co-channel
interference by using simple linear transmit process-
ing schemes e.g., zero-forcing transmission (ZFT) and
maximal-ratio transmission (MRT) [10], [11], and signif-
icantly improve the spectral efficiency. Massive MIMO
technology is also being incorporated in multi-pair full-
duplex relays to cancel the loop interference at the relay,
and inter-pair co-channel interference [7]–[9], [12], [13].
Reference [7] derived the achievable rate and a power
allocation scheme to maximize the ergodic sum-rate
for one-way decode and forward full-duplex massive
MIMO relaying. Zhang et al. in [8] proposed four
power scaling schemes for two-way full-duplex massive
MIMO relaying to improve its spectral and energy
efficiency. Reference [9] developed a power allocation
scheme to maximize the sum-rate for multi-pair two-way
full-duplex massive MIMO amplify-and-forward (AF)
relaying by using maximal-ratio combining (MRC)/MRT
processing at the relay, and by using least squares (LS)
channel estimation. Dai et al. in [12] considered a half-
duplex multi-pair two-way massive MIMO AF relay and
derived closed-form achievable rate expressions and a
power allocation scheme to maximize the sum-rate with
imperfect channel state information (CSI). The authors
in [13] developed power scaling schemes for half-duplex
massive MIMO one-way relay systems.
The authors in [9] have derived the spectral-efficiency
lower bound for MRC/MRT relay processing with LS
channel estimation. We extend the work done in [9], and
next list the main contributions of this paper.
∙ We derive closed-form lower bound for the spec-
tral efficiency of the multi-pair two-way AF full-
duplex massive MIMO relay for arbitrary num-
ber of relay antennas. We consider zero-forcing
reception (ZFR)/zero-forcing transmission (ZFT)
processing at the relay and minimum-mean-square-
error (MMSE) relay channel estimation. We note
that the bound obtained for MRC/MRT processing
based on LS channel estimation in [9] cannot be
trivially extended to the ZFR/ZFT processing with
MMSE channel estimation, considered herein. This
closed-form spectral-efficiency lower bound, with
arbitrary number of relay antennas, to the best of
 our knowledge, have not yet been derived in the
massive MIMO relaying literature.
∙ We also numerically demonstrate the considerable
spectral efficiency gains achieved due to MMSE
channel estimation and ZFR/ZFT processing.
II. S YSTEM M ODEL
We consider multi-pair two-way AF full-duplex re-
laying as shown in Fig. 1, where 𝐾 full-duplex user
pairs communicate via a single full-duplex relay on the
same time-frequency resource. We assume that the user
𝑆2𝑚−1 for 𝑚 = 1 to 𝐾 on one side of the relay, wants
to send as well as receive from the user 𝑆2𝑚 that is on
the other side of the relay. We also assume that there is
no direct link between the user-pairs (𝑆2𝑚−1 , 𝑆2𝑚 ) on
the either side of the relay due to large path loss and
heavy shadowing. Also the relay has 𝑁 transmit and
𝑁 receive antennas, while each user has one transmit
and one receive antenna. The users on either side of
the relay, due to full-duplex architecture, interfere with
each other; the interference caused is termed as inter-user
interference.
Fig. 1: Multi-pair two-way full-duplex AF massive MIMO relay
system.
At time instant 𝑛, each user 𝑆𝑘 , 𝑘 = 1 to 2𝐾,
√
transmits the signal 𝑝𝑘 𝑥𝑘 (𝑛) to the relay, and simulta-
neously the relay broadcasts x𝑅 (𝑛) ∈ ℂ𝑁 ×1 to all users.
Here 𝑝𝑘 denotes the transmit power of the 𝑘th user. The
received signal at the relay and the user 𝑆𝑘 are given as
∑2𝐾
√
y𝑅 (𝑛) = 𝑝𝑘 g𝑘 𝑥𝑘 (𝑛) + GRR x𝑅 (𝑛) + z𝑅 (𝑛)
𝑘=1
= G̃x(𝑛) + GRR x𝑅 (𝑛) + z𝑅 (𝑛), (1)
∑ √
𝑇
𝑦𝑘 (𝑛) = f𝑘 x𝑅 (𝑛) + Ω𝑘,𝑖 𝑝𝑖 𝑥𝑖 (𝑛) + 𝑧𝑘 (𝑛).
𝑖,𝑘∈𝑈𝑘
(2)
We denote the matrix G = [g1 , g2 , g3 , ⋅ ⋅ ⋅ , g2𝐾 ] ∈
ℂ𝑁 ×2𝐾 and the matrix (to be used later in the sequel)
F = [f1 , f2 , f3 , ⋅ ⋅ ⋅ , f2𝐾 ] ∈ ℂ𝑁 ×2𝐾 , where g𝑘 and f𝑘
denote the channels from the transmit antenna of the
𝑘th user to the relay receive antenna array, and from
the relay’s transmit antenna array to the receive antenna
of the 𝑘th{ user, respectively. Further,
} G̃ = GP where
√ √ √
P = diag 𝑝1 , 𝑝2 , ⋅ ⋅ ⋅ 𝑝2𝐾 with 0 ≤ 𝑝𝑖 ≤ 𝑃0 and
∑2𝐾
𝑖=1 𝑝𝑖 = 𝑃 . The receive signal at the relay and the
users are interfered by their own transmit signal, which
is called as the self-loop interference. Here G𝑅𝑅 and
Ω𝑘,𝑘 denote the self-loop interference at the relay and
the user 𝑆𝑘 . The entries of the matrix G𝑅𝑅 and the
scalar Ω𝑘,𝑘 are independent and identically distributed
2 2
(i.i.d.) with distribution 𝒞𝒩 (0, 𝜎𝐿𝐼𝑅 ) and 𝒞𝒩 (0, 𝜎𝑘,𝑘 ),
respectively. The term Ω𝑘,𝑖 (𝑘, 𝑖 ∈ 𝑈𝑘 , 𝑖 ∕= 𝑘) denote the
inter-user interference channel, which is modeled as i.i.d.
2
𝒞𝒩 (0, 𝜎𝑘,𝑖 ), where the set 𝑈𝑘 = [1, 3, 5, ⋅ ⋅ ⋅ , 2𝐾 − 1]
for odd 𝑘 and 𝑈𝑘 = [2, 4, 6, ⋅ ⋅ ⋅ , 2𝐾] for even 𝑘. The
𝑇
vector x(𝑛) =[ [𝑥1 (𝑛), 𝑥2 (𝑛),] 𝑥3 (𝑛), ⋅ ⋅ ⋅ , 𝑥2𝐾 (𝑛)] ∈
𝑁 ×1 𝐻
ℂ with 𝔼 x(𝑛)x (𝑛) = I2𝐾 . The vector z𝑅 (𝑛) ∈
ℂ𝑁 ×1 and the scalar 𝑧𝑘 (𝑛) are additive white Gaussian
noise (AWGN) at the relay and the user 𝑆𝑘 . The ele-
ments of z𝑅 and the scalar 𝑧𝑘 (𝑛) are modeled as i.i.d.
2
𝒞𝒩 (0, 𝜎𝑛𝑟 ) and 𝒞𝒩 (0, 𝜎𝑛2 ), respectively.
The channel matrices account for both small-scale and
1/2
large-scale fading; we therefore express G = H𝑢 D𝑢
1/2
and F = H𝑑 D𝑑 . Here the small-scale fading matrices
H𝑢 and H𝑑 have i.i.d. 𝒞𝒩 (0, 1) elements, while the
𝑘th element of large-scale diagonal fading matrices D𝑢
2 2
and D𝑑 are denoted as 𝜎𝑔,𝑘 and 𝜎𝑓,𝑘 , respectively. The
inter-element distance is assumed to be smaller than the
distance between the transmit array and the receive array
which leads to the channel between the transmit and
receive antennas to be independent.
In the first time slot (𝑛 = 1), the relay only receives
the signal and does not transmit. The signals received at
the relay and the user 𝑆𝑘 are given respectively as
y𝑅 (1) = G̃x(1) + z𝑅 (1)
∑ √
𝑦𝑘 (1) = Ω𝑘,𝑖 𝑝𝑖 𝑥𝑖 (1) + 𝑧𝑘 (1). (3)
𝑖,𝑘∈𝑈𝑘
At the 𝑛th time slot, the relay linearly precodes its
received signal y𝑅 (𝑛 − 1) using a matrix W such that
x𝑅 (𝑛) = 𝛼Wy𝑅 (𝑛 − 1), (4)
where 𝛼 is the scaling factor chosen to satisfy the relay
power constraint. The relay transmit signal x𝑅 , similar
to [8], can be re-expressed using (1) as
x𝑅 (𝑛) = 𝑠 (x(𝑛 − 𝜈) + x(𝑛 − 2𝜈) + ⋅ ⋅ ⋅
+ z𝑅 (𝑛 − 𝜈) + z𝑅 (𝑛 − 2𝜈) + ⋅ ⋅ ⋅ ) . (5)
Here 𝑠(⋅) is a function involving vector and matrix
operation, and 𝜈 is the relay processing delay (𝜈 = 1
in this paper). There are various solution proposed in
the relaying literature, e.g., [14], that significantly sup-
press the self-loop interference caused due to x𝑅 such
that the residual self-loop interference can be replaced
with x̃𝑅 (𝑛), an additional Gaussian noise source with
 { [ ]}
𝔼 x̃𝑅 (𝑛)x̃𝐻
𝑅 (𝑛) = 𝑃𝑁𝑅 I𝑁 [8]. Therefore, the relay lowing signal at the receive and transmit antenna array
√
receive signal in (1) can be re-expressed as Y𝑅,𝑅 = 𝜏 𝑃𝜌 G𝜑 + N𝑅,𝑅 , and
ỹ𝑅 (𝑛) = G̃x(𝑛) + GRR x̃𝑅 (𝑛) + z𝑅 (𝑛). (6) √
Y𝑅,𝑇 = 𝜏 𝑃𝜌 F𝜑 + N𝑅,𝑇 (12)
We re-write the relay transmit signal in (4), using (6), as √ 2𝐾×𝜏
where 𝜏 𝑃𝜌 𝜑 ∈ ℂ denotes the pilot symbols
x𝑅 = 𝛼Wỹ𝑅 (𝑛 − 1). (7) transmitted from all users with 𝑃𝜌 being the transmit
For the sake of brevity, we will drop the time labels. power of each pilot symbol. The matrices N𝑅,𝑅 and
Using (6), we re-write (4) as N𝑅,𝑇 denote the AWGN noise whose elements are i.i.d.
x𝑅 = 𝛼Wỹ𝑅 = 𝛼WG̃x + 𝛼WGRR x̃𝑅 + 𝛼Wz𝑅 . and distributed as 𝒞𝒩 (0, 1). The pilots are assumed to be
(8) orthogonal such that 𝜑𝜑𝐻 = I2𝐾 for 𝜏 ≥ 2𝐾 [15]. The
MMSE channel estimate of G and F are given by [16]
The relay transmit signal should satisfy its transmit
1 N𝑅,𝑅 𝜑𝐻
power constraint such that
{ [ ]} Ĝ = √ Y𝑅,𝑅 𝜑𝐻 D̄𝑢 = GD̄𝑢 + √ D̄𝑢
P𝑅 = Tr 𝔼 x𝑅 x𝐻 𝜏 𝑃𝜌 𝜏 𝑃𝜌
[ ] 𝑅 [ ] [ ] 1 N𝑅,𝑇 𝜑𝐻
=𝔼 ∥𝛼WG̃x∥2 +𝔼 ∥𝛼WGRR x̃𝑅 ∥2 +𝔼 ∥𝛼Wz𝑅 ∥2 , F̂ = √ Y𝑅,𝑇 𝜑𝐻 D̄𝑑 = FD̄𝑑 + √ D̄𝑑 .
𝜏 𝑃𝜌 𝜏 𝑃𝜌
(9) (13)
which leads to the following value of the scaling factor

The matrices[ Ĝ = [ĝ1 , ĝ2], ⋅ ⋅ ⋅ , ĝ2𝐾 ] ∈ ℂ𝑁 ×2𝐾
 𝑃𝑅
 and F̂ = f̂1 , f̂2 , ⋅ ⋅ ⋅ , f̂2𝐾 ∈ ℂ𝑁 ×2𝐾 , D̄𝑢 =
𝛼=⎷ [ ] .
𝔼 ∥WG̃x∥2 +𝔼 [∥WGRR x̃𝑅 ∥2 ]+𝔼 [∥Wz𝑅 ∥2 ] ( −1 )−1 ( −1 )−1
D𝑢 D
𝜏 𝑃𝜌 + I2𝐾 and D̄𝑑 = 𝜏 𝑃𝑑𝜌 + I2𝐾 .
(10)
We note that the elements of N𝐻 𝑅,𝑅 𝜑 and Ñ𝑅,𝑇 𝜑
𝐻
We next re-express the received signal after self-
are distributed as 𝒞𝒩 (0, 1). We therefore have Ĝ =
interference cancellation (SIC) at the user 𝑆𝑘 given in
G − E𝑔 and F̂ = F − E𝑓 , where E𝑔 and E𝑓 are
(2), using (8), as
√ √ estimation error matrices. The channel matrices Ĝ and
𝑦˜𝑘 = 𝛼f𝑘𝑇 W 𝑝𝑘′ g𝑘′ 𝑥𝑘′ + 𝛼 𝑝𝑘 𝜆𝑘 𝑥𝑘 F̂ are independent of the error matrices E𝑔 and E𝑓 ,

 
 
desired signal residual interference respectively [16]. The matrices Ĝ and F̂ are dis-
2𝐾
∑ √ tributed as 𝒞𝒩 (0, D̂𝑢 ) and
{ 2 𝒞𝒩 (0, D̂𝑑 ), respectively;
}
+ 𝛼f𝑘𝑇 W 𝑝𝑖 g𝑖 𝑥𝑖 + 𝛼f𝑘𝑇 WGRR x̃𝑅 the matrices D̂ 𝑢 = diag 𝜎
ˆ , 𝜎
ˆ 2
𝑔,2 ,}⋅ ⋅ ⋅ , 𝜎
2
ˆ𝑔,2𝐾 and

  { 𝑔,1
𝑖∕=𝑘,𝑘′ 2 2 2 2

  amplified loop interference D̂𝑑 = diag 𝜎
ˆ𝑓,1 ,𝜎
ˆ𝑓,2 , ⋅⋅⋅ , 𝜎
ˆ𝑓,2𝐾 , with 𝜎
ˆ𝑔,𝑘 =
4 4
inter-pair interference 𝜏 𝑃𝜌 𝜎𝑔,𝑘 2 𝜏 𝑃𝜌 𝜎𝑓,𝑘
∑ √ 2 +1
𝜏 𝑃𝜌 𝜎𝑔,𝑘
and 𝜎
ˆ𝑓,𝑘 = 2 +1 .
𝜏 𝑃𝜌 𝜎𝑓,𝑘
Hence, Eg ∼
+ 𝛼f𝑘𝑇 Wz𝑅 + Ω𝑘,𝑖 𝑝𝑖 𝑥𝑖 + 𝑧𝑘 .

 
 𝒞𝒩 (0, D𝑢 −D̂{𝑢 ) and Ef ∼ 𝒞𝒩 (0, D𝑑 −D̂𝑑 ), with}D𝑢 −
𝑖,𝑘∈𝑈 𝑘
amplified noise from relay
  AWGN at 𝑆𝑘 D̂𝑢 = diag 𝜎𝜉,𝑔,12
, 𝜎 2 , 𝜎 2 , ⋅ ⋅ ⋅ , 𝜎𝜉,𝑔,2𝐾
2
and
self loop interference { 𝜉,𝑔,2 𝜉,𝑔,3 }
2 2 2 2
and inter-user interference
D𝑑 − D̂𝑑 = diag 𝜎𝜉,𝑓,1 , 𝜎𝜉,𝑓,2 , 𝜎𝜉,𝑓,3 , ⋅ ⋅ ⋅ , 𝜎𝜉,𝑓,2𝐾
(11) 2 2
2 𝜎𝑔,𝑘 2 𝜎𝑓,𝑘
′
Here (𝑘, 𝑘 ) = (2𝑚 − 1, 2𝑚) or (2𝑚, 2𝑚 − 1), for with 𝜎𝜉,𝑔,𝑘 = 2 +1
𝜏 𝑃𝜌 𝜎𝑔,𝑘
and 𝜎𝜉,𝑓,𝑘 = 2 +1 .
𝜏 𝑃𝜌 𝜎𝑓,𝑘
𝑚 = 1, 2, 3, ⋅ ⋅ ⋅ , 𝐾 denotes the user pair which ex-
IV. R ELAY PRECODER DESIGN
change information with one another. The scalar 𝜆𝑘 =
f𝑘𝑇 Wg𝑘 − f̂𝑘𝑇 Wgˆ𝑘 is the residual self-interference. In We design relay precoder based on ZFR/ZFT pro-
this work we assume that the relay estimates channels cessing. The ZFR/ZFT matrix using estimated CSI is
G and F and uses them to design the precoder W. given by
The relay then transmits the SIC coefficient f̂𝑘𝑇 Wĝ𝑘 for W = F̄ ˆ 𝐻,
ˆ ∗ TḠ (14)
each user, where f̂𝑘 and ĝ𝑘 are the estimated channel ( )−1 ( )−1
ˆ
where F̄ = F̂ F̂ F̂𝐻 ˆ
and Ḡ = Ĝ Ĝ Ĝ𝐻
. We
coefficients. Before designing the relay precoder W, we
briefly digress to discuss the MMSE channel estima- next state the following proposition to simplify the relay
tion process. scaling factor 𝛼 in (10).
Proposition 1. For ZFR/ZFT precoder

 𝑃𝑅
III. C HANNEL E STIMATION 𝛼 =   ( 2𝐾 ,
)(15)
⎷ˆ ∑ 2 2 + 𝑃 𝜎2
𝜆 + 𝜂ˆ 𝑝𝑖 𝜎𝜉,𝑔,𝑖 + 𝜎𝑛𝑟 𝑅 𝐿𝐼𝑅
Assuming the coherence interval for transmission to 𝑖=1
be 𝑇 symbols, all users simultaneously transmit pilot 2𝐾
∑ 𝑝𝑖′
sequence of length 𝜏 ≤ 𝑇 symbols to the relay. During ˆ=
where 𝜆
(𝑁 − 2𝐾 − 1) 𝜎 2 ,
ˆ𝑓,𝑖
pilot transmission phase, the relay will receive the fol- 𝑖=1
 𝛼2 𝑝𝑘′ ∣f𝑘𝑇 Wg𝑘′ ∣
SNR𝑘 =
∑
2𝐾 ∑
𝛼2 𝑝𝑘 ∣𝜆𝑘 ∣2 + 𝛼2 ∣f𝑘𝑇 Wg𝑖 ∣2 + 𝛼2 ∥f𝑘𝑇 WG𝑅𝑅 ∥2 𝑃𝑁𝑅 + 𝛼2 ∥f𝑘𝑇 W∥2 𝜎𝑛2 𝑅 + 2 𝑝 + 𝜎2
𝜎𝑘,𝑖 𝑖 𝑛
′ 𝑖,𝑘∈𝑈 𝑘
𝑖∕=𝑘,𝑘
(20)
 [ ]2
𝛼2 𝑝𝑘′ 𝔼 f𝑘𝑇 Wg𝑘′ 
SNR𝑘,lower = 2 [ 𝑇 ] (25)
𝛼 𝑝𝑘′ var f𝑘 Wg𝑘′ + 𝛼2 𝑝𝑘 SI𝑘 + 𝛼2 IP𝑘 + 𝛼2 NR𝑘 + 𝛼2 LIR𝑘 + UI𝑘 + NU𝑘

2𝐾
∑ 1 fied as [ ]
𝜂ˆ = 2 2 𝜎 2
. ˆ 𝐻 z ∥2 = 𝜎 2 𝜂ˆ.
ˆ ∗ TḠ
𝑗=1 (𝑁 − 2𝐾 − 1) 𝜎
ˆ𝑓,𝑗 ˆ𝑔,𝑗 ′ 𝔼 ∥F̄ 𝑅 𝑛𝑟 (19)
[ Proof: To ] derive this result, we will first simplify By using (16), (18) and (19), we get (15).
𝔼 ∥WG̃x∥2 using (14).
[ ] [ ]
𝔼 ∥WG̃x∥2 = 𝔼 ∥F̄ ˆ ∗ TḠ
ˆ 𝐻 G̃x∥2 V. S PECTRAL EFFICIENCY OF ZFR/ZFT PRECODER
{ [ ]}
= Tr 𝔼 (F̄ ˆ ∗ TḠˆ 𝐻 (Ĝ + E )PP𝐻 (Ĝ𝐻 + E𝐻 )Ḡ ˆ T𝐻 F̄ ˆ𝑇 )
𝑔 𝑔 In this section, we calculate lower bound on the
{ [ ]}
(𝑎)
= Tr 𝔼 F̄ TPP TF̄ˆ ∗ 𝐻 ˆ𝑇 instantaneous spectral efficiency for ZFR/ZFT precoder.
The instantaneous SNR𝑘 at the user 𝑆𝑘 can be expressed
2𝐾
∑ { [ ]} using (11) as in (20) (shown at the top of this page).
+ 2
𝑝𝑖 𝜎𝜉,𝑔,𝑖 Tr 𝔼 Ḡ ˆ T𝐻 F̄ ˆ 𝑇 F̄
ˆ ∗ TḠ ˆ𝐻
The spectral efficiency of the system which includes the
𝑖=1
channel estimation overhead
{ 2𝐾 is }
(𝑏) ∑
2𝐾 [ ] ∑ 2𝐾 { [ ]} ( ) ∑
= ˆ 𝐻 ˆ
𝑝𝑖′ 𝔼 f̄𝑖 f̄𝑖 + 2
𝑝𝑖 𝜎𝜉,𝑔,𝑖 Tr 𝔼 Λ̂𝐹 TΛ̂𝐺 T ∗ 𝜏
𝑅= 1− 𝔼 log2 (1 + SNR𝑘 ) . (21)
𝑖=1 𝑖=1 𝑇
𝑘=1
2𝐾
∑
(𝑐)
=𝜆 ˆ+ 2
𝑝𝑖 𝜎𝜉,𝑔,𝑖 𝜂ˆ, (16) Next we derive a lower bound on the achievable rate
′
𝑖=1 using the method in [18], [19]. For the 𝑘 − 𝑘 pair, the
The equality in (𝑎) is obtained by [using the fact]that signal received by the 𝑘th user can be written as (see
ˆ 𝐻 Ĝ = Ĝ𝐻 Ḡ
Ḡ ˆ = I , and 𝔼 E PP𝐻 E 𝐻 = (11))
2𝐾 𝑔 𝑔 [ ]
√
∑
2𝐾 𝑦˜𝑘 = 𝛼 𝑝𝑘′ 𝔼 f𝑘𝑇 Wg𝑘′ 𝑥𝑘′ + 𝑛 ˜𝑘 , (22)
2
𝑝𝑖 𝜎𝜉,𝑔,𝑖 I𝑁 . In equality (𝑏), we define Λ̂𝐹 ≜
𝑖=1 ( ) where
ˆ 𝐻 F̄ˆ ) = (F̄𝐻 F̄)−1 and Λ̂ ˆ
ˆ 𝐻 Ḡ √ ( [ ]) √
(F̄ 𝐺 ≜ Ḡ = 𝑛 ˜ 𝑘 = 𝛼 𝑝𝑘′ f𝑘𝑇 Wg𝑘′ − 𝔼 f𝑘𝑇 Wg𝑘′ 𝑥𝑘′ + 𝛼 𝑝𝑘 𝜆𝑘 𝑥𝑘
( 𝐻 )−1 2𝐾
Ḡ Ḡ and use the fact that TPP𝐻 T = ∑ √
𝑇
diag{𝑝2 , 𝑝1 , ⋅ ⋅ ⋅ , 𝑝2𝐾 , 𝑝2𝐾−1 }. To derive equality in + 𝛼f 𝑘 W 𝑝𝑖 g𝑖 𝑥𝑖 + 𝛼f𝑘𝑇 WGRR x̃𝑅
(𝑐), we first note that the random matrices Λ̂𝐹 𝑖∕=𝑘,𝑘′
∑ √
and Λ̂𝐺 have inverse Wishart distribution, where + 𝛼f𝑘𝑇 Wz𝑅 + Ω𝑘,𝑖 𝑝(𝑘)𝑥𝑖 + 𝑧𝑘 . (23)
Λ̂𝐹 ∼ 𝒲 −1 (( D̂−1
𝑑 ,)2𝐾), Λ̂𝐺 ∼ (𝒲 ) (D̂𝑢 , 2𝐾),
−1 −1
𝑖,𝑘∈𝑈 𝑘
[ ]
with 𝑤 ˆ𝑓,𝑖,𝑗 = Λ̂𝐹 ,𝑤ˆ𝑔,𝑖,𝑗 = Λ̂𝐺 , ∀𝑖, 𝑗 = The value of 𝔼 f𝑘𝑇 Wg𝑘′ can be calculated from the
[𝑖,𝑗 ] 𝑖,𝑗 [ ] knowledge of channel distribution. We observe that
D̂−1
1, 2, 3, ..., 2𝐾 and 𝔼 Λ̂𝐹 = 𝑁 −2𝐾−1 𝑑
, 𝔼 Λ̂𝐺 = the desired signal and effective noise are uncorrelated.
[ ]
D̂−1
𝑢
[17]. We also have 𝔼 ˆf̄ 𝐻ˆf̄ = 𝔼 [𝑤 ˆ𝑓,𝑖,𝑗 ] = According[ to [20], ][21], we only exploit the knowledge
𝑁 −2𝐾−1 𝑖 𝑗
1 of the 𝔼 f𝑘𝑇 Wg𝑘′ in the detection, and treat uncor-
2 , ∀𝑖 = 𝑗 and 0, otherwise, and similarly,
(𝑁 −2𝐾−1)ˆ
[ 𝐻 ] 𝜎𝑓,𝑖 related additive noise 𝑛 ˜ (𝑘) as the worst-case Gaussian
𝔼 ḡ ˆ𝑗
ˆ 𝑖 ḡ = 𝔼 [𝑤 ˆ𝑔,𝑖,𝑗 ] = 1
2 , ∀𝑖 = noise when computing the spectral efficiency. We, con-
(𝑁 −2𝐾−1)ˆ
𝜎𝑔,𝑖
𝑗 and 0, otherwise. With the above ∑
equalities, we obtain sequently obtain lower bound on the achievable rate as
𝑝𝑖′
the equality in (𝑐), where 𝜆 ˆ = 2𝐾
(
2 .
𝜏 )∑ ( )
𝑖=1 (𝑁 −2𝐾−1)ˆ 2𝐾
𝜎𝑓,𝑖
{ [ ]} 𝑅 lower = 1 − log2 1 + SNR𝑘,lower , (24)
The expression Tr 𝔼 Λ̂∗𝐹 TΛ̂𝐺 T is simplified as 𝑇
{ [ ]} ∑ 𝑘=1
2𝐾
Tr 𝔼 Λ̂∗𝐹 TΛ̂𝐺 T = 𝑗=1 (𝑁 −2𝐾−1)1 2 𝜎ˆ 2 𝜎ˆ 2 ≜ 𝜂ˆ. where SNR𝑘,lower is given by (25) as shown at the top
𝑓,𝑗 ′
𝑔,𝑗 of this page. In (25), the residual self-interference after
On similar lines, we have
[ ] SIC (SI), the inter-pair interference (IP), the amplified
𝔼 ∥WG𝑅𝑅 x̃𝑅 ∥2 = 𝑃𝑅 𝜎𝐿𝐼𝑅
2
𝜂ˆ. (18) noise from the relay (NR), amplified loop interference
The last term in the denominator of (10) can be simpli- (LIR), self-loop interference and inter-user interference
 𝑢 𝑘 𝑝𝑘 ′
SNR𝑘zf (𝑝𝑘 , 𝑃𝑅 ) = ⎛ ⎞
2𝐾
∑ ∑ ( ) ∑ ( (1) )
⎝𝑑(1) + 𝑑(2) 𝑃 −1 + (3) (1) (2) (3)
𝑝𝑖 𝑃𝑅−1 𝑑𝑘,𝑖⎠𝑝𝑖 + 𝑣𝑘 + 𝑣𝑘 𝑃𝑅 + 𝑣𝑘 𝑃𝑅−1 +
(2)
𝑤𝑘,𝑖 + 𝑃𝑅−1 𝑤𝑘,𝑖 𝑝𝑖
𝑘,𝑖 𝑘,𝑖 𝑅
𝑖=1 𝑖,𝑘∈𝑈𝑘 𝑖,𝑘∈𝑈𝑘
(27)

(UI), and the noise at user (NU), are given as following. 60

[ ] ∑2𝐾
[ ]
SI𝑘 = 𝔼 ∣f𝑘𝑇 Wg𝑘 − f̂𝑘𝑇 Wĝ𝑘 ∣2 , IP𝑘 = 𝑝𝑖 𝔼 ∣f𝑘𝑇 Wg𝑖 ∣2 ,
50
𝑖∕=𝑘,𝑘′

Spectral efficiency (bps/Hz)

[ ] [ ]
NR𝑘 = 𝔼 ∣f𝑘𝑇 Wz𝑅 ∣ , LIR𝑘 = 𝔼
2
∣f𝑘 WG𝑅𝑅 x̃∣2 ,
𝑇
40 N = 256
∑ [ ] [ ]
UI𝑘 = 𝑝𝑖 𝔼 ∣Ω𝑘,𝑖 𝑥𝑖 ∣2 , NU𝑘 = 𝔼 ∣𝑧𝑘 ∣2 . (26)
𝑖,𝑘∈𝑈 𝑘 30
Theorem 1. The spectral efficiency for a finite N = 64
number of receive antenna at the relay with im- 20
perfect CSI based ZFR/ZFT{ processing is} lower
( )
bounded as 1 − 𝑇𝜏 log2 1 + SNR𝑘zf (𝑝𝑘 , 𝑃𝑅 ) , where 10 Monte−Carlo (MRC/MRT)
Lower bound (MRC/MFT)

SNRzf (𝑝 , 𝑃 ) is given by (27), shown at the top of this

Monte−Carlo (ZFR/ZFT)
𝑘 𝑘 𝑅 0 Lower bound (ZFR/ZFT)
page, with
( ) 0 10 −1020 30 40
2 2
(1) 1 𝜎𝜉,𝑓,𝑘 𝜎𝜉,𝑔,𝑖 2 2
SNR [dB]
𝑑𝑘,𝑖 = 2 + 2 + 𝜎𝜉,𝑓,𝑘 𝜎𝜉,𝑔,𝑖 𝜂ˆ Fig. 2: Spectral efficiency versus SNR for MRC/MRT and ZFR/ZFT,
(𝑁 − 2𝐾 − 1) 𝜎 ˆ𝑓,𝑖 ′ 𝜎
ˆ𝑔,𝑘′ where SNR𝜌 = 10 dB.
( )
(2) 1
𝑑𝑘,𝑖 = 𝜎𝑛2 2
2
+ 𝜂ˆ𝜎𝜉,𝑔,𝑖
(𝑁 − 2𝐾 − 1) 𝜎 ˆ𝑓,𝑖 ′
( ) 100
Monte−Carlo (MRC/MRT)
(3) 2 1 2 Lower bound (MRC/MFT)
𝑑𝑘,𝑖 = 𝜎𝑘,𝑖 2 + 𝜂
ˆ 𝜎 𝜉,𝑔,𝑖
(𝑁 − 2𝐾 − 1) 𝜎 ˆ𝑓,𝑖 ′
80
Monte−Carlo (ZFR/ZFT)
( )
Spectral efficiency (bps/Hz)

Lower bound (ZFR/ZFT)

(1) 2 1 2 2
𝑣𝑘 = 𝜎𝑛𝑟 2 + 𝜎𝜉,𝑓,𝑘 𝜂ˆ + 𝜂ˆ𝜎𝐿𝐼𝑅 𝜎𝑛2
(𝑁 − 2𝐾 − 1)ˆ 𝜎𝑔,𝑘 ′ 60
( )
(2) 2 1 2
𝑣𝑘 = 𝜎𝐿𝐼𝑅 2 + 𝜎𝜉,𝑓,𝑘 𝜂ˆ
(𝑁 − 2𝐾 − 1)ˆ 𝜎𝑔,𝑘 ′ 40
SNR=−10 dB
(3) 2
𝑣𝑘 = 𝜂ˆ𝜎𝑛𝑟 𝜎𝑛2 , SNR=10 dB
(1) 2 2 (2)
2 2 20
𝑤𝑘,𝑖 = 𝜂ˆ𝜎𝑘,𝑖 𝜎𝐿𝐼𝑅 , 𝑤𝑘,𝑖 = 𝜂ˆ𝜎𝑘,𝑖 𝜎𝑛𝑟 , 𝑢𝑘 = 1.
Proof: Refer to Appendix A.
0
0 200 400 600 800 1000
VI. S IMULATION R ESULTS Number of relay antennas (N)
Fig. 3: Spectral efficiency versus the number of relay antennas for
We investigate the performance of the multi-pair two- MRC/MRT and ZFR/ZFT processing, where SNR𝜌 = 10 dB.
way full duplex AF relay system by using Monte-
Carlo simulations. We will validate the lower-bound 𝜏 = 2𝐾. We begin by comparing the analytical lower
expression derived for the spectral efficiency in Theorem bound for the spectral efficiency, obtained in Theorem 1
1. In [22], we have derived the lower-bound expression with their exact expression in (21) using Monte-Carlo
for MRC/MRT processing considering MMSE based simulations. We compare the bound for 𝑁 = 64 and
channel estimation for the system under consideration. 𝑁 = 256 relay antennas and set 𝐾 = 10 user pairs,
2 2
In this paper, we have used the results from [22] to 𝜎𝑔,𝑖 = 𝜎𝑓,𝑖 = 𝜎 2 , for 𝑖 = 1, 2, 3, ⋅ ⋅ ⋅ , 2𝐾, 𝜎𝐿𝐼𝑅
2
=
2 2 2 2
compare the performance of ZFR/ZFT with MRC/MRT 𝜎 , 𝜎𝑈 𝐼 ≜ 𝜎𝑘,𝑗 = 𝜎 for 𝑘, 𝑗 = 1, 2, 3, ⋅ ⋅ ⋅ , 2𝐾,
processing. For this study, we choose, noise variances SNR𝜌 = 10 dB and all users are allocated equal power
as 𝜎𝑛2 = 𝜎𝑛𝑟 2
= 𝜎 2 , and the SNR is defined as i.e., 𝑝𝑖 = 𝑃𝑅 /2𝐾, ∀𝑖 = 1, 2, 3, ⋅ ⋅ ⋅ , 2𝐾. We see from
2
SNR = 𝑃𝑅 /𝜎 . We define the pilot signal to noise Fig. 2 that the derived lower bound and exact expression
ratio as SNR𝜌 = 𝑃𝜌 /𝜎 2 , and set the length of the overlap for ZFR/ZFT processing for 𝑁 = 256 relay
coherence interval 𝑇 = 200 symbols, the training length antennas. For MRC/MRT, the lower bound marginally
 50
and full-duplex system with ZFR/ZFT processing. As
2
Monte−Carlo (MRC/MRT) we increase the value of self-loop interference 𝜎𝐿𝐼𝑅
Lower bound (MRC/MFT) 2
Monte−Carlo (ZFR/ZFT)
and inter-user interference 𝜎𝑈 𝐼 , the spectral efficiency
40 of full-duplex system decreases. For 𝑁 < 650, the half-
Spectral efficiency (bps/Hz)
Lower bound (ZFR/ZFT)
2 2
SNR=10dB duplex relay with 𝜎𝐿𝐼𝑅 = 0 dB, 𝜎𝑈 𝐼 = 5 dB performs
30 better than full-duplex relay. We also observe that with
the increase in the the number of relay antennas the rate
20
of increase of spectral efficiency in case of full-duplex
relay is higher as compared to half-duplex relay.

10 SNR=0dB

VII. C ONCLUSION
0
SNR=−10dB
5 10 15 20 25 30 35 40
Number of user pairs (K) We considered a multi-pair AF full-duplex massive
Fig. 4: Spectral efficiency versus number of user pairs for MRC/MRT MIMO two-way relay with full-duplex users with single
and ZFR/ZFT processing, where 𝑁 = 128, SNR𝜌 = 10 dB. transmit and receive antenna. We derived closed-form
spectral efficiency expression for ZFR/ZFT relay pro-
cessing with MMSE channel estimation, and for arbitrary
140
2 2
number of relay antennas, which have not yet been
σLIR = −10 dB, σUI = −10 dB
derived in the literature. We showed the accuracy of these
120
lower bounds for different number of relay antennas,
Spectral efficiency (bps/Hz)

100 user pairs and relay transmit power. We also numerically

2 2
σLIR = 0 dB, σUI = 0 dB investigated the loop and inter-user interference values
80 for which the full-duplex relay outperforms a half-
duplex relay.
60

40 2 2
σLIR = 0 dB, σUI = 5 dB A PPENDIX A

20 σ2 = 5 dB, σ2 = 0 dB Half−duplex
LIR UI
Full−duplex Starting with the numerator of (25), we have
0 [ ] [( )𝑇 ( )]
200 400 600 800 1000
𝔼 f𝑘 𝑇 Wg𝑘′ = 𝔼 f̂𝑘 + e𝑓,𝑘 F̄ ˆ 𝐻 ĝ ′ + e ′
ˆ ∗ TḠ
Number of relay antennas (N) 𝑘 𝑔,𝑘
Fig. 5: Spectral efficiency versus number of relay antennas for (𝑎)
[ ] (𝑏)
ZFR/ZFT processing comparing half-duplex and full-duplex systems, = 𝔼 1𝑇𝑓,𝑘 T1𝑔,𝑘′ = 𝔼[1] = 1, (28)
where SNR = 10 dB, SNR𝜌 = 10 dB. Here the values of 𝜎𝐿𝐼𝑅2 and
2 are with respect to 𝜎 2 .
𝜎𝑈 𝐼
The equality in (𝑎) is obtained by using the following
results: ĝ𝑘𝐻′ Ḡ ˆ 𝐻 ĝ ′ = 1 ′ , f̂ 𝑇 F̄
ˆ = 1𝑇 ′ , Ḡ ˆ ∗ = 1𝑇 ,
𝑔,𝑘 𝑘 𝑔,𝑘 𝑘 𝑓,𝑘
differs from the exact expression. We also observe that
ˆ 𝑇 f̂ ∗ = 1 . Equality in (𝑏) is because 1𝑇 T1 ′ = 1.
F̄
the spectral efficiency, for high SNR values, saturates for 𝑘 𝑓,𝑘 [ ] 𝑓,𝑘 𝑔,𝑘
both MRC/MRT and ZFR/ZFT. This is because the loop The expression for var f𝑘 𝑇 Wg𝑘′ is given by (36)
interference also increases proportionally with increase (solved at the top of this page). The equality in (𝑎)
in SNR. therein is because F̂𝑇 WĜ = T, i.e., f̂𝑘𝑇 Wĝ𝑗 =
Fig. 3 compares the spectral efficiency versus 𝑁 for ′
1, ∀𝑗 = 𝑘 and 0, otherwise. Equalities in (𝑏) are ob-
MRC/MRT and ZFR/ZFT processing with SNR = 10 dB tained by substituting the value of W from (14) and
and SNR = −10 dB. The performance of MRC/MRT by simple manipulations. The equality [in (𝑐) is] be-
and ZFR/ZFT processing is almost same for SNR = 1
cause 𝔼 [𝑤 ˆ𝑓,𝑘,𝑘 ] = (𝑁 −2𝐾−1)ˆ 2
𝜎𝑓,𝑘
, 𝔼 𝑤 ˆ𝑔,𝑘′ ,𝑘′ =
−10 dB. The spectral-efficiency versus 𝐾 for different ∑
1 2𝐾 1
value of SNR is shown in Fig. 4. As the number of (𝑁 −2𝐾−1)ˆ 𝜎2
, and 𝜂ˆ ≜ 𝑗=1 (𝑁 −2𝐾−1)2 𝜎ˆ 2 𝜎ˆ 2 . Re-
′ 𝑓,𝑗 ′
𝑔,𝑘 𝑔,𝑗
multi-pairs increases the SNR of each user decreases and member that there is no need to perform SIC in the case
hence noise dominates. The ZFR/ZFT neglects the effect ZFR/ZFT processing, and hence the self-interference
of noise which degrades the spectral-efficiency as 𝐾 term can be re-written (
as )
increases. In contrast MRC/MRT works well at low SNR 2 2
1 𝜎𝜉,𝑓,𝑘 𝜎𝜉,𝑔,𝑘 2 2
as it maximizes the received SNR while neglecting the SI𝑘 = 2 + 2 +𝜎𝜉,𝑓,𝑘 𝜎𝜉,𝑔,𝑘 𝜂ˆ.
(𝑁 − 2𝐾 − 1) 𝜎 ˆ𝑓,𝑘 ′ 𝜎
ˆ𝑔,𝑘′
inter-pair interference. Fig. 5 compares the spectral effi-
ciency versus number of relay antennas for half-duplex (30)
 [ ] [ 2 ]  [ ]2
   
var f𝑘 𝑇 Wg𝑘′ = 𝔼 f𝑘 𝑇 Wg𝑘′  − 𝔼 f𝑘 𝑇 Wg𝑘′ 
(𝑎) 2 [ ] [ ] { [ ]}
= 𝜎𝜉,𝑓,𝑘 𝔼 ĝ𝑘𝐻′ W𝐻 Wĝ𝑘′ + 𝜎𝜉,𝑔,𝑘
2 𝑇
′ 𝔼 f̂𝑘 WW
𝐻 ∗ 2
f̂𝑘 + 𝜎𝜉,𝑓,𝑘 2
𝜎𝜉,𝑔,𝑘 ′ Tr 𝔼 WW𝐻
(𝑏) 2
[ ] [ ] { [ ]}
= 𝜎𝜉,𝑓,𝑘 𝔼 1𝑇𝑔,𝑘′ TΛ̂∗𝐹 T1𝑔,𝑘′ + 𝜎𝜉,𝑔,𝑘
2 𝑇
′ 𝔼 1𝑓,𝑘 TΛ̂𝐺 T1𝑓,𝑘
2
+ 𝜎𝜉,𝑓,𝑘 2
𝜎𝜉,𝑔,𝑘 ′ Tr 𝔼 Λ̂∗𝐹 TΛ̂𝐺 T
2
(𝑐)
2
𝜎𝜉,𝑓,𝑘 𝜎𝜉,𝑔,𝑘 ′
2 2
= 2 + 2 + 𝜎𝜉,𝑓,𝑘 𝜎𝜉,𝑔,𝑘 ′𝜂
ˆ. (36)
(𝑁 − 2𝐾 − 1)ˆ
𝜎𝑓,𝑘 (𝑁 − 2𝐾 − 1)ˆ 𝜎𝑔,𝑘 ′

Similarly, the other terms in the denominator of (25) can [10] T. L. Marzetta, “Noncooperative cellular wireless with unlimited
be written as[ follows ( numbers of base station antennas,” IEEE Trans. Wireless Com-
2𝐾
) ] mun., vol. 9, no. 11, pp. 3590–3600, 2010.
∑ 1 2
𝜎𝜉,𝑓,𝑘 𝜎𝜉,𝑔,𝑖2
[11] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and
2 2
IP𝑘 = 𝑝𝑖 2 + 2 +𝜎𝜉,𝑓,𝑘 𝜎𝜉,𝑔,𝑖 𝜂ˆ , spectral efficiency of very large multiuser MIMO systems,” IEEE
𝑁 −2𝐾−1 𝜎 ˆ𝑓,𝑖 ′ 𝜎
ˆ𝑔,𝑘′
𝑖∕=𝑘,𝑘′ Trans. Commun., vol. 61, no. 4, pp. 1436–1449, 2013.
(31) [12] Y. Dai and X. Dong, “Power allocation for multi-pair massive
( ) MIMO two-way AF relaying with linear processing,” IEEE
2 1 2
Trans. Wireless Commun., vol. 15, no. 9, pp. 5932–5946, 2016.
NR𝑘 = 𝜎𝑛𝑟 2 + 𝜎𝜉,𝑓,𝑘 𝜂ˆ , (32) [13] H. Cui, L. Song, and B. Jiao, “Multi-pair two-way amplify-and-
(𝑁 − 2𝐾 − 1)ˆ 𝜎𝑔,𝑘 ′ forward relaying with very large number of relay antennas,” IEEE
( ) Trans. Wireless Commun., vol. 13, no. 5, pp. 2636–2645, May
2 1 2 2014.
LIR𝑘 = 𝑃𝑅 𝜎𝐿𝐼𝑅 2 + 𝜎𝜉,𝑓,𝑘 𝜂ˆ , [14] T. Riihonen, S. Werner, and R. Wichman, “Spatial loop interfer-
(𝑁 − 2𝐾 − 1)ˆ 𝜎𝑔,𝑘′ ence suppression in full-duplex MIMO relays,” in Conference
(33) Record of the Forty-Third Asilomar Conference on Signals,
∑ Systems and Computers,. IEEE, 2009, pp. 1508–1512.
2
UI𝑘 = 𝑝𝑖 𝜎𝑘,𝑖 , NU𝑘 = 𝜎𝑛2 . (34) [15] M. Biguesh and A. B. Gershman, “Training-based MIMO chan-
𝑖,𝑘∈𝑈 𝑘 nel estimation: a study of estimator tradeoffs and optimal training
signals,” IEEE Trans. Signal Process., vol. 54, no. 3, pp. 884–
Substituting the values obtained from (34)-(36) in (25), 893, March 2006.
we obtain (27). [16] S. M. Kay, “Fundamentals of statistical signal processing, volume
I: estimation theory,” 1993.
R EFERENCES [17] P. Graczyk, G. Letac, and H. Massam, “The complex wishart
distribution and the symmetric group,” Annals of Statistics, pp.
[1] L. Sanguinetti, A. A. D’Amico, and Y. Rong, “A tutorial on
287–309, 2003.
the optimization of amplify-and-forward MIMO relay systems,”
[18] T. L. Marzetta, “How much training is required for multiuser
IEEE J. Sel. Areas Commun., vol. 30, no. 8, pp. 1331–1346,
MIMO?” in ACSSC’06 Fortieth Asilomar Conference on Signals,
2012.
Systems and Computers. IEEE, 2006, pp. 359–363.
[2] K. J. Lee, H. Sung, E. Park, and I. Lee, “Joint optimization
[19] J. Jose, A. E. Ashikhmin, T. L. Marzetta, and S. Vishwanath,
for one and two-way MIMO AF multiple-relay systems,” IEEE
“Pilot contamination and precoding in multi-cell TDD systems,”
Trans. Wireless Commun., vol. 9, no. 12, pp. 3671–3681, Decem-
IEEE Trans. Wireless Commun., vol. 10, no. 8, pp. 2640–2651,
ber 2010.
2011.
[3] A. Nadh, J. Samuel, A. Sharma, S. Aniruddhan, and R. K.
[20] B. Hassibi and B. M. Hochwald, “How much training is needed
Ganti, “A taylor series approximation of self-interference channel
in multiple-antenna wireless links?” IEEE Trans. Inf. Theory,
in full-duplex radios,” IEEE Trans. Wireless Commun., vol. 16,
vol. 49, no. 4, pp. 951–963, 2003.
no. 7, pp. 4304–4316, 2017.
[21] M. Médard, “The effect upon channel capacity in wireless com-
[4] A. Sabharwal, P. Schniter, D. Guo, D. Bliss, S. Rangarajan, munications of perfect and imperfect knowledge of the channel,”
and R. Wichman, “In-band full-duplex wireless: Challenges and IEEE Trans. Inf. Theory, vol. 46, no. 3, pp. 933–946, 2000.
opportunities,” IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. [22] E. Sharma, R. Budhiraja, K. Vasudevan, and L. Hanzo, “Full-
1637–1652, Sep. 2014. duplex massive MIMO multi-pair two-way AF relaying: Energy
[5] Y. Y. Kang, B.-J. Kwak, and J. H. Cho, “An optimal full-duplex efficiency optimization,” CoRR, vol. abs/1705.09043, 2017.
AF relay for joint analog and digital domain self-interference
cancellation,” IEEE Trans. Commun., vol. 62, no. 8, pp. 2758–
2772, Aug. 2014.
[6] Z. Zhang, Z. Ma, Z. Ding, M. Xiao, and G. K. Karagiannidis,
“Full-duplex two-way and one-way relaying: Average rate, out-
age probability, and tradeoffs,” IEEE Trans. Wireless Commun.,
vol. 15, no. 6, pp. 3920–3933, 2016.
[7] H. Q. Ngo, H. A. 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.
1721–1737, 2014.
[8] Z. Zhang, Z. Chen, M. Shen, and B. Xia, “Spectral and energy
efficiency of multipair two-way full-duplex relay systems with
massive MIMO,” IEEE J. Sel. Areas Commun., vol. 34, no. 4,
pp. 848–863, 2016.
[9] Z. Zhang, Z. Chen, M. Shen, B. Xia, W. Xie, and Y. Zhao,
“Performance analysis for training-based multi-pair two-way
full-duplex relaying with massive antennas,” IEEE Trans. Veh.
Technol., 2017, available in IEEExplore under Early Access with
DOI: 10.1109/TVT.2016.2644986.