arxiv_2311.15158

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/375959818
Angular-Distance Based Channel Estimation for Holographic MIMO
Preprint in IEEE Journal on Selected Areas in Communications · November 2023

DOI: 10.1109/JSAC.2024.3389116
CITATIONS READS
8 122
4 authors, including:
Yuanbin Chen Ying Wang

Nanyang Technological University Beijing University of Posts and Telecommunications
40 PUBLICATIONS 1,086 CITATIONS 879 PUBLICATIONS 25,551 CITATIONS
SEE PROFILE SEE PROFILE
Zhaocheng Wang
Tsinghua University
377 PUBLICATIONS 17,788 CITATIONS
SEE PROFILE
All content following this page was uploaded by Yuanbin Chen on 28 November 2023.
The user has requested enhancement of the downloaded file.

1
Angular-Distance Based Channel Estimation for

Holographic MIMO
Yuanbin Chen, Ying Wang, Member, IEEE, Zhaocheng Wang, Fellow, IEEE,
and Zhu Han, Fellow, IEEE
Abstract—Leveraging the concept of the electromagnetic sig- tionalities shared among plethora of newly encountered par-
nal and information theory, holographic multiple-input multiple- ticularities, with critical performance metrics to be placated
output (MIMO) technology opens the door to an intelligent and including but not subject to ultra-wideband, ultra-massive
endogenously holography-capable wireless propagation environ-
ment, with their unparalleled capabilities for achieving high access, ultra-reliability, and low latency [1]–[3]. As a “holly
spectral and energy efficiency. Less examined are the important grail” of wireless technologies, massive multi-input and multi-
issues such as the acquisition of accurate channel information output (mMIMO) has non-trivial capabilities of achieving
by accounting for holographic MIMO’s peculiarities. To fill this unprecedented spectral efficiency, the critical ingredients of
knowledge gap, this paper investigates the channel estimation
arXiv:2311.15158v1 [cs.IT] 26 Nov 2023
which have been commercially rolled out and approved for

for holographic MIMO systems by unmasking their distinctions
from the conventional one. Specifically, we elucidate that the inclusion in 5G next radio (NR) standards [4]. While looking
channel estimation, subject to holographic MIMO’s electromag- back from MIMO, mMIMO, to the more recent extremely
netically large antenna arrays, has to discriminate not only large-scale (XL) MIMO, one may notice that all three adhere
the angles of a user/scatterer but also its distance information, to the same philosophy, namely to adapt to an uncontrol-
namely the three-dimensional (3D) azimuth and elevation angles lable wireless environment. Their beamforming functionalities
plus the distance (AED) parameters. As the angular-domain
representation fails to characterize the sparsity inherent in conform to the beam-space model in the specific domain
holographic MIMO channels, the tightly coupled 3D AED pa- based upon a number of ideal assumptions, e.g., predefined
rameters are firstly decomposed for independently constructing antenna array geometry with explicit calibrations, along with
their own covariance matrices. Then, the recovery of each propagation in the absence of near-field scattering and mutual
individual parameter can be structured as a compressive sensing coupling [5]–[9]. This would be fruitless as the array aperture
(CS) problem by harnessing the covariance matrix constructed.
This pair of techniques contribute to a parametric decomposition becomes larger and electromagnetically manipulated, shaped
and compressed deconstruction (DeRe) framework, along with in arbitrary geometries and covered with dense antenna ele-
a formulation of the maximum likelihood estimation for each ments. Therefore, we are highly anticipating the evolution of
parameter. Then, an efficient algorithm, namely DeRe-based the classical signal processing methodologies to the electro-
variational Bayesian inference and message passing (DeRe-VM), magnetic (EM) level using EM signal and information theory,
is proposed for the sharp detection of the 3D AED parameters
and the robust recovery of sparse channels. Finally, the proposed allowing for greater degrees of freedom in the context of
channel estimation regime is confirmed to be of great robustness 6G [10].
in accommodating different channel conditions, regardless of the Benefiting from the advancement of meta-materials and
near-field and far-field contexts of a holographic MIMO system,
as well as an improved performance in comparison to the state-
meta-surfaces, along with their widespread uses in wireless
of-the-art benchmarks. communications, holographic MIMO has been crafted as
an excellent candidate to enable holographic radios. The
Index Terms—Holographic MIMO, electromagnetic signal,
channel estimation, near-field communications. fundamentals of holographic MIMO have been expounded
in [11]–[13], and herein we will shine the spotlight on the
qualitative leaps from the conventional MIMO to its holo-
I. I NTRODUCTION graphic counterpart for the following twofold aspects. Firstly,
The paradigm shift from the fifth-generation (5G) to the regarding physical fundamentals, holographic MIMO is con-
sixth-generation (6G) communications catalyzes pillar func- sidered to have a nearly continuous aperture with antenna
element spacing being much less than half a wavelength of
This work was supported in part by the National Key R&D Program of incident EM waves. This contrasts with the half-wavelength
China under grant 2022YFA1003901, in part by the Beijing Natural Science condition typical in conventional MIMO. Such a distinction
Foundation under Grant 4222011, and in part by NSF CNS-2107216, CNS-
2128368, CMMI-2222810, ECCS-2302469, US Department of Transporta- leads to transformative changes in hardware structure, notably
tion, Toyota and Amazon. (Corresponding authors: Ying Wang; Zhaocheng in terms of the array aperture size (from small to extremely
Wang.) large) and the number/density of antenna elements (from
Yuanbin Chen and Ying Wang are with the State Key Laboratory of Net-
working and Switching Technology, Beijing University of Posts and Telecom- sparse to dense). Intriguingly, the densely packed antenna
munications, Beijing 100876, China (e-mail: chen [email protected]; elements may lead to enhancement in spectral efficiency [14].
[email protected]). Furthermore, the attendant mutual coupling among antenna
Zhaocheng Wang is with the Beijing National Research Center for Infor-
mation Science and Technology, Department of Electronic Engineering, Ts- elements, albeit deemed detrimental in conventional MIMO
inghua University, Beijing 100084, China (e-mail: [email protected]). designs, fundamentally transforms, and is potential to be
Zhu Han is with the Department of Electrical and Computer Engineering leveraged in achieving the super-directivity, a phenomenon
in the University of Houston, Houston, TX 77004, USA, and also with the
Department of Computer Science and Engineering, Kyung Hee University, that describes the significantly higher array gains obtained in
Seoul, South Korea, 446-701 (e-mail: [email protected]). holographic MIMO [15], [16]. Secondly, from the perspec-
2
tive of propagation channels, another qualitative change in naturally transitioning the discrete antenna aperture to its
holographic MIMO results from its extremely large aperture continuous counterpart, for which EM waves can be manipu-
size. Unlike mMIMO communications that typically employ lated by continuous-to-discrete antenna elements that are sub-
planar wavefront in far-field cases, holographic MIMO can wavelength spaced for a holographic array. Secondly, while
naturally draw the far-field region to its near-field counterpart the sparse representations presented in [26], [28], and [29]
(also known as the Fresnel region) as the aperture size indeed tackle the energy spread effect exposed in the near-
increases. Therefore, holographic MIMO near-field communi- field scenario, their applicability is confined to the uniform
cations have to discriminate not only the angle of a scatterer linear array (ULA) case, in the sense that their extensions
but also its distance information, in contrast to the angle-only- to more complicated uniform planar array (UPA) regimes
aware counterpart in mMIMO systems. are problematic. This is attributed to the intricate coherence
Given that holographic MIMO just plays a basic physical between the two-dimensional (2D) azimuth-elevation angle
role, associated physical-layer technologies are required for pair and distance parameters as well as a more stringent
further implementation of holographic communications. To restrict isometry property (RIP) condition. In [27], although
this end, significant efforts have been devoted to the chan- the UPA case is considered for the near-field codebook design,
nel modeling [17]–[21], channel estimation [19], [22], and a simplified assumption is adopted, in which the UPA case
holographic beampattern design [23]–[25]. In these literatures, is approximatedly pushed to an asymptotic ULA case when
the holographic channel models employed might not arrive at conducting coherence analysis. We thus intend to revisit
a good consensus, which can be roughly classified into two this pressing issue within the context of holographic UPA
categories: i) EM channel based on Fourier planar wavefront configurations in this treatise. As a third consideration, given
representation [17], [20], [21], and ii) parametric physical the beneficial sparsity in facilitating holographic channel es-
channel based on multi-path representation [22], [26]. More timation, it is imperative to account for the overhead induced
precisely, the former goes more deeply into characterizing by channel sparsification and the complexity of algorithm
what kind of channel distributions that could appear in the design. The modeling approaches in [26] and [29] facilitate
radiative near-field and what cannot. The latter is more the relief of heavy overhead to different degrees, but suffer
concentrated on the imposed sparsity in the sense that the from the multiplicative gridding complexity, i.e., the product
received signal contains a sum of a few spherical waves, based of individual complexity associated with discretizing the az-
upon which the compressive sensing (CS) method can be imuth angle, elevation angle, and the distance of holographic
harnessed for facilitating high-efficiency parameter recovery. MIMO channels. This significantly undermines the efficiency
In spite of this, we believe that this pair of somewhat different of pursuing the optimum in the CS procedure, thus in need
channel models for characterizing holographic MIMO may be of more efficient techniques for searching through this three-
aligned to the same ingredients but using different methods dimensional (3D) azimuth and elevation angles plus the dis-
for analyzing the holographic MIMO channels. tance (AED) grid. In a nutshell, a channel estimation strategy
Holographic MIMO’s intriguing features contribute signifi- for holographic MIMO systems has to be tailored in response
cantly to the achievement of holographic radios, yet just a few to these non-trivial challenges, with the goal of encapsulating
fledgling efforts on channel estimation in the true sense of the the potential sparsities for a robust and accurate estimate
term are available to date, e.g., [19], [22]. Even so, we are performance at reduced complexity and overhead.
fortunate to observe, as more attention is drawn to related Geared towards the challenges outlined, the current work
topics, the emergence of pioneering studies with a focus set out with the aim of filling the knowledge gap in the state-
on near-field channel estimation based on the assumption of of-the-art by following contributions:
spherical wavefront utilizing various techniques [22], [26]– • We investigate the uplink channel estimation in the
[29], such as CS-based approaches [26], [29], and learning- holographic MIMO system. For an electromagnetically
based means [28]. In particular, the potential sparsity intrinsic large-scale UPA, a unified array model is exploited in
to the near-field channel has been unveiled in [26], where order for EM waves to be manipulated by the antenna
a polar-domain representation is proposed in lieu of the elements that are sub-wavelength spaced across the holo-
angular-domain counterpart in the far field. Base upon the graphic array of interest. Furthermore, the employed
polar-domain philosophy, the near-field codebook solution is model portrays the variations of signal amplitude, phase,
devised in [27] for multiple access. In [29], a joint dictionary and projected aperture across array positions, based upon
learning and sparse recovery empowered channel estimation which a parametric channel model is presented. By re-
is presented for near-field communications. Furthermore, in a vealing the invalidity of the employment of conventional
distinct approach, a model-based deep learning framework for planar wavefront in holographic MIMO channels, we
near-field channel estimation is showcased in [28], in which evince that the channel estimation has to discriminate
the channel estimation problem is attributed to a CS problem not only the angle of a user/scatterer but also its distance
that is solved by deep learning-based techniques. information, with to-be-determined 3D AED parameters
Despite these impressive initiatives, we evince that the involved.
channel estimation for the holographic MIMO particularly for • To facilitate a sharp detection for the 3D AED pa-
the near-field regime remains still an open issue, as evidenced rameters, we conceive a parametric decomposition and
by the following facts. Firstly, owing to the extremely larger compressed reconstruction (DeRe) framework for holo-
array size and denser element spacing, the conventional ar- graphic MIMO systems. To elaborate, the tightly coupled
ray modeling, which treats the antenna elements as sizeless 3D AED parameters are decomposed by constructing the
points, makes no sense. A generic model is required for one-dimensional (1D) covariance matrix associated with
3
the azimuth angle, the elevation angle, and the distance, evaluations, along with some useful insights, and finally the
for facilitating their independent estimates. Furthermore, paper is concluded in Section VII.
an off-grid basis is utilized to mine potential sparsities
inherent to the covariance matrix constructed, and then
the 3D AED parameters can be reconstructed as their II. S YSTEM M ODEL
respective 1D compressive sensing (CS) problems. To
this end, it is sufficient to estimate the offset vectors A. Scenario
associated with each dimension. Additionally, we evince
that the proposed DeRe paradigm may be a curse when We consider an uplink holographic MIMO system where
concerning the attendant angular index misinterpreta- multiple single-antenna users are served by a base station (BS)
tion induced by the improper stitching of the obtained having N = Ny × Nz antenna elements in the form of a
three 1D parameters, but a blessing when embraced UPA. NRF (NRF ≪ N ) feeds are embedded in the substrate
with regard to significantly reduced gridding complexity, in order to generate reference waves carrying user-intended
followed by the angular index correction as a remedy. signals, with each feed connected to a radio frequency (RF)
• Due to the ill-conditioned sensing matrix (containing off- chain for signal processing. As illustrated in Fig. 1, the UPA
set vectors) and the sparse vectors with imperfect priors, coincides with the y-z plane and centered at the origin, where
the formulated 1D CS problem may differ somewhat Ny and Nz denote the number of antenna elements along
from the standard one, demonstrating a high level of the y- and z-axis, respectively. For notational convenience,
intractability. To tackle this issue, a layered probability Ny and Nz are assumed to be odd numbers. The central
model for unknown parameters is firstly established to location of the (ny , nz )th antennan element is pny ,n oz =
capture the specific structure of sparse vectors, with T
[0, ny δ, nz δ] , in which ny = 0, ±1, ±2, ..., ± y2
N −1
and
the capability of addressing the uncertainty of imperfect Nz −1
prior information. Based on this, the maximum like- nz = 0, ±1, ±2, ..., ± 2 . Along the y- and z-axis, the
lihood estimation (MLE) of the offset vectors can be physical measurements of the UPA are approximately Ny δ
acquired, along with the minimum mean square error and Nz δ, respectively, with δ representing the inter-element
(MMSE) estimation of the sparse vectors. Then, we spacing.
propose an efficient algorithm, namely DeRe-based vari- In contrast to the conventional array modeling that treats
ational Bayesian inference and message passing (DeRe- each antenna element√as a sizeless
√ point, the size of an antenna
VM) in pursuit of a stable solution in a turbo-like way. element, denoted by A× A, should be explicitly addressed.
The proposed DeRe-VM algorithm is featured by a It is possible to theoretically unify the modeling of continuous
computational complexity that scales proportionally with holographic surface and discrete antenna arrays by adjusting
the number of antenna elements raised to the power of the element size A√ and their inter-element spacing δ [11],
1.5. This is markedly more efficient when contrasted with [17], where δ ≥ A. Particularly, let δA2 ≤ 1 denote the
state-of-the-art schemes, where the complexity exhibits array occupation ratio that indicates the portion of the entire
a quadratic relationship. UPA plate area being occupied by the antenna elements.
• Simulation results underscore the role of DeRe frame- Regarding the extreme case of δA2 = 1, the UPA can be
work as a critical instrument in mitigating false alarms treated as a contiguous surface. Therefore, such a modeling
(misinterpreting the genuine power of the non-significant is imperative to account for the projected aperture of each
paths as their significant counterparts) in the process of antenna element when signals are impinged from various
parameter detection. The proposed DeRe-VM algorithm directions, particularly in the presence of extremely large
exhibits a notable superiority over the state-of-the-art number of antenna elements. Furthermore, we assume in this
λ2
angular-domain solutions, primarily attributable to its paper that the effective antenna aperture is A = 4π for the
use of a more accurate Fresnel approximation, without isotropic antenna element.
noticeable increase in gridding complexity. Furthermore,
we demonstrate that the proposed DeRe-VM algorithm
is featured by significant robustness, not only due to B. Channel Model and Communication Protocol
its ability to accommodate various channel conditions
– such as the level of channel sparsity and the near-field We model the channel spanning from the BS to the user
or far-field contexts of the holographic MIMO system – based on path parameters of the system [22], [26]. The as-
but also in its insensitivity agasint perturbations induced sumption of the planar wavefront fails to describe the features
by imperfections in practical implementations, thereby of the link distance falling within the range of Rayleigh
ensuring the accuracy of the parameter recovery. distance as EM waves do not uniformly radiate across dif-
The remainder of this paper is organized as follows. Sec- ferent antenna elements. In this case, the generic spherical-
tion II introduces the system model and the challenges to wavefront model is essential to encapsulate phase variations
be resolved. In Section III, we show the way to achieve the across different
h antenna
√
elements.
√ i h
In particular,
√
we denote
√ i
parametric decomposition, and in Section IV, we reconstruct Any ,nz = ny δ − 2A , ny δ + 2A × nz δ − 2A , nz δ + 2A
the parameters in each dimension into their own CS problems. by the surface region of the (ny , nz )th antenna element, and
T
Then, an efficient DeRe-VM algorithm is proposed in Sec- denote ex = [1, 0, 0] by the UPA normal vector, i.e., the
tion V for attaining stable solutions to the formulated CS prob- unit vector along the x direction. By taking into account
lems. Section VI provides numerical results for performance the basic free-space propagations for a theoretical analysis,
4
N Antenna
(n , n )
y z
N RF RF chains N RF Feeds Waveguide elements
N z Elements
RF chain rl
( n y , nz )
J
Digital RF chain ql
Data
rl
...
beamforming ( 0, 0 )
stream
...
...
...
N RF ´ N
Fp Î jl
RF chain y
Holographic beamforming M p Î
N ´N
Ny
x
d
d
Fig. 1. Schematic diagram of holographic MIMO transmission and UPA.
the channel power gain between the user and the (ny , nz )th in (1), i.e.,
antenna element is given by [30] T
(ny ,nz ) 1 s − pny ,nz ex
f (θl , φl , rl ) ≈ 2 A
4π s − pny ,nz s − pny ,nz
(ny ,nz )
f (θl , φl , rl ) | {z }
Projected aperture
T
(s − p̄) ex
Z
1 Arl sin θl cos φl
= 2 dp̄, (1) = 3. (3)
Any ,nz 4π∥s − p̄∥ ∥s − p̄∥ 4π s − pny ,nz
| {z } | {z }
Free-space path loss Projection to signal direction
Accordingly, the holographic channel h between the BS and
the user can be modeled as [22]
where rl represents the distance between the user (or scat- L
r
1X
terer) and the reference antenna element of the UPA (i.e., h= βl a (θl , φl , rl ), (4)
L
(ny , nz ) = (0, 0)), and (θl , φl ) denotes the azimuth-elevation l=1
angle pair of the lth path, with l = 1 being the line-of- where βl denotes the channel coefficient of the lth path and
sight (LoS) path and l ≥ 2 being non-line-of-sight (NLoS) L is the path number. p Note that for the LoS path, i.e., l = 1,
paths. The user location s is determined by the LoS path, we denote βl = f (ny ,nz ) (θl , φl , rl ), since the LoS path
T
i.e., s = [rl sin θl cos φl , rl sin θl sin φl , rl cos θl ] , l = 1. is distance-determinate [9], [22]. For the NLoS path, i.e.,
The examination in (1) is applied to each individual antenna l ≥ 2, the channel coefficient βl follows a complex Gaussian
element, taking explicit account of the element size. More distribution, i.e., βl ∼ CN 0, f (ny ,nz ) (θl , φl , rl ) . The array
explicitly, the model in (1) accounts for the projected aperture response vector a (θl , φl , rl ) ∈ CN ×1 is structured by the
of each antenna element, as reflected by the projection of the following entry
UPA normal vector ex to the wave propagation direction at
each local point p̄, in contrast to the conventional model for (ny ,nz ) 2π (ny ,nz )
a (θl , φl , rl ) = exp −ȷ rl − rl , (5)
free-space path loss. In [31]–[33], similar models incorporat- λ
ing the projected aperture of the antenna element have also (n ,n ) (n ,n )
where ∆rl y z ≜ rl y z − rl represents the extra-distance
been considered for communicating with continuous surfaces.
traveled by the wave to arrive at the (ny , nz )th antenna
Hence, a two-dimensional integration over the surface of each
element with respect to the reference one, and it can be
element is required for a precise evaluation of (1).
unfolded as
Owing to the fact that each antenna element (not the entire "
2 1/2 #
array) in A is on the wavelength scale, there may be typically (n ,n ) r (ny ,nz ) (ny ,nz ) (n ,n )
∆rl y z = rl 1+ rl − 2 r rl gl y z −1 ,
marginal variations in the wave propagation distance ∥s − p̄∥
s−p̄
and projection direction ∥s−p̄∥ across different points p̄ ∈ q (6)
(ny ,nz )

Any ,nz . Hence, in light of the fact in [30], we have where r ≜ δ + n2y
herein represents the n2z
distance between the (ny , nz )th antenna element and the
(n ,n )
reference one, and gl y z = cos ϑ(ny ,nz ) sin θl sin φl +
1 1 sin ϑ(ny ,nz ) cos θl = √ 21 2 (ny sin θl sin φl + nz cos θl ) is
2 ≈ 2, ny +nz
4π∥s − p̄∥ 4π s − pny ,nz n
T a geometry term, with ϑ(ny ,nz ) = arctan nyz indicating the
T
(s − p̄) ex s − pny ,nz ex azimuth angle of the (nz , ny )th antenna element with respect
≈ , ∀p̄ ∈ Any ,nz , (2) to the reference one.
∥s − p̄∥ s − pny ,nz
For uplink channel estimation, it is assumed that the users
transmit mutual orthogonal pilot sequences to the BS, e.g.,
which facilitates the approximation of the channel power gain by employing orthogonal frequency or time resources for
5
C. Challenges
Most channel estimation schemes, being based on the

classical planar-wavefront assumption, exploit the sparsity in
Scatterer 2 the angular domain to achieve efficient estimation. However,
Scatterer 3
this may result in severe energy spread effect in holographic
MIMO counterparts. To be explicit, Fig. 2 portrays the power
Scatterer 1 scattered in relation to the angle pair (θ, φ) in Fig. 2(a), and
the power scattered in relation to the distance r in Fig. 2(b),
respectively, in which three scatterers are assumed in the prop-
Angular energy spread
agation environment. As transpired in Fig. 2(a), the normal-
ized power scattered typically results in the proliferation of
angles, and fluctuates rapidly across the holographic antenna
array. Despite the detachment of distance information r, the
(a) true azimuth-elevation angle pair (θ, φ) cannot be precisely
distinguished. Additionally, as portrayed in Fig. 2(b), the
Distance energy spread
power scattered versus the distance reveals a fluctuating trend
from peak to trough. In this case, it is difficult to distinguish
Scatterer 2 the genuine scattered power of the exact distance, since the
energy may spread farther away with increased distance.
Actually, the angular-domain transform matrix just samples
Scatterer 3 the angular information but ignores the distance information,
which is implausible for the holographic near-field context
due to the angles being functioned by the distance. Thus, the
angular-domain transform matrix fails to align the sampling
points to the genuine scattered power of the significant path
Scatterer 1 corresponding to a user/scatterer. This always leads to a false
alarm, i.e., the genuine channel power corresponding to the
(b) non-significant path is misinterpreted as a significant one, thus
eroding the detection performance of the 3D AED parameters.
Fig. 2. (a) Angular energy spread: the normalized power scattered results
in the proliferation of a multitude angles. (b) Distance energy spread: To tackle this issue, we intend to achieve the effect decom-
the normalized power scattered in relation to the distance experiences a position of the 3D AED parameters by constructing their
fluctuating from peak to tough. respective covariance matrices, for which each covariance
matrix constructed retains its self-contained information while
canceling out the potential impact of other parameters.
different users, hence allowing the channel estimation to be
independent of one another. Without loss of generality, we
examine the channel estimation process of an arbitrary user.
III. D ECOMPOSITION OF 3D AED PARAMETERS θ, φ,
Denote P by the pilot number and xp by user’s information-
AND r
bearing signal, respectively, and the received signal yp ∈
CNRF ×1 at the BS is given by
Given that the bare minimum of channel structure can
be used when doing channel estimation, let us define V ∈
yp = Fp Mp hxp + Fp Mp np , p = 1, ..., P, (7) CNy ×Nz as the matrix involving some appropriately selected
entries from the covariance matrix of the exact channel, whose
where Fp ∈ CNRF ×N represents the combining matrix at the (ny , nz )th entry can be given by
BS, Mp ∈ CN ×N is the holographic pattern that records the
radiation amplitude of each antenna element, np ∈ CN ×1 ∼
n o
2
vny ,nz ≜ E hny ,nz h∗n′y ,n′z
CN 0, σ IN denotes the additive white Gaussian noise
(AWGN) with power σ 2 , and h ∈ CN ×1 characterizes the L
L P

(n′y ,n′z ) (ny ,nz )
E {βl βl∗′ } exp ȷ 2π 2π
, ∀l ̸= l′ ,
P
= λ ∆rl ′ − ȷ λ ∆rl
channel spanning from the BS to the user. T The aggregated l=1 l′ =1
received pilot sequence y = y1T , ..., yPT ∈ CP NRF ×1 of all L
(a) X

2π

(n′y ,n′z )

(ny ,nz )
the P uplink pilots from the user is given by = β̄l2 exp ȷ ∆rl − ∆rl , (9)
λ
l=1
y = Fh + n, (8)
where β̄l2 ≜ E βl2 and (a) holds owing to the orthog-
h
T T
iT onal assumption between two independent paths. Although
in which n = nT1 (F1 M1 ) , ..., nTP (FP MP ) ∈ V represents a part of the covariance matrix, we refer to
P NRF ×1 it as the covariance matrix throughout of this paper. As
hC is the noise i matrix, and F =
T
T
(F1 M1 ) x1 , ..., (FP MP ) xP
T
∈ CP NRF ×N delivers observed in (9), vny ,nz is dependent on the distance difference
(n′ ,n′ ) (n ,n )
the overall observation matrix. ∆rl y z − ∆rl y z with respect to the antenna elements
6
(ny , nz ) and (n′y , n′z ). Using the Fresnel approximation [34], We let the antenna element spacing be δ = λ4 , and the
(6) can be further recast to (ny , nz )th entry of the θ-based covariance matrix Wθ takes
the form of
2 q
(n2y +n2z )δ2

(ny ,nz ) (ny ,nz ) 2 2 (ny ,nz ) L
∆rl ≈ 1 − gl − ny + nz δgl , X
2rl
wnθ y ,nz = 2β̄l2 cos (πny sin θl sin φl ) exp (ȷπnz cos θl ) .
(10) l=1
in which the first term involves both the angular and distance (13)
information, while the second term pertains exclusively to the Note that δ = λ4 conveys double implications, that is, it not
(n ,n )
angular information, i.e., gl y z . This property is paramount only indicates a denser antenna spacing in the holographic
since it inspires the subsequent parametric deconstruction. MIMO of interest, but it also complies with the Nyquist
In the subsections that follows, tailored tricks are presented sampling theorem under the DeRe framework for acquiring
for effectively decomposing the tightly coupled 3D AED statistical characteristics of θ (i.e., θ-based covariance ma-
parameters. trix Wθ ). To be explicit, it is evident from (12) that the
estimation of θ is equivalent to identifying all L spectral
power peaks with the frequency index {cos θl , l = 1, ..., L}
A. Angular Decomposition given the sequences wnθ y ,nz with the time index nz . The phase
term in (12), i.e., exp ȷ 4π λ δn z cos θ l , can be structured as a
2δ
The structure of (10) inspires a potential technique to select standard Fourier-like form, namely exp ȷ2πn z · λ cos θ l .
the index of the n′y , n′z th entry of the conjugate channel The frequency component 2δ

λ cos θl has to follows Nyquist
h∗n′y ,n′z with extra care for ensuring that vny ,nz is determined sampling theorem, yielding that 2δ 2δ
λ cos θl ≤ λ ≤ 2 , and
1
exclusively by the angular information. The index selection we thus have δ ≤ λ4 . Additionally, as for the constructed
(n′ ,n′ ) (n ,n ) 2 θ Ny ×Nz
strategy needs to satisfy gl y z = −gl y z and n′y + θ-based covariance θ matrix W ∈ C , the row vectors
′ 2 2
(nz ) = ny + nz . This pair2 thereof, i.e., W , are independent of one another. Each
′ ′
of restrictions can be readily independent vectornydelivers ,:
an observation indexed by ny
addressed by letting ny , nz = (−ny , −nz ), and thus vny ,nz
in (9) can be trimmed to for the estimate of θ, i.e., θ̂ ny
, in which these N y estimates
θ̂ny , ny = 1, ..., Ny , contribute in collaboration to a faithful
vny ,nz (θl , φl ) θ̂, as elaborated on later.
L 2) Extraction of φ: The same technique can be applied
X
2 4π q 2 2 (ny ,nz ) to construct the φ-based covariance matrix Wφ , but using a
= β̄l exp ȷ ny + nz δgl
l=1
λ different index-selection strategy, i.e., n′y , n′z = (ny , −nz ),
L hence yielding
X
2 4π
= β̄l exp ȷ δ (ny sin θl sin φl + nz cos θl ) . (11) wnφy ,nz = vny ,nz (θl , φl ) + vn′y ,n′z (θl , φl )
λ
l=1
X L
The physical significance behind this index selection trick is = 2β̄l2 cos (πnz cos θl ) exp (ȷπny sin θl sin φl ) .
l=1
to allow the indices of the antenna elements corresponding
∗ (14)
to the channel hny ,nz and its conjugate copy hn′y ,n′z to be
symmetrical to the origin, as illustrated in Fig. 3. As a result, As shown in (14), both the amplitude and phase of φ
vny ,nz (θl , φl ) retains only the angular information of the contains the information of θ, which implies that an estimate
original channel structure. To facilitate the reduced searching of θ̂ is a prerequisite for the determination of φ. The φ-
space in the estimation process, the decomposition can be based covariance matrix W φ
∈ CNy ×Nz incorporates Nz
further refined to trim the 2D model at hand into a pair of 1D independent vectors with vector size Ny . In other words, the
φ
quantities consisting exclusively of θ or φ. n z th column vector of W serves as an observation that can
be used for the CS process to attain φ̂nz , nz = 1, ..., Nz ,
1) Extraction of θ: We aim to retain the θ-associated term
the Nz estimates of which can be appropriately fused by
in the phase of vny ,nz (θl , φl ), i.e., nz cos θl , by constructing
θ a sophisticated design for the robust recovery of φ. Further
a θ-based covariance matrix W . Particularly, the (ny , nz )th
θ θ details appear in Sec. V.
entry in W can be calculated bywny ,nz = vny ,nz (θl , φl ) +
vn′y ,n′z (θl , φl ). By letting n′y , n′z = (−ny , nz ), we have B. Distance Decomposition
The distance decomposition can be achieved by construct-
wnθ y ,nz = vny ,nz (θl , φl ) + v−ny ,nz (θl , φl ) ing the r-based covariance matrix Wr that strips out the
X L
4π
angular information for retrieving the distance information.
= β̄l2 exp ȷ δ (ny sin θl sin φl + nz cos θl ) Specifically, we use the same trick when constructing Wr .
λ
l=1 As illustrated in Fig. 3, we employ the channel hny ,nz and
L its conjugate copy corresponding to the origin of the UPA,
X 4π
+ β̄l2 exp ȷ δ (−ny sin θl sin φl + nz cos θl ) namely h∗0,0 , as a pair of inputs for wnr y ,nz , i.e.,
λ
l=1
wnr y ,nz ≜ E hny ,nz h∗0,0

L
X 4π 4π
= 2β̄l2 cos δny sin θl sin φl exp ȷ δnz cos θl . L
λ λ X
2 2π (ny ,nz ) (0,0)
l=1 = β̄l exp ȷ ∆rl − ∆rl . (15)
(12) λ
l=1
7
z Construct covariance Construct θ(φ)-based covariance matrix
Angular decomposition
matrix V
hny ,nz Index
conversion
vny ,nz + v- ny ,nz = wnqy ,nz
(n , n )
y z
éë Wq ûù
y
{ hny ,nz h-*ny ,- nz } vny ,nz
n y , nz
vny ,nz + vny ,- nz = wnjy ,nz

( -n , -n )
y z
1j
Index
h-*ny ,- nz conversion éë Wj ùû
n y , nz
z
Construct r-based cova-
Distance decomposition
Coordinate transformation
hny ,nz riance matrix
(n , n )
y z
nz > 0
rl ( nz > 0 )
( 0, 0 )
y {h n y , nz
*
h0,0 } wnr y ,nz d nz
rl ( nz = 0 )
éë W r ùû nz = 0
n y , nz
*
1j h0,0 rl ( nz < 0 )
nz < 0
Fig. 3. Schematic illustration of parametric decomposition.
Recall that the distance difference takes

the form of vectors shown in (13). Furthermore, a wide measurement
(n2y +n2z )δ2
2
(ny ,nz )
∆rl = 1 − gl
(ny ,nz )
−δ× matrix Bθny ∈ CNRF ×Nz (NRF ≪ Nz ) can be introduced for a
2rl
(n ,n )
lower-dimension observation ynθ y , thus yielding
(ny sin θl sin φl + nz cos θl ), with gl y z =
cos ϑ (ny ,nz )
sin θl sin φl + sin ϑ (ny ,nz )
cos θl = √ 1 ynθ y = Bθny wnθ y , (18)
n2y +n2z
× (ny sin θl sin φl + nz cos θl ). By substituting them into where the entries in Bθny are allowed to follow a complex
(15), we have (16) presented at the top of the next page. p
Gaussian distribution CN 0, 1/ Ny .
For simplicity, we firstly consider a special case of nz = 0,
Given that there are only L significant angles of arrival
and hence (16) can be trimmed to (17), shown at the top
(AoAs) θ = {θ1 , ..., θL } in the physical propagation
of the next page. A closed examination in (17) indicates
environment, it is beneficial to mine the sparse characteristics
that the recovery of the distance r from Wr relies upon
concealed behind wnθ y for achieving sparse estimates of
the acquisition of the azimuth-elevation angle pair (θ, φ),
L AoAs at reduced complexity and overhead. For this
which has a direct impact on the algorithm flow, as will be
purpose, a uniform grid of Ñz AoAs θ over [−π/2, π/2)
detailed in Sec. V. Additionally, for a more general instance
are prescribed to take values from the discreteo sets
of nz ̸= 0, the same form as (17) can be attained by shifting n
θ̃nz : sin θ̃nz = Ñ2 nz − Ñz2−1 , nz = 0, ..., Ñz − 1 .
the reference system by δnz (m) along the z-axis using the z
coordinate transformation. Nevertheless, the true AoAs do not coincide exactly with
Thus far, we have achieved the decomposition of the 3D the grid points. To capture the attributes of mismatches
AED parameters (θ, φ, r) by constructing their associated between the true angles and the grid points, the off-
covariance matrices, each of which retains fully knowledge grid basis is harnessed for its sparse representation [35].
that can be explored. In light of the closed structures in (13), More precisely, we introduce the θ-based offset vector
(14) and (17), we are inspired to conduct the CS procedure on ∆θ = ∆θ1 , ..., ∆θÑz , in which ∆θnz is given by
each parameter to ensure their accurate and robust recoveries. (
θnz − θ̃nz,q , if nz = nz,q , q = 1, ...L,
Therefore, the following of this treatise moves on to describe ∆θnz = (19)
in great detail the CS-based reconstruction, along with an ef- 0, otherwise,
ficient algorithm proposed for resolving CS-related setbacks. in which nz,q indicates the index of the AoA grid point nearest
to the AoA of the lth path. Thus, wnθ y can be structured as
IV. CS- BASED R ECONSTRUCTION
A. CS Formulation of θ wnθ y = Ψθny (∆θ) uθny + nθny , (20)
θ Ny ×Nz
The constructed θ-based covariance matrix W ∈ C
where uθny ∈ CÑz ×1 is an L-sparse vector, nθnz ∈ CÑz ×1
encompasses Ny vectors with vector size Nz , in which the
ny th vector can be drawn out from Wθ given by wnθ y ≜ is the noise vector, and Ψθny ∈ CNz ×Ñz is the dictionary
T matrix. The design of the dictionary matrix is required to
Wθ ny ,: ∈ CNz ×1 . The vector wnθ y is actually an ob- satisfy the RIP condition in order to facilitate the accurate
servation interpreted as a linear combination of all L steering sparse signal recovery. As a necessity for accurate sparse
8
L
" #!
X ȷπ n2y + n2z δ
(ny ,nz )
2
wnr y ,nz = β̄l2 exp 1 − gl − 2 (ny sin θl sin φl + nz cos θl ) . (16)
4 rl
l=1
L
" #!
πn2y λ
2
X
(n ,n ) π
wnr y ,nz = β̄l2 exp ȷ 1− gl y z − ny sin θl sin φl . (17)
16rl 2
l=1
h i
T
signal recovery, the dictionary matrix has to satisfy the RIP (Wr ) ∈ CNz ×1 . To facilitate the gridding over dis-
:,ny
condition [35] for eliminating the grid coherence, which tance, we likewise introduce a uniform grid of discrete points
adversely impacts the estimation performance. To be specific, of r, i.e.,
the RIP condition can be interpreted as two arbitrary columns 2
′

(n ,n )
h the (nz , inz )th entry
being (approximately) orthogonal, where r̃nz : r̃nz = nZz +1
∆
1 − gl y z , nz = 0, ..., Ñz − 1 ,
θ Nz ×Ñz θ
of Ψny ∈ C is crafted by Ψny (∆θ) =
(nz ,n′z ) (25)
n o λ
Ny2 δ= 4 Ny2
exp ȷπnz cos θ̃ny + ∆θny . Thus, (18) can be recast to where Z∆ = 2 →
is the minimum value to ensure
2 λ
8β∆ δ 2β∆
a CS form with uncertain dictionary matrix (containing the the approximate orthogonality of the dictionary matrix with
to-be-determined ∆θ) and sparse vector uθny , i.e., β∆ representing the orthogonality controlling factor in order
to make sure the grid coherence being below 12 [26]. Taking
ynθ y = Bθny Ψθny (∆θ) uθny + Bθny nθny . (21)
into account the
upper
q and lower bounds of the Rayleigh
D 3 2D 2
distance, i.e., 0.5 λ , λ (D is the array aperture), the
B. CS Formulation of φ
The CS reconstruction for φ can be executed using similar number of the grid points can be determined in accordance
2
Z∆ (ny ,nz )
tactics to those employed in θ’s. However, the phase term in with r̃nz = nz +1 1 − gl ≥ Dmin (Dmin =
(14) includes the θ-associated part, which may preclude the q
3
direct gridding over φ. To resolve this hiccup, we introduce 0.5 Dλ ), and the maximum of Ñz takes the value of
an intermediate variable ζ = {ζ1 , ..., ζL } for satisfying $ %
Ny2
2
sin ζl = sin θl sin φl , l = 1, ..., L. Next, an arbitrary vector

(ny ,nz )
Ñz = 2 λD 1 − gl . (26)
indexed by nz can be extracted from the φ-based covariance max 2β∆ min
matrix Wφ , i.e., wnζ z ≜ [Wφ ]:,nz ∈ CNy ×1 . A wide
measurement matrix Bζnz ∈ CNRF ×Ny is left-multiplied by Furthermore,
the r-based offset vector ∆r =
the vector wnζ z for confining its size with the number of RF ∆r1 , ..., ∆rÑz is specified with its ñz th entry given by
chains, thus yielding ynζ z = Bζnz wnζ z . To perform gridding (
rnz − r̃nz , if nz = nz,q , q = 1, ...L,
over ζ, a uniform grid of Ñz angles ζ is introduced,o i.e., ∆rnz = (27)
n
2 Ñy −1
0, otherwise,
ζ̃ny : sin ζ̃ny = Ñ ny − 2 , ny = 0, ..., Ñy − 1 .
y
Regarding the mismatches between the true angles and the where nz,q indicates the index of the distance grid point
n grid points,o ζ-based offset vector is defined as
prescribed nearest to the distance of the lth path. Then, denoting ynr y
∆ζ = ∆ζ1 , ..., ∆ζÑy , with ∆ζny given by and urny by the r-based observation vector and the sparse
( vector, respectively, we can likewise structure wnr y as its CS
ζny − ζ̃ny,q , if ny = ny,q , q = 1, ...L, form
∆ζny = (22)
0, otherwise,
ynr y = Brny Ψrny (∆r) urny + Brny nrny , (28)
in which ny,q indicates the index of the AoA grid point nearest
to the AoA of the lth path. Thus, wnζ z can be structured as a where Brny ∈ CNRF ×Nz is the size-reducing measurement
CS form matrix, nrny is the noise vector, and Ψrny (∆r) ∈ CNz ×Ñz ,
wnζ z = Ψζnz (∆ζ) uζnz + nζnz , (23) Ψrny (∆r) ∈ CNz ×Ñz is the dictionary matrix, whose
in which uζnz ∈ CÑy ×1 is an L-sparse vector, nζnz ∈ (nz , ñz )th entry is given by
CÑy ×1 is the noise vector, and Ψζnz ∈ CNy ×Ñy represents
r
Ψl ∆rny nz ,ñz
the dictionary matrix, whose ny , n′y th entry isogiven by " #!
ζ n n2y 1
Ψnz (∆φ) n ,n′ = exp ȷπnz sin ζ̃nz,q + ∆ζnz . There- = exp ȷπ − ny sin θl cos φl . (29)
y y r̃nz + ∆rnz 2
fore, the ζ-based observation ynζ z can be recast to
ynζ z = Bζnz Ψζnz (∆ζ) uζnz + Bζnz nζnz . (24) D. Discussion
Our goal is to obtain the offset vectors ∆θ, ∆φ and
C. CS Formulation of r ∆r from the 3D AED parameter-based CS problem given
An arbitrary vector indexed by nz can be drawn out in (21), (24), and (28), respectively. Then, the 3D AED
from the r-based covariance matrix Wr , i.e., wnr y = parameter can be recovered by θ̂l = θ̃nz,q + ∆θnz , φ̂l =
9
sin(ζ̃ny,q +∆ζny )
arcsin sin θ̂l
, and r̂l = r̃nz,q + ∆rnz , q = 1, ..., L.
The way to accurately acquire the offset vectors will be
elaborated in the next section. Let us look first at a pair of two
issues, in which how the proposed DeRe paradigm is a curse
when concerned in terms of angular index misinterpretations
but a blessing when embraced with regard to significantly
reduced gridding complexity.
1) Angular Index Correction: The outputs of the DeRe
framework, i.e., θ, φ, and r, just play separate roles on
a significant path, which can be referred to as the partial
information. When recovering the original holographic chan-
nel using the obtained independent partial information, there
typically exist the angular index misinterpretations which
significantly deteriorate the estimation performance. To elab-
(a)
orate, in examination of case L = 3, Fig. 4(a) portrays the
genuine angular power scattered in relation to the azimuth-
elevation angle pair (θ, φ) after DeRe filtering, as well as
the misinterpretation between the attained angular indices and
the true indices of the significant angles when used with
partial information.h There i exist three zero-entries in each of

the sparse vectors uθny and uζnz ny,q . By retrieving the
nz,q
indices of the non-zero entries for this pair of sparse vectors,
the true index of the lth significant angle pair (θl , φl ) may
be adversely misinterpreted for that of other significant angle
pairs. For example, having retrieved the index q = 2 for
θ̃nz ,2 and ζ̃ny ,2 , we may obtain a significant angle pair θ̃nz ,2
and ζ̃ny ,2 (with red marker □ in Fig. 4(b)). However, this
results in a false index for the genuine significant angle pair
associated with the path whose true index is l = 3. Such a
(b)
misinterpretation significantly erodes the estimation accuracy.
To address this hiccup, it is crucial to conduct the angular Fig. 4. (a) Normalized power of the azimuth-elevation angle pair (θ, φ), and
index correction after attaining the estimates of θ̃l and ζ̃l . their independent partial information projected on the xoz and yoz plane. (b)
Misinterpretations between the attained angular indices and the true indices
Regarding the nz th entry in the θ-based sparse vector uθny , of the significant angles.
we denote q by its index of the non-zero entry, whose qth
non-zero entry is given by
1
Nz [uζnz ]ny,q
eζ
P
h i = , q = 1, ..., L.
uθny = 2β̄l2 cos (πny sin θl sin φl ) , q = 1, ..., L. (30) ny,q Nz
nz =1
2 cos(πnz cos(θ̃nz,q +∆θnz ))
nz,q
(33)
Similarly, the non-zero entry in the φ-based sparse vector uζnz The idea of the angular index correction is straightforward,
indexed by q is given by that is, identifying the two entries with the closest power, and
ζ matching their indices to the index of the path. Therefore,
unz ny,q = 2β̄l2 cos (πnz cos θl ) , q = 1, ..., L. (31)
we can accurately retrieve the exact index of the lth path
Since there are L paths in total, there will be L! misinter- associated with the azimuth-elevation angle pair (θl , φl )
pretation combinations. We are trying to identify the true 2
eθ − eζ

indices of the azimuth and elevation angles associated with q̂ = arg min nz,q ny,q
. (34)
the lth path. To this end, a pair of variables eθ and eζ are 2
Having identified the true index q̂,
the power gain β̄l can be
introduced to represent the power gain of the lth path (i.e., β̄l2 ) 2 θ ζ
obtained by β̄l = e n = e n , q̂ = 1, ..., L.
associated with the decomposed θ and ζ, respectively. Given z,q̂ y,q̂
sin(ζ̃ny,q +∆ζny ) 2) Gridding Complexity: The use of the DeRe framework

θ̂l = θ̃nz,q + ∆θnz and φ̂l = arcsin sin θ̂l
,q = is capable of significantly reducing the gridding complexity.
1, ..., L, we have In particular, gridding across the coupled
3D AED parameters
h
θ
i has a complexity order of O N 3 , in conjunction with
Ny uny N = Ny × Nz . Consequently, the estimation process would
1 X nz,q
eθ n =

z,q

Ny n =1 2 cos πn sin θ̂ sin φ̂
be less efficient owing to the resultant extremely large search
y y l l space. By contrast, 1D gridding over θ has a complexity
order of O (Nz ), and gridding over both φ and r shares
h i
N
Py uθny
1 nz,q
= Ny 2 cos(πny sin(ζ̃ny,q +∆ζny ))
, q = 1, ..., L, the same complexity order of O (Ny ), which leads to a total
ny =1 gridding complexity order of O (Nz + 2Ny ). Therefore, it is
(32) intuitive that the original multiplicative gridding complexity
10
is relaxed to its additive counterpart, significantly lowering implementation and the intended design. Accordingly, (37)
the computational burden in practice. can be recast as
ynθ y = E F̄ + ∆F̄ diag (hτ ) Eh→V h∗,τ

E. Robustness of DeRe Framework = E F̄ + ∆F̄ E {diag (hτ ) Eh→V h∗,τ }

Given the potential adverse effects of hardware imperfec- = F̄ + E ∆F̄ E {diag (hτ ) Eh→V h∗,τ } . (38)

tions, such as deviations between the practical and theoretical
designs of the combining matrix F and holographic pattern The form in (38) is appealing since it is analytically tractable.
M, this subsection aims for examining the robustness of the This implies that the resultant error can be mitigated through
DeRe framework against such perturbations in the context skillful designs of the combining matrix and holographic pat-
of channel estimation. Note that ynθ y is actually a fictitious tern. Specific design details are beyond the scope of this paper,
vector for ease of analysis (as are wnθ y and V), so before and interested readers may refer to [36], [37] for inspiration
proceeding to unveiling DeRe’s robustness, let us establish in transceiver design within the context of holographic scope.
a critical bridge between the physical signal received from The robustness performance of the DeRe framework will be
the BS and the observation ynθ y utilized in the CS procedure. evaluated later in Sec. VI.
Subsequently, we shall take θ as an illustrative instance.
V. D E R E -VM A LGORITHM
Let yτ , τ ∈ {1, ..., T } denote an arbitrary sample
of the received signal at the BS. We define fa→b (a) In pursuit of an accurate and robust channel estimation
as the mapping from a to b. For an arbitrary obser- for holographic MIMO systems, we intend to recover the
vation sampled at τ , we obtain ynθ,τ = Bθ,τ wnθ,τ , sparse vectors uθny , uζny and urny , as well as their associated
y ny y
offset vectors ∆θ, ∆ζ and ∆r, as presented in (21), (24),
whose mapping relation is given by fw→y wnθ,τ =
y and (28), respectively. Albeit the seeming straightforwardness
fw→y (fV→w (V)) = fw→y (fV→w (fh→V (hτ ))). The of that kind of parameter-recovery problem, it poses arduous
mapping fw→y (fV→w (Vτ )) can be recast by means of a challenges due to a pair of peculiarities: i) the uncertainty of
matrix manipulation given by the sensing matrices including the offsets to be determined,
fV→w (Vτ ) = CV→w (IN + EV→w ) Vτ , (35) i.e., Ψθny (∆θ), Ψζny (∆ζ), and Ψrny (∆r), as well as ii)
the imperfect priors of the sparse vectors induced by the
where EV→w = ẽ1 ; ...; ẽTN ∈ CN ×N , with the nth vector
T
unavailability of the exact distributions for uθny , uζny , and urny .
ẽn given by ẽn = en′ and en′ ∈ RN ×1 representing the basis Therefore, targeting the twofold issues, a layered probability
vector whose n′ th entry is non-zero. The matrix CV→w ∈ model is leveraged to provide appropriate prior distributions
CNz ×N is constituted by [c1 , ..., cN ], whose nth vector is for the sparse vectors, followed by an efficient and robust
crafted by algorithm design.
(
Nz ×1 ei , n = (ny , i) , i = 1, ..., Nz , A. Layered Probability Model
cn ∈ R = (36)
0, others. For the sake of brevity, we take θ as an example to
It thus holds that fw→y (fV→w (fh→V (h ))) τ
= explicate the way to harness the structured sparsity un-
θ
θ,τ τ
Bny CV→wθ (IN + EV→wθ ) (diag (h )) Eh→V h , ∗,τ
in derlying w ny . Concerning that Ny estimates of θ can be
which the implementation of Eh→V is similar to EV→w ’s, attained by resolving Ny θcopies θof problem θ
(21), we can
just using different index selection strategy shown in (9). group the sparse vectors u 1 , ..., u n y
, ..., u Ny into Ñz blocks
As a result, the received signal at the BS that serves as an h with block size N y , in which the n z th block is given by
i h iT
θ θ θ θ θ θ θ
observation yny can be given by u1,nz , ..., uny ,nz , ..., uNy ,nz . Let ρ = ρ1 , ..., ρNy rep-
o resent the precision vector, in which 1/ρθ
ny ,nz denotes the
n
∗,τ
ynθ y = E Bθ,τ τ
ny CV→wθ (IN + EV→wθ ) diag (h ) Eh→V h variance of uθny ,nz . Furthermore, a support vector denoted by
= Bθny CV→wθ (IN + EV→wθ ) E {diag (hτ ) Eh→V h∗,τ } . αθ ∈ {0, 1}Ñz is introduced, whose nz th entry αθ indicates
nz
(37) whether the nz th block is active (αnθ = 1) or not (αnθ = 0).
z z
θ θ θ
In practical systems, the linear transformation in (37), i.e., Therefore, the joint distribution of u , ρ , and α is in the
Bθny CV→wθ (IN + EV→wθ ) can be achieved by appropri- form
of

ately designing the combining matrix and the holographic pat- p uθny , ρθny , αθ = p αθ p ρθny |αθ p uθny |ρθny , (39)
tern. The remnant expectation term E {diag (hτ ) Eh→V h∗,τ } | {z } | {z }| {z }
Support Precision Sparse vector
can be realized by an output of the Hadamard product whose
inputs take the signal received at the BS and its conjugate Then let us move on to describe in greater detail the prior
counterpart. distribution of each layer.
The structure in (37) indicates that the signal received 1) Layer I: Probability Model for Support Vector αθ :
θ
at the BS can be crafted to attain the observation yny Due to the limited scatterers in the physical propagation
embracing θ-associated statistical characteristics. Regard- environment, the birth-death process of αθ can be modeled
ing the error induced by hardware imperfections, we let as a Markov prior
Bθ,τ
ny CV→wθ (IN + EV→wθ ) = F̄ + ∆F̄, where F̄ denote Ñz
the intended design of the combining matrix and holographic p αθ = p αnθ z
Y
p αnθ z |αnθ z −1 ,

(40)
pattern, while ∆F̄ captures the deviation between the actual nz =2
11

with the transition probability p αnθ z |αnθ z −1 given by Layer III Layer II Layer I
1−αθnz αθ T
U1,1 T
F
( u
θ θ
(1 − p0→1 ) (p0→1 ) nz , αnθ z −1 = 0, p u1T | y1T u1,1 f 1,1
p αnz |αnz −1 =
...
...
...
1−αθnz αθ unTy ,1 f nuy ,1 U nT ,1 f1ȡ D1T
(p1→0 ) (1 − p1→0 ) nz , αnθ z −1 = 1, y
...
...
...
(41) uTN y ,1 f Nuy ,1 U TN
A group of N y
y ,1 estimates of Tˆ
where p0→1 represents the transition probability of entries in
...
T T
p u ny | y ny
u1,T nz f1,unz U1,T n
the sparse vector uθny shifting from 0 to 1, and vice versa for z
...
...
...
p0→1 . T
u n y , nz f u
n y , nz
U nT ,n y z
f nȡz D nT z
2) Layer II: Probability Model for Precision Vector ρθ : A
...
...
...
uTN y ,nz f Nuy ,nz UN T
Gamma prior distribution is imposed on the precision vector y , nz
...
ρθ of the support vector αθ u1,T N z f1,uN z U1,T N z
...
...
...
Ñz Ny Ñz p uTN y | yTN y T
un y , N z f u
Un T
f Nȡz D NT

Y Y Y n y , N z y ,Nz
p ρθ |αθ = p ρθnz |αnθ z = p ρθny ,nz |αnθ z

z
...
...
...
T
uN y , N z f u
U TN
nz =1 ny =1 nz =1 N y , N z y ,Nz
Spasrse vector uT Precision vector ȡT Support vector ĮT

| {z }
ρ
fn z
Ny Ñz h iαθn Fig. 5. Factor graph of the joint distribution p yθ , v θ ; ∆θ .

Y Y z
= Γ ρθny ,nz ; aθny ,nz , bθny ,nz
ny =1 nz =1 h i
h i1−αθn E uθny ,nz |yθ ; ∆θ , with the expectation taken over the con-
z
× Γ ρθny ,nz ; āθny ,nz , b̄θny ,nz , (42)
ditional marginal posterior
where Γ (ρ; a, b) is a Gamma hyperprior with the shape
p uθny ,nz |yθ ; ∆θ
parameter a and rate parameter b. For the case of αnθ z = 1,
the shape and rate parameters aθny ,nz and bθny ,nz are selected XZ
p yθ , uθ , ρθ , αθ ; ∆θ

∝
aθn ,n
h i
to satisfy bθ y z = E ρθny ,nz , since the variance 1/ρθny ,nz of −uθny ,nz
αθ
ny ,nz h i XZ
p uθ , ρθ , αθ ; ∆θ p yθ |uθ ; ∆θ ,

uθny ,nz is equivalent to E ρθny ,nz and uθny ,nz deviates from = (45)
−uθny ,nz
zero with a high probability when it is active. On the contrary, αθ
we let the shape and rate parameters āθny ,nz and b̄θny ,nz be where −uθny ,nz represents the entry removal with respect to
āθn ,n
h i
chosen such that b̄θ y z = E ρθny ,nz ≫ 1, conditioned on uθny ,nz over the sparse vector u and ∝ refers to the equality
ny ,nz
the event of αnθ z ̸= 1, in which case the variance of uθny ,nz after scaling. Furthermore, the uncertain grid offset ∆θ̂ can
tends to be zero when the associated path is inactive. be attained by MLE as
3) Layer III: Probability Model for Sparse Vector uθ :
Z
∆θ̂ = arg max ln p yθ ; ∆θ = arg max ln p yθ , v θ ; ∆θ ,

The conditional probability for sparse vector uθ is considered ∆θ ∆θ vθ
to follow a non-stationary Gaussian prior distribution with (46)
distinct precision ρθ
in which v θ = uθ , ρθ , αθ represents the collection of
Ny Ñz
Y Y variables associated with the probability model. We thus attain
p uθ |ρθ = p uθny ,nz |ρθny ,nz

the MMSE estimate of uθny ,nz by figuringout the MLE of∆θ̂
ny =1 nz =1
| {z } and the conditional marginal posterior p uθny ,nz |yθ ; ∆θ .
u
fny ,nz Recall that Ny estimates of θ can be attained by solving
Ny
Y Ñz
Y Ny copies of problem (21), and the associated sparse vectors
= CN uθny ,nz ; 0, 1/ρθny ,nz , (43) uθ1 , ..., uθny , ..., uθNy have been grouped into Ñz blocks with
ny =1 nz =1 block size Ny , as illustrated in the factor graph in Fig. 5. Such
Additionally, in the presence of the complex Gaussian noise, a grouping facilitates an appropriate fusion of Ny beliefs,
we have i.e., messages passing via the path uθny ,nz → fnuy ,nz →
ρθny ,nz → fnρz , ny = 1, ..., Ny . To be explicit, in light of the
p ynθ y |uθny = CN ynθ y ; Bθny Ψθny uθny , κθ INRF , (44) belief propagation [38], the message νfnρz →αθn is calculated
z
by taking the product of all messages received by the factor
where κθ = σ −2 represents the noise precision
that follows a node fnρz on all other edges, i.e.,
Gamma hyperprior p κθ = Γ κθ ; aκ , bκ . The tunable aκ
νfnρz →αθn αnθ z

and bκ tend to zero, i.e., aκ , bκ → 0, for approximating a
z
more general hyperprior. X YNy
ρ θ θ θ
4) Channel hRecovery Problem Formulation: Given obser- ∝ θ
f nz αnz , ρ1,nz , . . . , ρ Ny ,nz ρθny ,nz .
i −αnz ny =1
Ny Nz ×1
vations yθ = y1θ ; ...; yN
θ
y
∈ C , our objective is to (47)
h i
estimate the sparse vector u = uθ1 ; ...; uθNy ∈ CNy Nz ×1 This implies that each block nz contains partial priors of
and the grid offsets ∆θ. More precisely, we intend to the Ny sparse vectors, and priors of Ñz blocks function
compute the MMSE estimates of uθny ,nz , i.e., ûθny ,nz = collaboratively on the channel support vector αθ .
12
It is still arduous to obtain the precise estimates of ûθny ,nz approximate message passing (AMP)-based algorithms, e.g.,
and ∆θ̂, since the conditional marginal posterior in (45) is Turbo-AMP [40] and Turbo-CS [41], easily get entrapped into
unavailable due to the complex integration, and the dense their poor local optimum due to the ill-conditioned dictionary
loops in the factor graph exacerbate matters. Additionally, the matrix Ψθ (∆θ) induced by the indeterminate ∆θ. To facil-
non-concave nature of problem (46) hinders the pursuit of its itate the practical implementation, two entities are specified
stationary solution even if one uses some plain optimization in the DeRe-VM-E Step, namely Entity A and Entity B as
approaches. Therefore, the subsection that follows proceeds exemplified in Fig. 6, in support of the sparse VBI estimates
to present beneficial techniques for clearing up the above and forward-backward message passing. Furthermore, this
hindrances, as well as an effective algorithm for facilitating pair of entities need to interact with each other to eliminate
the efficient estimation of θ. the self-reinforcement during the update procedure.
• Entity A: By capitalizing sparse VBI [42], this entity
B. DeRe-VM Algorithm intends to calculate the approximate conditional marginal

posteriors of latent variables, i.e., ψ v θ . We denote
As a key feature of DeRe-VM, an approximate variational
νχ→αθn αnθ z , nz = 1, ..., Ñz by the input of Entity A,
distribution ψ v θ is employed to refine the intractable z
which is actually the message passed from Entity B and
posterior p v θ |yθ ; ∆θ in the DeRe-VM-E Step, while up-
incorporates the Markov priors of the structured sparsity
dating the indeterminate grid offset θ using the gradient-based
associated with the sparse vector uθny , ny = 1, ..., Ny .
technique in the DeRe-VM-M Step.
The sparse VBI estimates implemented in Entity A will
1) DeRe-VM-M Step: Inexact Majorization-Minimization
be elaborated on later.
(MM): The likelihood function
• Entity B: Let νfn ρ θ (·) denote the message passed
ln p yθ ; ∆θ is indeed intractable owing to the absence of z →αnz
from Entity A, which constitutes the input of Entity
closed-form expressions induced by the prohibitive integrals
B. To facilitate exposition, the input and the output
of the joint distribution with regard to indeterminate v θ . An al-
of Entity B are defined as νninz ≜ νfnρz →αθn (·) and
ternative is to construct its surrogate function based on ∆θ (r) z
in the rth iteration in place of the objective in problem (46), νnout

z
≜ ν χ→α θ
nz
(·) , n z = 1, ..., Ñz , respectively. In
out θ

providing that the exact posterior p v θ |yθ ; ∆θ (r) has been each algorithm iteration, the message ν nz α nz ≜
identified. Unfortunately, the exact posterior p v θ |yθ ; ∆θ (r) νχ→αθn αnθ z can be obtained in accordance with the
z
is also highly unavailable in our examined problem owing belief propagation rule [38], [39]
to the intricate loops inherent in the factor graph. We thus ψ αnθ z

νnout θ

need to leverage the variational Bayesian inference (VBI) and α nz ∝ , ∀nz , (51)
z
νninz αnθ z
message passing techniques to attain an alternative ψ v θ
for a faithful approximation
of the conditional posterior which intuitively equals the approximate posterior with
p v θ |yθ ; ∆θ (r) . The to-be-determined posterior ψ v θ has

respect to αnθ z , since νninz αnθ z plays the role of Entity
a factorized form, i.e., ψ v θ = θ θ
A’s prior with respect to αnθ z . The inference also holds
Q
j∈J ψ vj , with vj
θ for Entity B’s output, and this pair of entities would inter-
representing an individual
in the variable collection v , and
we have J = j|∀vjθ ∈ v θ . Then, an equivalent of the act iteratively until their convergence has been achieved.
original MLE problem in (46) can be obtained based on the Having established the method for carrying out message
approximate posterior, which is formulated as exchange between the twin entities, let us move on to an
Z
p yθ , v θ ; ∆θ (r)
elaboration of the sparse VBI estimation in Entity A for
θ
max ψ v ln dv θ . (48) closed-form updates of the associated posteriors.
ψ(v θ ),∆θ ψ (v θ )
| {z } 3) Entity A: Sparse VBI Estimation: Firstly, to determine
F (∆θ;∆θ (r) ) the approximate conditional marginal posterior ψ αθ , the
sparse VBI estimation can be outlined as a minimization
In the (r + 1)th iteration in the DeRe-VM-M Step, ∆θ (r+1) of the Kullback-Leibler divergence (KLD) with respect to
can be updated by
p v θ |yθ ; ∆θ and ψ αθ , thus yielding that

∆θ (r+1) = arg max F ∆θ; ∆θ (r) . (49)

ψ vθ
Z
∗ θ θ

∆θ ψ v = arg min ψ v ln (52a)
ψ (vjθ ),∆θ
p (v θ |yθ ; ∆θ)
The non-concave nature of the objective, nevertheless, im-
pedes the pursuit of the global optimum for problem (49). s.t. ψ v θ =
Y
ψ vjθ .

(52b)
We thus exploit the following gradient update in an effort to j∈J
obtain a stationary solution
∗

The stationary point ψ vjθ to problem (52a) needs to satisfy
(r)
(r+1) (r) (r) ∂F ∆θ; ∆θ
∆θ = ∆θ + ι , (50) ψ ∗ vjθ

∂∆θ
∗

vjθ ψ vjθ
Q
j̸=l ψ
Z Y
where ι(r) is the step size determined by the Armijo rule [35]. ∗ θ θ
dvjθ ,

= arg min ψ vj ψ vl ln θ |yθ ; ∆θ)
2) DeRe-VM-E Step: A Pair of Entities: Due to myriads ψ (vj ),∆θ
θ p (v
j̸=l
of loops in a dense factor graph depicted in Fig. 5, a direct (53)
employment of the classical sum-product message passing
(SPMP) [39] fails to traverse the factor and variable nodes which can be attained via alternating updates of ψ vjθ , ∀j ∈
J . Specifically, for given ψ vl , l ̸= j, the optimal ψ ∗ vjθ
θ

over the entire factor graph. As potential solutions, the
13
DeRe-VM-E Step DeRe-VM-M Step

Message passing
q
Entity A Entity B y ( vq )
r1,n z
nf ρ
®a nqz Dθ ( r +1) = Dθ ( r )
q
nz
qˆl = q n + Dq n ,
W r nq ,n
y z f ρ
nz anq
z
n out
nz n in
nz anq
c ¶ ( Dθ; Dθ ( ) )
r z ,q z
z
Dθ +i(
r)
q = 1,..., L
n c ®a q ¶Dθ
r qN y , nz
nz
Forward-backward
Sparse VBI estimation message passing
Fig. 6. Schematic diagram of the proposed DeRe-VM algorithm by taking θ as an example.
Algorithm 1 Parametric Decomposition and Compressed

• Update sparse vector uθ : ψ uθ can be updated by a
Reconstruction-based Variational Bayesian Inference complex Gaussian distribution
Message Passing (DeRe-VM) Algorithm
1: % Parametric Decomposition Ny
Y
θ
CN uθny ; µθny , Σθny .

2: Decompose 3D AED parameters θ, φ, and r by constructing ψ u = (55)
their respective covariance matrix according to (13), (14), and ny =1
(17).
3: % Compressed Reconstruction Σθny and µθny can be calculated by
4: Initialize the offset vector ∆θ = 0, transition probability p0→1 ,
p0→1 , hyper-parameters aθny ,nz , bθny ,nz , āθny ,nz , b̄θny ,nz , κθ , −1
H
Armijo step size ι, the maximum iteration number rmax , and Σθny = diag ρny ,1 , ..., ρny ,Ñz + κθ Ψθny (∆θ) Ψθny (∆θ) ,
the convergence threshold ϵ, ϵ′ . H
5: repeat = Λθny − κθ Λθny Ψθny (∆θ)
6: DeRe-VM-E Step:
H
−1
7: % Entity A: Sparse VBI Estimation × I + κθ Ψθny (∆θ) Λθny Ψθny (∆θ) Ψθny (∆θ) Λθny ,
8: repeat (56)
Update posteriors ψ uθ , ψ ρθ , and ψ αθ in compli-

9:
ance with (55)-(59), respectively.
10: until The convergence criteria of DiLuS-STPL-E Step ismet. H
11: Obtain the message from Entity A, i.e., νninz αn θ
≜ µθny = κθ Σθny Ψθny (∆θ) ynθ y , (57)
z
θ

νfnz →αθn αnz , and then pass it to Entity B.
ρ
z ãθn aθκ
12: % Entity B: Forward-backward Message Passing where ρny ,nz = b̃θ y z , κθ
,n
= , and Λθny =
Update the message νnout θ θ bθκ

13: z
αn z
≜ νχ→αθn αn z
, and then n
y ,nz
z
b̃θn ,Ñ
θ
send it to Entity A for the next iteration. b̃ny ,1 y z
diag ãθ , ..., ãθ .
14: DeRe-VM-M Step: ny ,1 ny ,Ñz
Construct the surrogate function F ∆θ; ∆θ (r) in (48)

15: • Update precision vector ρθ : ψ ρθ is given by
using the output of Entity A in DeRe-VM-E Step, i.e., ψ v t .

16: Update ∆θ (r+1) according to (50). Ny

Y Ñz
Y
ψ ρθ = Γ ρθny ,nz ; ãθny ,nz , b̃θny ,nz , (58)

θ,(r−1) θ,(r)
17: until The convergence criteria µny − µny ≤ ϵ and
θ,(r−1) θ,(r) ny =1 nz =1
Σny − Σ ny ≤ ϵ′ are met or the maximum iteration
number rmax is exceeded. where ãθny ,nz = ν̃nout aθny ,nz + 1 − ν̃nout
θ
āny ,nz + 1,
Obtain the optimal posterior ψ ∗ v θ and the offset vector ∆θ.

18: z z
θ 2
The estimate of θ̂l can be determined by θ̂l = θ̃nz,q +∆θnz , q = b̃ny ,nz = ν̃nz bny ,nz + 1 − ν̃nz b̄ny ,nz + µθny +
θ out θ out
1, ..., L.
19: Regarding the estimates of φ and r, repeat Step 3-Step 18,
Σθny ,nz , and ν̃nout
z
represents the posterior probability
and we just use φ and r in place of the superscript of asso- of αnθ z = 1, also known as the message passed from
ciated variables, such as yφ , v φ , yr , and v r , etc. Next, we Entity A.
obtain two other estimates of φ̂l and r̂l , as given by φ̂l = • Update support vector αθ : We obtain the posterior
sin(ζ̃ny,q +∆ζny )
arcsin sin θ̂l
, and r̂l = r̃nz,q + ∆rnz , q = 1, ..., L. ψ αθ in the form of
20: Conduct angular index correction in accordance with (30)-(34).
21: Output the index q̂ indicating the lth significant path associated Ñz
3D AED parameters (θl , φl , rl ).
Y
ψ αθ =
out
ν̃nz δ αnθ z − 1 + 1 − ν̃nout δ αnθ z .

z
nz =1
(59)
that satisfies the KLD minimization in problem (52a) is Likewise, the foregoing procedures are also applicable to
specified as both φ and r for obtaining their CS-based solutions. When
establishing the layered processing probability model and
ψ ∗ vjθ ∝ exp ln p yθ , v θ ; ∆θ Q ψ(vθ ) ,

(54) conducting sparse estimation techniques, we just use φ and
j̸=l l
R r in place of the superscript of associated variables, e.g., yφ

where ⟨f (z)⟩h(z) = f (z)h (z) dz. and yr , etc., and the corresponding phases have been omitted

Secondly, the posteriors ψ uθ , ψ ρθ and ψ αθ can be for conciseness. In brief, the proposed DeRe-VM algorithm
updated as follows. have been summarized in Algorithm 1.
14
C. Computational Complexity -5
(a)
-5
(b)
Following the overall flow of DDS-VBIMP summarized

-10 -10
in Algorithm 1, the main computationalburden lies in the DeRe-VM
updates of ψ uθ , ψ ρθ , and ψ αθ in Step 9 using

-15 -15
(55)-(59), respectively. The total number of multiplications
Convergence points

θ
to update ψ u is Ny 3NRF Ñz2 + 2Ñz NRF 2
. Upon the
NMSE (dB)
NMSE (dB)
-20 -20
utilization of the matrix inversion lemma on (56) in order to

lower complexity, the computational complexity with regard -25 -25
3
to the matrix inversion in (56) is O Ny NRF . Thus, the
-30 -30
total computational complexity
order of the proposed
DeRe-
VM algorithm is O NRF Ny Ñz2 + Ny Ñz NRF 2
per iteration, -35 -35
concerning that it typically holds that Ny , Ñz ≫ NRF . This

implies that for an N -element UPA (N = Ny × Nz ), the -40
1 2 3 4 5 6 7 8 9
-40
1 2 3 4 5 6 7 8 9
complexity of the proposed DeRe-VM algorithm just scales Iteration Iteration
proportionally with the number of antenna elements raised

Fig. 7. Convergence behavior of DeRe-VM. (a) NMSE performance of 3D
to the power of 1.5. Additionally, our proposed DeRe-VM AED parameters θ, φ, and r. (b) NMSE performance of the auto-correlation
algorithm functions within the off-grid basis and is capable matrix R.
of eliminating on-grid errors, sharing the similarities with
the methodology presented in [26], namely polar-domain 0
simultaneous gridless weighted (P-SIGW). Albeit the evident
superiority of P-SIGW, there are two issues that should -5
be taken with a grain of salt: i) P-SIGW fails to capture

the coherence between the 2D elevation-azimuth angle pair -10 Rayleigh distance
(θ, φ) and the distance r for a robust recovery of 3D AED
NMSE (dB)
parameters, and ii) the computational complexity of P-SIGW -15
increases quadratically with the number of antenna elements,

-20
making it less favorable for a large N . In the section that
follows, a comprehensive performance evaluation will be
-25
illustrated, showcasing their distinct advantages in relation to
various cases.
-30
VI. S IMULATION R ESULTS -35

5 10 15 20 25 30 15 40 25 50
A. Configuration Distance (m)
In this section, we conduct performance evaluation of

Fig. 8. NMSE performance versus the distance.
the proposed DeRe-VM algorithm for channel estimation
in the holographic MIMO system of interest. The antenna
elements at the BS are equipped in the form of Ny = 129
key system parameters are set as follows unless otherwise
and Nz = 65. The considered holographic MIMO sys-
specified: NRF = 16, P = 32, Pp = 33 dBm, β∆ = 1.2,
tem operates at 30 GHz with a bandwidth of 100 MHz. p0→1 i i i
It is assumed that the number of paths is L = 3. The p0→1 +p1→0 = 0.06, any ,nz = āny ,nz = bny ,nz = 0.99,
i
simulation results are averaged over 500 random channel b̄ny ,nz = 0.01, i ∈ {θ, φ, r}.
realizations. For each realization, the distance between the BS
and the user/scatterer is sampled from a uniform distribution
rl ∼ U (Dmin , 0.5Dmax ) , l = 1, ..., L, and based on the B. Convergence Behavior
fixed distance, the associated angular parameters of the LoS
path are randomly sampled from the uniform distributions We demonstrate the convergence behavior of the proposed
θl ∼ U (0, π) and φl ∼ U (−π/2, π/2) , l = 1, ..., L, DeRe-VM algorithm by plotting the NMSE curves of 3D
respectively. Furthermore, the complex gain of each NLoS AED parametersh θ, φ, and r iin Fig. 7(a), which is cal-
2 2
path is modeled to follow a complex Gaussian distribution
culated by E ∥x − x̂∥2 / ∥x∥2 , ∀x ∈ {θ, φ, r} [44]. By
with zero mean and variance σ NL , i.e., CN 0, σ NL , in which contrast, Fig. 7(b) shows the NMSE performance of the
2
σ NL is 20 dBm weaker than that of the LoS paths [43]. auto-correlation matrix, obtained by E R − R̂ / ∥R∥F .
2
In the process of the CS reconstruction, the number of the F
uniform grid is set to Ñz = Nz for θ, and Ñy = Ny for φ, A steep decline can be spotted in the first few iterations,
respectively. We concentrate on the normalized mean square followed by a gradual settling at a value after around seven
error (NMSE) performance of the auto-correlation matrix R, iterations for both scenarios. The NMSE gap in accuracy
between r and θ(φ) can be attributed to the presence of a

which can be reconstructed by R ≜ E hhH after we
obtain the estimates of the 3DAED parameters θ̂, φ̂, and residual correlation inside the distance dictionary matrix. Al-
2
2 beit this non-orthogonality, DeRe-VM still exhibits favorable
r̂. The NMSE is defined as E R − R̂ / ∥R∥F . Other estimation precision in terms of R’s NMSE performance.
F
15
0 -5
-5
-10
-10
-15
-15
Angular-domain representation
NMSE (dB)
NMSE (dB)
-20 15.48 dB -20 Angular-domain representation
-25
-25
-30
-30
-35
-40 -35
-10 -5 0 5 10 15 20 16 24 32 40 48 56 64
SNR (dB)
(a) (a)
0 -10
-5
-15
-10
-15 NMSE (dB) -20
NMSE (dB)
-20
-25
-25
-30 4.03 dB
-30
-35
3.32 dB
Performance gap
-40 -35
-10 -5 0 5 10 15 20 16 24 32 40 48 56 64
SNR (dB)
(b) (b)
Fig. 9. NMSE performance versus SNR with different BS-user distances. Fig. 10. NMSE performance versus the pilot number P with different BS-
(a) The BS-user distance is taken from rl ∼ U (1 m, 20 m) for the near-field user distances. (a) The BS-user distance is taken from rl ∼ U (1 m, 20 m)
region. (b) The BS-user distance is taken from rl ∼ U (30 m, 50 m) for the for the near-field region. (b) The BS-user distance is taken from rl ∼
far-field region. U (30 m, 50 m) for the far-field region.
C. NMSE Performance of Ny = 129 and Nz = 65, the Rayleigh distance can

be determined to be 26 meters. As shown in Fig. 8, the
For comparison, the exploited benchmark schemes are
proposed DeRe-VM algorithm consistently reveals the best
listed as follows:
NMSE performance, whereas that of the scheme Without AIC
• P-SIGW [26]: The off-grid polar-domain simultaneous swings erratically as the distance increases from the near-
iterative gridless weighted (P-SIGW) algorithm exploits field to the far-field. The poor NMSE performance attained
different transform matrix constructed for the indepen- by the scheme Without AIC results from the ambiguity of
dent decomposition of angle and distance parameters. R imposed by the thoughtlessly stitching of the 3D AED
• DeRe-VBI [42]: In lieu of Markov prior in (41), DeRe- parameters, as the angular index misinterpretation frequently
VBI assumes the independent and identically distributed occurs. Furthermore, a NMSE performance leveling off is
(i.i.d.) sparse support vector priors. spotted with respect to the schemes incorporating distance
• Off-grid scheme [35]: The holographic MIMO channel estimation, e.g., DeRe-VM, P-SIGW, and DeRe-VBI, regard-
is treated in the angular domain using VBI approach. less of fluctuations in distance. In contrast, both on-grid
• On-grid scheme [45]: The holographic MIMO channel and off-grid schemes exhibit poor estimating performance,
is treated in the angular domain using a discrete Fourier particularly in the near-field region. This is primarily due
transform (DFT)-based on-grid method proposed in [45]. to the fact that distance-estimation schemes are capable of
• Without AIC: The angular index correction in Step 20 structuring distance grids in a sense that accommodates to
of Algorithm 1 is removed. various channel environments. More precisely, when doing
We begin the NMSE performance evaluation by investigat- gridding over the distance r, a uniform grid of discrete
ing their corresponding curves versus the BS-to-user distance points embraces the far-field case, i.e., nz = 0 in (25),
ranging from 5 m to 50 m. In accordance with the carrier allowing the far-field information to be retrieved from the
frequency of 30 GHz and the number of antenna elements r-based covariance matrix. Therefore, the proposed DeRe-
16
-5 10-2
Imperfections
-10
NMSE (dB)
-15
NMSE
-20
10-3
-25
-30
-35 0 100 200 300 400 500

2 3 4 5 6 7 8
Fig. 11. NMSE performance versus the number of NLoS paths (L − 1). Fig. 12. NMSE performance at different error ratios εF̄ .
VM reveals significant robustness in accommodating diverse significant paths. Interestingly, both cases demonstrate a slight
channel conditions at varying distances in the context of a decline in NMSE when P > 40, in contrast to a precipitous
holographic MIMO system. drop when P < 40, which implies a potential compromise
In Fig. 9, we investigate the impact of varying SNR on between the pilot number and NMSE, that is, favorable
the NMSE performance for different BS-to-user distance NMSE performance can be achieved given an appropriate P ,
configurations, where the 26-meter Rayleigh distance is taken while a large P may incur heavy pilot overhead for systems.
as the dividing line for the near-field and far-field regions. Additionally, the performance gap can still be spotted in
Firstly, as expected, an increase in SNR boosts the NMSE per- Fig. 10(b), e.g., 3.32 dB when P = 40, between DeRe-
formance and the proposed DeRe-VM algorithm is superior VM and On-grid schemes, owing to a more accurate Fresnel
to the benchmark schemes adopted, regardless of the distance approximation in (10) than its far-field counterpart.
configurations. This is made possible by the fact that DeRe- Fig. 11 demonstrates the variation of NMSE performance
VM offers the capability of capturing, via their independent with respect to the number of NLoS paths among different
decompositions, potential structured sparsities in relation to schemes. Since L = 1 represents the LoS path, the number
3D AED parameters in holographic MIMO channels. On the of NLoS paths is denoted by L − 1. As spotted in Fig. 11,
other hand, the scatterers in the physical environment can be the NMSE performance achieved by all the schemes being
characterized by Markov prior for the channel support αθ , considered exhibits a slight degradation with an increased
which also explains the reason why DeRe-VM is superior number of NLoS paths, albeit to different degrees. As the
to DeRe-VBI. Secondly, the NMSE performance achieved number of NLoS paths increases, an augmentation occurs in
by on-grid and off-grid schemes significantly enhances when the quantity of non-zero entries within each sparse vector. This
the examined settings shift to the far-field case. For instance, without a doubt increases the dimensionality of the solution
as SNR = 5 dB, the performance gap resulted from DeRe- space, rendering the CS-based algorithm more susceptible
VM and off-grid schemes falls from 15.48 dB in the near- to becoming ensnared in local optima in the pursuit of the
field case to 4.03 dB in the far-field case. Additionally, one optimal value. As a result, figuring out the LoS path becomes
interesting finding is that in the far-field configuration, higher more challenging. Fortunately, due to the “turbo” nature
SNR draws the NMSE performance achieved by Without AIC of DeRe-VM, with which the updates of parameters and
to a great level, in stark contrast to the pale efforts made by the posteriors mutually reinforce each other, it can be ensured
increased SNR for the marginal improvement of NMSE in the that the algorithm will eventually converge to a stationary
near-field context. This is because a high SNR can compensate point [46], [47].
for the erroneous estimation caused by the absence of the To evaluate the robustness of the proposed DeRe frame-
angular index correction when just estimating the 2D azimuth- work, Fig. 12 plots the NMSE curves versus the sample
elevation angle pair. number T . In conjunction with (38), we define εF̄ =
r
Fig. 10 plots the NMSE curves in relation to the pilot n
2
o n
2
o
length P for various distance cases, as in Fig. 9. With more E ∆F̄ /E F̄ + ∆F̄ as the error ratio to quan-
pilots inserted, holographic MIMO channels can be better tify the level of hardware imperfections induced by for
characterized, resulting in a diffing degree of improvement example, the limited resolution of phase shifters and am-
for NMSE with various schemes. To be specific, regarding plitude controllers. As it transpires, the NMSE performance
rl ∼ U (1 m, 20 m) showcased in Fig. 10(a), an increment in is expected to improve with an increase in the number of
P boost NMSE performance marginally in the near-field case samples. An eye-catching observation pertains to the existence
while considerably in the far-field case. This can be attributed of gaps in NMSE performances across various εF̄ in the
to the fact that on- and off-grid schemes based on the angular- presence of a small number of training samples, while the
domain representation suffer from a performance bottleneck NMSE performance achieved by all cases, as T grows larger,
induced by a severe energy spread and erroneous detection of tends to approach the best-case bound. This verifies that the
17
DeRe framework has the capability to learn the statistical [13] T. Gong, P. Gavriilidis, R. Ji, C. Huang, G. C. Alexandropoulos, L. Wei,
properties of the 3D AED parameters by harnessing the Z. Zhang, M. Debbah, H. V. Poor, and C. Yuen, “Holographic MIMO
communications: Theoretical foundations, enabling technologies, and
samples adequately despite the presence of hardware imper- future directions,” IEEE Commun. Surv. Tutor., to appear, 2023.
fections, showcasing the great robustness of the proposed [14] A. A. D’Amico and L. Sanguinetti, “Holographic MIMO communica-
DeRe framework. tions: What is the benefit of closely spaced antennas?” arXiv preprint:
2307.13467, 2023.
[15] T. L. Marzetta, “Super-directive antenna arrays: Fundamentals and new
VII. C ONCLUSION perspectives,” in 2019 53th Asilomar Conf. Signals, Sys., Comp., Pacific
In this work, we have investigated the channel estimation Grove, CA, USA, Nov. 2019, pp. 1–4.
[16] L. Han, H. Yin, M. Gao, and J. Xie, “A superdirective beamforming
for holographic MIMO systems in face of a few non-trivial approach with impedance coupling and field coupling for compact
challenges posed by the holographic MIMO nature. A so- antenna arrays,” arXiv preprint: 2302.08203, 2023.
phisticated DeRe framework is devised for deep decouplings [17] A. Pizzo, L. Sanguinetti, and T. L. Marzetta, “Fourier plane-wave
series expansion for holographic MIMO communications,” IEEE Trans.
of the 3D AED parameters, followed by an efficient DeRe- Wireless Commun., vol. 21, no. 9, pp. 6890–6905, Sep. 2022.
VM algorithm for resolving the resultant CS problems. We [18] A. Pizzo, T. L. Marzetta, and L. Sanguinetti, “Spatially-stationary model
evince that the implementation of the proposed DeRe frame- for holographic mimo small-scale fading,” IEEE J. Sel. Areas Commun.,
vol. 38, no. 9, pp. 1964–1979, Sep. 2020.
work facilitates sharp detection of the parameters for each [19] Ö. T. Demir, E. Björnson, and L. Sanguinetti, “Channel modeling and
dimension, but it may adversely results in misinterpretation of channel estimation for holographic massive MIMO with planar arrays,”
the angular index if these quantities obtained are not properly IEEE Wireless Commun. Lett., vol. 11, no. 5, pp. 997–1001, May 2022.
structured together. As a remedy, angular index correction [20] L. Wei et al., “Multi-user holographic MIMO surfaces: Channel mod-
eling and spectral efficiency analysis,” IEEE J. Sel. Topics Signal
is carried out for the accurate estimation. Our simulation Process., vol. 16, no. 5, pp. 1112–1124, Aug. 2022.
results substantiate that the proposed DeRe-VM algorithm [21] L. Wei, C. Huang, G. C. Alexandropoulos, Z. Yang, J. Yang, W. E. I.
exhibits great robustness to the distance fluctuations in the Sha, Z. Zhang, M. Debbah, and C. Yuen, “Tri-polarized holographic
MIMO surfaces for near-field communications: Channel modeling and
holographic MIMO system of interest, irrespective of whether precoding design,” IEEE Trans. Wireless Commun., to appear, 2023.
the scenario involves near-field or far-field. In contrast to [22] M. Ghermezcheshmeh, V. Jamali, H. Gacanin, and N. Zlatanov, “Para-
the conventional far-field solutions based on the angular metric channel estimation for LoS dominated holographic massive
MIMO systems,” arXiv preprint: 2112.02874, 2022.
domain, the proposed DeRe-VM paradigm still shows slightly [23] R. Deng, B. Di, H. Zhang, and L. Song, “HDMA: Holographic-pattern
improved NMSE performance owing to its employment of division multiple access,” IEEE J. Sel. Areas Commun., vol. 40, no. 4,
a more accurate Fresnel approximation, without noticeable pp. 1317–1332, Apr. 2022.
[24] R. Deng, B. Di, H. Zhang, Y. Tan, and L. Song, “Reconfig-
increase in sparsifying complexity. urable holographic surface-enabled multi-user wireless communica-
tions: Amplitude-controlled holographic beamforming,” IEEE Trans.
R EFERENCES Wireless Commun., vol. 21, no. 8, pp. 6003–6017, Aug. 2022.
[25] H. Zhang, H. Zhang, B. Di, M. D. Renzo, Z. Han, H. V. Poor, and
[1] C.-X. Wang et al., “On the road to 6G: Visions, requirements, key
L. Song, “Holographic integrated sensing and communication,” IEEE
technologies, and testbeds,” IEEE Commun. Surv. Tutor., vol. 25, no. 2,
J. Sel. Areas Commun., vol. 40, no. 7, pp. 2114–2130, Jul. 2022.
pp. 905–974, Second-quarter, 2023.
[26] M. Cui and L. Dai, “Channel estimation for extremely large-scale
[2] J. M. Jornet, E. W. Knightly, and D. M. Mittleman, “Wireless commu-
MIMO: Far-field or near-field?” IEEE Trans. Commun., vol. 70, no. 4,
nications sensing and security above 100 GHz,” Nat. Commun., vol. 14,
pp. 2663–2677, Apr. 2022.
no. 1, p. 841, Feb. 2023.
[3] M. Matthaiou, O. Yurduseven, H. Q. Ngo, D. Morales-Jimenez, S. L. [27] Z. Wu and L. Dai, “Multiple access for near-field communications:
Cotton, and V. F. Fusco, “The road to 6G: Ten physical layer challenges SDMA or LDMA?” IEEE J. Sel. Areas Commun., vol. 41, no. 6, pp.
for communications engineers,” IEEE Commun. Mag., vol. 59, no. 1, 1918–1935, Jun. 2023.
pp. 64–69, Jan. 2021. [28] X. Zhang, Z. Wang, H. Zhang, and L. Yang, “Near-field channel
[4] “Technical specification group services and system aspects; Release 17 estimation for extremely large-scale array communications: A model-
description; summary of Rel-17 work items,” 3rd Generation Partner- based deep learning approach,” IEEE Commun. Lett., vol. 27, no. 4, pp.
ship Project, Tech. Rep. 3GPP TR 21.917 V0.7.0, Jun. 2022. 1155–1159, Apr. 2023.
[5] A. R. Flores and R. C. de Lamare, “Robust and adaptive power [29] X. Zhang, H. Zhang, and Y. C. Eldar, “Near-field sparse channel
allocation techniques for rate splitting based MU-MIMO systems,” representation and estimation in 6G wireless communications,” arXiv
IEEE Trans. Commun., vol. 70, no. 7, pp. 4656–4670, Jul. 2022. preprint: 2212.13527, 2022.
[6] A. M. Elbir, K. V. Mishra, M. R. B. Shankar, and S. Chatzinotas, [30] H. Lu and Y. Zeng, “Communicating with extremely large-scale ar-
“The rise of intelligent reflecting surfaces in integrated sensing and ray/surface: Unified modeling and performance analysis,” IEEE Trans.
communications paradigms,” IEEE Netw., to appear, 2022. Wireless Commun., vol. 21, no. 6, pp. 4039–4053, Jun. 2022.
[7] C.-Y. Yeh and E. W. Knightly, “Eavesdropping in massive MIMO: New [31] E. Björnson and L. Sanguinetti, “Power scaling laws and near-field
vulnerabilities and countermeasures,” IEEE Trans. Wireless Commun., behaviors of massive MIMO and intelligent reflecting surfaces,” IEEE
vol. 20, no. 10, pp. 6536–6550, Oct. 2021. Open J. Commun. Soc., vol. 1, pp. 1306–1324, Sep. 2020.
[8] E. Shtaiwi, H. Zhang, S. Vishwanath, M. Youssef, A. Abdelhadi, and [32] S. Hu, F. Rusek, and O. Edfors, “Beyond massive MIMO: The potential
Z. Han, “Channel estimation approach for RIS assisted MIMO systems,” of positioning with large intelligent surfaces,” IEEE Trans. Signal
IEEE Trans. on Cogn. Commun. Netw., vol. 7, no. 2, pp. 452–465, Jun. Process., vol. 66, no. 7, pp. 1761–1774, Apr. 2018.
2021. [33] D. Dardari, “Communicating with large intelligent surfaces: Fundamen-
[9] Y. Chen, Y. Wang, Z. Wang, and P. Zhang, “Robust beamforming for tal limits and models,” IEEE J. Sel. Areas Commun., vol. 38, no. 11,
active reconfigurable intelligent omni-surface in vehicular communica- pp. 2526–2537, Nov. 2020.
tions,” IEEE J. Sel. Areas Commun., vol. 40, no. 10, pp. 3086–3103, [34] J. Sherman, “Properties of focused apertures in the fresnel region,” IRE
Oct. 2022. Trans. Anntenas Propag., vol. 10, no. 4, pp. 399–408, Jul. 1962.
[10] G. Kramer, “Information rates for channels with fading, side informa- [35] Y. Chen, Y. Wang, and Z. Wang, “Reconfigurable intelligent surface
tion and adaptive codewords,” arXiv preprint: 2302.02903, 2023. aided high-mobility millimeter wave communications with dynamic
[11] C. Huang, S. Hu, G. C. Alexandropoulos, A. Zappone, C. Yuen, dual-structured sparsity,” IEEE Trans. Wireless Commun., vol. 22, no. 7,
R. Zhang, M. D. Renzo, and M. Debbah, “Holographic MIMO surfaces pp. 4580–4599, Jul. 2023.
for 6G wireless networks: Opportunities, challenges, and trends,” IEEE [36] E. Björnson, J. Hoydis, M. Kountouris, and M. Debbah, “Massive
Wireless Commun., vol. 27, no. 5, pp. 118–125, Oct. 2020. MIMO systems with non-ideal hardware: Energy efficiency, estimation,
[12] R. Deng, B. Di, H. Zhang, D. Niyato, Z. Han, H. V. Poor, and and capacity limits,” IEEE Trans. Inf. Theory, vol. 60, no. 11, pp. 7112–
L. Song, “Reconfigurable holographic surfaces for future wireless 7139, Nov. 2014.
communications,” IEEE Wireless Commun., vol. 28, no. 6, pp. 126– [37] H. Zhang, N. Shlezinger, F. Guidi, D. Dardari, M. F. Imani, and Y. C. El-
131, Dec. 2021. dar, “Beam focusing for near-field multiuser MIMO communications,”
18
IEEE Trans. Wireless Commun., vol. 21, no. 9, pp. 7476–7490, Sep. Zhaocheng Wang (IEEE Fellow) received the B.S.,
2022. M.S., and Ph.D. degrees from Tsinghua University,
[38] D. Baron, S. Sarvotham, and R. G. Baraniuk, “Bayesian compressive in 1991, 1993, and 1996, respectively.
sensing via belief propagation,” IEEE Trans. Signal Process., vol. 58, From 1996 to 1997, he was a Post-Doctoral Fel-
no. 1, pp. 269–280, Jan. 2010. low with Nanyang Technological University, Singa-
[39] F. Kschischang, B. Frey, and H.-A. Loeliger, “Factor graphs and the pore. From 1997 to 1999, he was a Research Engi-
sum-product algorithm,” IEEE Trans. Inf. Theory, vol. 47, no. 2, pp. neer/ Senior Engineer with the OKI Techno Centre
498–519, Feb. 2001. (Singapore) Pte. Ltd., Singapore. From 1999 to
[40] L. Chen, A. Liu, and X. Yuan, “Structured turbo compressed sensing 2009, he was a Senior Engineer/Principal Engineer
for massive MIMO channel estimation using a markov prior,” IEEE with Sony Deutschland GmbH, Germany. Since
Trans. Veh. Technol., vol. 67, no. 5, pp. 4635–4639, May 2018. 2009, he has been a Professor with the Department
[41] S. Som and P. Schniter, “Compressive imaging using approximate of Electronic Engineering, Tsinghua University, where he is currently the
message passing and a Markov-tree prior,” IEEE Trans. Signal Process., Director of the Broadband Communication Key Laboratory, Beijing National
vol. 60, no. 7, pp. 3439–3448, Jul. 2012. Research Center for Information Science and Technology (BNRist). He has
[42] D. G. Tzikas, A. C. Likas, and N. P. Galatsanos, “The variational authored or coauthored two books, which have been selected by IEEE Series
approximation for bayesian inference,” IEEE Signal Process. Mag., on Digital and Mobile Communication and published by Wiley-IEEE Press.
vol. 25, no. 6, pp. 131–146, Nov. 2008. He has authored/coauthored more than 200 peer-reviewed journal articles.
[43] W. Wang and W. Zhang, “Joint beam training and positioning for in- He holds 60 U.S./EU granted patents (23 of them as the first inventor).
telligent reflecting surfaces assisted millimeter wave communications,” His research interests include wireless communications, millimeter wave
IEEE Trans. Wireless Commun., vol. 20, no. 10, pp. 6282–6297, Oct. communications, and optical wireless communications. He is a fellow of the
2021. Institution of Engineering and Technology. He was a recipient of the ICC2013
[44] Y. Li, C. Tao, G. Seco-Granados, A. Mezghani, A. L. Swindlehurst, Best Paper Award, the OECC2015 Best Student Paper Award, the 2016 IEEE
and L. Liu, “Channel estimation and performance analysis of one-bit Scott Helt Memorial Award, the 2016 IET Premium Award, the 2016 National
massive MIMO systems,” IEEE Trans. Signal Process., vol. 65, no. 15, Award for Science and Technology Progress (First Prize), the ICC2017 Best
pp. 4075–4089, Aug. 2017. Paper Award, the 2018 IEEE ComSoc Asia-Pacific Outstanding Paper Award,
[45] P. Wang, J. Fang, H. Duan, and H. Li, “Compressed channel estimation and the 2020 IEEE ComSoc Leonard G. Abraham Prize. He was an Associate
for intelligent reflecting surface-assisted millimeter wave systems,” Editor of IEEE T RANSACTIONS ON W IRELESS C OMMUNICATIONS from
IEEE Signal Process. Lett., vol. 27, pp. 905–909, May 2020. 2011 to 2015 and IEEE C OMMUNICATIONS L ETTERS from 2013 to 2016.
[46] A. Liu, L. Lian, V. Lau, G. Liu, and M.-J. Zhao, “Cloud-assisted He is currently an Associate Editor of IEEE T RANSACTIONS ON C OM -
cooperative localization for vehicle platoons: A turbo approach,” IEEE MUNICATIONS , IEEE S YSTEMS J OURNAL , and IEEE O PEN J OURNAL OF
Trans. Signal Process., vol. 68, pp. 605–620, Jan. 2020. V EHICULAR T ECHNOLOGY.
[47] Y. Chen, Y. Wang, X. Guo, Z. Han, and P. Zhang, “Location tracking
for reconfigurable intelligent surfaces aided vehicle platoons: Diverse
sparsities inspired approaches,” IEEE J. Sel. Areas Commun., vol. 41,
no. 8, pp. 2476–2496, Aug. 2023.
Yuanbin Chen received the B.S. degree in com-

munications engineering from the Beijing Jiaotong Zhu Han (S’01–M’04-SM’09-F’14) received the
University, Beijing, China, in 2019. He is currently B.S. degree in electronic engineering from Tsinghua
pursuing the Ph.D. degree in information and com- University, in 1997, and the M.S. and Ph.D. degrees
munication systems with the State Key Laboratory in electrical and computer engineering from the
of Networking and Switching Technology, Beijing University of Maryland, College Park, in 1999 and
University of Posts and Telecommunications. His 2003, respectively.
current research interests are in the area of holo- From 2000 to 2002, he was an R&D Engineer
graphic MIMO, reconfigurable intelligent surface of JDSU, Germantown, Maryland. From 2003 to
(RIS), and radio resource management (RRM) in 2006, he was a Research Associate at the Univer-
future wireless networks. He was the recipient of sity of Maryland. From 2006 to 2008, he was an
the National Scholarship in 2020 and 2022. assistant professor at Boise State University, Idaho.
Currently, he is a John and Rebecca Moores Professor in the Electrical
and Computer Engineering Department as well as in the Computer Science
Department at the University of Houston, Texas. Dr. Han’s main research
targets on the novel game-theory related concepts critical to enabling efficient
Ying Wang (IEEE Member) received the Ph.D. and distributive use of wireless networks with limited resources. His other
degree in circuits and systems from the Beijing Uni- research interests include wireless resource allocation and management,
versity of Posts and Telecommunications (BUPT), wireless communications and networking, quantum computing, data science,
Beijing, China, in 2003. In 2004, she was invited smart grid, carbon neutralization, security and privacy. Dr. Han received an
to work as a Visiting Researcher with the Com- NSF Career Award in 2010, the Fred W. Ellersick Prize of the IEEE Com-
munications Research Laboratory (renamed NiCT munication Society in 2011, the EURASIP Best Paper Award for the Journal
from 2004), Yokosuka, Japan. She was a Research on Advances in Signal Processing in 2015, IEEE Leonard G. Abraham Prize
Associate with the University of Hong Kong, Hong in the field of Communications Systems (best paper award in IEEE JSAC)
Kong, in 2005. She is currently a Professor with in 2016, IEEE Vehicular Technology Society 2022 Best Land Transportation
BUPT and the Director of the Radio Resource Paper Award, and several best paper awards in IEEE conferences. Dr. Han
Management Laboratory, Wireless Technology In- was an IEEE Communications Society Distinguished Lecturer from 2015 to
novation Institute, BUPT. Her research interests are in the area of the 2018 and ACM Distinguished Speaker from 2022 to 2025, AAAS fellow
cooperative and cognitive systems, radio resource management, and mobility since 2019, and ACM distinguished Member since 2019. Dr. Han is a 1%
management in 5G systems. She is active in standardization activities of highly cited researcher since 2017 according to Web of Science. Dr. Han is
3GPP and ITU. She took part in performance evaluation work of the Chinese also the winner of the 2021 IEEE Kiyo Tomiyasu Award (an IEEE Field
Evaluation Group, as a Representative of BUPT. She was a recipient of Award), for outstanding early to mid-career contributions to technologies
first prizes of the Scientific and Technological Progress Award by the China holding the promise of innovative applications, with the following citation:
Institute of Communications in 2006 and 2009, respectively, and a second “for contributions to game theory and distributed management of autonomous
prize of the National Scientific and Technological Progress Award in 2008. communication networks.”
She was also selected in the New Star Program of Beijing Science and
Technology Committee and the New Century Excellent Talents in University,
Ministry of Education, in 2007 and 2009, respectively. She has authored over
100 papers in international journals and conferences proceedings.
View publication stats

arxiv_2311.15158

Uploaded by

Copyright:

Available Formats

arxiv_2311.15158

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

arxiv_2311.15158

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Angular-Distance Based Channel Estimation for Holographic MIMO

Preprint in IEEE Journal on Selected Areas in Communications · November 2023

Yuanbin Chen Ying Wang

SEE PROFILE SEE PROFILE

The user has requested enhancement of the downloaded file.

Angular-Distance Based Channel Estimation for

which have been commercially rolled out and approved for

Fig. 1. Schematic diagram of holographic MIMO transmission and UPA.

Most channel estimation schemes, being based on the

z Construct covariance Construct θ(φ)-based covariance matrix

vny ,nz + vny ,- nz = wnjy ,nz

Fig. 3. Schematic illustration of parametric decomposition.

Recall that the distance  difference takes

sin(ζ̃ny,q +∆ζny ) 2) Gridding Complexity: The use of the DeRe framework

2) Layer II: Probability Model for Precision Vector ρθ : A

Gamma prior distribution is imposed on the precision vector y , nz

p ρθ |αθ = p ρθnz |αnθ z = p ρθny ,nz |αnθ z

Spasrse vector uT Precision vector ȡT Support vector ĮT

Ny Ñz h  iαθn Fig. 5. Factor graph of the joint distribution p yθ , v θ ; ∆θ .

B. DeRe-VM Algorithm intends to calculate the approximate conditional marginal

in the rth iteration in place of the objective in problem (46), νnout

DeRe-VM-E Step DeRe-VM-M Step

Fig. 6. Schematic diagram of the proposed DeRe-VM algorithm by taking θ as an example.

Algorithm 1 Parametric Decomposition and Compressed

16: Update ∆θ (r+1) according to (50). Ny

R r in place of the superscript of associated variables, e.g., yφ

Following the overall flow of DDS-VBIMP summarized

updates of ψ uθ , ψ ρθ , and ψ αθ in Step 9 using

utilization of the matrix inversion lemma on (56) in order to

concerning that it typically holds that Ny , Ñz ≫ NRF . This

proportionally with the number of antenna elements raised

be taken with a grain of salt: i) P-SIGW fails to capture

parameters, and ii) the computational complexity of P-SIGW -15

increases quadratically with the number of antenna elements,

VI. S IMULATION R ESULTS -35

In this section, we conduct performance evaluation of

-20 15.48 dB -20 Angular-domain representation

C. NMSE Performance of Ny = 129 and Nz = 65, the Rayleigh distance can

-35 0 100 200 300 400 500

Yuanbin Chen received the B.S. degree in com-

View publication stats

You might also like

Recall that the distance difference takes

Ny Ñz h iαθn Fig. 5. Factor graph of the joint distribution p yθ , v θ ; ∆θ .