Bayesian Channel Estimation for Massive MIMO
Communications
Chengzhi Zhu, Zhitan Zheng, Bin Jiang, Wen Zhong, and Xiqi Gao
National Mobile Communications Research Laboratory, Southeast University, Nanjing 210096, P. R. China
Email: {bjiang, xqgao}@seu.edu.cn
Abstract—In this paper, we derive the Bayes-Optimal estimator
based on approximate message passing (AMP) algorithm in
massive multiple-input multiple-output (MIMO) systems, which
requires statistical channel state information (CSI). According to
the analysis of channel model in beam domain, the convariance
matrix is derived for CSI acquisition. With the aid of statistical
CSI, the convergence of the proposed algorithm has significant
improvement in comparison with which use the expectation-
maximization (EM) algorithm to fit the statistical CSI. Sim-
ulations show great mean squared error (MSE) performance
that approximates the Minimum Mean Square Error (MMSE)
estimator, and better convergence performance than other AMP
algorithm can be achieved. Besides, the results prove that
performance of the random pilot in this algorithm is close to
that of the orthogonal pilot based on Zadoff-Chu sequences.
I. INTRODUCTION
Massive MIMO systems, which employ a large number of
antennas at the base station (BS) to simultaneously serve a
relatively large number of users [1], are believed to be one of
the key candidate technologies for forthcoming 5G wireless
networks [2], [3] with the potential large gains in spectral
efficiency and energy efficiency.
Channel state information which is typically obtained with
the assistance of the periodically inserted pilot signals [4],
plays a significant role in massive MIMO transmission. CSI
makes it possible to adapt transmissions to current channel
conditions, which is crucial for achieving robust communica-
tion in massive MIMO systems. Due to the fact that statistical
CSI varies over much longer time scales than instantaneous
CSI, we use statistical CSI instead of instantaneous CSI. And
more importantly, statistical CSI requires much less over-
head. MIMO channel estimation based on Gaussian-Mixture
Bayesian learning has been investigated in [5], [6]. Instead
of using expectation-maximization (EM) [5]–[7] algorithm to
learn the channel properties, we use statistical CSI as known
properties to reduce the complexity and improve the perfor-
mance of AMP algorithm. In massive MIMO systems, accurate
statistical CSI is required not only in channel estimation, but
also in other aspects [8], [9], such as user scheduling .
In this work, we model each channel element in the beam
domain as a Gaussian variable. This model enables a more
accurate learning of AMP algorithm, in comparison with
Gaussian-Mixture. We can reconstruct the channel components
with great mean-squared error (MSE) performance which is
close to LMMSE channel estimation.
Throughout this paper, we use the following notation: C
denotes the set of complex numbers. We use a
i,j
to denote
the (i, j)th element of matrix A. A
T
denotes the transpose of
A and A
H
denotes the conjugate transpose of A. Identity
matrix is denoted by I
K
. E{·} represents the expectation
operation. x ∼ N
C
(µ, σ
2
) denotes a random complex variable
x comply with the complex Gaussian distribution with mean
µ and variance σ
2
, where
f
X
(x) =
1
πσ
2
exp
−
|x − µ|
2
σ
2
.
II. SYSTEM MODEL
We consider single-cell TDD massive MIMO wireless trans-
mission scheme which consists of one BS equipped with N
antennas and K single-antenna users. Assume that the BS is
equipped with a uniform linear array (ULA), and the antennas
are separated by half wavelength.
We assume that the uplink pilot S ∈ C
L×K
where L
denotes the pilot length. We use h
k,n
to represent the channel
coefficient of kth user and nth beam in the beam domain [8].
With these definitions, the received signal of the lth symbol
in the nth beam can be written as
y
l,n
=
K
k=1
s
l,k
h
k,n
+ z
l,n
= s
T
l
h
n
+ z
l,n
, (1)
where s
l,k
represents lth symbol of the kth user’s pilot signal,
z
l,n
is the Gaussian noise in the beam domain with zero mean
and variance σ
2
z
, s
l
= [s
l,1
, s
l,2
, ...s
l,K
]
T
, and h
n
∈ C
K×1
is
the channel vector of all users in the nth beam.
Let g
k
= [h
k,1
, h
k,2
, ..., h
k,N
]
T
. According to [8]–[10], the
uplink channel of kth user can be modeled as
g
k
= v
k
A
a(θ)r
k
(θ)dθ, (2)
where a (θ) = [1, exp(−jπ sin(θ)), ..., exp(−jπ(N − 1) sin(θ))]
T
is the ULA response vector [10], A = [−π/2, π/2] is the
angle of arrival (AOA), v
k
∼ N
C
(0, I
N
) and r
k
(θ)
denotes the channel gain function. We assume that the
channel phases are uniformly distirbuted, thus E{g
k
} = 0,
and different beams of channels are uncorrelated, i.e.,
E
{r
k
(θ)r
H
k
(θ
′
)} = S
k
(θ)δ(θ − θ
′
) [9]. Let R
k
denotes
channel covariance matrix:
R
k
= E{g
k
g
H
k
} =
A
a(θ)a
H
(θ)S
k
(θ)dθ.
(3)