1
Spatially Non-Stationary XL-MIMO Channel
Estimation: A Three-Layer Generalized
Approximate Message Passing Method
Anzheng Tang∗ , Jun-Bo Wang∗ , Yijin Pan∗ , Wence Zhang† , Xiaodan Zhang‡ ,
Yijian Chen§ , Hongkang Yu§ , and Rodrigo C. de Lamare¶
Abstract—In this paper, channel estimation problem for ex-
arXiv:2403.02633v1 [cs.IT] 5 Mar 2024
tremely large-scale multi-input multi-output (XL-MIMO) sys-
tems is investigated with the considerations of the spherical
wavefront effect and the spatially non-stationary (SnS) property.
Due to the diversities of SnS characteristics among different
propagation paths, the concurrent channel estimation of multiple
paths becomes intractable. To address this challenge, we propose
a two-phase channel estimation scheme. In the first phase, the
angles of departure (AoDs) on the user side are estimated,
and a carefully designed pilot transmission scheme enables the
decomposition of the received signal from different paths. In
the second phase, the subchannel estimation corresponding to
different paths is formulated as a three-layer Bayesian inference
problem. Specifically, the first layer captures block sparsity in
the angular domain, the second layer promotes SnS property
in the antenna domain, and the third layer decouples the
subchannels from the observed signals. To efficiently facilitate
Bayesian inference, we propose a novel three-layer generalized
approximate message passing (TL-GAMP) algorithm based on
structured variational massage passing and belief propagation
rules. Simulation results validate the convergence and effective-
ness of the proposed algorithm, showcasing its robustness to
different channel scenarios.
Index Terms—XL-MIMO systems, SnS property, channel es-
timation, approximate message passing.
I. I NTRODUCTION
With the advancement of wireless communication technol-
ogy and the growing demand for higher data rates, it is antic-
ipated that the number of antennas and the array aperture will
significantly exceed those used in existing massive multiple-
input-multiple-output (MIMO) systems [1], [2]. This trend has
given rise to the concept of extremely large-scale MIMO (XL-
MIMO) [3], [4], which involves deploying an exceptionally
large number of antennas in a compact space with a discrete
Anzheng Tang, Jun-bo Wang and Yijin Pan are with the National Mobile
Communications Research Laboratory, Southeast University, Nanjing 210096,
China. (email: {anzhengt, jbwang, and panyj}@seu.edu.cn)
Wence Zhang is with the School of Computer Science and Communication
Engineering and the Jiangsu Key Laboratory of Security Technology for
Industrial Cyberspace, Jiangsu University, Zhenjiang 212013, China. (e-mail:
wencezhang@ujs.edu.cn)
Xiaodan Zhang is with the School of Management, Shenzhen Insti-
tute of Information Technology, Shenzhen 518172, China. (e-mail: Dan-
nyzxd@163.com)
Yijian Chen and Hongkang Yu are with the Wireless Product Research
and Development Institute, ZTE Corporation, Shenzhen 518057, China. (e-
mail:{yu.hongkang, chen.yijian}@zte.com.cn)
Rodrigo C. de Lamare is with the Centre for Telecommunications Studies,
Pontifical Catholic University of Rio de Janeiro, Rio de Janeiro 22451-900,
Brazil, and also with the Department of Electronic Engineering, University
of York, York YO10 5DD, U.K. (e-mail: delamare@puc-rio.br).
or even continuous aperture [5], [6]. Thanks to the substantial
beamforming gain and vast spatial degrees of freedom (DoF),
XL-MIMO is considered a promising technology for beyond
5G and 6G communications.
In XL-MIMO systems, the significant increase in the array
aperture results in near-field transmission [4], [5], [7]–[9].
Consequently, various channel characteristics, not typically
considered in conventional massive MIMO systems, require
exploration, for example, the spherical wavefront effect and
SnS property. Specifically, the large-aperture antenna array
leads to a non-negligible Rayleigh distance. As a result, the
plane wavefront assumption becomes inapplicable to near-field
XL-MIMO channels, and instead, the spherical wavefront ef-
fects should be considered [10], [11]. Moreover, the extremely
large array aperture introduces the SnS property along the ar-
ray. Intuitively, different portions of the array could observe the
propagation environment from different perspectives, implying
that the array elements can capture the signal from a certain
propagation path but with varying powers. To quantitatively
characterize the SnS property, the visibility region (VR)-based
channel modeling method is widely adopted in XL-MIMO
systems [12]–[16]. Considering both the near-field effect and
SnS property, the problem of channel estimation for XL-
MIMO systems becomes much more difficult.
A. Related Works
While a number of methods have been proposed to estimate
the channels in XL-MIMO systems, most of them only apply
to spatially stationary channel models [17]–[21]. To address
the SnS estimation problem, [22] proposed subarray-wise and
scatterer-wise estimation schemes. However, these schemes
were heuristic and did not fully exploit the inherent structure
of the XL-MIMO channels, such as potential characteristics
in the antenna or angular domains. Moreover, based on the
assumption of subarray-wise VRs, [23] proposed a group
time block code (GTBC) based signal extraction scheme
for each subarray in [23]. Subsequently, the SnS channel
estimation was transformed into several spatially stationary
estimation tasks. Nevertheless, the method overlooked the
spatial correlation among subarrays, and the subarray-wise
VR assumption may be idealistic, as users or scatterers might
only have visibility to a portion of array elements in each
subarray. Additionally, in the context of SnS reconfigurable
intelligent surface (RIS) cascaded channels, a three-step chan-
nel estimation and VR detection scheme was proposed in [12],

where the VR detection method, relying on receive signal
power, is contingent on the accuracy of channel estimation
and is sensitive to noise levels. Particularly, in low signal-to-
noise ratio (SNR) scenarios, this approach may suffer severe
performance degradation. Furthermore, the method becomes
impractical when considering multipath propagation between
the RIS and users.
Considering the inherent antenna-domain or delay-domain
sparsity in SnS channels, various Bayesian inference-based
methods have been proposed. Exploiting delay-domain spar-
sity, [24] introduced an adaptive grouping sparse Bayesian
learning (AGSBL) scheme for uplink channel estimation.
Utilizing antenna-domain sparsity resulting from the SnS
property, [25] characterized the XL-MIMO channel with a
subarray-wise Bernoulli-Gaussian distribution. Subsequently,
a bilinear message passing (MP) algorithm was developed
for joint user activity detection and channel estimation in
the presence of the SnS property. To simultaneously capture
antenna- and delay-domain sparsity, [26] proposed a structured
prior model with the hidden Markov model (HMM) to cap-
ture the characteristics of VR and delay domain clustering.
While these contributions primarily focused on characterizing
antenna-domain or delay-domain channels using statistical
distributions such as mean and variance, they overlooked
essential propagation characteristics, which potentially lead to
a degradation in estimation performance.
To address the aforementioned issues, [27] proposed to
jointly utilize the sparse properties of channels in the polar
domain and VR vectors and developed a joint channel esti-
mation and VR detection algorithm based on the fast sparse
Bayesian learning (FSBL) framework. However, the FSBL-
based method involved matrix inversion operations in each
iteration, resulting in extremely high computational complex-
ity. Furthermore, in our previous work [28], we proposed
a two-stage VR detection and channel estimation scheme
based on message passing (MP) and orthogonal matching
pursuit (OMP), considering the antenna-domain and angular-
domain characteristics. Nervelessness, a notable drawback of
this method is the manual separation of the two stages, leading
to inadequate utilization of angular-domain information during
the VR detection stage. Additionally, it is worth mentioning
that the aforementioned estimation methods are developed in
the context of a multiple-input-single-output (MISO) scenario,
and direct extension to the MIMO scenario is not feasible.
B. Main Contributions
In this paper, we study the XL-MIMO channel estimation
problem, considering the user equipped with multiple anten-
nas and accounting for spherical wavefront effects and SnS
property. Given the distinct characteristics inherent in various
propagation paths, the concurrent estimation of multiple paths
poses a significant challenge. To this end, we present a two-
phase channel estimation scheme. The main contributions can
be summarized as follows:
• We propose a two-phase channel estimation scheme tai-
lored for the SnS XL-MIMO channel estimation problem.
Specifically, in the first phase, the AoDs on the user side
2
are estimated, and a carefully designed pilot transmis-
sion scheme enables the decomposition of the received
signal from different SnS paths. In the second phase, the
subchannel estimation corresponding to different paths is
estimated.
• To solve the cascaded estimation of VRs and channels
coefficients, we formulate the subchannels estimation as a
three-layer Bayesian inference problem. To be specific, in
the first layer, the goal is to capture the block sparsity of
the SnS channel within the angular domain. The second
layer focuses on capturing the SnS property in the antenna
domain. Moving forward, the third layer facilitates the
decoupling of subchannels from the observed signal.
• To efficiently achieve Bayesian inference, we pro-
pose a computationally efficient three-layer generalized
message-passing (TL-GAMP) algorithm by combining
structured variational message passing and belief prop-
agation rules. Additionally, in the first and third layers,
we simplify the message calculations by omitting certain
infinitesimal terms, thereby significantly reducing com-
putational complexity.
• In comparison to existing baselines, the proposed chan-
nel estimation algorithm demonstrates superior perfor-
mance due to the joint exploitation of angular-domain
and antenna-domain characteristics. Meanwhile, the al-
gorithm demonstrates remarkable robustness in handling
different channel scenarios.
C. Organization and Notations
Organization: The rest of this paper is organized as follows.
In Section II, we introduce the system model. Section III
presents a two-phase channel estimation scheme. In Section
IV, the subchannel estimation problem is formulated as a
three-layer generalized bilinear inference problem, and a com-
putationally effective algorithm is proposed. Simulations are
carried out in Section V. Finally, conclusions are drawn in
Section VI.
Notations: lower-case letters, bold-face lower-case let-
ters, and bold-face upper-case letters are used for scalars,
T
vectors and matrices, respectively; The superscripts (·)
H
and (·) stand for transpose and conjugate transpose, re-
spectively; diag (x1 , x2 , . . . , xN ) denotes a diagonal ma-
trix with {x1 , x2 , . . . , xN } being its diagonal elements;
blkdiag (X1 , X2 , . . . , XN ) denotes a block diagonal matrix
with {X1 , X2 , . . . , XN } being its diagonal elements; CM ×N
denotes an M ×N complex matrix. In addition, a random vari-
able x ∈ C drawn from the complex Gaussian distribution with
mean m and variance v is characterized by
n the probability
o den-
2
sity function (PDF) CN (x; m, v) = exp −|x − m| /v /πv;
a random variable γ ∈ R from Gamma distribution with
mean a/b and variance a/b2 is characterized by the PDF
Ga(γ; a, b) = γ a−1 exp(−γb); mna →nb (x) indicates a mes-
sage passed from node na to node nb .
II. S YSTEM M ODEL
Consider a narrowband millimeter-wave (mmWave) XL-
MIMO system, where the time-division duplexing (TDD)
mode is adopted. As depicted in Fig. 1, a BS with NR antennas

BS Side NR
RF
Analog Beam
chain
Digital Beam
Ă
forming
Ă
forming
RF
chain
Ă
RF
chain
Fig. 1. Illustration of mmWave XL-MIMO system with a hybrid beamforming architecture.
and NRF radio frequency (RF) chains serves a user1 with
NT ≪ NR antennas. Without loss of generality, we assume
that the transceivers are arranged with uniform linear arrays
(ULAs), and the antenna spacing is denoted by d = λ/2,
where λ indicates the carrier wavelength. The set of BS
antenna elements is given by N = {1, 2, · · · , NR }.
The large-aperture antenna array on the BS side results
in a more significant Rayleigh distance, placing the user in
the near-field region of the BS. Consequently, the assump-
tion of the plane wavefront, commonly employed in far-field
communications, becomes inapplicable [7]. Furthermore, with
the increase in array aperture, the presence of obstacles and
incomplete scattering in wireless propagation environments
introduces the SnS property along the array on the BS side
[12], [14]–[16].
To account for the spherical wavefront effect and the
SnS property, we employ a single-side near-field model to
characterize the uplink channel H ∈ CNT ×NR , which can be
represented as [14], [26]
p L
X ing that the channels are block-fading, the channel parameters
H= NT NR gl (sl ⊙ aR (ϑl , rl )) aH
T (ψl ),
l=1
where L is the number of propagation paths, and when l = 1,
it refers to the line-of-sight (LoS) path, while l > 1 indicates
non-line-of-sight (NLoS) paths; gl represents the l-th complex
path gain; sl = [sl,1 , sl,2 , · · · , sl,NR ]T ∈ ZNR ×1 is the visibil-
ity indicator vector of the l-th path; ⊙ indicates the Hadamard
product. If the l-th propagation path is spatially stationary, sl is
represented as an all-one vector, i.e., sl = 1NR ×1 . Conversely,
if the l-th path is SnS, the n-th entry of sl is modeled as [12],
[14], [28] ( time slots, as depicted in Fig. 2.
1, if n ∈ φl ,
sl,n = (2)
0, otherwise.
where set φl ⊆ N denotes the VR for the l-th path, i.e., the
l-th path is visible to elements in φl , and is invisible to the
elements outside φl . Moreover, we define ϕl = |φl |/NR , indi-
cating the proportion of visible elements to the l-th path, where
|φl | denotes the cardinality of the set φl . The parameters ϑl
and rl denote the angle of arrival (AOA) on the BS side and the
distance between the reference antenna element on the BS side
1 Notably, considering the orthogonal pilots for different user, the proposed
estimation scheme can be straightforwardly extended to multi-user scenario.
3
H User Side
VR 1 NT RF
Digital Beam
Analog Beam
chain
forming
forming
Ă
Ă
Ă
RF
chain
VR L Scatterers
and the scatterers or user in the l-th path, respectively. Without
loss of generality, we consider the first antenna element as
the reference element. Additionally, ψl represents AoD of
the user in the l-th path. Finally, aT (ψl ) ∈ CNT ×1 and
aR (ϑl , rl ) ∈ CNR ×1 are respectively the far-field and near-
field array steering vectors at the user and BS sides, which
are respectively given by
1 2π
aT (ψl ) = √ [1, · · · , e−j λ (NT −1)d sin ψl ]T , (3)
NT
1 2π 2π
aR (ϑl , rl ) = √ [ej λ ∆l,1 , · · · , ej λ ∆l,NR ]T , (4)
NR
where ∆l,n = −d(n − 1) sin ϑl + d2 (n − 1)2 cos2 ϑl /(2rl ) is
the wave path difference between the n-th element and the
reference element for the l-th path.
III. T WO -P HASE E STIMATION S CHEME
Due to the channel reciprocity in TDD mode, the estimation
of the uplink channel is sufficient [29]. Meanwhile, consider-
(1) remain constant within each coherence block. Assume that
each coherence block is divided into several consecutive
subframes, and P consecutive subframes are utilized for the
channel estimation task with each subframe further divided
into K time slots. In each uplink subframe, the user only
activates a single RF chain to transmit the pilot signal on one
beam while the BS combines the received pilot signal using all
RF chains associated with different beams [30]. Specifically,
the BS utilizes K time slots to traverse all received beams.
Subsequently, the user changes the transmit beam every K
Given that each propagation path exhibits different SnS
property, the concurrent estimation of multipath presents a
significant challenge. To tackle this issue, we propose a two-
phase estimation scheme in this paper. In the first phase, the
AoDs can be initially estimated. According to the obtained
AoDs, the user can select a suitable transmit beam aligned
with the AoDs. This allows the BS to effectively decouple the
received signals from multipaths. Moving to the second phase,
we proceed to estimate the remaining path parameters. Finally,
through the superposition of all paths, we obtain the complete
channel information. In the following, we will elaborate on
the details of the two-phase estimation scheme.
 4

Coherence Block where G = diag(g1 , · · · , gL )∈ CL×L is the path gain.

P subframes (9), we have Y0 = CAH
By combining
√ (8) with T F + N,
Uplink channel estimation Downlink or uplink data transmission
where C = NT NR W AR G. Define M = AT CH , the re-
H

covery of ψl based on Y0H can be formulated as a semidefinite

program (SDP) problem [31], i.e.,
K time slots K time slots K time slots µ µ
min tr(Z) + tr (Toep(u))
UE f1 f2 Ă fP u, M, Z 2M 2NT
1 2
+ Y0H − FH M F (10)
BS W1 W2 Ă WK W1 W2 Ă WK Ă W1 W2 Ă WK 2 
Toep(u) M
s.t. ⪰ 0,
P0 subframes P  P0 subframes MH Z
Phase I: Phase II:
AoDs estimation Subchannels estimation
where µ is a regularization factor, u and Z are two auxiliary
variables; Toep(u) is a Toeplitz matrix with u being its
Fig. 2. Two-phase uplink channel estimation scheme. first row. Notably, the problem in (10) is convex and can
be directly solved using CVX [32]. The recovery of ψ̂l is
1) Stage I: AoDs Estimation: In the first phase, as the
then based on the solution of Toep(u) by the Multiple Signal
channel parameters are unknown, both the user and BS ran-
Classification (MUSIC) or Estimating Signal Parameter via
domly select transmit and combiner beams from predefined
Rotational Invariance Techniques (ESPRIT) [33], [34].
codebooks. Let fp ∈ CNT ×1 represent the transmit beam
2) Stage II: Subchannels Estimation: Once the AoD param-
in the p-th subframe, and Wk ∈ CNR ×NRF denotes the
eters are obtained, we can appropriately design the transmit
combiner matrix for the k-th time slot in the p-th subframe.
beams in the subsequent subframes to align with the estimated
Mathematically, the received pilot signal yp,k ∈ CNRF ×1 at
AoDs. In particular, we have fp = aT (ψ̂l ) for P0 +1 ≤ p ≤ P
the BS in the k-th time slot of the p-th subframe is given by
and 1 ≤ l ≤ L with P − P0 ≥ L. In this manner, the received
yp,k = WkH Hfp sp,k + WkH np,k , (5) signal in the (P0 + l)-th subframe can be expressed as
where sp,k ∈ C and np,k ∼ CN (np,k ; 0, β −1 INR ) ∈ CNRF ×1 yP0 +l ≈ WH HaT (ψ̂l ) + nP0 +l . (11)
denote the transmit pilot symbol and noise vector with mean
Due to the asymptotic orthogonality of the transmit steering
zero and covariance matrix β −1 INR , respectively.
vector, we have aHT (ψl )aT (ψ̂l ) ≈ 1, if ψ̂l = ψl , otherwise 0.
Without loss of generality, we assume sp,k = 1 for all p
Therefore, (11) can be rewritten as
and k, since the pilot sequence is known at the BS and can
be readily eliminated [29]. By collecting the received pilot yP0 +l ≈ WH hl + nP0 +l , (12)
signals from different receive beams corresponding to the p-th √
T T T where hl ≜ NT NR gl Sl aR (ϑl , rl ) is defined as the l-th
transmit beam, the pilot signal yp = [yp,1 , yp,2 , · · · , yp,K ]T ∈
M ×1 subchannel with Sl = diag(sl ). For notation simplification,
C with M = KNRF in the p-th subframe can be
in the subsequent discussion, we simplify yP0 +l and nP0 +l as
expressed as
yl and nl .
yp = WH Hfp + np , (6)
In the second stage, our aim is to estimate the subchannel
where W = [W1T , W2T , · · · , WK T
] ∈ CNR ×M ; np = hl and extract the corresponding visibility vector sl according
[(W1 np,1 )T , (W2 np,2 )T , · · · , (WK np,K )T ]T ∈ CM ×1 is the to the received signal yl . To enable efficient estimation, it
equivalent noise vector with covariance matrix is crucial to leverage the sparsity of channels. From the
perspective of sparsity level, the polar domain emerges as the
R = blkdiag(β −1 W1H W1 , · · · , β −1 WK
H
WK ). (7) preferred choice for near-field channels [17]–[19]. However, it
Assuming the utilization of P0 subframes in the first phase2 , necessitates sampling in both the distance and angle domains,
thus, we have resulting in a significantly larger codebook size and increased
Y0 = WH HF + N, (8) computational complexity. Consequently, we propose to utilize
the angular-domain sparsity as an alternative, as demonstrated
where F = [f1 , · · · , fP0 ] ∈ CNT ×P0 , Y0 = [y1 , · · · , yP0 ] ∈ in our previous works [7], [28].
CM ×P0 and N = [n1 , · · · , nP0 ] ∈ CM ×P0 are the stacked
Theorem 1. The angular-domain representation of the SnS
transmit beam matrix, received pilot matrix and noise matrix,
subchannel hl is given by
respectively.
Additionally, stacking all transmitted steering vectors and hl ≈ Sl Dcl , (13)
received steering vectors as AT = [aT (ψ1 ), · · · , aT (ψL )] ∈
CNT ×L and AR = [s1 ⊙ aR (ϑ1 , r1 ), · · · , sL ⊙ aR (ϑL , rL )] ∈ where cl ∈ CQ×1 denotes the block-sparse angular channel
CNR ×L , the channel in (1) can be compactly expressed as vector; D = [a(θ1 ), a(θ2 ), · ·n· , a(θQ )] ∈ CNR ×Q is angular- o
p domain codebook, where θq : sin(θq ) = 2q−1−Q Q ,q ∈ Q
H = NT NR AR GAH T, (9) √
and a(θq ) = 1/ NR [1, e−jπ sin(θq ) · · · , e−jπ(NR −1) sin(θq ) ]T .
2 We intentionally choose P ≫ L to ensure a super-resolution estimation
0
Q ≥ NR indicates the discrete grid number of angular space
in the first stage. and Q = {1, 2, · · · , Q}.
 5

Proof: We first consider the inverse Discrete Fourier 1st layer 2nd layer 3rd layer
transformation (IDFT) of hl in the angular domain, thus, we D Sl A nl
have hl = Dcl . Then, utilizing the fact that sl,n ∈ {0, 1}, we rl
can easily obtain hl =Sl hl = Sl Dcl . Therefore, the proof is cl xl tl zl
completed. xl  Dcl tl  Sl xl zl  Atl rl  zl nl
Remark 1. Notably, through the IDFT operation about hl ,
Fig. 3. The three-layer generalized bilinear inference problem is to
the SnS property has been implicitly captured in cl . However, estimate the angular-domain channel cl and visibility indicator
accurately extracting SnS information directly from cl is chal- matrix Sl from the received pilot signal yl .
lenging. An alternative approach is to reconstruct the subchan-
nel ĥl based on the estimated ĉl . Then, SnS information can be TABLE I. Factor and Distribution in (20)
obtained using energy detection-based methods [12]. However, Factor Distribution Function
these methods depend on the accuracy of the reconstruction
fβ (β) p(β) β −1
of ĥl and are sensitive to noise levels. Particularly, in low fγ l (γ l ) p(γ l ) Ga(γl,q ; ξ, η)
SNR scenarios, this approach might suffer severe performance fcl (cl , γ l ) p(cl |γ l ) (17)
degradation. Related simulation results can be found in our fsl (sl ) p(sl ) (18)
fxl (xl , cl ) p(xl |cl ) δ(xl − Dcl )
previous work [28]. To address this issue, we propose explicitly ftl (tl , sl , xl ) p(tl |sl , xl ) δ(tl − Sl xl )
characterizing SnS characteristics, as shown in (13). This fzl (zl , tl ) p(zl |tl ) δ(zl − Atl )
provides an opportunity to directly estimate sl . frl (rl , zl , β) p(rl |zl , β) CN (rl ; zl , β −1 IM )

Utilizing Theorem 1, the received pilot signal in (12) can
be rewritten as
yl ≈ WH Sl Dcl + nl , (14)
According to (14), the estimation of subchannel hl is trans-
formed into the joint estimation of Sl and cl . Since the
equivalent noise nl is colored, we consider to perform a pre-
whitening procedure. Assume that the noise covariance matrix
R in (7) can be decomposed by the Cholesky factorization as
R = β −1 BBH , where B ∈ CM ×M is a lower triangular
matrix. In this manner, we can obtain the pre-whitening
transformation matrix B−1 . Then, the whitened received pilot
signal from the l-th path can be rewritten as
el = B−1 yl ≈ PSl Dcl + n
y e l = Φl c l + n
el, (15)
where P = B−1 WH , n e l = B−1 nl and Φl = PSl D.
After whitening, the covariance matrix of the noise n e l is
B−1 RB−H = β −1 IM ×M , thus, n e l becomes white noise.
The SnS property introduces a dependency between the
equivalent measurement matrix Φl and Sl , yet during the
channel estimation phase, Sl remains unknown. While it seems
intuitive to estimate Φl and cl through bilinear inference,
establishing a statistical prior model for Φl proves challeng-
ing due to its dependence on P, Sl , and D. Consequently,
existing bilinear inference methods become inapplicable. This
highlights the pressing need for novel channel estimation
techniques capable of addressing the challenges posed by the
SnS property in XL-MIMO systems.
IV. P ROPOSED S UBCHANNEL E STIMATION A LGORITHM
In this section, we formulate the subchannel estimation
as a three-layer generalized bilinear inference problem, and
propose a computationally effective algorithm based on belief
propagation and variational message passing.
A. Probability Model and Factor Graph Representations
To address the potential divergence problem caused by the
generic measurement matrix P, we first perform the unitary
transformation on P, i.e., P = UΛVH [35], [36]. Thus, the
received pilot signal can be further rewritten as
rl ≈ ASl Dcl + nl , (16)
where rl = UH y el , A = ΛVH , and nl = UH n e l . Note that,
since U is a unitary matrix, nl is still a zero mean Gaussian
noise vector with the same covariance matrix as n el.
The estimation of the SnS channel, as depicted in (16), can
be formulated as a three-layer Bayesian inference problem
with a single measurement vector (SMV) [37], wherein the
simultaneous estimation of the visibility indicator matrix Sl
and the angular channel cl is required. More intuitively, we
introduce the intermediate variables xl ≜ Dcl , tl ≜ Sl xl , and
zl ≜ Atl , the three-layer Bayesian inference framework is
shown in Fig. 3, where blank circles denote the realizations of
the random variables that need to be estimated and gray boxes
indicate the function operations. The measurement matrix of
the first and third layers D and A are known and the second-
layer measurement matrix Sl needs to be estimated.
To capture the angular-domain block sparsity, we assume
that the widely used two-layer hierarchical prior for cl [38].
Specifically, each element cl,q has a conditionally independent
distribution expressed as
Q Q
Y Y
p(cl |γ l ) = p(cl,q |γl,q ) = −1
CN (cl,q ; 0, γl,q ), (17)
q=1 q=1
where γl,q ∼ Ga(γl,q ; ξ, η) is modeled as a Gamma prior with
ξ and η as the shape parameters to ensure a positive variance.
−1
It is evident that cl,q converges to zero as γl,q approaches
zero, capturing the angular-domain sparsity.
In terms of visibility indicator vector sl , since the VR
exhibits spatial-correlated property, the non-zero elements of
the VR indicator vector sl concentrate on a specific subset of
the whole array. To capture the spatial correlation, the prior
distribution of visibility indicator vector sl can be character-
ized with a one-order Markov chain as
N
Y
p(sl ) = p(sl,n |sl,n−1 ), (18)
n=1
 6

p(γ l , cl , xl , sl , tl , zl , β|rl ) ∝ p(rl |zl , β)p(β)p(zl |tl )p(tl |sl , xl )p(sl )p(xl |cl )p(cl |γ l )p(γ l )
≜ fxl (xl , cl )fcl (cl , γ l )fγ l (γ l ) ftl (tl , sl , xl )fsl (sl ) frl (rl , zl , β)fzl (zl , tl )fβ (β), (20)
| {z }| {z }| {z }
the first layer the second layer the third layer

fH fc fx s1 ft fz fr The third
1 1 1 1 1 1

H1 c1 x1 t1 layer
z1
fH Ă fc Ă fx Ă s ft Ă fz Ă fr fC
q q n n n n
n
Hq cq xn tn zm C
fH Ă fc Ă fx Ă sN ft Ă fz Ă fr
Q Q N N N N

HQ cQ xN tN zM

The first layer The second fs1 ft

1
fs ,s
1 2
ft
2
fs
N 1
,sN
ft
N
layer s1 s2 Ă sN

Fig. 4. Illustration of factor graph representations of (20).

where p(sl,1 |sl,0 ) = (1 − ϕl )δ(1 − sl,1 ) + ϕl δ(sl,1 ), where mf lcq

(cq ) m f lcq
(cq ) m f lHq
(H q )
xn xn cq

ϕl reflects the sparsity level of SnS channels. The transition xn fx cq fc

q
Hq fH
q
n

probability of Markov chain is given by mf (x n ) mc (xn ) m fc (cq )

xn
lxn q
l fx lcq
n q
( s
(1 − p01 )1−sl,n p01l,n , sl,n−1 = 0, Extrinsic Input Backward Passing
p(sl,n |sl,n−1 ) = 1−s (19)
p10 l,n (1 − p10 )sl,n , sl,n−1 = 1, Extrinsic Output Forward Passing

where p01 = p(sl,n = 0|sl,n−1 = 1) and other three transition Fig. 5. Illustration of message passing in the first layer.
probabilities are defined similarly. By the steady-state assump- with s\n and c\q denotes the all elements of s and c except
tion, the Markov chain can be completely characterized by two sn and cq , respectively.
parameters ϕl and p10 . The other three transition probabilities Given the substantial number of antenna elements in XL-
can be easily obtained as p01 = ϕl p10 /(1−ϕl ), p00 = 1−p01 , MIMO systems, both MMSE estimators (21) and (22) in-
and p11 = 1 − p10 , respectively. volve high-dimensional integrals, making them impractical
Based on the prior distribution of cl and sl provided in to compute. Recently, low-complexity techniques such as
(17) and (18), the joint posterior distribution of γ l , cl , xl , GAMP [37]–[41] have been extensively employed for solving
sl , tl , zl , and β given rl can be factorized as (20), as the maximum a posterior (MAP) estimation problem. For
shown in the top of this page, where the involved probability instance, building on the GAMP method [39], [40], the work
distributions are listed in Table I. For brevity, we omit the in [41] introduced a novel bilinear message-scheduling GAMP
subscript l in the subsequent derivation. Moreover, the factor method, which jointly addresses device activity detection,
graph representation of (20) is depicted in Fig. 4, where the channel estimation, and data decoding in a grant-free massive
gray squares represent the factor nodes, and the blank circles MIMO scenario. Furthermore, in [37], the bilinear GAMP
represent the variable nodes. approach was extended to the multi-layer case to handle
Based on the probabilistic model provided in (20), the cascaded problems in relay communication systems. Motivated
optimal estimators of s and c under the minimum mean square by the success of the multi-layer bilinear GAMP in cascaded
error (MMSE) principle can be derived as estimation problems, we propose a computationally efficient
Z three-layer GAMP algorithm to solve the MMSE estimation
ŝn = sn p(sn |r)dsn , ∀n ∈ N , (21) problem in (21) and (22). In the following, we will elaborate
Z on the details of message passing in each layer.
ĉq = cq p(cq |r)dcq , ∀q ∈ Q, (22)
B. Message Passing in the First Layer
where the posterior probabilistic distribution p(sn |r) and In the first layer, the sparsity of angular channel c is
p(cq |r) are respectively given by captured, and corresponding coefficients cq , ∀q ∈ Q are
Z estimated by exploiting the prior distribution information and
p(sn |r) = p(γ, c, x, s, t, z, β|r)dγdcdt dzdβds\n , (23) the likehood messages from the second layer. Meanwhile, the
Z prior distribution information associated with x is output for
p(cq |r) = p(γ, c, x, s, t, z, β|r)dγdsdtdzdβdc\q , (24) the second layer. Fig. 5 illustrates the message passing in the
first layer.
 7

ms (tn ) mf (tn )
Denote the message from xn to fxn and the belief Znf n
l ft
n tn
ltn

of cq as mxn →fxn (xn ) ∝ CN (xn ; ⃗xn , ⃗νnx ) and b(cq ) ∝ fx ,x n 1

sn ft tn
n
CN (cq ; ĉq , ν̂qc ), respectively. Based on the variational message
n

mf lsn
(sn )
passing rule [36], [38], the message from fcq to γq can be tn
fx Znb mf (x n ) xn
derived as n 1
,xn
tn
lxn
Z 
Extrinsic Input Backward Passing
mfcq →γq (γq ) ∝ exp log fcq (cq , γq )b(cq )dcq
n  o (25) Extrinsic Output Forward Passing
2 c
∝ γq exp −γq |ĉq | + ν̂q . Fig. 6. Illustration of message passing in the second layer.
Combining the likehood message mfcq →γq (γq ) in (25) and Combining the messages mxn →f (xn ) (xn ) from the second
the prior message fγq (γq ) associated with γq , the belief of γq layer and mci →fxn (ci ) with i ̸= q, the message from fxn to
is denoted as cq can be given by
b(γq ) ∝ mfcq →γq (γq )fγq (γq ) Z
n 
2
o (26) mfxn →cq (cq ) ∝ fxn (xn , c)mxn →f (xn ) (xn )
= γqξ exp −γq η + |ĉq | + ν̂qc .
Q
From (26), it can be seen that b(γq ) obeys the Gamma
Y
mci →fxn (ci ) dc\q dxn (36)
2
distribution with the shape parameters ξ +1 and η +|ĉq | + ν̂qc . i̸=q
Hence, the approximate posterior mean of γq is given by c
=CN (cm ; ⃗cn,q , ⃗νn,q ),
ξ+1 c
γ̂q = 2 . (27) where the ⃗cn,q and ⃗νn,q are respectively given by
η + |ĉq | + ν̂qc PQ
With the belief γq and variational message passing rule, the ⃗xn − i̸=q dn,i c⃗i,n ⃗xn − x⃗n + dn,q c⃗q,n
⃗cn,q = = , (37)
message from fcq to cq is given by dn,q dn,q
PQ 2
⃗νnx + i̸=q |dn,i | ν⃗ci,n 2
⃗νnx + ν⃗xn − |dn,q | ν⃗cq,n
Z 
c
mfγq →cq (cq ) ∝ exp log fcq (cq , γq )b(γq )dγq ⃗νn,q = 2 = 2 .
(28) |dn,q | |dn,q |
= CN (cq ; 0, γ̂q−1 ). (38)
Combining mfγq →cq (cq ) in (28) and the message Then, the message from cq to fcq is calculated as
mcq →fcq (cq ) derived in (39), the belief of cq is given by NR
Y
b(cq ) ∝ mcq →fcq (cq )mfcq →cq (cq ) = CN (cq ; ĉq , ν̂qc ), (29) mcq →fcq (cq ) ∝ mfxn →cq (cq ) = CN (cq ; ⃗cq , ⃗νqc ), (39)
n=1
where ν̂qc and ĉq are respectively given by
where the variance ⃗νqc and mean ⃗cq are respectively defined as

−1
ν̂qc = 1/⃗νqc + γ̂q , ĉq = ⃗cq / 1 + γ̂q ⃗νqc .


(30) NR
!−1 NR
X X
c c
, ⃗cq = ν̂qc c



Denote the message from fxn to cq as mfxn →cq (cq ) ∝ ⃗νq = 1/⃗νn,q ⃗cn,q /⃗νn,q . (40)
c
CN (cq ; ⃗cn,q , ⃗νn,q ) derived in (36). According to the belief n=1 n=1
propagation rule [42], the message from cq to fxn can be
derived as C. Message Passing in the Second Layer
b(cq ) In the second layer, the SnS property is captured, and
mcq →fxn (cq ) ∝ = CN (cq ; c⃗q,n , ν⃗cq,n ), (31)
mfxn →cq (cq ) the visibility indicator vector sn for n = 1, 2, · · · , NR are
where ν⃗cq,n and c⃗q,n are respectively given by estimated by exploiting the messages from the first and
−1
third layers. Meanwhile, the likehood and prior information
ν⃗cq,n = (1/ĉq − 1/⃗cn,q ) , (32) associated with xn and tn for the first and third layers
ν⃗cq,n c


c⃗q,n = ĉq /ν̂q − ⃗cn,q /⃗νn,q . (33) are output. The message passing is shown in Fig. 6, where
fxn ,xn−1 ≜ p(sn |sn−1 ).
Given the message mcq →fxn (xn ) in (31), the message
Denote the message from the third layer as mtn →ftn (tn ) ∝
mfxn →xn (xn ) of the first layer from fxn to xn is given by
CN (tn ; ⃗tn , ⃗νnt ) derived in (65). Thus, the message from ftn
Q
Z Y to sn can be given by (41), as shown in the top R of next page,
mfxn →xn (xn ) ∝ fxn (xn , c) mcq →fxn (cq ) dc where (a) is obtained according to the fact δ(x)f (x)dx =
(34)
q=1 f (0) and sn ∈ {0, 1}. In addition, πnout is defined as
= CN (xn ; x⃗n , ν⃗xn ),
CN (0; ⃗tn + x⃗n , ⃗νnt + ν⃗xn )
Q
where dn ∈ C denotes the n-th row of D; x⃗n and ν⃗xn are πnout = . (42)
CN (0; ⃗tn , ⃗νnt ) + CN (0; ⃗tn + x⃗n , ⃗νnt + ν⃗xn )
respectively calculated as
Q Q
With the message mftn →sn (sn ) in (41), the message from
x⃗n =
X
dn,q c⃗q,n , ν⃗xn =
X
|dn,q |
2
ν⃗cq,n . (35) sn to ftn is given by
q=1 q=1 msn →ftn (sn ) ∝ (1 − πnin )δ(sn ) + πnin δ(1 − sn ), (43)
 8

Z
mftn →sn (sn ) ∝ ftn (tn , sn , xn )mtn →ftn (tn )mxn →fxn (xn ) dxn dtn

=CN (0; ⃗tn , ⃗νnt )δ(sn ) + CN (0; ⃗tn + x⃗n , ⃗νnt + ν⃗xn )δ(1 − sn ) (41)
(a)
∝ (1 − πnout )δ(sn ) + πnout δ(1 − sn ),

where πnin is defined as m C l f (rm ) m f lzm

(zm ) m f lz m
(zm ) m f ltn
(tn )
rm rm zm zm

fC C fr zm fz tn
ψnf ψnb m m

πnin = . (44) mf (C ) m z (rm ) m f (zm ) mt (zm )

ψnf ψnb + (1 − ψnf )(1 − ψnb ) rm
lC m
l fr
m zm
lzm n
l fz
m

Extrinsic Input Backward Passing

Here, ψnf and ψnb are the forward and backward messages Extrinsic Output Forward Passing
along the Markov chain, and are respectively defined as [28]
Fig. 7. Illustration message passing in the third layer.
f
p01 (1 − ψn−1 out
)(1 − πn−1 ) + p11 λfn−1 πn−1
out
D. Message Passing in the Third Layer
ψnf = f
, (45)
(1 − ψn−1 out ) + ψ f
)(1 − πn−1 out
n−1 πn−1 In the third layer, the variables tn , ∀n ∈ N , are decoupled
from r and estimated by combining the messages from the
b out b out
p10 (1 − ψn+1 )(1 − πn+1 ) + p11 ψn+1 πn+1 second layer. Meanwhile, the likehood information associated
ψnb = . (46)
b
p0 (1 − ψn+1 out ) + p ψ b
)(1 − πn+1 out
1 n+1 πn+1 with tn is output for the second layer. The message passing
is shown in Fig. 7.
where p0 = p10 + p00 , p1 = p11 + p01 . Denote the message from fzm to tn and the belief of tn as
Combining the message mftn →sn (sn ) in (42) and mfzm →tn ∝ CN (tn ; ⃗tm,n , ⃗νm,n
t
) and b(tn ) ∼ CN (tn ; t̂n , v̂nt ),
msn →ftn (sn ) in (44), the belief of sn can be given by respectively. According to the belief propagation rule, the
message from tn to fzm can be given by
ψnf ψnb πnout
b(sn ) = , (47) b(tn )
ψnf ψnb πnout + (1 − ψnf )(1 − ψnb )(1 − πnout ) mtn →fzm (tn ) ∝ = CN (tn ; t⃗n,m , ν⃗tn,m ), (52)
mfzm →tn (tn )
which indicates the visibility probability of the n-th antenna. If
where ν⃗tn,m and t⃗n,m are respectively given by
b(sn ) is greater than a threshold, the n-th antenna is regarded
ν⃗tn,m = 1/ 1/ν̂nt − 1/⃗νm,n
t


as being visible, i.e., , (53)
t t t
t⃗n,m = ν⃗n,m t̂n /ν̂n − ⃗tm,n /⃗νm,n


( . (54)
1, b(sn ) > ψth ,
ŝn = (48)
0, b(sn ) ≤ ψth , With the message mtn →fzm (tn ), the message from fzm to
zm is given by
where ψth is a predetermined threshold. Using the obtained Z NR
b(sn ), ψ0f in (45) and is updated in each iteration as [43]
Y
mfzm →zm (zm ) = fzm (zm , t) mtn →fzm (tn ) dt
n=1
(55)
NR
ψ0f =
1 X
b(sn ). (49) ∝ CN (zm ; z⃗m , ν⃗zm ).
NR n=1
where am ∈ CN denotes the m-th row of A; z⃗m and ν⃗zm are
With the message msn →ftn (sn ) in (43) and mtn →ftn (tn ) respectively calculated as
in (65), the output message mftn →xn (xn ) from ftn to xn is NR
X NR
X 2
given by z⃗m = am,n t⃗n,m , ν⃗zm = |am,n | ν⃗tn,m . (56)
Z n=1 n=1
z
mftn →xn (xn ) ∝ ftn (tn , sn , xn )msn →ftn (sn ) Denote the belief of zm as b(zm ) ∝ CN (zm ; ẑm , ν̂m )
(50) provided in (61). Similar to (25), the message from frm to
mtn →ftn (tn ) dsn dtn β is given by
∝CN (xn ; ⃗tn , ⃗νnt ). Z 
mfrm →β (β) ∝ exp log frm (rm , β)b(zm )dzm
Furthermore, since xn is only connected with ftn and fxn in n  o (57)
the factor graphs, the message from xn to fxn also satisfies 2 z
∝ β exp −β |rm − ẑm | + ν̂m .
mxn →fxn (xn ) ∝ CN (xn ; ⃗tn , ⃗νnt ). Thus, we have ⃗xn = ⃗tn and
⃗νnx = ⃗νnt . Similarly, the output message mftn →tn (tn ) from ftn Combining the prior message fβ (β) and mfrm →β (β) in
to tn is given by (57), the belief of β is given by
( M 
)
mftn →tn (tn ) = (1 − πnin )δ(tn ) + πnin CN (tn ; x⃗n , ν⃗xn ). (51)
X 
M −1 2 z
b(β) ∝ β exp −β |rm − ẑm | + ν̂m . (58)
m=1
 9

It is observed that the belief b(β) obeys the According to (67), the posterior mean t̂n and variance ν̂nt
Gamma distribution with
PM  shape parameters M and are respectively given by
2 z
|r − ẑ | + ν̂ m . Thus, the approximate posterior
Z
m=1 m m
t̂n = tn p(tn |r) dtn = ωn ttmp
n (68)
mean β̂ is given by Z
2
β̂ = P 
M
. (59) ν̂nt = tn − t̂n p(tn |r) dtn (69)
M 2 z
m=1 |rm − ẑm | + ν̂m  2

= ωn (1 − ωn ) t̂n + νntmp .
With b(β), the message from frm to zm is given by
Z  Note that due to the fact that t is equal to SDc, the posterior
mfrm →zm (zm ) ∝ exp log frm (rm , β)b(β) dβ mean t̂ can be equivalently seen as the approximate posterior
(60) estimation of SnS subchannel h.
−1
∝ CN (zm ; rm , β̂ ). E. Messages Approximation in Factor Nodes fzm and fxn
Furthermore, since zm is only connected with frm and fzm From the factor graphs shown in Fig. 4, it is clear that
in the factor graph, the message from zm to is given by there are M N edges between fzm and tn for all m and
mzm →fzm (zm ) ∝ CN (zm ; rm , β̂ −1 ). n. Similarly, there are N Q edges between fxn and cq for
Combining mzm →fzm (zm ) and mfzm →zm (zm ) in (60) and all n and q. Therefore, 2M N + 2N Q messages have to be
(55), the belief of zm is given by updated for forward and backward message passing in each
z
iteration. To further reduce the computational complexity of
b(zm ) ∝ CN (zm ; ẑm , ν̂m ), (61) message passing, we can approximate the means and variances
z
where ν̂m and ẑm are respectively given by of Gaussian messages by omitting some small terms.
 −1   1) Messages approximation for mtn →fzm (tn ) and
z
ν̂m = β̂ + 1/ν⃗zm z
, ẑm = ν̂m rm β̂ + z⃗m /ν⃗zm . (62) mcq →fxn (cq ): According to (66), it can be seen that
t
⃗νnt ≪ ⃗νm,n when N is large enough. Meanwhile, according
With the messages mzm →fzm (zm ) and mti →fzm (ti ) with to the Bayesian theory, we have ν̂nt
< ⃗νnt [38]. Thus, the
i ̸= n, the message from fzm to tn is given by variance ν⃗tn,m = 1/ 1/ν̂nt − 1/⃗νm,n t
≈ ν̂nt . Utilizing (64))
t
and ν⃗n,m ≈ ν̂n , the mean t⃗n,m can be approximated as
t
Z
mfzm →tn (tn ) = fzm (zm , t)mzm →fzm (zm )
!
t t̂ n a∗m,n rm − z⃗m + am,n t⃗n,m
NR t⃗n,m ≈ ν̂n − 2
(63) ν̂nt β̂ −1 + ν⃗zm − |am,n | ν⃗tn,m
Y
mti →fzm (ti ) dzm dt\n (70)
i̸=n (a) rm − z⃗m
t ≈ t̂n − a∗m,n ν̂nt ,
∝CN (tn , ⃗tm,n , ⃗νm,n ), β̂ −1 + ν⃗zm
where ⃗tm,n and ⃗νm,n
t
are respectively given by where (a) is obtained utilizing the fact that z⃗m ≫ am,n t⃗n,m
2
and ν⃗zm ≫ |am,n | ν⃗tn,m from (56). Similarly, the variance
⃗tm,n = rm − z⃗m + t⃗n,m , c
ν⃗q,n and mean c⃗q,n provided in (32) can be approximated as
am,n
ν⃗cq,n = 1/ 1/ν̂qc − 1/⃗νn,q c


2 (64) ≈ ĉq , (71)
t
β̂ −1 + ν⃗zm − |am,n | ν⃗tn,m
⃗νm,n = 2 . ⃗
x n − x⃗ n
|am,n | c⃗q,n ≈ ĉq − d∗n,q ν̂qc x . (72)
⃗νn + ν⃗xn
Combining the message mfzm →tn (tn ) for all m, the output
2) Messages approximation for mfzm →zm (zm ) and
message mftn →tn (tn ) from tn to ftn is given by
mfxn →xn (xn ): With the approximated ν⃗tn,m and t⃗n,m , the
M
Y variance ν⃗zm and mean z⃗m provided in (56) can be simplified
mtn →ftn (tn ) = mfzm →tn (tn ) ∝ CN (tn ; ⃗tn , ⃗νnt ), (65) as
m=1 NR NR
z
X 2 t
X 2
where the variance ⃗tn and mean ⃗νnt are respectively given by ν⃗m = |am,n | ν⃗n,m ≈ |am,n | ν̂nt , (73)
!−1 n=1 n=1
M M NR N R
rm − z⃗m
X X
t t
, ⃗tn = ⃗νnt t
⃗tm,n /⃗νm,n



⃗νn = 1/⃗νm,n . (66)
X X
z⃗m = am,n t⃗n,m ≈ am,n t̂n − ν⃗zm . (74)
m=1 m=1
n=1 n=1 β̂ −1 + ν⃗zm
With mftn →tn (tn ) and mtn →ftn (t
n ), we have b(tn ) ∝ Similarly, the variance ν⃗xn and mean x⃗n can be also approxi-
(1 − πnin )δ(tn ) + πnin CN (tn ; x⃗n , ν⃗xn ) CN (tn ; ⃗tn , ⃗νnt ). Thus, mated as
the approximate posterior probability distribution of tn is Q Q
given by (67), as shown in the top of next page, where
X 2
X 2
ν⃗xn = |dn,q | ν⃗cq,n ≈ |dn,q | ν̂qc , (75)
the auxiliary variables Sn , wn , νntmp , and ttmp n are re- q=1 q=1
spectively given by Sn = CN (0; ⃗tn − x⃗n ,
⃗νnt + ν⃗xn ), Q Q
ωn = πnin Sn / (1 − πnin )CN (0; ⃗tn , ⃗νnt ) + πnin Sn , νntmp =
X X ⃗xn − x⃗n
x⃗n = dn,q c⃗q,n ≈ dn,q ĉq − ν⃗xn . (76)
−1
(1/⃗νnt + 1/ν⃗xn ) , and ttmp = νntmp ⃗tn /⃗νnt + x⃗n /ν⃗xn .


q=1 q=1
⃗νnx + ν⃗xn
n
 10

(1 − πnin )δ(tn ) + πnin CN (tn ; x⃗n , ν⃗xn ) CN (tn ; ⃗tn , ⃗νnt )

p(tn |r) = R
((1 − πnin )δ(tn ) + πnin CN (tn ; x⃗n , ν⃗xn )) CN (tn ; ⃗tn , ⃗νnt )dtn
(1 − πnin )CN (tn ; ⃗tn , ⃗νnt )δ(tn ) + πnin CN (tn ; x⃗n , ν⃗xn )CN (tn ; ⃗tn , ⃗νnt ) (67)
=
(1 − πnin )CN (0; ⃗tn , ⃗νnt ) + πnin CN (tn ; x⃗n , ν⃗xn )CN (tn ; ⃗tn , ⃗νnt )dtn
R

= (1 − ωn )δ(tn ) + ωn CN (tn ; ttmp tmp

n , νn ),

3) Messages approximation for mtn →ftn (tn ) and Algorithm 1 Proposed TL-GAMP for subchannel estimation
mcq →fcq (cq ): Input: received vector r, measurement matrix A and code-
!−1 book D;
M 2 ⃗tn , ⃗νnt , t̂n , ν̂nt , ∀n ,
 c 
(a) X |am,n | Initialize: ⃗νq , ⃗cq , ∀q ,
⃗νnt ≈ , (77) z f b
m=1 β̂ −1 + ν⃗zm {z⃗m , ν⃗m , ∀m}, ξ, η, p01 , ψ0 , ψNR and β̂.
M 1: while the stopping criterion is not met do
(c) X rm − z⃗m /*The first layer*/
⃗tn ≈ t̂n + ⃗νnt a∗m,n , (78)
m=1 β̂ −1 + ν⃗zm 2: Update the posterior estimate of γq according to (27),
2 ∀q ∈ Q;
where (a) and (b) are obtained with ν⃗zm ≫ |am,n | ν⃗tn,m ; 3: Update the posterior estimate of cq according to (30),
2
(c) is obtained with |am,n | /(β̂ −1 + ν⃗zm ) ≪ 1/ν̂nt , since ∀n ∈ N ;
2 z
|am,n | /(β̂ −1 + ν⃗m ) ≪ 1/⃗νnt for large M from (77) and 4: Update the output message mfxn →xn (xn ) of the first
⃗νnt ≥ ν̂nt . Similarly, the variance ⃗νqc and mean ⃗cq in (40) can layer according to (75) and (76), ∀q ∈ Q;
be approximated as 5: Update the message from cq to fcq according to (79)
NR 2
!−1 and (80), ∀q ∈ Q;
c
X |dn,q | /*The second layer*/
⃗νq ≈ , (79)
⃗ν x + ν⃗xn
n=1 n 6: Update the forward and backward messages of Markov
NR chain according to (45) and (46);
X ⃗xn − x⃗n
⃗cq ≈ ĉq + ⃗νqc d∗n,q . (80) 7: Update the output messages of Markov chain accord-
n=1
⃗νnx + ν⃗xn ing to (44);
8: Update the belief of sn according to (47), ∀n ∈ N ;
F. Overall Algorithm and Complexity Analysis 9: Update ψ0f of Markov chain according to (49);
The proposed three-layer generalized approximate message 10: Update the output message of the second layer to xn
passing (TL-GAMP) algorithm can be organized in a more given by ⃗xn = ⃗tn , ⃗νnx = ⃗νnt , ∀n ∈ N ;
succinct form, which is summarized in Algorithm 1 and it can /*The third layer*/
be terminated when it reached a maximum number of iteration 11: Update the message from zm to fzm according to (73)
or the difference between the estimates of two consecutive and (74), ∀m ∈ M;
iterations is less than a threshold. Once the subchannels t̂l for 12: Update the posterior estimate of β according to (59);
l = 1,P 2, · · · , L are obtained, the final channel is estimated as 13: Update the posterior estimate of zm according to (62),
L ∀m ∈ M;
Ĥ = l=1 t̂l aH T (ψ̂l ).
In the following, we provide the computational complexity 14: Update the message from tn to ftn according to (77)
analysis for the TL-GAMP algorithm. The TL-GAMP algo- and (78), ∀n ∈ N ;
rithm requires pre-processing, i.e., performing an economic 15: Update the posterior estimation of tn according to (68)
SVD for A and unitary transformation, and the complexity and (69).
is O(M 2 NR ). It is noted that the pre-processing can be 16: end while
carried out offline, thus, the computational complexity can 17: Calculate ŝn according to (48).
be overlooked. Examining the TL-GAMP algorithms, it is Output: t̂n , ŝn , and ĉq for all n and q.
evident that there is no matrix inversion involved, and the
most computationally intensive parts only involve matrix- which is O(I1 M NR + I2 NR Q), where I1 and I2 denote the
vector products. Specifically, in the first layer, the complexity number of iterations in stages I and II, respectively. It can be
of the proposed algorithm is dominated by the computation seen that the two algorithms have comparable computational
of ⃗cq and x⃗n for all q and n, which requires O(NR Q). complexity. However, it is important to note that the novel TL-
In the second layer, the complexity is dominated by the GAMP algorithm exhibits significant performance superiority
computation of πnout , πnin , ψnf , ψnb , and b(sn ), which requires compared to the Turbo-BOMP algorithm, as will be verified
O(NR ). In the third layer, the complexity is dominated by the in Section VI.
computation of ⃗tn and z⃗m , given by O(M NR ). Therefore, the V. S IMULATION R ESULTS
complexity of TL-GAMP per iteration is O(NR Q + M NR ), In this section, we evaluate the performance of the pro-
which linearly increases with M , NR , and Q. For comparison, posed channel estimation scheme under various simulations.
we also provide the overall complexity of Turbo-BOMP [28], The system parameters are shown in Table II. In particular,
 11

0 5 5
M=64,SNR=5dB
-2 M=96,SNR=5dB
M=64,SNR=15dB 0 0
-4
M=96,SNR=15dB
-6
NMSE(dB)

-5 -5

NMSE(dB)

MSE(dB)
-8

-10
-10 -10
-12
LS LS
-14 PD-OMP PD-OMP
-15 AD-GAMP -15 AD-GAMP
-16 Turbo-BOMP Turbo-BOMP
Proposed TL-GAMP Proposed TL-GAMP
-18 -20 -20
5 10 15 20 25 30 35 40 45 50 0 2.5 5 7.5 10 12.5 15 0 2.5 5 7.5 10 12.5 15
Iteration SNR SNR

Fig. 8. Convergence performance Fig. 9. ψl = 0.25 for all l. Fig. 10. ψl ∈ (0,1) for all l

TABLE II. Simulation Parameters

Notations
Number of BS antennas NR
Number of User antennas NT
Number of RF chains NRF
Carrier frequency fc
Number of channel paths L
Angle of arrival ϑl
Angle of departure ψl
Distance between BS and UE or scatterers rl
Proportion of visible antenna elements ψl
SNR
Number of channel realizations
we consider the normalized mean square error (NMSE) as
the performance metrics, which is defined as NMSE ≜
∥Ĥ − H∥2F /∥H∥2F , where H and Ĥ are the true channel
and estimated channel, respectively. To effectively evaluate
subchannel estimation performance, we compare our proposed
TL-GAMP algorithm with the following benchmarks:
• LS: Least squares estimator based on the formulation
(12). More specifically, the SnS subchannel is estimated
as ĥl,LS = (WWH )−1 WH yl for all l.
• PD-OMP: On-grid polar-domain simultaneous orthogo-
nal matching pursuit algorithm for near-field channels
proposed in [17] without considering the SnS property.
• AD-GAMP: Antenna-domain generalized approximate
message passing algorithm without utilizing the angular-
domain statistical characteristics, which is similar to the
methods proposed in [25], [26].
• Turbo-BOMP. Two-stage VR detection and channel es-
timation scheme proposed in our previous work [28],
where the angular-domain information can not be utilized
in VR detection stage.
We first examine the average convergence behavior of the
TL-GAMP algorithm. As shown in Fig. 8, the NMSE ex-
hibits a consistent monotonic reduction across iterations. The
convergence of the proposed TL-GAMP algorithm is verified
for all the considered simulation scenarios. Considering both
performance and computational complexity, the maximum
iteration can be safely set as 20 for subsequent simulations.
Then, the NMSE performance versus SNR with fixed size
of VR, i.e., fixed ϕl , is examined, as shown in Fig. 9. It
is clear that the conventional near-field channel estimation
approaches, such as LS and PD-OMP encounter pronounced
performance degradation since the SnS property is not con-
sidered. Meanwhile, compared to the existing AD-GAMP and
Turbo-BOMP schemes, our proposed TL-GAMP algorithm
Parameters exhibits remarkable superiority. To elaborate, in contrast with
256
the AD-GAMP approach, which only utilizes the antenna-
16 domain channel characteristics, our approach concurrently
8 harnesses both antenna-domain and angular-domain channel
30GHz characteristics, yielding enhanced estimation accuracy. In con-
6
U (−π/2, π/2) trast to the Turbo-BOMP approach, the collaboration between
U (−π/2, π/2) VR detection and angular channel estimation in our proposed
[10, 50]m TL-GAMP algorithm effectively prevents error propagation.
(0,1]
β Consequently, the NMSE performance of the TL-GAMP al-
2000 gorithm improves significantly.
Furthermore, we present the NMSE performance versus
SNR with ϕl ∈ (0, 1], depicted in Fig. 10. Notably, compared
to ϕl = 0.25, a notable observation emerges: the AD-GAMP
method experiences a substantial performance drop, while the
proposed TL-GAMP algorithm’s performance remains nearly
unchanged. This divergence in performance can be attributed
to the AD-GAMP algorithm’s reliance on the sparsity of the
antenna domain. Consequently, as the VR expands, the antenna
domain’s sparsity diminishes, leading to a consequential per-
formance degradation. In contrast, the proposed TL-GAMP
algorithm shows commendable robustness to varying ϕl . To
further substantiate this observation, Fig. 11 plots the NMSE
performance against ϕl with M = 96 and SNR = 10dB, it is
clear that the the performance of the proposed TL-GAMP only
exhibits slight fluctuation as ϕl , which means that the joint
utilization of statistical characteristics in antenna and angular
domain significantly improve the algorithm robustness for the
size of VR. As a result, the proposed TL-GAMP is suitable
for both spatially stationary and SnS scenarios.
Furthermore, the performance comparison of the proposed
TL-GAMP algorithm and the benchmarks with different pilot
length M is evaluated in Fig. 12 with SNR = 10dB and
ϕl ∈ (0, 1] for all l. The pilot length M increases from 32
to 192. It is observed that the performance of all algorithms
keep improving when pilot length M increases. However, the
performance improvement diminishes as M keeps increasing,
and eventually the performance saturates. As such, we can set a
moderate value M to take a compromise between performance
and estimation delay. Furthermore, we can also see that the
proposed TL-GAMP algorithms significantly outperform the
existing algorithms, especially when the pilot length is small.
This indicates that the proposed algorithms can significantly
reduce the overhead for SnS XL-MIMO channel estimation.
Fig. 13 depicts the NMSE performance concerning the
 12

0 0 0

-5
-5 -5

NMSE(dB)
NMSE(dB)
NMSE(dB)

-10
-10 -10

-15
LS
LS
-15 AD-GAMP, M = 96 -15 NF-OMP
PD-OMP
AD-GAMP, M = 128 AD-GAMP AD-GAMP
Proposed TL-GAMP, M=96 Turbo-BOMP Turbo-BOMP
-20
Proposed TL-GAMP, M=128 Proposed TL-GAMP Proposed ML-GAMP
-20 -20
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 32 64 96 128 160 192 10 20 30 40 50 60 70 80 90 100
The proportion of visibility region Pilot length Distance(m)

Fig. 11. NMSE performance versus ϕl Fig. 12. NMSE performance versus M . Fig. 13. NMSE performance versus rl

1 4 1.5
True Region
Estimated Region
3
0.8 1

2
0.5
0.6
1
0
0.4 0

-0.5
-1
0.2
-2 -1

0 -3 -1.5
10 20 30 40 50 60 70 80 90 160 170 180 190 200 210 10 20 30 40 50 60 70 80 90
Antenna Index Index Antenna Index

(a) VR detection probability b(sl ). (b) Angular-domain coefficients ĉl (c) Antenna-domain coefficients t̂l

Fig. 14. Near-field SnS scenarios with M = 96 and SN R = 15dB.

2
10 1.5
True Region
Estimated Region
8 1
1.5

6 0.5

1 4 0

2 -0.5
0.5
0 -1

0 -2 -1.5
50 100 150 200 250 160 170 180 190 200 210 50 100 150 200 250
Antenna Index Index Antenna Index

(a) VR detection probability b(sl ). (b) Angular-domain coefficients ĉl (c) Antenna-domain coefficients t̂l

Fig. 15. Far-field spatially stationary scenarios with M = 96 and SN R = 15dB.

distance between the BS and either the scatterers or the user, the visible region and accurately estimate the corresponding
with fixed values of M = 96 and SNR = 10dB. Notably, the channel coefficients.
results indicate robustness across all algorithms in response to
varying distances. However, it is noteworthy that the proposed VI. C ONCLUSION AND F UTURE W ORKS
TL-GAMP algorithm significantly outperforms the benchmark In this work, we have studied the SnS channel estimation
methods in the considered configurations. Furthermore, the problem for the XL-MIMO systems and proposed a two-phase
demonstrated robustness across varying distances highlights channel estimation scheme capable of efficiently recovering
a key attribute of our proposed channel estimation scheme: its channels. Initially, to decouple multiple propagation paths, we
feasibility in both near-field and far-field settings. first estimate the AoDs of the user, followed by the estima-
Finally, the estimation performance of the involved inter- tion of individual subchannels. The problem of subchannel
mediate quantities are presented including the VR detection estimation is formulated as a three-layer generalized bilinear
probability denoted as b(sl ), the estimated angular channel inference problem. To achieve the inference with low computa-
coefficients ĉl , and the estimated subchannel coefficients t̂l . tional complexity, we propose a computationally efficient TL-
From Fig. 14 and Fig. 15, it can be observed that across various GAMP algorithm by exploiting both the antenna-domain and
conditions, including far-field and near-field scenarios, as well angular-domain characteristics of SnS channels. Simulation
as both spatial stationary and SnS scenarios, the proposed results show that the proposed algorithm can achieve much
algorithm consistently exhibits the capability to robustly detect better NMSE performance than the existing algorithms with

reduced pilot overhead. Furthermore, the algorithm exhibits ro-
bustness to different channel scenarios. Looking ahead, future
research may involve concurrent estimation of multiple paths
to further reduce training overhead and the optimization of
combining matrices. Additionally, there is potential to extend
the proposed TL-GAMP method to double-side XL-MIMO
channel estimation with SnS property.
R EFERENCES
[1] X. You, C.-X. Wang, J. Huang, X. Gao, Z. Zhang, M. Wang, Y. Huang,
C. Zhang, Y. Jiang, J. Wang et al., “Towards 6G wireless communication
networks: Vision, enabling technologies, and new paradigm shifts,” Sci.
China Inf. Sci, vol. 64, pp. 1–74, 2021.
[2] H. Tataria, M. Shafi, A. F. Molisch, M. Dohler, H. Sjland, and F. Tufves-
son, “6G wireless systems: Vision, requirements, challenges, insights,
and opportunities,” Proc. IEEE, vol. 109, no. 7, pp. 1166–1199, 2021.
[3] Z. Wang, J. Zhang, H. Du, W. E. I. Sha, B. Ai, D. Niyato, and
M. Debbah, “Extremely large-scale MIMO: Fundamentals, challenges,
solutions, and future directions,” IEEE Wirel. Commun., pp. 1–9, 2023.
[4] M. Cui, Z. Wu, Y. Lu, X. Wei, and L. Dai, “Near-field MIMO
communications for 6G: Fundamentals, challenges, potentials, and future
directions,” IEEE Commun. Mag., vol. 61, no. 1, pp. 40–46, 2023.
[5] D. Dardari and N. Decarli, “Holographic communication using intelli-
gent surfaces,” IEEE Commun. Mag., vol. 59, no. 6, pp. 35–41, 2021.
[6] Z. Zhang and L. Dai, “Pattern-division multiplexing for multi-user
continuous-aperture MIMO,” IEEE J. Sel. Areas Commun., vol. 41,
no. 8, pp. 2350–2366, 2023.
[7] A. Tang, J.-B. Wang, Y. Pan, W. Zhang, Y. Chen, H. Yu, and R. C.
de Lamare, “Line-of-sight extra-large MIMO systems with angular-
domain processing: Channel representation and transceiver architecture,”
IEEE Trans. Commun., vol. 72, no. 1, pp. 570–584, 2024.
[8] Y. Pan, C. Pan, S. Jin, and J. Wang, “RIS-aided near-field localization
and channel estimation for the terahertz system,” IEEE J. Sel. Topics
Signal Process., vol. 17, no. 4, pp. 878–892, 2023.
[9] K. T. Selvan and R. Janaswamy, “Fraunhofer and fresnel distances:
Unified derivation for aperture antennas,” IEEE Antennas Propag. Mag.,
vol. 59, no. 4, pp. 12–15, 2017.
[10] H. Lu and Y. Zeng, “Communicating with extremely large-scale ar-
ray/surface: Unified modeling and performance analysis,” IEEE Trans.
Wireless Commun., vol. 21, no. 6, pp. 4039–4053, 2022.
[11] E. Bjrnson and L. Sanguinetti, “Power scaling laws and near-field
behaviors of massive MIMO and intelligent reflecting surfaces,” IEEE
open J. Commun. Soc., vol. 1, pp. 1306–1324, 2020.
[12] Y. Han, S. Jin, C.-K. Wen, and T. Q. S. Quek, “Localization and channel
reconstruction for extra large RIS-assisted massive MIMO systems,”
IEEE J. Sel. Topics Signal Process., vol. 16, no. 5, pp. 1011–1025,
2022.
[13] K. Zhi, C. Pan, H. Ren, K. K. Chai, C.-X. Wang, R. Schober, and X. You,
“XL-MIMO with near-field spatial non-stationarities: Low-complexity
detector design,” in 2023 IEEE Glob. Commun. Conf., 2023, pp. 7194–
7199.
[14] Z. Yuan, J. Zhang, Y. Ji, G. F. Pedersen, and W. Fan, “Spatial non-
stationary near-field channel modeling and validation for massive MIMO
systems,” IEEE Trans. Antennas Propag., vol. 71, no. 1, pp. 921–933,
2023.
[15] E. D. Carvalho, A. Ali, A. Amiri, M. Angjelichinoski, and R. W. Heath,
“Non-stationarities in extra-large-scale massive MIMO,” IEEE Wirel.
Commun., vol. 27, no. 4, pp. 74–80, 2020.
[16] J. Flordelis, X. Li, O. Edfors, and F. Tufvesson, “Massive MIMO
extensions to the COST 2100 channel model: Modeling and validation,”
IEEE Trans. Wireless Commun., vol. 19, no. 1, pp. 380–394, 2020.
[17] M. Cui and L. Dai, “Channel estimation for extremely large-scale
MIMO: Far-field or near-field?” IEEE Trans. Commun., vol. 70, no. 4,
pp. 2663–2677, 2022.
[18] W. Liu, H. Ren, C. Pan, and J. Wang, “Deep learning based beam
training for extremely large-scale massive MIMO in near-field domain,”
IEEE Commun. Lett., vol. 27, no. 1, pp. 170–174, 2023.
[19] Y. Zhang, X. Wu, and C. You, “Fast near-field beam training for
extremely large-scale array,” IEEE Wireless Commun. Lett., vol. 11,
no. 12, pp. 2625–2629, 2022.
[20] X. Wei and L. Dai, “Channel estimation for extremely large-scale
massive MIMO,” IEEE Commun. Lett.
[21] Z. Hu, C. Chen, Y. Jin, L. Zhou, and Q. Wei, “Hybrid-field channel
estimation for extremely large-scale massive MIMO system,” IEEE
Commun. Lett., vol. 27, no. 1, pp. 303–307, 2023.
13
[22] Y. Han, S. Jin, C.-K. Wen, and X. Ma, “Channel estimation for extremely
large-scale massive MIMO systems,” IEEE Wireless Commun. Lett.,
vol. 9, no. 5, pp. 633–637, 2020.
[23] Y. Chen and L. Dai, “Non-stationary channel estimation for extremely
large-scale MIMO,” IEEE Trans. Wireless Commun., pp. 1–15, 2023,
early access, doi: 10.1109/TWC.2023.3343740.
[24] X. Cheng, K. Xu, J. Sun, and S. Li, “Adaptive grouping sparse bayesian
learning for channel estimation in non-stationary uplink massive MIMO
systems,” IEEE Trans. Wireless Commun., vol. 18, no. 8, pp. 4184–4198,
2019.
[25] H. Iimori, T. Takahashi, K. Ishibashi, G. T. F. de Abreu, D. Gonzlez G.,
and O. Gonsa, “Joint activity and channel estimation for extra-large
MIMO systems,” IEEE Trans. Wireless Commun., vol. 21, no. 9, pp.
7253–7270, 2022.
[26] Y. Zhu, H. Guo, and V. K. N. Lau, “Bayesian channel estimation in
multi-user massive MIMO with extremely large antenna array,” IEEE
Trans. Signal Process., vol. 69, pp. 5463–5478, 2021.
[27] X. Yu, W. Shen, R. Zhang, C. Xing, and T. Q. S. Quek, “Channel
estimation for XL-RIS-aided millimeter-wave systems,” IEEE Trans.
Commun., vol. 71, no. 9, pp. 5519–5533, 2023.
[28] A. Tang, J.-B. Wang, Y. Pan, W. Zhang, Y. Chen, H. Yu, and
R. C. De Lamare, “Joint Visibility Region and Channel Estima-
tion for Extremely Large-scale MIMO Systems,” arXiv e-prints, p.
arXiv:2311.09490, Nov. 2023.
[29] J. Gao, C. Zhong, G. Y. Li, J. B. Soriaga, and A. Behboodi, “Deep
learning-based channel estimation for wideband hybrid mmwave mas-
sive MIMO,” IEEE Trans. Commun., vol. 71, no. 6, pp. 3679–3693,
2023.
[30] J. Wu, S. Kim, and B. Shim, “Parametric sparse channel estimation for
RIS-assisted terahertz systems,” IEEE Trans. Commun., pp. 1–1, 2023.
[31] J. He, H. Wymeersch, and M. Juntti, “Channel estimation for RIS-aided
mmwave MIMO systems via atomic norm minimization,” IEEE Trans.
Wireless Commun., vol. 20, no. 9, pp. 5786–5797, 2021.
[32] Z. Yang and L. Xie, “Exact joint sparse frequency recovery via opti-
mization methods,” IEEE Trans. Signal Process., vol. 64, no. 19, pp.
5145–5157, 2016.
[33] T. Ma, Y. Xiao, and X. Lei, “Channel reconstruction-aided MUSIC
algorithms for joint AoA and AoD estimation in MIMO systems,” IEEE
Wireless Commun. Lett., vol. 12, no. 2, pp. 322–326, 2023.
[34] A. Liao, Z. Gao, H. Wang, S. Chen, M.-S. Alouini, and H. Yin, “Closed-
loop sparse channel estimation for wideband millimeter-wave full-
dimensional MIMO systems,” IEEE Trans. Commun., vol. 67, no. 12,
pp. 8329–8345, 2019.
[35] Z. Yuan, Q. Guo, and M. Luo, “Approximate message passing with
unitary transformation for robust bilinear recovery,” IEEE Trans. Signal
Process., vol. 69, pp. 617–630, 2021.
[36] Y. Guo, P. Sun, Z. Yuan, C. Huang, Q. Guo, Z. Wang, and C. Yuen,
“Efficient channel estimation for RIS-aided MIMO communications with
unitary approximate message passing,” IEEE Trans. Wireless Commun.,
vol. 22, no. 2, pp. 1403–1416, 2023.
[37] Q. Zou, H. Zhang, and H. Yang, “Multi-layer bilinear generalized
approximate message passing,” IEEE Trans. Signal Process., vol. 69,
pp. 4529–4543, 2021.
[38] L. Mo, X. Lu, J. Yuan, C. Zhang, Z. Wang, and P. Popovski, “Gener-
alized unitary approximate message passing for double linear transfor-
mation model,” IEEE Trans. Signal Process., vol. 71, pp. 1524–1538,
2023.
[39] S. Rangan, “Generalized approximate message passing for estimation
with random linear mixing,” in 2011 IEEE Int. Symp. Inf. Theor. Proc.,
2011, pp. 2168–2172.
[40] P. Schniter and S. Rangan, “Compressive phase retrieval via generalized
approximate message passing,” IEEE Trans. Signal Process., vol. 63,
no. 4, pp. 1043–1055, 2015.
[41] R. B. Di Renna and R. C. de Lamare, “Joint channel estimation, activity
detection and data decoding based on dynamic message-scheduling
strategies for mMTC,” IEEE Trans. Commun., vol. 70, no. 4, pp. 2464–
2479, 2022.
[42] S. Rangan, P. Schniter, and A. K. Fletcher, “Vector approximate message
passing,” IEEE Trans. Inf. Theory, vol. 65, no. 10, pp. 6664–6684, 2019.
[43] W. Zhu, M. Tao, X. Yuan, and Y. Guan, “Message passing-based joint
user activity detection and channel estimation for temporally-correlated
massive access,” IEEE Trans. Commun., vol. 71, no. 6, pp. 3576–3591,
2023.