Professional Documents
Culture Documents
Near-Field Sparse Channel Estimation For Extremely Large-Scale RIS-Aided Wireless Communications
Near-Field Sparse Channel Estimation For Extremely Large-Scale RIS-Aided Wireless Communications
4
Université Paris-Saclay, CNRS, CentraleSupélec, Laboratoire des Signaux et Systèmes,
3 Rue Joliot-Curie, 91192 Gif-sur-Yvette, France
5
University of Southampton, Southampton SO17 1BJ, U.K.
Abstract—A significant increase in the number of reconfig- extremely large-scale RIS (XL-RIS). Nonetheless, further chal-
urable intelligent surface (RIS) elements results in a spherical lenges arise. Firstly, a significant increase in the number of RIS
wavefront in the near field of extremely large-scale RIS (XL- elements exacerbates the estimation of cascaded channels, since
RIS). Although the channel matrix of the cascaded two-hop link
may become sparse in the polar-domain representation, their the number of cascaded channel coefficients to be determined
accurate estimation of these polar-domain parameters cannot will be given by the number of RIS elements. Secondly, the
be readily guaranteed. To tackle this challenge, we exploit the resultant Rayleigh distance expansion renders the commonly
sparsity inherent in the cascaded channel. To elaborate, we first used planar-wavefront assumption invalid due to the spherical
estimate the significant path-angles and distances corresponding wavefront shape experienced in the near-field region.
to the common paths between the BS and the XL-RIS. Then,
the individual path parameters associated with different users are To mitigate the excessive complexity and overhead of the
recovered. This results in a two-stage channel estimation scheme, RIS channel estimation, significant efforts have been devoted
in which distinct learning-based networks are used for channel to exploiting the potential sparsities of the cascaded channel
training at each stage. More explicitly, in stage I, a denoising to achieve pilot overhead savings [4]–[9]. Specifically, due to
convolutional neural network (DnCNN) is employed for treating the limited number of significant paths induced by the sparsity
the grid mismatches as noise to determine the true grid index
of the angles and distances. By contrast, an iterative shrinkage of scatterers in the environment, the cascaded channel exhibits
thresholding algorithm (ISTA) based network is proposed for sparsity in the angular domain. These sparse channels may be
adaptively adjusting the column coherence of the dictionary efficiently estimated by employing compressive sensing (CS)
matrix in stage II. Finally, our simulation results demonstrate based techniques, such as the popular orthogonal matching
that the proposed two-stage learning-based channel estimation pursuit (OMP) [4], [5] and the variational Bayesian Inference
outperforms the state-of-the-art benchmarks.
Index Terms—XL-RIS, channel estimation, near-field, power (VBI) algorithms [6], [7]. However, these investigations are
leakage, power drift. predicated on the assumption of a planar wavefront in the
far-field region, since the Rayleigh distance is typically a few
I. I NTRODUCTION meters and may be negligible in practice. However, in the XL-
Recent academic and industrial interest in reconfigurable RIS regime relying on an extremely large number of elements,
intelligent surfaces (RISs) has been fueled by its potential to the near-field effect can no longer be disregarded due to the
improve the spectral and/or energy efficiency of communi- resultant energy spread effect. More explicitly, the transmit
cation systems [1], [2]. However, they also pose challenges. power may spread across multiple directions, instead of being
For example, when the communication link between the base concentrated in the direction of the desired destination [10]. To
station (BS) and the user is adequate, their performance gain tackle this issue, a polar-domain representation that simultane-
becomes negligible. This is primarily due to the double fading ously accounts for both the angle and distance information has
experienced by the transmitted signal in the two-hop RIS-aided been proposed in [10].
system, which is determined by the multiplicative cascaded However, there are two major issues if polar-domain sparsity
channel gain model [3]. To mitigate the deleterious effects is assumed for the cascaded channel in the XL-RIS regime.
of multiplicative twin-hop channel gain, the number of RIS Firstly, in the presence of the polar-domain representation,
elements may be increased, thus resulting in the concept of the the nonzero elements of the cascaded channel matrix (whose
physical meanings correspond to the angles and distances
* Zixing Tang and Yuanbin Chen contributed equally to this work. associated with significant paths) may not fall precisely on
the uniform grids prescribed by the dictionary matrix, leading
y -axis
to grid mismatches. As such, the power of nonzero elements M -ULA XL-RIS
in the cascaded channel matrix may leak to other elements, x -axis
hence resulting in a power leakage effect. Secondly, due to 2
1
the inherent column coherence of the dictionary matrix, the 2 1 d2
d 2( m )
single power peak corresponding to the one and only significant 1 s1
s2 d1( m )
path may drift to different grid indices, which is referred to d1
scatterer
as the “power drift” phenomenon. Therefore, given this pair of 2
challenges, a tailor-made channel estimation scheme is required r2 …
scatterer
for the XL-RIS regime .
N -ULA BS
Inspired by the dual-structured sparsity of the cascaded
channel presented in [6], we propose a two-stage learning-based blocked direct link
channel estimation approach, with the aim of embracing the UE
challenges inherent in near-field XL-RIS aided wireless com-
munications. Specifically, upon assuming polar-domain sparsity Fig. 1. XL-RIS system in the near field.
for the near-field cascaded channel matrix, the resultant sparse
cascaded channel matrix shares the same nonzero rows, but
it has different nonzero columns for each user. To this end, satisfies E ssH = 1. Specifically, the signal received at the
the goal of our two-stage channel estimation is to estimate the BS in slot t, 1 ≤ t ≤ τ is given by
significant angles and distances corresponding to the common √
paths of all users (contained in the nonzero rows) spanning from Y(t) = Hdiag(et )h ps + N(t), (1)
the BS to the XL-RIS in its stage I. By contrast, in stage II, where p represents the transmit power of the user, while
we recover the distances and angles at the RIS, as well as the N(t) ∈ CM ×K is the additive white Gaussian noise (AWGN)
complex gains associated with each user. More explicitly, in following the distribution vec(N(t)) ∼ CN (0, σ 2 I). Next, we
stage I, a denoising convolutional neural network (DnCNN) is right-multiply the received signal by the pilot sequence, yielding
employed, which treats grid mismatches as noise, when aiming √
for retrieving the true grid index of the angles and distances Y(t)sH = pHdiag(et )h + N(t)sH
√ (2)
corresponding to the common paths. This effectively mitigates = pHdiag(h)et + N(t)sH .
the power leakage effect induced by the grid mismatch. For
Furthermore, let us denote the cascaded channel by G =
stage II, we propose a regime termed as the iterative shrinkage
Hdiag(h) ∈ CN ×M . Upon stacking τ signals, the overall
thresholding algorithm (ISTA) based network for adaptively
received signal matrix Y = [Y(1)sH , . . . , Y(τ )sH ] ∈ CN ×τ
adjusting the column coherence of the dictionary matrix to
is expressed as
mitigate the power drift effect. We will also demonstrate that √
Y = pGE + N, (3)
the proposed two-stage channel estimation scheme has the
potential to significantly reduce the gridding complexity and the where E = [e1 , . . . , eτ ] and N = [N(1)sH , . . . , N(τ )sH ].
estimation overhead. Finally, our simulation results demonstrate
that the proposed approach achieves accurate estimation of the B. Near-field Cascaded Channel Sparsity Model
near-field cascaded channel in XL-RIS aided wireless systems. It is assumed that there are LBR propagation paths between
II. S YSTEM AND CHANNEL MODEL the XL-RIS and the BS, as well as LRU paths between the user
and the XL-RIS. The near-field channel between the XL-RIS
A. System Model and the BS, as well as the one between the user and the XL-RIS
We consider the narrow-band XL-RIS uplink communication are modeled as [11]
system, where K single-antenna users are served by an N -
antenna uniform linear array (ULA) assisted BS. The XL-RIS H = AN (θ, r)diag(ρ)AH
M (ϕ, s), (4)
is equipped with M passive reflecting elements in the form h = AM (φ, d)β, (5)
of a ULA. As illustrated in Fig. 1, the direct link is blocked
by obstacles. Therefore, we only focus on the uplink channel where AN (θ, r) ∈ CN ×LBR , AM (ϕ, s) ∈ CM ×LBR and
estimation of the XL-RIS assisted link. AM (φ, d) ∈ CM ×LRU represent the near-field steering ma-
Let us consider an arbitrary user and assume that a coherence trices having angles of arrival/departure (AoAs/AoDs) and dis-
block contains T time slots. Let H ∈ CN ×M denote the tances. Furthermore, ρ ∈ CLBR ×1 and β ∈ CLRU ×1 represent
near-field channel spanning from the XL-RIS to the BS and the complex gains of the paths between the XL-RIS and the
h ∈ CM ×1 denote the near-field channel impinging from BS, as well as those between the user and the XL-RIS. As
the user upon the XL-RIS. Moreover, denote the phase shift illustrated in Fig. 1, θ denotes the AoAs at the BS, while ϕ
vector in time slot t by et ∈ CM ×1 . In the first τ slots, the and φ are the AoDs and AoAs at the XL-RIS, respectively.
user transmits an orthogonal pilot sequence s ∈ C1×K , which Still referring to (4), r represents the distances between the BS
and the scatterers (or XL-RIS), while s and d represent the 8
10-3
6
AN (θ, r) = [aN (θ1 , r1 ), . . . , aN (θLBR , rLBR )] , (6)
5
Power
4
Power leakage
1 h (m)
iT
aM (φl , dl ) = √ . . . , e−jk(dl −dl ) , . . . , (9)
M Fig. 2. An example of the row sparsity of YP olar , where LBR = 1 and
N = 64.
(m) m2 δ 2 cos2 φ
dl ≈ dl − mδ sin φ + , (10)
2dl III. T WO - STAGE CHANNEL ESTIMATION
(m)
where dl represents the distance between the mth reflecting In this section, we aim for reducing the multiplicative grid-
(0)
element and the lth scatterer, and dl = dl . δ is the reflecting ding complexity order of O(NG MG ), to its additive coun-
2πfc
element spacing, while k = c denotes the wavenumber at terpart, namely to O(NG + MG ), by conceiving a two-stage
the central carrier f c . Moreover, m satisfies − M −1
≤ m ≤ procedure. This beneficially reduces the estimation complexity.
2
M −1
, m ∈ Z. Given that all users share the common channel spanning from
2
According to [10], the near-field steering matrices exhibit the XL-RIS to the BS, we decompose the common parameters
beneficial sparsities in the presence of the polar-domain repre- at the BS from the user-specific parameters at the XL-RIS.
sentation. Let us denote the overcomplete dictionary matrices Firstly, both the arrival steering matrix corresponding to the
in the polar domain by FN ∈ CN ×NG and FM ∈ CM ×MG , common AoAs θ and the distances r at the BS are estimated
respectively, with NG and MG indicating the number of grids. by exploiting the row sparsity of the received signal. Next,
The steering matrices in (6)-(8) can be recast as AN (θ, r) = given the knowledge of the estimated arrival steering matrix,
FN XN , AM (ϕ, s) = FM XBR RU the estimation of both the cascaded matrix corresponding to
M , and AM (φ, d) = FM XM ,
respectively, where XN ∈ C NG ×LBR BR
, XM ∈ C MG ×LBR the individual AoDs/AoAs ϕ, φ and the distances s, d at the
and XRU ∈ C MG ×LRU
are sparse matrices. Explicitly, each XL-RIS as well as the cascaded complex gains is formulated as
M
column vector has exactly one nonzero element. Hence, the a CS problem exploiting the column sparsity. This formulates
cascaded channel G can be characterized by the sparse matrix the proposed two-stage channel estimation procedure. A pair
′
Λ ∈ CNG ×MG , yielding of learning-based networks are harnessed. On the one hand,
this approach exhibits robustness to grid-modeling mismatch
G = FN ΛF̃H
M, (11) and achieves faster inference speed than traditional iterative
algorithms. On the other hand, the learning-based approach
where Λ has LBR LRU nonzero elements corresponding to the performs well in encoding and decoding the signal, where
cascaded complex gains {ρl βp }. It can be formally shown that the substitution of neural networks for the sensing matrix is
′
there are only MG distinct column vectors in FM ◦ F∗M ∈ beneficial in order to make up for the shortcomings of the
2
M ×MG
C , where ◦ represents the transposed Khatri-Rao prod- dictionary matrix in the polar domain [13], [14].
′
uct. Therefore, F̃M ∈ CM ×MG is the combination of MG ′
distinct column vectors. Accordingly, the gridding complexity A. Stage I: Common BS Steering Matrix Estimation
NG ×τ
order for a single user is proportional to O(NG MG ), with their Based on (11), we observe that YP olar = FH NY ∈ C
typical values being NG = N and MG = M , respectively. Such is a row-sparse and column-full matrix. The indices of nonzero
a high-dimensional input renders channel estimation inefficient. rows are related to the arrival angles and distances at the BS.
Again, based on (11), the cascaded channel exhibits dual- Accordingly, the estimated arrival steering matrix ÂN (θ, r) can
structured sparsity of [6], which is a unique property in RIS- be constructed.
aided systems. This property inspires us to separately estimate However, the angles θ and distances r are continuous values,
the common parameters at the BS corresponding to the rows which may not fall on the discrete grids of the dictionary
and the parameters at the XL-RIS corresponding to the columns matrix FN . Fig. 2 is an example of the row-sparse structure
at a reduced gridding complexity. of YP olar and the y-axis represents the power of each row
FN Y
H N G
Stage I: Common BS Steering Matrix Estimation
Algorithm 1 Training for the DnCNN-based row recovery.
0
C r = [c r , , cr ]
N G LBR
residual noise 1: Initialize network parameters ϖ1 , learning rate lr,1 ;
Conv + ReLU
Conv + BN +
Conv + BN +
1
1 cr = F H
Y
2: for episodes = 1,...,Nep do
N :,i
ReLU
ReLU
2 i =1
…
Conv
3
4 3: Generate mini-batch training samples {(Cir , XiN )}Ii=1 1
;
5
6 64 filters 64 filters 64 filters 2 filters 4: Set L1 (ϖ1 ) = 0;
7 (3 3 2) (3 3 64) (3 3 64) (3 3 64)
XN 5: for i = 1, P. . . , I1do
A N = FN X N τ
cir = τ1 j=1 FH i
i
i i
6: N Y :,j , Cr = cr , . . . , cr , ;
1 H
p
AN Y Stage II: Individual Steering Matrix and Complex Gain Estimation 7: Update loss function L1 (ϖ 1 )2 = L1 (ϖ1 ) +
pl
1 i i i
2I1 ∥R(Cr ; ϖ1 ) − Cr − XN ∥F ;
H
1 p1
H H , M M G
,
4 p2 V
M
(0) F
M G M
F
6 H
p3 8: end for
of YP olar . It can be seen that the grid mismatch in distance where ∥ · ∥F denotes the Frobenius norm and D1 =
leads a power peak to spread, which is again referred to as {(Cir , XiN )}Ii=1
1
represents I1 clean training pairs. Algorithm 1
power leakage in this paper. To avoid the resultant performance summarizes the detailed offline training process of the proposed
degradation, we propose a sparse row recovery algorithm based DnCNN.
on the DnCNN [13]. Remark 1: It is worth noting that the power drift may
Briefly, the DnCNN excels at learning the residual noise also exist in the common angle and distance estimation during
and separating the noise from a noisy image by harnessing stage I. Due to the inherent column coherence1 of the dictionary
feed-forward convolutional neural networks (CNNs). The image matrix, the single power peak corresponding to the significant
denoising is treated as a plain learning problem, implying path may drift to different grid indices, which is referred
that for a propagation path component, both the other path to as the “power drift” effect. According to Eq. (15) (i.e.,
components and the AWGN are treated as residual noises rs = 1s Z△ (1 − θ2 )) in [10], the sampling points versus distance
contaminating the desired image (i.e., the path of interest). The are sufficient in the case of s = 1. Given that in conventional
process of estimating the steering vector for each propagation systems, the number of BS antennas N is not large, the power
path is equivalent to the associated noise removal. The DnCNN- drift can be neglected in the steering matrix associated with
based recovery scheme is capable of beneficially harnessing the the BS. Therefore, we only focus our attention on the grid
leaked power, while removing the noise. mismatch in stage I. However, the power drift effect becomes
As shown in Fig. 3, the DnCNN consists of Lc convolutional significant for the cascaded steering matrix at the XL-RIS, and
layers. The first convolutional layer is followed by a rectifier the amplitude of the power drift may degrade the estimation
linear unit (ReLU). The succeeding Lc − 2 convolutional layers accuracy of the complex gain in stage II, as it will be elaborated
are followed by batch normalization (BN) and a ReLU. Due to on later.
the feed-forward layer’s inability to process complex numbers,
we split the real and imaginary parts of the input and output B. Stage II: Individual Steering Matrix and Complex Gain
into two channels. Specifically, as only the row features have Estimation
to be emphasized, we sum the column vectors of YP olar and
define the column vector cr ∈ CNG ×1 as Based on the orthogonality of the steering vector of different
τ angles and distances, we have ÂH N AN (θ, r) ≈ ILBR . Hence,
1 X H
cr = F Y , (12) the received signal matrix Y can be projected onto the common
τ i=1 N :,i arrival steering matrix subspace as
where [·]:,i represents the ith column of the matrix. To ensure 1 1 H
that the input dimension matches the output dimension, we √ ÂH Y ≈ diag(ρ)AH
M (ϕ, s)diag(h)E + √ ÂN N. (14)
p N p
create LBR identical column vectors as the input matrix Cr =
[cr , . . . , cr ] ∈ CNG ×LBR . Let us denote the parameters of the This subsection aims for estimating the cascaded
DnCNN by ϖ1 and its output by R(Cr ; ϖ1 ) ∈ CNG ×LBR . matrix HRIS = [hRIS,1 , . . . , hRIS,LBR ]H =
Instead of learning a mapping from a noisy image directly to diag(ρ)AH (ϕ, s)diag(h) in (14) for each user, which
M
a denoised one, learning the residual noise is more beneficial. contains the individual angles, distances and complex gains.
Therefore, the loss function used for training is given by
I1 1 The column coherence can be calculated by multiplying any two columns of
1 X
∥R(Cir ; ϖ 1 ) − Cir − XiN ∥2F ,
L1 (ϖ1 ) = (13) the dictionary matrix. If the product is non-zero, we refer to these two columns
2I1 i=1 as being coherent.
10-3
Algorithm 2 Training for the ISTA-based network with an
6
adaptive dictionary matrix.
1: Initialize network parameters ϖ2 , learning rate lr,2 ;
5
Significant path power
2: for episodes = 1,...,Nep do
4
3: Generate mini-batch training pairs;
4: Set L2 (ϖ2 ) = 0;
5: for i = 1, . . . , I2 do
Power
3 Power drift
6: for l = 1, . . . , LBR do
7: pil = [ √1p ÂH i H
N Y ]:,l ;
2
8: Update loss function L2 (ϖ2 ) = L2 (ϖ2 ) +
1 ∥N (pil ;ϖ2 ,F̃)−hil ∥22
1 2I2 LBR ∥hil ∥22
;
9: end for
0 10: end for
0 100 200 300 400 500 600 700 800 900
11: Calculate the gradient ∇L2 (ϖ2 );
12: Update the network ϖ2 = ϖ2 + lr ∇L2 (ϖ2 );
Fig. 4. Sparsity of hRIS,l , where LRU = 1 and M = 128. 13: end for
14: return ϖ2 , F̃.
tth layer, λ(t) and κ(t) represent the “learnable” parameters 2 Layer=10
2
Layer=2
of the tth layer, while V ∈ CM ×τ represents the “learn- Layer=20
0
Layer=5
0
able”
(t) parameter shared by all layers. Let us define ϖ2 = Layer=30 Layer=8
1 XI2 L
X BR ∥N pil ; ϖ2 , F̃ − hil ∥22 -8
-10
L2 (ϖ2 , F̃) = ,
2I2 LBR i=1 ∥hil ∥22 -12
-10
l=1
(22) -12
-14
where N pil ; ϖ2 , F̃ is the output of the proposed network.
-16 -14
The network is trained offline as shown in Algorithm 2 100 200 200 400 1000 2000 3000 4000
and the trained network may then be used online at a low Episodes Episodes
(0) (0)
complexity. To elaborate, hl is initialized by hl = 0,
Fig. 5. The loss function of two trained networks with the learning rate of
while F̃ is initialized by F̃ = F̃M for faster convergence. 5 × 10−5 and 1 × 10−4 respectively.
Additionally, the learnable parameters avoid the manual ad-
justment of hyperparameters such as λ and κ. The adaptive IV. S IMULATION RESULTS
dictionary matrix reduces the coherence between the column
vectors during the training. In this section, simulation results are provided for the perfor-
Remark 2: As a result of the grid mismatch, the power mance evaluation of the proposed scheme. As for the simulation
leakage continues to be evident in the process of recovering parameters, we set fc = 30 GHz, N = 64, M = 128, LBR = 3,
the distances and angles at the RIS as well as the complex LRU = 3, lr,1 = 5 × 10−5 , lr,2 = 1 × 10−4 . Fig. 5 plots
gains associated with each user in stage II. Fortunately, this the loss function of the two networks for different numbers of
issue would be addressed by the learned dictionary matrix and layers. Upon increasing the number of training episodes, both
we focus only on the power drift in stage II. networks tend to converge. Increasing the number of layers is
conducive to the improvement of system performance, but the
C. Complexity Analysis benefits obtained gradually saturate when the number of layers
This subsection firstly analyzes the computational complex- is increased. It is observed in Fig. 5 that the performance of
ity of the proposed two-stage channel estimation scheme. the ISTA-based network is slightly degraded when the number
Due to the offline nature of the proposed network’s train- of layers increases to 8. This is caused by using insufficient
ing, we examine only the complexity of online predic- network training.
tion.
Specifically, the complexity order of the DnCNN is Furthermore, two benchmarks are adopted for comparison
PLc (l) (l) (l) (l−1) (l)
O NG LBR l=1 Dx Dy Dz c c [15], where NG with our proposed scheme. (i) OMP: The OMP algorithm is
and LBR correspond to the output tensor dimension, Lc is the employed to solve problem (3) based on the sparsity in (11),
(l) (l) (l)
number of the DnCNN layers, Dx , Dy and Dz represent the resulting in the gridding complexity order of O (NG MG ). (ii)
tensor dimension of the convolutional kernel in the lth layer, and DnCNN-OMP: The common AoAs and distances at the BS
finally, c(l) represents the number of convolution kernels in the are recovered by the proposed DnCNN in stage I, while the
lth layer. In stage II, the complexity order of the ISTA network individual angles and distances at the XL-RIS are estimated
is O (2τ M LBR Ld + M LBR Ld + 2M MG LBR Ld ), where Ld by OMP in stage II. (iii) DnCNN-ISTANET: Our proposed
denotes the number of layers. Since the computational com- near-field cascaded channel estimation for the XL-RIS regime.
plexity of the XL-RIS system is dependent on N and M , Fig. 6 quantifies the normalized mean n square2 error
o (NMSE)
∥Ĝ−G∥F
the proposed algorithm achieves reduced complexity com- performance (which is defined by E ∥G∥2F
) versus the
pared to the OMP algorithm with the complexity order of signal-to-noise ratio (SNR). Observe that the NMSE decreases,
O (τ N MG NG LBR LRU ) [4]. when the SNR increases. Although the performance of OMP is
Additionally, the gridding complexity order of the two stages better than that of DnCNN-OMP at low SNRs, the OMP scheme
is O(NG ) and O(MG ), respectively. The two-stage approach entails excessive estimation overhead for XL-RIS systems. As
reduces the gridding complexity order from O(NG MG ) to the SNR increases, the curve of the OMP scheme no longer
O(NG + MG ). Accordingly, the input layer dimension of the decays beyond 15 dB, which is attributed to the power leakage.
DnCNN is proportional to NG and the hidden layer dimension By contrast, the NMSE of the DnCNN-OMP scheme improves,
of the ISTA network is proportional to MG . Thus, we have an implying that the proposed DnCNN successfully mitigates the
efficient channel estimation process for the XL-RIS regime of power leakage. The higher the SNR, the better the DnCNN
interest. performs. Furthermore, the DnCNN-ISTANET scheme outper-
ACKNOWLEDGMENT
-1
DnCNN-ISTANET This work was supported in part by the Beijing Natural
-2 DnCNN-OMP Science Foundation under Grant 4222011. Q. Wu’s work
OMP
is supported by NSFC 62371289 and FDCT under Grant
-3
0119/2020/A3. The work of M. Di Renzo was supported
NMSE (dB)