210 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 9, NO.

2, FEBRUARY 2020

Cascaded Channel Estimation for Large Intelligent

Metasurface Assisted Massive MIMO
Zhen-Qing He and Xiaojun Yuan , Senior Member, IEEE

Abstract—In this letter, we consider the problem of channel potential application of the metamaterial in wireless commu-
estimation for large intelligent metasurface (LIM) assisted mas- nications, the LIM with integrated electronics retains almost
sive multiple-input multiple-output (MIMO) systems. The main all the advantages of massive MIMO such as allowing for
challenge of this problem is that the LIM integrated with a an unprecedented focusing of energy that enables highly
large number of low-cost metamaterial antennas can only pas- efficient wireless charging and remote sensing. Although tra-
sively reflect the incident signals by certain phase shifts, and
does not have any signal processing capability. To deal with
ditional reflecting surfaces have a variety of applications in
this, we introduce a general framework for the estimation of radar and satellite communications, their application in terres-
the transmitter-LIM and LIM-receiver cascaded channel, and trial wireless communication was not possible earlier. This is
propose a two-stage algorithm that includes a sparse matrix fac- because these reflecting surfaces only had fixed phase shifters
torization stage and a matrix completion stage. Simulation results that could not adapt the induced phases with time-varying
illustrate that the proposed method can achieve accurate channel channels which generally constitute the wireless propagation
estimation for LIM-assisted massive MIMO systems. environments.
Index Terms—Bilinear factorization, channel estimation, large To achieve full potential of the LIM-assisted massive MIMO
intelligent metasurface, massive MIMO, matrix completion. systems, the channel state information (CSI) between the
base station (BS) and the LIM and between the LIM and
the receiver is essential in reflect beamforming [2], energy-
I. I NTRODUCTION efficient design [3], as well as in simultaneous passive beam-
forming and information transfer [5]. The main challenge of
ASSIVE multiple-input multiple-output (MIMO), as
M a promising technology for future wireless systems,
has attracted growing research interest in both academia and
the CSI acquisition of the LIM-assisted MIMO systems is
that the LIM, unlike the BS or the receiver, only passively
reflects the electromagnetic waves, and does not have any
industry over recent years. Although massive MIMO exhibits signal processing capability. By leveraging the programmable
huge potentials to support a significantly large amount of property of the LIM and the rank-deficient structure of the
mobile data traffic and wireless connections, implementing massive MIMO channel, we formulate the BS-LIM and LIM-
this system with large-scale antenna arrays in practice remains receiver cascaded channel estimation as a combined sparse
very challenging due to high hardware cost and increased matrix factorization and matrix completion problem. To solve
power consumption. To achieve green and sustainable wireless the problem, we present a two-stage algorithm which includes
networks, researchers have started looking into energy effi- the sparse matrix factorization stage and the matrix completion
cient techniques to improve the system performance, ranging stage. The proposed two-stage algorithm includes the bilin-
from the utilization of energy efficient hardware components to ear generalized approximate message passing (BiG-AMP) [6]
the design of green resource allocation and transceiver signal for sparse matrix factorization and the Riemannian manifold
processing algorithms. gradient-based algorithm for matrix completion [7]. Numerical
To reduce energy consumption and enhance communi- results demonstrate that our algorithm is able to attain accu-
cation quality in wireless networks, the large intelligent rate channel estimation for the LIM-assisted massive MIMO
metasurface (LIM) [1], also known as the intelligent reflect- systems.
ing surface [2] or the reconfigurable intelligent surface [3], To the best of our knowledge, this is the first attempt to
has been recently proposed as an innovative technology that tackle the cascaded channel estimation problem for the LIM-
conceptually goes beyond contemporary massive MIMO com- assisted massive MIMO systems with all passive elements in
munications. Metamaterials, as an emerging technology known the LIM.1 Notice that the authors in [8] proposed a compres-
for its flexibility in manipulating electromagnetic waves, have sive sensing and training based deep learning approach for the
found applications such as in radar imaging [4], etc. As a LIM-assisted channel estimation. Nevertheless, in [8] a few
Manuscript received August 8, 2019; revised October 10, 2019; accepted
active antenna elements of the LIM are required to circum-
October 14, 2019. Date of publication October 22, 2019; date of current ver- vent the challenging cascaded channel estimation problem by
sion February 7, 2020. This work was supported in part by the National Key utilizing conventional channel estimation techniques. In addi-
Research and Development Program of China under Grant 2018YFB1801105, tion, it should be noted that the BiG-AMP algorithm has been
and in part by the National Natural Science Foundation of China under
previously exploited for blind signal detection in conventional
Grant 61801084. The associate editor coordinating the review of this article
and approving it for publication was V. Raghavan. (Corresponding author: massive MIMO systems [9], where the transmitter does not
Xiaojun Yuan.) need to send pilot signals to the receiver, and the channel and
The authors are with the Center for Intelligent Networking and
Communications, National Key Laboratory of Science and Technology 1 The LIM-based MIMO technology is still immature in terms of practical
on Communications, University of Electronic Science and Technology implementations and a number of follow-up works are necessary to prove its
of China, Chengdu 611731, China (e-mail: [email protected]; practicality and utility. At this stage, this letter can be seen as an academic
[email protected]). exercise for channel estimation by assuming the practicality of the LIM-based
Digital Object Identifier 10.1109/LWC.2019.2948632 MIMO technology.
2162-2345 c 2019 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.

Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 23,2022 at 02:15:14 UTC from IEEE Xplore. Restrictions apply.
 HE AND YUAN: CASCADED CHANNEL ESTIMATION FOR LIM ASSISTED MASSIVE MIMO 211

on/off state3 of the n-th LIM reflect element at time t, respec-

tively. By summarizing all the T samples in a transmission
block, the received signal can be recast as
Y = G(S  (H X )) + W (2)
where Y  [y [1], . . . , y [T ]] ∈ CL×T , W  [w [1], . . . , w [T ]]
∈ CL×T , S  [s[1], . . . , s[T ]] ∈ CN ×T , and X 
[x [1], . . . , x [T ]] ∈ CM ×T . In general, a smart programmable
controller is built in the LIM to adaptively adjust the states
Fig. 1. A LIM-assisted massive MIMO system.
and phases of the LIM based on environmental changes,
which is usually referred to as reflect beamforming or passive
beamforming [2], [5].
The BS-LIM channel matrix H and the LIM-user chan-
the data signal can be simultaneously estimated by factoriz-
nel matrix G are assumed to be rank-deficient, i.e.,
ing the received signal matrix. In contrast, this letter aims to
rank(H) < min{N, M} and rank(G) < min{L, N}. The
factorize the cascaded channels of the LIM-assisted MIMO
rank-deficiency property commonly arises in millimeter
systems by using pilot signals, which is a totally different
wave MIMO channels under far-field and limited-scattering
problem from the one considered in [9].
assumptions.
Notations: E{·}, Var{·}, δ(·), C, (·)T , (·)H , (·)∗ , and
·F denote the expectation operator, the variance operator,
the Dirac delta function, the space of complex number, the III. P ROBLEM S TATEMENT
transpose, the conjugate transpose, the conjugate, and the The objective of this letter is to estimate the channel matri-
Frobenius norm, respectively. The (i, j)-th entry, the i-th row, ces H and G based on the observation Y in (2) by assuming
and the j-th column of a matrix A, are denoted by ai,j , that the pilot symbols X and S are known to the receiver. We
aTi , and a j , respectively. We use  and CN (a, C ) to stand
note that there exists an ambiguity problem in the estimates of
for the Hadamard product and the circularly-symmetric com- G and H, since the following equality holds for any full-rank
plex Gaussian distribution with mean vector a and covariance diagonal matrix Φ ∈ CN ×N :
matrix C, respectively.  
G(S  (H X )) = G  S  (H  X ) (3)
II. S YSTEM M ODEL where G   GΦ and H   Φ−1 H . Nevertheless, it is suffi-
Consider a LIM-assisted massive MIMO system, as shown cient to acquire the knowledge of the alternative G  and H  ,
in Fig. 1, where the LIM consists of N low-cost passive rather than the ground-truth G and H. This is because G  and
elements and the BS is equipped with M transmit antennas H  can be viewed as effective channels in the sense that they
to serve a number of user terminals, each equipped with L would not affect the reflect/passive beamforming design (see
receive antennas. In particular, the LIM is deployed to assist Remark 1 below).
the BS in communicating with the users. Without loss of Remark 1: In reflect/passive beamforming [2], [3], [5], the
generality, we consider the communication from the BS to phase shift vectors {s[t]} are generally set as a constant vector
a reference user. The baseband equivalent channels from the in each transmission block,4 i.e.,
BS to the LIM and from the LIM to the user are respec-
tively denoted by H ∈ CN ×M and G ∈ CL×N . The channel s[t] = s, t = 1, . . . , T .
component of the direct link between the BS and the user is Then, the signal model in (2) reduces to
neglected due to unfavorable propagation conditions or can
be estimated (and cancelled from the model) via conventional Y = Gdiag{s}H X + W .
massive MIMO channel estimation methods by turning off the
LIM. We assume a block-fading2 channel with coherence time Clearly, we have
T, i.e., the channel remains unchanged within each transmis- Gdiag{s}H = G  diag{s}H  .
sion block of length T and varies from block to block. Then,
the received signal of the reference user can be expressed as This implies that the design of s for reflect beamforming based
on G and H yields the same result as that based on G  and H  .
y [t] = G(s[t]  (H x [t])) + w [t], t = 1, . . . , T (1)
where x [t] ∈ CM and w [t] ∈ CL are respectively IV. T WO -S TAGE C HANNEL E STIMATION A PPROACH
the transmitted signal and the additive noise drawn from In this section, we propose a two-stage channel estima-
CN (0, σ 2 I ) at time t with σ 2 being the noise power. The tion approach for the LIM-assisted massive MIMO systems.
phase shift vector s[t] of the LIM is defined √ as s[t]  We first present the outline of the proposed scheme and then
[s1,t e jθ1,t , . . . , sN ,t e jθN ,t ]T where j  −1, with θn,t ∈ elaborate the implementation details.
(0, 2π] and sn,t ∈ {0, 1} representing the phase shift and the
3 The state “off” of a LIM reflect element means that there is only structure-
2 The block-fading channel model naturally arises when any two of the three mode reflection generated as if the element is a conducting object, whereas
nodes (i.e., the BS, the LIM, and the user) are in a moving state. In practice, the state “on” means that there are both structure-mode reflection and antenna-
both the BS and the LIM may be fixed at known positions. In this situation, mode reflection [11]. Note that the structure-mode can be absorbed into the
the block fading assumption is still valid when considering moving scatterers direct link in channel modelling.
or considering wireless communications over certain millimeter wave bands 4 In reflect/passive beamforming, there is no need to change the value of
in bad weather conditions (e.g., heavy rain, downpour, and monsoon) [10]. the phase shift vector s[t] unless the channel changes.

Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 23,2022 at 02:15:14 UTC from IEEE Xplore. Restrictions apply.
 212 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 9, NO. 2, FEBRUARY 2020

A. Outline of the Proposed Scheme Algorithm 1 : JBF-MC Algorithm

Based on the discussions in Section III, we need to estimate Input: Y, S, X, prior distributions p(G) and p(Z)
G and H up to an ambiguity in (3) by appropriate design of % sparse matrix factorization via BiG-AMP
g
1: Initialization: ∀l , n, t: generate ĝl,n (1) from p(gl,n ), vl,n (1) =
the training signals {s[t]} and {x[t]}. To begin with, let us
recast the received signal (2) as a bilinear matrix factorization νg2 , ẑn,t (1) = E{zn,t }, vn,tz (1) = λν 2 , and û (0) = 0
z l,t
model 2: for i = 1, . . . , Imax % outer iteration
3: for j = 1, . . . , J
max % inner iteration
p g
4: ∀l , t: v̄l,t (i) = N 2 z
n=1 |ĝl,n (i)| vn,t (i) + vl,n (i)|ẑn,t (i)|
2
Y = GZ + W (4) N
5: ∀l , t: p̄l,t (i) = n=1 ĝl,n (i)ẑn,t (i)
p p  g
6: ∀l , t: vl,t (i) = v̄l,t (i) + N z
n=1 vl,n (i)vn,t (i)
where Z  S  (H X ). This motivates us to propose a two- p
stage scheme: first estimate G and Z, and then determine H 7: ∀l , t: p̂l,t (i) = p̄l,t (i) − ûl,t (i − 1)v̄l,t (i)
based on the estimated Z. It follows from the theories of matrix 8: b (i) = σ 2 v p (i)/v p (i) + σ 2 
∀l , t: vl,t l,t l,t
factorization and matrix completion that the proposed two- p p
9: ∀l , t: b̂l,t (i) = vl,t (i)(yl,t − p̂l,t (i))/(vl,t (i) + σ 2 ) + p̂l,t (i)
stage scheme is workable when Z is a sparse matrix and H is
10: u (i) = 1 − v b (l )/v p (i)/v p (i)
∀l , t: vl,t
a low-rank matrix. The sparse structure of Z can be guaranteed l,t l,t  l,t
 p
when the pilot symbol S of the LIM is designed to be a sparse 11: ∀l , t: ûl,t (i) = b̂l,t (i) − p̂l,t (i) /vl,t (i)
matrix with a large number of zero elements. The low-rank q  T 2 u −1
12: ∀l , n: vl,n (i) = t=1 |ẑn,t (i)| vl,t (i)
structure of H is assumed in the system model.  q  
We now discuss the design of the pilot signals S and X. As 13: ∀l , n: q̂l,n (i) = ĝl,n (i) 1 − vl,n (i) T z u
t=1 vn,t (i)vl,t (i)
q 
mentioned above, the pilot signal S of the LIM should be set + vl,n (i) T ∗
t=1 ẑn,t (i)ûl,t (i)
to a sparse matrix. In this way, we generate {st,n } indepen- r (i) =   L |ĝ (i)|2 v u (i)−1
dently from a Bernoulli distribution Bernoulli(λ) with (small) 14: ∀n, t: vn,t l=1 l,n l,t
∀n, t: r̂n,t (i) = ẑn,t (i) 1 − vn,t r (i)  L v g (i)v u (i)
λ being the probability of taking the value of 1. The trans- 15: l=1 l,n l,t
mitted pilot signal X is designed to be a full-rank matrix, i.e., r (i)  L ĝ ∗ (i)û (i)
+ vn,t l=1 l,n l,t
rank(X) = M (M < T), which ensures the successful recovery 16:
q
∀l , n: ĝl,n (i + 1) = E gl,n |q̂l,n (i), vl,n (i)
of BS-LIM channel matrix H via matrix completion and will g q
become clear in Section IV-C. 17: ∀l , n: vl,n (i + 1) = Var gl,n |q̂l,n (i), vl,n (i)
18: ∀n, t: ẑn,t (i + 1) = E zn,t |r̂n,t (i), vn,t r (i)
We now describe the algorithm design at the receiver side
z
∀n, t: vn,t (i + 1) = Var zn,t |r̂n,t (i), vn,t r (i)
via two separate stages, i.e., the matrix factorization stage 19:
and the matrix completion stage. In the matrix factorization 20: if a certain stopping criterion is met, stop
stage, we adopt the state-of-the-art BiG-AMP algorithm [6] 21: end for g g
22: ∀l , n, t, ĝl,n (i) = ĝl,n (i + 1), vl,n (i) = vl,n (i + 1),
to recover G and Z from Y. Then, based on the estimated Z  z z
ẑn,t (i) = ẑn,t (1), vn,t (i) = vn,t (1)
in the first stage and the pilot signal S, the matrix comple- 23: end for
tion stage is to retrieve H by using the available Riemannian 24: G ← G(i  + 1), Z  ←Z  (i + 1)
manifold gradient-based algorithm [7] (based on the fact that % matrix completion via RGrad
H is rank-deficient). Such a two-stage channel estimation pro- 25: Initialization: A(0) = 0
cedure is referred to as the joint bilinear factorization and 26: for k = 1, . . . , Kmax 
matrix completion (JBF-MC) algorithm which is presented 27: Q(k ) = S ∗  Z  − A(k )
in Algorithm 1. The details of the JBF-MC algorithm are PS(k ) (Q(k ))2F
28: α(k ) = ∗
elaborated in the following subsections. S (PS(k ) (Q(k )))2F
29: W (k ) = A(k ) + αk PS(k ) (Q(k ))
 
30: A(k + 1) = Hr W (k )
B. Sparse Matrix Factorization 31: if a certain stopping criterion is met, stop
Based on the observation Y, we use the BiG-AMP algo- 32: end for
33: H ← AX  † with A  = A(k + 1)
rithm [6] to approximately calculate the minimum mean-  and H
squared error estimates of G and Z, i.e., the means Output: G
of the marginal posteriors {p(gl,n |Y )} and {p(zn,t |Y )}.
Specifically, the BiG-AMP implements the sum-product loopy
belief propagation over a factor graph induced by the posterior We choose independent Gaussian priors for G and independent
distribution Bernoulli-Gaussian priors for Z, i.e.,

 N
L  L 
 N
p(G, Z |Y ) ∝ p(Y |B)p(G)p(Z ) (5)
p(G) = p(gl,n ) = CN (gl,n ; 0, νg2 ) (7)
l=1 n=1 l=1 n=1
where B  GZ and the likelihood p(Y|B) is given as
N 
 T N 
 T
p(Z ) = p(zn,t ) = sn,t CN (zn,t ; 0, νz2 )
L 
 T
n=1 t=1 n=1 t=1
p(Y |B) = p(yl,t |bl,t ) 
l=1 t=1 + (1 − sn,t )δ(zn,t ) (8)
L  T  
= exp −|yl,t − bl,t |2 /σ 2 . (6) where νg2 and νz2 are respectively the average variances of the
l=1 t=1 LIM-user channel matrix G and the non-zero elements of the

Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 23,2022 at 02:15:14 UTC from IEEE Xplore. Restrictions apply.
 HE AND YUAN: CASCADED CHANNEL ESTIMATION FOR LIM ASSISTED MASSIVE MIMO 213

sparse matrix Z; the on/off state sn,t ∈ {0, 1} at time t for D. Computational Complexity
the n-th reflect element of the LIM is known by the receiver We now offer a brief discussion on the computational com-
because S is a pilot signal. plexity of the proposed JBF-MC algorithm. Note that the
To achieve computational efficiency, the BiG-AMP lever- total computational complexity of the JBF-MC consists of
ages the central limit theorem and the Gaussian approximation implementations of the BiG-AMP for matrix factorization
for the involved messages. Details of the BiG-AMP algorithm and the RGrad for matrix completion and the computation
can be found from Lines 1 to 24 of Algorithm 1. At the i-th involved in (13). We thus sketch the respective complex-
iteration of the BiG-AMP, the means and variances of {gl,n } ity as follows. First, the complexity of the BiG-AMP is
and {zn,t } from Lines 16 to 19 are respectively calculated with dominated by basic matrix multiplications in Lines 4–6 and
respect to the approximate marginal posterior distributions: Lines 12–15 of Algorithm 1, requiring O(LNT ) flops per
 
q iteration. Consequently, the complexity of the BiG-AMP is
p̂ (i) (gl,n ) ∝ p(gl,n ) CN gl,n ; q̂l,n (i ), vl,n (i ) (9) at most Imax Jmax O(LNT ) flops, where Imax and Jmax are
 
the maximum numbers of the outer and inner iterations of
p̂ (i) (zn,t ) ∝ p(zn,t ) CN zn,t ; r̂n,t (i ), vn,t
r (i ) (10)
the BiG-AMP, respectively. Second, the complexity of the
where p(gl,n ) and p(zn,t ) are the prior distributions defined RGrad algorithm is dominated by the calculations in Lines
in (7) and (8), respectively. 27 and 29, requiring O(rNT ) and O(r 3 ) flops, respec-
It is worth noting that the K-SVD (K-means singular tively. Therefore, the total cost of the RGrad is at most
value decomposition) [12] and the SPAMS (SPArse Modeling Kmax O(rNT ) flops by noticing that r < min{N , T }, where
Software) [13] can be exploited to solve the bilinear sparse Kmax denotes the maximum number of iterations of the
matrix factorization problem (4). However, these approaches RGrad. Finally, the computation of (13) in Line 33 requires
perform much worse than the BiG-AMP algorithm, as will O(M 2 T ) + O(M 3 ) + O(MNT ) flops.
be seen from the numerical results presented in Section V.
Additionally, note that the likelihood p(Y | B) and the pri- V. N UMERICAL S TUDIES
ors p(G) and p(Z) are characterized by the parameters σ 2 , We now carry out numerical experiments to corroborate
νg2 , and νz2 , respectively. These parameters can be esti- the effectiveness of the proposed JBF-MC algorithm for the
mated by utilizing the standard expectation-maximization cascaded channel estimation of the LIM-assisted massive
methodology [6], [14]. In this letter, we do not elaborate on MIMO systems. We assume that the antenna elements form
how to estimate these parameters due to space limitation. a half-wavelength uniform linear array (ULA) configuration
at the BS, the LIM, and the receiver side. Following the
C. Matrix Completion superposition principle of different paths in the prorogation
To facilitate the estimate of H, we now recover the missing environment [9], [14], the baseband BS-LIM and LIM-user
 by using the rank-deficient property of the original
entries of Z channel matrices H and G are respectively modeled as
H. We employ the Riemannian gradient (RGrad) algorithm [7] Kh Kg
 
to solve the matrix completion problem, which is summarized H = αk a L (ϑk )a H
T (ωk ), G = βk a R (ψk )a H
L (ϑk )
from Lines 25 to 32 of Algorithm 1. The RGrad algorithm is
k =1 k =1
to solve the following matrix completion problem:
where Kh and Kg stand for the number of propagation paths
1 
 )2 subject to rank(A) = r . (11)
min S ∗  (A−Z F
of radio signals in the BS-LIM channel and the LIM-user
A 2 channel, respectively; a L (ϑk ) ∈ CN , a T (ωk ) ∈ CM , and
In Lines 28 and 29 of the JBF-MC algorithm, PS(k ) (·) stands a R (ψk ) ∈ CL are the steering vectors of the ULA at the
for the projection operation to the left singular vector subspace LIM, the BS, and the receiver side, respectively. In each trial,
(denoted by S(k )) of the current estimate A(k), corresponding the angular parameters ψr , ϑr , and ωk independently follow
to the first r eigenvalues of A(k). In Line 30 of Algorithm 1, from the uniform distribution within (0, 1], and the path gain
Hr (W ) is the hard-thresholding operator for the best rank-r coefficients {αk } and {βk } are independently drawn from
approximation of the associated SVD, i.e., CN (0, 1). The number of paths in the channel matrix H, i.e.,
   Kh , is set to be enough small such that it has a low-rank struc-
H Σ(i , i ), i ≤ r ture to facilitate its estimation in the matrix completion stage.
Hr W  U Σr V , Σr (i , i ) = (12)
0, i >r In addition, our empirical experiments suggest that choosing
Kg to be a positive integer more than half of the receive
where W  U ΣV H is the SVD of W and Σ(i , i ) is the antenna number L or the passive antenna number N in the
(i, i)-th entry of Σ. Finally, the estimate of the channel matrix LIM can lead to a better performance. The sampling matrix
H can be computed as S, i.e., the on/off states of the elements in the LIM, are set to
 †
H = AX (13) be a random 0-1 matrix.
The pilot symbols in X are generated from CN (0, 1) and
where X † = (X X H )−1 X is the Moore-Penrose inverse and the signal-to-noise ratio (SNR) is defined as 10log10 (1/σ 2 ) dB.
A is output of the RGrad algorithm. Here, we assume that the The estimation performance is evaluated in terms of the nor-
pilot length T is no less than the number of transmit antennas malized mean-square-error (NMSE). All the simulation results
M and rank (X) = M, so as to ensure the existence of X † . It is are obtained by averaging 200 independent trials. Notice that
 and H have a diagonal
worth noting that the final estimates G the outputs G  and H of the JBF-MC algorithm still con-
ambiguity resulting from the sparse matrix factorization stage. tain diagonal ambiguities. These ambiguities are eliminated
Nevertheless, recall the discussions from Section III that there based on the true values of G and H in the calculation of the
is no need to eliminate the diagonal ambiguity. NMSEs. In the sparse matrix factorization stage, to benchmark

Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 23,2022 at 02:15:14 UTC from IEEE Xplore. Restrictions apply.
 214 IEEE WIRELESS COMMUNICATIONS LETTERS, VOL. 9, NO. 2, FEBRUARY 2020

matrix completion fails if λ is too small; the sparse matrix fac-

torization fails if λ is too large. This is because when a larger
number of samples (i.e., large λ) are available for matrix com-
pletion, the performance of the BiG-AMP algorithm becomes
worse as a large number of non-zero variables are needed to
be estimated in the sparse matrix factorization stage.

VI. C ONCLUSION
In this letter, we considered the problem of cascaded
channel estimation of the LIM-assisted MIMO systems. We
introduced a general framework for this problem by lever-
aging a combined bilinear spare matrix factorization and
matrix completion, and presented a two-stage algorithm that
includes the bilinear generalized message passing algorithm
for sparse matrix factorization and the Riemannian manifold
gradient-based algorithm for matrix completion. We provided
experimental evidences that the proposed approach achieves
Fig. 2. NMSEs of G (left subplots) and H (right subplots) versus the SNR
and the number of pilots with N = 70, M = L = 64, and λ = 0.2.
an accurate channel estimation for the LIM-assisted MIMO
systems.

R EFERENCES
[1] C. Liaskos, S. Nie, A. Tsioliaridou, A. Pitsillides, S. Ioannidis,
and I. Akyildiz, “A new wireless communication paradigm through
software-controlled metasurfaces,” IEEE Commun. Mag., vol. 56, no. 9,
pp. 162–169, Sep. 2018.
[2] Q. Wu and R. Zhang, “Intelligent reflecting surface enhanced wireless
network via joint active and passive beamforming,” IEEE Trans. Wireless
Commun., to be published.
[3] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and
C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency in
wireless communication,” IEEE Trans. Wireless Commun., vol. 18, no. 8,
pp. 4157–4170, Aug. 2019.
[4] T. Sleasman et al., “Experimental synthetic aperture radar with
dynamic metasurfaces,” IEEE Trans. Antennas Propag., vol. 65, no. 12,
pp. 6864–6877, Dec. 2017.
[5] W. Yan, X. Kuai, and X. Yuan. (May 2019). Passive Beamforming
and Information Transfer via Large Intelligent Metasurface. [Online].
Available: https://arxiv.org/abs/1905.01491
[6] J. T. Parker, P. Schniter, and V. Cevher, “Bilinear generalized approxi-
mate message passing—Part I: Derivation,” IEEE Trans. Signal Process.,
Fig. 3. Phase transitions (in NMSE) of G (left subplots) and H (right sub- vol. 62, no. 22, pp. 5839–5853, Nov. 2014.
plots) versus the sparsity level and the number of pilots. The upper subplots [7] K. Wei, J.-F. Cai, T. F. Chan, and S. Leung, “Guarantees of Riemannian
are for SNR = 10 dB and the lower subplots for SNR = 20 dB. optimization for low rank matrix recovery,” SIAM J. Matrix Anal. Appl.,
vol. 37, no. 3, pp. 1198–1222, Sep. 2016.
[8] A. Taha, M. Alrabeiah, and A. Alkhateeb. (Apr. 2019). Enabling Large
Intelligent Surfaces With Compressive Sensing and Deep Learning.
our JBF-MC algorithm for the estimation of G, we adopt the [Online]. Available: https://arxiv.org/abs/1904.10136
K-SVD [12] and the SPAMS [13] for comparison. In all sim- [9] J. Zhang, X. Yuan, and Y.-J. A. Zhang, “Blind signal detection in massive
ulations, we set N = 70, M = L = 64, Kh = r = 4, and MIMO: Exploiting the channel sparsity,” IEEE Trans. Commun., vol. 66,
Kg = 35. For comparison, in the matrix completion stage no. 2, pp. 700–712, Feb. 2018.
we also adopt the IHT (iterative hard thresholding) and the [10] Z. Pi and F. Khan, “An introduction to millimeter-wave mobile broad-
band systems,” IEEE Commun. Mag., vol. 49, no. 6, pp. 101–107,
IST (iterative soft thresholding) [15] to estimate the BS-LIM Jun. 2011.
channel H. [11] R. B. Green, “The general theory of antenna scattering,” Ph.D. disserta-
The NMSEs versus the SNR (with T = 300) and the num- tion, Elect. Comput. Eng. Dept., Ohio State Univ., Columbus, OH, USA,
ber of pilots (with SNR = 10 dB) are depicted in Fig. 2 under 1963.
the sparsity level (sampling rate) λ = 0.2. It is observed that [12] M. Aharon, M. Elad, and A. M. Bruckstein, “K-SVD: An algorithm
for designing overcomplete dictionaries for sparse representation,” IEEE
the proposed JBF-MC algorithm has a significant performance Trans. Signal Process., vol. 54, no. 11, pp. 4311–4322, Nov. 2006.
gain over the baseline methods, especially for the estimation [13] J. Mairal, F. Bach, J. Ponce, and G. Sapiro, “Online learning for
of G. We also observe that the NMSE of H is smaller than matrix factorization and sparse coding,” J. Mach. Learn. Res., vol. 11,
that of G. This is because Z is a sparse matrix with a much pp. 19–60, Mar. 2010.
smaller number of non-zero variables to be estimated. Fig. 3 [14] C. Huang, L. Liu, C. Yuen, and S. Sun, “Iterative channel estimation
using LSE and sparse message passing for mmWave MIMO systems,”
shows the phase transitions of the NMSEs of the channel esti- IEEE Trans. Signal Process., vol. 67, no. 1, pp. 245–259, Jan. 2019.
mation versus the sparsity level and the number of pilots. It is [15] J.-F. Cai, E. J. Candès, and Z. Shen, “A singular value threshold-
seen from Fig. 3 that there is a performance tradeoff between ing algorithm for matrix completion,” SIAM J. Optim., vol. 20, no. 4,
the sparse matrix factorization and the matrix completion: The pp. 1956–1982, Mar. 2010.

Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on May 23,2022 at 02:15:14 UTC from IEEE Xplore. Restrictions apply.