Swindlehurst Channel Estimation With Reconfigurable Intelligent SurfacesA General Framework

Channel Estimation With
Reconfigurable Intelligent
Surfaces—A General
Framework
This article provides an overview of the identifiability of wireless channels for
reconfigurable intelligent surfaces by considering the available pilot data and the setup
of the surfaces during training.
By A. L EE S WINDLEHURST , Fellow IEEE, G UI Z HOU , Graduate Student Member IEEE,
R ANG L IU , Graduate Student Member IEEE, C UNHUA PAN , Member IEEE,
AND M ING L I , Senior Member IEEE
ABSTRACT | Optimally extracting the advantages available orthogonal frequency-division multiplexing (OFDM) scenarios,
from reconfigurable intelligent surfaces (RISs) in wireless availability of direct links between the users and BS, exploita-
communications systems requires estimation of the channels tion of prior information, as well as a number of other special
to and from the RIS. The process of determining these cases. We further conduct simulations of representative
channels is complicated when the RIS is composed of passive algorithms and comparisons of their performance for various
elements without any sensing or data processing capabilities, channel models using the relevant Cramér–Rao bounds.
and thus, the channels must be estimated indirectly by a
KEYWORDS | Channel estimation; direction of arrival (DOA)
noncolocated device, typically a controlling base station (BS).
estimation; intelligent reflecting surface; massive multi-input–
In this article, we examine channel estimation for passive
multi-output (MIMO); reconfigurable intelligent surface (RIS).
RIS-based systems from a fundamental viewpoint. We study
various possible channel models and the identifiability of the
models as a function of the available pilot data and behavior of I. I N T R O D U C T I O N
the RIS during training. In particular, we will consider situations There has been an explosion of interest in the use of
with and without line-of-sight propagation, single-antenna and reconfigurable metasurfaces for wireless communication
multi-antenna configurations for the users and BS, correlated systems in the last few years [1], [2]. Such reconfig-
and sparse channel models, single-carrier and wideband urable intelligent surfaces (RISs) provide tunable degrees
of freedom for adjusting the propagation characteristics
of problematic channels (e.g., sparse channels with fre-
Manuscript received 1 October 2021; revised 20 February 2022; accepted quent blockages) that make them a valuable resource for
17 April 2022. Date of publication 9 May 2022; date of current version
19 September 2022. This work was supported in part by the U.S. National maintaining and enhancing the quality of service (QoS) for
Science Foundation under Grant ECCS-2030029 and Grant CNS-2107182 and in users (UEs) in the network. However, most techniques that
part by the National Natural Science Foundation of China under Grant
61971088. (Corresponding author: A. Lee Swindlehurst.) exploit this ability require channel state information (CSI)
A. Lee Swindlehurst is with the Center for Pervasive Communications and to and from the elements of the RIS, which is a challenge
Computing, University of California at Irvine, Irvine, CA 92697 USA (e-mail:
swindle@uci.edu). since the number of RIS elements may be very large, and
Gui Zhou is with the School of Electronic Engineering and Computer Science, more importantly, they are often constructed only as pas-
Queen Mary University of London, London E1 4NS, U.K.
Rang Liu and Ming Li are with the School of Information and Communication
sive devices without active RF receivers or computational
Engineering, Dalian University of Technology, Dalian 116024, China. resources. Consequently, channel estimation for RIS-based
Cunhua Pan is with the National Mobile Communications Research Laboratory,
systems has been a subject of intense study.
Southeast University, Nanjing 211189, China.
When the RIS is composed of passive elements, the
Digital Object Identifier 10.1109/JPROC.2022.3170358 CSI must be estimated by devices—most often a base
0018-9219 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
1312 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022
Authorized licensed use limited to: Chung-ang Univ. Downloaded on January 06,2023 at 13:51:58 UTC from IEEE Xplore. Restrictions apply.
Swindlehurst et al.: Channel Estimation With Reconfigurable Intelligent Surfaces—General Framework
station (BS) or access point—that are not colocated with RIS with active RF and digital processing will be provided
the RIS. For example, training signals transmitted by the at the end.
UEs are received by the BS after reflection from the RIS After stating our general assumptions and notational
and possibly also over a direct path to the BS, and these conventions in Section II, we begin with a discussion of
known signals are exploited for CSI estimation. In order to CSI estimation for unstructured channels in Section III. We
estimate the RIS-based channel components, the reflection will first consider the narrowband single-user multi-input–
coefficients of the RIS must be varied as well, at least multi-output (MIMO) case and the corresponding least-
during a portion of the training period. However, even with squares (LS) and minimum mean-squared error (MMSE)
variable training from the UEs and RIS, the fact that the solutions, and then, we will examine extensions to the
impact of the RIS is only indirectly viewed in the data wideband and multiuser cases, as well as special cases
means that the complete structure of the channel is not involving a single-antenna BS and UEs and methods for
identifiable. In particular, while the cascaded or composite reducing the training overhead. Then, in Section IV, we
channel from the UEs to the BS can be determined, the focus on the estimation of geometric channel models and
individual components of the channel involving the RIS follow the same format of beginning with the narrowband
cannot. Fortunately, this is typically not a problem for single-user MIMO case and then considering the same
designing beamforming algorithms at the BS or optimizing generalizations and special cases as in Section III. We
the RIS reflection properties, since ultimately, the QoS only summarize some of the tradeoffs associated with chan-
depends on the composite channel. nel estimation for unstructured and geometric models in
A large amount of work on CSI estimation for passive Section V, and we include some representative numerical
RIS-based systems has been published recently. Initially, examples in Section VI to illustrate their relative perfor-
this work focused on estimating unstructured models, mance. Several additional topics will be briefly considered
where the channels are simply described using complex in Section VII, including the use of some active trans-
gains [3]–[14]. Such models are simple and lead to ceivers at the RIS, scenarios with more than one RIS,
straightforward algorithms, but the required training over- and machine learning approaches. Finally, some conclu-
head is very large and may render such approaches imprac- sions and suggestions for future research are offered in
tical. Methods for reducing the training overhead, for Section VIII.
example, based on grouping the RIS elements or exploiting
the common BS-RIS channel among the users, have been II. G E N E R A L A S S U M P T I O N S
proposed, but larger reductions are possible when the A N D N O TAT I O N
channels are sparse if parametric or geometric channel In this article, we primarily consider scenarios with a
models are used instead [15]–[29]. In these models, the single BS, a single RIS, and potentially multiple cochannel
channels are parameterized by the angles of arrival (AoAs), UEs. Various assumptions are made about the number of
angles of departure (AoDs), and complex gains of each antennas at the BS and UEs, and the number of UEs that
propagation path. As long as the number of multipaths is are active. We assume that the BS and UEs employ fully
not large, then the total number of parameters to be esti- digital rather than hybrid digital/analog architectures,
mated can be 1–2 orders of magnitude smaller than in the which to date is the most common case considered in the
unstructured case, and the amount of training can be corre- literature (see [20], [24], and [25] for approaches using
spondingly reduced. On the other hand, geometric models a hybrid RF network). We also assume a standard time-
require knowledge of the array calibration and RIS element division duplex protocol in which pilot symbols transmit-
responses, as well as the model order; errors in the model- ted by the UEs in the uplink are exploited by the BS to
ing assumptions will degrade some of this advantage. obtain a channel estimate, which is then used for down-
Many CSI estimation techniques have been proposed link beamforming or multiplexing. This assumes reciprocal
under a wide array of assumptions, from Rayleigh fading to uplink and downlink channels between the BS, RIS, and
line-of-sight (LoS) propagation, single-antenna-to-multi- users, which in turn typically requires some type of RF
antenna configurations, single carrier and multi carrier transceiver calibration and RIS elements whose behavior
modulation, scenarios with and without a direct link is independent of the angle of incidence. Pilots could also
between BS and UEs, and a variety of other special be embedded in the downlink for channel estimation at the
cases. In this article, we take a systematic approach to UEs, but this is similar to the uplink problem and thus is
the problem and organize the various approaches that not explicitly considered.
have been proposed—as well as some that have not— Matrices and vectors are denoted by boldface capital
under a common framework. In this way, the advantages and lowercase letters, respectively. In some cases, the kth
and disadvantages of different assumptions and solution column or row of a matrix A will be denoted by A:k or
approaches become clearer, and avenues for future work Ak: , respectively. The transpose, conjugate transpose, and
are elucidated. Consistent with most work that has been conjugate are denoted by (·)T , (·)H , and (·)∗ , respectively.
published to date, we will focus on channel estimation for The Kronecker, Khatri–Rao, and Hadamard products of two
scenarios involving a single passive RIS, although a brief matrices are indicated by C = A ⊗ B, C = A B, and
discussion about extensions to cases with multiple RIS or C = A B, respectively. An N × N identity matrix is
Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1313
represented as IN , and N × 1 vectors composed of all

ones or zeros are denoted by 1N and 0N , respectively.
A circular complex multivariate Gaussian distribution with
mean μ and covariance R is denoted by CN (μ, R). The
function vec(A) creates a vector from matrix A by stacking
its columns. The function a creates an integer from real
number a by truncating its decimal part, and a mod b is
the modulo operator that returns the integer remainder
of a/b. A diagonal matrix with elements of vector c on
the diagonal is indicated by diag(c), and a block diag-
onal matrix with block entries C1 , C2 , . . . is written by
blkdiag([C1 C2 , . . . , ]).
The reflective properties of an RIS with N elements is
described by the N × N diagonal matrix Φ = {diag(φ)}, Fig. 1. Generic scenario involving an M-antenna BS, an N-element
where φ = [β1 ejα1 · · · βN ejαN ]T . There are a number of RIS, and a K-element user.
practical issues associated with φ that are important for
RIS performance optimization, such as the dependence of
the gains β on the phases α, the fact that the phases are
typically discrete and frequency-dependent, and so on. For
the most part, these issues are not directly relevant to the the transmit power, and the signal-to-noise ratio (SNR)
generic channel estimation problem, which only requires is defined as P/σ 2 . The channels are all assumed to be
that φ is to be known and sufficiently controllable. How- block flat fading and constant over a coherence interval
ever, certain simplifying assumptions about φ are made sufficiently long to permit channel estimation and subse-
below for performance analyses or purposes of illustration. quent data transmission. On the other hand, the reflection
coefficients of the RIS, Φt , can vary synchronously with
the UE uplink transmission. Some prior work ignores the
III. E S T I M A T I O N O F U N S T R U C T U R E D direct channel component Hd , assuming that it is either not
CHANNEL MODELS present (e.g., due to a blockage) or that it was estimated
We begin with models where the channel between individ- in a previous step and its contribution has
ual network elements is described by a complex coefficient √ been removed
from the received data, i.e., yt −→ yt − P Hd xt .
in the case of a narrowband single-carrier signal or a It is important to note that not all of the components
complex-valued impulse response for wideband transmis- of the channel-related term Hd + HΦt GH are individually
sion. Such unstructured or nonparametric channel models identifiable. In particular, for any invertible N ×N diagonal
are appropriate for situations with rich multipath scat- matrix Λ, we have
tering (e.g., sub-6-GHz systems), where it is difficult to
describe the aggregate characteristics of the propagation H
environment. We initially focus on the narrowband single- HΦt GH = HΛΦt Λ−1 GH = H̃Φt G̃ (2)
user scenario and then examine cases involving wideband
signals or multiple users. As will become clear, the limiting where H̃ = HΛ and G̃ = G(Λ∗ )−1 . Thus, there is a scaling
factor with unstructured CSI estimation is the large train- ambiguity between each pair of the N columns {hk , gk } of
ing overhead that is required. Approaches for reducing the H and G that cannot be resolved using data obtained as
training overhead are discussed at the end of the section. in (1). Most methods for beamforming, precoding, or RIS
reflection optimization do not require this ambiguity to be
resolved, although as briefly discussed later, with certain
A. Narrowband Single-User MIMO
additional information, the individual channel components
The scenario assumed here is shown in Fig. 1, with an can be identified. For this reason, channel estimation in
M -antenna BS, an N -element RIS, and a single UE with the context of RIS-aided communication systems focuses
K antennas. The geometries of the RIS and the arrays at primarily on determination of the M K ×(N +1) composite
the BS and UE are arbitrary. If the UE transmits the K × 1 or cascaded channel Hc , defined using properties of the
vector xt at time t, the signal received at the BS is given by Khatri–Rao product
√
yt = P Hd + HΦt GH xt + nt (1) ∗ 1
vec(Hd + HΦt G ) = [hd
H
G H] ≡ Hc φ̃t (3)
φt
where Hd , H, G are, respectively, the channels between the
BS and UE, the BS and RIS, and the RIS and UE, and where hd = vec(Hd ), φt is a vector containing the diagonal
nt denotes additive noise or interference. Assuming that elements of Φt and φ̃t is defined implicitly in the last
2
E{xt xH
t } = IK and nt ∼ CN (0, σ IM ), P represents equality.
Given an estimate of Hc , if one wishes to obtain a noting that for any positive definite matrix B, we have
separate estimate of H and G, then a method is needed
to resolve the ambiguity in (2). For example, Hc could 1
B−1 ≥ (9)
be uniquely decomposed if one fixes the first row of H ii Bii
or G to be some known vector (e.g., all ones). Another
method to resolve the ambiguity will be discussed later with equality for all i only if B is diagonal. Thus, a good
in the context of geometric channel models. Equation (3) choice for Z would make (8) diagonal. Such a choice may
together with further use of the Kronecker product allows not be optimal in general, but a diagonal covariance matrix
us to rewrite (1) in a compact form also greatly simplifies the computation of ĥc,LS in (7).
The most common training approach that meets the
√ above design goals breaks the training interval T into T /K
yt = P xTt ⊗ IM Hc φ̃t + nt (4a)
√ T subblocks of length K , where T /K is assumed to be an
= P φ̃t ⊗ xTt ⊗ IM hc + nt (4b) integer. For each subblock, b = 1, . . . , T /K , φt = φ̄b is held
√ constant, while the pilots xt are chosen as an orthonormal
≡ P Zt hc + nt (4c)
sequence that repeats itself for each subblock. For example,
the subblock sequence for the UE is X = [x1 , . . . , xK ],
where hc = vec(Hc ) and the M × M K(N + 1) matrix Zt is where XXH = K IK , which is then repeated T /K times
implicitly defined.
The composite channel hc is clearly underdetermined
in (4), and thus, multiple pilot symbols must be transmit-
ted in order for it to be uniquely estimated. Combining the
xt pilots = X · · · X
X (10a)

repeated T /K times
data from T such pilots together, we have

φt pilots = φ̄1 · · · φ̄1 · · · φ̄ T · · · φ̄ T . (10b)
y1 Z1 K

K

.
√ .

√ repeated K times
y= .. = P .. hc + n ≡ P Zhc + n. (5) repeated K times
yT ZT
Using this approach, we have
Provided that T ≥ K(N + 1) and Z is full rank, there are T

∗

two common ways to estimate hc , as discussed next. ZH Z =
T
φ̃t φ̃t ⊗ x∗t xTt ⊗ IM (11a)

t=1
T
∗ T
1) Least Squares: The simplest approach for estimating = φ̃t φ̃t ⊗ x∗t xTt ⊗ IM (11b)
hc is to use the standard deterministic LS method t=1

˜ ∗˜ T
T /K ∗
√ 2 = φ̄b φ̄b ⊗ XXH ⊗ IM (11c)
ĥc,LS = arg min y− P Zhc (6) b=1
hc
∗
= K ΨH Ψ ⊗ IM K (11d)
whose solution is given by
˜ H = [1 φ̄H ] and
where φ̄
1 b b
ĥc,LS = √ Z† y (7)
P H
1 φ̄1
† −1 H
where Z = (Z Z) Z . Assuming again that nt ∼
H
. ..
Ψ = .. . . (12)
CN (0, σ 2 IM ) and that the noise is temporally uncorrelated,
H
the LS channel estimate is unbiased and equivalent to 1 φ̄T /K
the maximum likelihood (ML) estimate, and its covariance
matrix corresponds to the the Cramér–Rao bound (CRB)
To achieve a diagonal ΨH Ψ, the columns of the (T /K) ×
H (N +1) matrix Ψ must be orthogonal, with (T /K) ≥ N +1.
Rĥc,LS = E ĥc,LS − hc ĥc,LS − hc (8a) If ΨH Ψ can be made proportional to an identity matrix,
H −1 then ZH Z is also a scaled identity matrix.
1 σ2
= E Z† nnH Z† = ZH Z . (8b) For the above training protocol, the general solution
P P in (7) is implemented by taking data from the bth pilot
subblock
Ideally, xt and φt should be designed to optimize the
CSI estimation performance. While such an optimization is √
Yb = P Hd + HΦ̄b GH X + Nb (13)
generally intractable, a good choice can be found [3] by
√
and multiplying on the right by XH /(K P ) to obtain of requiring only two phase states for each RIS ele-
ment (1-bit resolution). In addition, a diagonal ΨH Ψ
1 ˜ + n̄ only requires that the RIS gains are equal at these
yb ≡ √ vec(Yb XH ) = Hc φ̄ b b (14)
K P two phase values. In this approach, T /K must be a
multiple of 4 for the Hadamard matrix to exist, but
√
where Φ̄b = diag(φ̄b ) and n̄b = vec(Nb XH )/(K P ). The this is not a significant issue for large N .
result yb from each of the T /K subblocks then forms a
column of the following combined equation: 2) Linear MMSE : The LS approach assumes a deter-
ministic channel with no prior information. On the other

Yc = Hc φ̄ ˜T
˜ · · · φ̄ + N̄ = Hc ΨH + N̄ (15) hand, the MMSE estimator assumes a stochastic model
1 K
for {Hd , G, H}, usually in terms of correlated Rayleigh
fading with prior information of the second-order statistics.
where N̄ = [n̄1 · · · n̄T /K ], from which an estimate of the However, the composite channel is composed of products
composite channel is obtained by multiplying Yc by Ψ on of the Gaussian elements in H and G, which makes the
the right, assuming ΨH Ψ ∝ IN+1 . MMSE estimate E{hc |y} difficult to compute, although
Several methods have been proposed to choose the RIS message-passing algorithms have been proposed for this
training sequence to satisfy ΨH Ψ ∝ IN+1 . problem [5], [32]–[34]. Instead, the linear MMSE, or
1) When the direct path is absent (i.e., the first col- LMMSE, estimate given by ĥc,LM = Wy can be found by
umn of Ψ is removed), a simple approach is to set solving [35], [36]
(T /K) = N and “turn on” one RIS element at a time
for each K -sample pilot subblock, with all other ele-
2
W = arg min E W̃y − hc . (17)
ments “turned off”1 [7], [30]. This results in ΨH Ψ = W̃
diag{β12 , . . . , βN
2
}. If the RIS elements are identical
and each active element is tuned to the same phase, Assuming spatially and temporally white Gaussian noise
then βi = β , which results in ZH Z = β 2 K IM KN . This uncorrelated with hc , the LMMSE estimate is given by
in turn leads to an estimate variance of σ 2 /(β 2 P K)
for each element of hc . √ −1
ĥc,LM = P Rhc ZH P ZRhc ZH + σ 2 IM T y (18)
2) Better performance is achieved by activating all RIS
elements over the entire training interval, in order
to benefit from the RIS array gain. One approach where Rhc = E{hc hHc } and we have assumed E{hc } = 0.
for doing so assigns the RIS phase shifts such that Using orthogonal pilot and RIS reflection sequences like
the N + 1 columns of Ψ equal the columns of the those discussed above also simplifies computation of the
(T /K) × (T /K) matrix that defines the (T /K)-point LMMSE estimate. For example, let I = IM K(N+1) and
discrete Fourier transform (DFT) [3], [7] assume the Hadamard reflection pattern so that ZH Z = T I.
Then, the LMMSE estimate simplifies to
[Ψ]mn = ej2π(m−1)(n−1)/(T /K) (16)
−1
1 σ2
ĥc,LM = √ Rhc R hc + I ZH y. (19)
for m = 1, . . . , (T /K) and n = 1, . . . , N + 1. If the RIS PT PT
gains are assumed to be phase-independent and sat-
isfy βi = β , then this leads to ΨH Ψ = (T β 2 /K)IN+1 The matrices in (19) involving Rhc are data independent
and the variance of the channel coefficient estimates and can be computed and stored offline since Rhc changes
is σ 2 /(β 2 P T ), a factor of T /K ≥ N +1 smaller than in relatively slowly. The resulting error covariance is given by
the first approach. In addition to the need for phase-
−1
independent RIS element gains, which is difficult to σ2
Re,LM = Rhc − Rhc Rhc + I Rhc . (20)
achieve in practice, the RIS phase shifts would have to PT
be tunable with at least log2 (T /K) bits of resolution,
which may be problematic for large N . For the above training protocol, Re,LM converges to
3) An alternative that achieves the same performance Re,LS = (σ 2 /(P T ))I for high SNR (i.e., σ 2 /P → 0) or long
is to choose the columns of Ψ from among the training intervals (T → ∞).
columns of a T /K -dimensional Hadamard matrix, A bigger issue than the computational complexity
whose entries are constrained to be ±1 [6], [31]. This of (19) is how to determine the composite channel covari-
achieves orthogonality for Ψ and has the advantage ance Rhc . In theory, the covariance could be estimated by
taking sample statistics of ĥc obtained over a long period
1 “Turning off” an RIS element assumes that it becomes a perfect
of time or using simulations involving detailed propagation
absorber of RF energy, which in practice is not possible. Thus, such
elements will still reflect a small amount of energy and thus degrade the models of the environment. However, the size of hc means
orthogonality assumption. that such procedures would require a large amount of
data. At first glance, a more reasonable approach might be common uniform planar array (UPA) architecture assumed
to determine Rhc based on covariance information about for RISs [37]. Furthermore, assumptions of uncorrelated
its constituent parts. For MIMO channels, it is commonly fading are hard to justify in RIS-aided wireless systems,
assumed that the multipath scattering at the source is which are often motivated by propagation environments
uncorrelated with the scattering at the destination, which with sparse propagation paths and frequent blockages.
would normally lead to the following descriptions: In these environments, the BS and RIS installations are
envisioned to be in elevated positions away from nearby
1
2 2
H
RF scatterers. This leads to low-rank channel correlation
H = RHB H̃RHR (21a)
1 H matrices and consideration of geometric models, as dis-
2 2
G = RGU G̃RGR (21b) cussed in Section IV.
1 H
2 2
Hd = RHd B H̃d R Hd U (21c)
B. Wideband Single-User MIMO

where the subscripts {B, R, U }, respectively, correspond
to BS, RIS, and UE, and indicate which side of the In wideband scenarios where the channel is frequency
link the correlation matrix is associated with (e.g., RHB selective, we assume that the UE transmits an orthogonal
is the correlation matrix for the BS side of the channel H). frequency-division multiplexing (OFDM) signal composed
The matrices {H̃, G̃, H̃d } are of the same dimensions as of Nc subcarriers from each of its K antennas. The symbols
{H, G, Hd }, respectively, and are composed of uncorrelated are given by the rows of the K × Nc matrix XF t =
CN (0, 1) elements. Under this model, it can be shown that [xFt,1 · · · xF
t,Nc ] in the frequency domain, where t is the
the composite channel covariance matrix would have the OFDM symbol index. Prior to transmission, the data XF t are
following form: first converted to the time domain using the Nc ×Nc matrix
FH that denotes the Nc -point inverse DFT: Xt = XF t F
H
and then is appended with a cyclic prefix of length Lcp
RTHd U ⊗ RHd B 0T
Rhc = (22) that is longer than the maximum delay spread of the
0 RR ⊗ RTGU ⊗ RHB
channel, L. At the BS, the cyclic prefix is removed, and the
data are converted back to the frequency domain through
where we define RR = RGR RTHR . multiplication by the DFT matrix F. This generates a model
However, there are problems with the use of the so- essentially identical to (1) for each subcarrier n
called Kronecker product channel correlation model for
H and G in (21a) and (21b). First, it has been shown √
yF d,n + Hn Φt,n Gn
P HF F FH
t,n = xF F
t,n + nt,n (23)
that the model is not accurate for RIS channels, especially
when N is large or the spacing between RIS elements
is small [37]. Second, the inherent ambiguity illustrated where {HF d,n , Hn , Gn } represent the DFT at subcarrier
F F
in (2) poses a difficulty for describing H or G as in (21a) n for the UE-BS, RIS-BS, and UE-RIS channel impulse
and (21b). Even if the actual channels followed a Gaussian responses, respectively. Thus, one can employ the same
model described by (21a) and (21b), we do not have access estimation methods discussed above on a per-subcarrier
to the actual channels (or even estimates thereof), only basis, although to exploit the channel correlation in fre-
versions for which the ambiguity in (2) has been removed. quency and reduce the training overhead, pilot data are
For example, resolving the ambiguity by fixing the first normally transmitted only on a subset of the subcarriers,
row of H to be a vector of ones corresponds to dividing and interpolation is used to construct channel estimates
each column of the BS-RIS channel by its first element, for others [4]. An alternative approach proposed in [13] is
rendering an H—and also a G—that will no longer be to use shorter OFDM symbols during the training period.
Gaussian. Thus, the estimation of the covariance matrices Note that most prior work on RIS channel estimation
in (21a) and (21b) is problematic. Various assumptions with OFDM signals has assumed that the RIS reflection
could be made to simplify the direct calculation of Rhc , properties are frequency independent, i.e., Φt,n = Φt , but
such as assuming, for example, that the multipath scatter- this is generally true only for relatively narrow bandwidths
ing at the RIS is isotropic, in which case RGR and RHR [38], [39]. If one sets φt,n to have desirable properties
could be taken as identity matrices and Rhc would be (e.g., Z with orthogonal columns) at a particular subcar-
block diagonal with identical block entries except for the rier n, then in general, those properties will not be inher-
block associated with Hd . This would greatly simplify the ited at other subcarriers. This issue motivates the design
computation of Rhc and the estimate in (19). However, of RIS circuit architectures that have invariant properties
the fact that the RIS multipaths are confined to only the across wider frequency bands.
half-space associated with the active side of the RIS means An alternative to estimating the channels in the fre-
that, even with isotropic scattering in that half-space, an quency domain and using interpolation is to directly esti-
identity covariance for RGR and RHR only holds for certain mate the channel impulse response. In the time domain,
special RIS geometries, none of which correspond to the we represent the data received for sample s of OFDM
symbol t as where the composite channel is now Hc = gH H =

H diag(g∗ ). The training overhead in this case is reduced
√
L−1 to N + 1 samples.
yt,s = P Hd (k) + H(k)Φt,s−k GH (k) xt,s−k + nt,s
k=0
2) Single-Antenna BS and UE: When both the BS and UE
(24) have only a single antenna, we denote the RIS-BS channel
as the N × 1 row vector hT and write the BS output and
composite channel as
where {Hd (k), H(k), G(k)}L−1k=0 represent the channel
impulse responses and L is the maximum number √
of taps. Defining hd (k) = vec(Hd (k)) and hc (k) = yt = P hTc φ̃t xt + nt (28a)
vec([hd (k) G∗ (k)H(k)]), after removal of the cyclic prefix, T

hTc = hd gH h T
= hd h̄c (28b)
we can write
√
L−1
T
where only h̄c is identifiable.
yt,s = P φ̃t,s−k ⊗ xTt,s−k ⊗ IM hc (k) + nt,s (25a)
k=0
D. Multiple-User Scenarios
√
L−1
= P Zt,s−k hc (k) + nt,s (25b) The models and approaches discussed above are easily
√
k=0 generalized to the multiple UE case. Assuming that UE u
= P [Zt,s Zt,s−1 · · · Zt,s−L+1 ] hc + nt,s (25c) has Ku antennas for u = 1, . . . , U , then the model in (1)
!
yt,1 holds if we simply set K = u Ku and all UE antennas
√
yt = ...
transmit orthogonal pilot sequences. Some prior work has
= P Zt hc + nt (25d)
proposed that the users take turns transmitting pilots, in
yt,Nc which case there is no change to the algorithms described
above, but this only makes sense if one exploits the fact
where hc = [hTc (0) · · · hTc (L − 1)]T is the LM K(N + that each user’s composite channel shares a common RIS-
1) × 1 vector containing all unknown channel coefficients BS component H. This idea will be explored further in
and Zt is an M Nc × LM K(N + 1) block-circulant matrix Section III-C. For multicarrier signals, a scheme is required
with first block row [Zt,1 Zt,Nc · · · Zt,Nc −L+2 ]. Finally, to allocate the pilot subcarriers to the UEs, but otherwise,
assuming that the channel is stationary over To total OFDM the channel estimation is the same. One implication for
symbols, we have the LMMSE approach is that, assuming that the channels
for different UEs are uncorrelated, the matrices RGU and
RHc U will be block-diagonal.
Z1
√
.
√
y= P .. hc + n = P Zhc + n. (26) E. Reducing the Complexity and
ZTo Training Overhead
As noted already above, one of the key hurdles to over-
The time-domain approach assumes that only pilot data come in CSI estimation for RIS-aided systems is the large
are transmitted first, followed by payload data. The total required training overhead. Consequently, recent work has
number of pilot symbols required is T = To Nc ≥ focused on a variety of methods to reduce this overhead,
KL(N + 1). While more OFDM symbols are likely required some of which are described next. The use of geometric
for the frequency domain method to obtain the same channel models to reduce pilot overhead is reserved for
channel estimation accuracy, this is offset by the fact that Section IV.
data and pilots can be transmitted together.
1) RIS Element Grouping: A simple approach to reduce
the number of pilots and estimation complexity is to assign
identical phases to RIS elements with highly correlated
C. Single-Antenna Scenarios
channels [4], [6]. High channel correlation occurs when
1) Single-Antenna UE: The single-antenna UE case is adjacent RIS elements are closely spaced; retaining the
often considered in the literature since it simplifies the flexibility of arbitrary phase shifts for such elements does
notation and reduces the algorithm complexity, but there not provide a significant increase in degrees of freedom
is fundamentally little difference with the general multi- for beamforming design since the designed phases would
antenna UE case described above. The channel G becomes likely be nearly identical. Suppose that groups of size J
a 1 × N row vector that we denote by gT , while the direct are identifed, and assume for simplicity that N = N/J is
channel Hd is simply an M × 1 vector hd . The pilot data an integer and no direct channel Hd is present. Then, we
received at the BS are given by define φt = φt ⊗ 1J , where φt is N × 1, and write
√
yt = P (hd + H diag(g∗ )φt ) xt + nt (27) Hc φt = Hc (φt ⊗ 1J ) = Hc (IN ⊗ 1J )φt = Hc φt (29)
where the effective composite channel Hc is now estimated in the normal way, while the other users do not
M K × N . Each column of Hc is thus a unit-coefficient transmit. Then, in the second step, the other users transmit
linear combination of the columns of Hc corresponding and the estimate of the RIS-BS channel obtained in the first
to a given group of RIS elements. The revised model is step is exploited to reduce the training required for the
identical in form to the general case, and thus, the methods remaining channels.
described above can be implemented to estimate Hc with Assume that the users are ordered such that the user
a reduction in the required training overhead by a factor corresponding to the first row of G, denoted by gT1 , is
of J . the one selected for the first step. The LS method is
A generalization of this idea presented in [6] succes- used to estimate the composite channel H diag(g1 ) and the
sively reduces the size of the groups over multiple blocks direct channel hd,1 , which requires at least T1 = N + 1
of pilot and payload data in order to eventually resolve the training samples. Recall that only the product H diag(g1 ) is
channels for all of the RIS elements. estimated and not the individual terms H and g1 . In fact,
we can treat diag(g1 ) as Λ in (2), so step 1 provides us with
2) Low-Rank Channel√ Covariance: We see from the noise- H
an estimate of H̃, and we can set the first row of G̃ to 1TK .
free part of (5), y = P Zhc , that in the general case, ˆ
the M T × M K(N + 1) data matrix Z should be full With the estimate H̃, during step 2, the training data model
rank M K(N + 1) since, otherwise, components of hc in is approximately given by
the nullspace of Z could not be identified. Like the LS √ ˆ H

approach, this requires T ≥ K(N + 1) training sam- yt P H̃d + H̃Φt G̃ xt + nt (32a)

ples. However, if Rhc is rank deficient, then it would be
enough for the column span of Rhc to lie within the √ ˆ h̃d
P xTt ⊗ IM H̃Φt + nt (32b)
column span of ZT . In particular, suppose that Rhc is rank g̃∗
r < M K(N + 1) and thus can be factored as Rhc = UUH , M ×(K−1)(M +N)
where U has r columns. Then, in principle, it would be √
sufficient to choose P Z̃t h̃c + nt (32c)
ZT = UV (30) where we drop the first column of Hd to create H̃d and we

drop the first row of ones in G̃ since UE 1 does not transmit.
T T T
We also have defined g̃ = vec(G̃ ) and h̃c = [h̃d g̃H ].
for some full rank r × M T matrix V, and thus, theoreti-
Stacking T2 of these training vectors together, we get an
cally, it would be sufficient that T ≥ r/M . Unfortunately,
equation analogous to (5), where in this case, Z is M T2 ×
due to constraints on the possible values for φt , finding V
(M +N )(K −1). Assuming that linearly independent pilots
that exactly satisfies (30) is generally not possible if T <
xt and RIS reflection vectors φt are chosen, we can solve
K(N + 1). It may, however, be possible to approximately ˆ √
for the remaining channel parameters using h̃c = Z† y/ P ,
solve (30) for larger values of T that are still much smaller
provided that M T2 ≥ (K − 1)(M + N ) or, equivalently,
than K(N + 1), provided that r is not too large. In addition
T2 ≥ (K − 1)((N/M ) + 1). Given the N + 1 samples
to reducing the training overhead, the low-rank channel
needed for step 1, the minimum required training time is
covariance can be exploited to significantly reduce the
thus
amount of data required to estimate Rhc as well as the
cost of computing the LMMSE solution in (18) since only
N
an r × r inverse rather than an M T × M T inverse is Tmin = (N + 1) + + 1 (K − 1) (33)
M
required
√ " # which for large M is significantly less than the value

P σ 2 −1 H H
ĥc = 2 U Ir − W(W + Ir ) U Z y (31) K(N + 1) required by the standard LS method.
σ P
where W = UH ZH ZU.
3) Exploiting Common Channels: The LS method in IV. E S T I M A T I O N O F S T R U C T U R E D
Section III-A1 ignores the Kronecker product structure of CHANNELS
the composite channel, which can be exploited to reduce The large training overhead required for unstructured
the training overhead. The key observation is that in the channel estimation motivates the consideration of channel
uplink, the composite channel for each user shares the models that are described by fewer parameters. Such mod-
same RIS-BS channel H [10]. To explain how this infor- els are often used in millimeter-wave or higher frequency
mation can be exploited, assume without loss of generality bands, where multipath scattering is sparse and propaga-
a scenario with K single-antenna users. The approach is tion is often dominated by strong specular components.
divided into two steps [9], [12]. In the first, one of the In such cases, the channels can be described by a small
users is selected and the composite channel for this user is number of propagation paths defined by path gains, AoAs,
and AoDs.2 The resulting number of parameters is often and relatively little pilot data with the RIS switched off.
more 1–2 orders of magnitude less than that required Consequently, we will not explicitly consider the estima-
in the unstructured case, and the training overhead is tion of Hd in this section. We will further assume that
correspondingly reduced. the BS and the UEs (when they have multiple antennas)
Parametric channels are described by the array response employ ULAs so that their array response depends on a
or “steering” vectors associated with the angle of an incom- single angle/spatial frequency, and we assume that the RIS
ing (AoA) or outgoing (AoD) signal. For example, the elements are arranged as a UPA, so its spatial response
response of an Mx -element uniform linear array (ULA) to depends on two spatial frequencies. Generalizations to
a signal arriving with azimuth angle θaz is described by the arbitrary array geometries are straightforward.
Vandermonde vector
T A. Parametric Estimation
ax (ωx ) = 1 ejωx ej2ωx · · · ej(Mx −1)ωx (34) The structured channel estimators that we will consider
assume parametric channel models of the following form,
where the spatial frequency ωx is defined by ωx = which we describe first for the RIS-BS channel:
2πΔx sin(θaz ) and Δx is the distance in wavelengths
between the antennas.3 For an Mx × My UPA with antenna
dH
H= γH,k aB (χ BH,k ) aH
U (χ RH,k ) (37a)
separations of Δx and Δy in the x- and y -directions, the
k=1
array response vector can be written as
= AB (χ BH ) ΓH AH
R (χ RH ) (37b)
a(χ) = ax (ωx ) ⊗ ay (ωy ) (35)

where the columns of
where the vertical array response component is similar

AB (χ BH ) = [aB (ωBH,1 ) · · · aB (ωBH,dH )] (37c)
to (34)
AR (χ RH ) = [aR (χ RH,1 ) · · · aR (χ RH,dH )] (37d)

T
−1)ω2
ay (ωy ) = 1 ejω2 ej2ω2 · · · e j(My
(36)
respectively, represent the steering vectors for the prop-
agation paths with AoA spatial frequencies χ BH =
but defined by ωy = 2πΔy sin(θel ) cos(θaz ) with elevation [ωBH,1 · · · ωBH,dH ]T at the BS and AoD spatial frequen-
AoA θel . The vector χ = [ωx ωy ]T corresponds to a 2-D cies χ RH = [χ TRH,1 · · · χ TRH,dH ]T from the RIS. The
spatial frequency. For either a ULA or UPA, there is a diagonal matrix ΓH = diag{γ H } = diag{γH,1 · · · γH,dH }
one-to-one correspondence between the angles and spatial contains the complex path gains γ H = [γH,1 · · · γH,dH ]T .
frequencies as long as both Δx and Δy are no more than The RIS AoDs for path k, denoted by χ RH,k , are written as
one-half wavelength. This is important for applications vectors since the RIS spatial frequencies are 2-D:
involving localization since the angles provide useful infor-
mation for locating a signal source. However, from the
ωRH,k,x
viewpoint of channel estimation, it is enough to know χ , χ RH,k = . (38)
and any ambiguities in determining the angles need not be ωRH,k,y
resolved.
In this section, we focus on the estimation of structured Parametric models such as (37) are usually employed
or geometric channel models. To simplify the discussion, when the number of paths dH is smaller than the array
we assume that the direct UE-BS channel is absent. This dimensions M and N , and thus, the channel H is low-rank.
assumption is often made for scenarios with low-rank near- Parametric CSI estimation involves finding the spatial
specular propagation at high frequencies, where blockages frequencies of signals collected by an array. For example,
are common. However, such blockages may not be perma- suppose that n observations are available from an arbitrary
nent, and given the lower path loss that a direct path signal M -element array receiving signals from d directions
would likely undergo, the direct channel Hd can make
a significant contribution to the received signal. In such
Y = A(χ )S + N (39)
cases, Hd could be estimated using standard algorithms
2 For very large RIS, where the BS or UEs are in the Fresnel region of
where Y is M × n, S is d × n, N is the noise, and
the RIS, the channel parameterization must also include range or the 3-D A = [a(χ 1 ) · · · a(χ d )] is the M × d array response
coordinates of the various devices, and the large-scale fading becomes
antenna-dependent. However, here, we focus on the more common far- matrix. The matrix S is not typically assumed to be known.
field scenario. This is the classical model assumed for AoA estimation,
3 Note that we assume a narrowband propagation model here where
and many methods have been developed to estimate χ =
time delays can be represented by phase shifts. For large arrays,
ignoring the frequency dependence of the model leads to the beam- [χ 1 · · · χ d ]T from Y . The simplest method is based on
squint effect [26]. (matched filter) beamforming, which involves searching
for d peaks in the spectrum representation
pB (χ) = aH (χ)RY a(χ) (40) AR (χ RH ) = ARD QRH (44)
where RY is the sample covariance matrix where ARD is N × NRD and QRH is NRD × dH . Because
the RIS AoDs are 2-D, a 2-D grid is necessary to specify the
1 H AoD pairs. Substituting (43) and (44) into (37), we can
RY = YY . (41)
n write
Alternatively, one can employ higher resolution algorithms

such as MUSIC [40] or ESPRIT [41], which require com- H = ABD QBH ΓH QTRH AH
RD (45a)
putation of the eigendecompositon of RY . If N is spa- h = vec (H) = (A∗RD ⊗ ABD ) γ HD (45b)
tially and temporally white, the (deterministic) ML method
[42] finds the AoA estimates from the d-dimensional (or where γ HD = vec(QBH ΓH QTRH ) is a dH -sparse vector of
2d-dimensional for azimuth/elevation angles) problem length NBD NRD whose dH nonzero elements are equal
to γ H . The M N × NBD NRD matrix A∗RD ⊗ ABD can be
χ̂ M L = arg min trace P⊥
A (χ)RY (42) thought of as an overcomplete dictionary for the vectorized
ω
channel h.
Consider an overcomplete representation of the data
where P⊥ A (χ) = IM − A(χ)[A (χ)A(χ)]
H −1 H
A (χ). The cor-
in (39)
responding ML estimate of S is given by Ŝ = A† (χ̂ M L )Y ,

where (·)† represents the pseudoinverse, although esti-

mates of χ̂ from other algorithms can be substituted Y = AD Q S + N = AD C + N (46)
for χ M L to estimate S . MUSIC and ESPRIT require that
rank(S ) = d and thus that n ≥ d, but the beamforming and where the dictionary AD is M × ND and the matrix Q
ML methods are theoretically viable for smaller values of n. is ND × d, with one nonzero element per column equal
The theoretical limit for identifiability is d < (M + n)/2, to one. As before, S is d × n. The matrix C = Q S
so for large enough M , n can be as small as 1 [43]. exhibits common row sparsity; only d rows have nonzero
For a massive antenna array where M is large, all the elements, and these rows are equal in some order to the
above methods provide asymptotically efficient AoA esti- rows of S . If C can be estimated, then the location of
mates [44], and thus, a simple technique such as beam- its nonzero rows would correspond to the AoAs, and the
forming is preferred due to its low computational load and entries in those rows to the rows of S . Provided that
minimal assumptions. A key requirement for all the above M ≥ O(d log(ND /d)), the problem of finding the sparse
methods is that the value of d is to be known or estimated entries in C can be solved by a number of different
from the data. CS approaches, such as iterative hard thresholding (IHT)
[49], orthogonal matching pursuit (OMP) [50], alternating
B. Compressive Sensing
direction method of multipliers (ADMM) [51], or using
Compressive sensing (CS) formulations of the geo-
AMP-based algorithms [52], [53]. As mentioned above,
metric CSI estimation problem are also possible, using
the accuracy of dictionary-based approaches is limited by
sparse representations from an overcomplete dictionary
the resolution of the sampled grid, which cannot be made
[45]–[48]. For example, we can represent the array
infinitely fine due to numerical and computational issues.
response vectors for the BS side as
Off-grid AoAs create a basis mismatch that leads to leakage
of energy into adjacent rows of C . One approach to refine
AB (χ BH ) = ABD QBH (43) the grid-based AoA estimates is to apply a small angular
rotation to each selected column of AD that maximizes the
where ABD is an M × NBD matrix whose columns are correlation with the received data [21]
BS array response vectors sampled on an NBD grid of

frequencies corresponding to the possible BS AoAs and χ k = arg max diag(a(χ))aD,k Y H (47)
ω
QBH is an NBD ×dH matrix whose kth column has a single 2
1 in the position corresponding to ωBH,k , assuming that

it is one of the grid points. If ωBH,k is not on the grid, for k = 1, . . . , d, where aD,k is the kth column of AD Q̂ . The
then the model in (43) is an approximation. While the final estimate of χ k is the spatial frequency corresponding
error can be made small by making NBD large, increas- to aD,k plus the estimated rotation χ k . A more fundamen-
ing the coherence of the dictionary eventually leads to tal approach to solve the basis mismatch problem for 1-D
computational and numerical issues. More will be said on spatial frequencies is to recast the sparse recovery problem
this topic shortly. The RIS also has a similar overcomplete using the atomic norm [22], [54].
C. Single-User MIMO Single Carrier Table 1 Composite Channel Parameterization for Single-User MIMO
In this section, we consider structured channel estima- Single-Carrier Case. The Elements of These Vectors Form the Elements
of the Parameter Vector η for Geometric Channel Models
tion for the case involving a single multiantenna UE. As
with the RIS-BS channel, we can define parametric and
overcomplete representations of the RIS-UE channel as
follows:
G = AU (χ U G ) ΓG AH
R (χ RG ) (48a)
= AU D QU G ΓG QTRG AH
RD (48b)
∗ As in the nonparametric case, not all of the parameters
g = vec (G) = (ARD ⊗ AU D ) γ GD (48c)
or decompositions shown in (51) are identifiable. To see
this, let
where we assume dG paths with gains γ G =
[γG,1 · · · γG,dG ]T and ΓG = diag{γ G }. In this case, $ %
Λ = diag aR (χ ) (53)
γ GD = vec(QU G ΓG QTRG ) is a dG -sparse vector of length
NU D NRD whose nonzero elements correspond to γ G .
Recall the general model in (5), where the BS data from be a diagonal matrix formed from an arbitrary RIS array
the T training samples are stacked together in a single response vector for 2 × 1 direction of arrival (DOA) χ .
vector Then,
√ ∗
y= P Zhc + n (49) ATR (χ RG ) AH
R (χ RH ) = AR (χ RG ) Λ AR (χ RH ) Λ
T H
$ % $ %
= ATR χ RG AH
R χ RH (54)
where
T where the elements of χ RH and χ RG are defined as
φ1 ⊗ xT1 ⊗ IM χ RH,k = χ RH,k +χ and χ RG,k = χ RG,k +χ , respectively.

Z= ..
. (50) Thus, if the RIS AoA and AoD spatial frequencies are
shifted by the same amount, there is no change to the
φTT ⊗ xTT ⊗ IM
composite channel response. In addition, since the channel
gains for H and G always appear together as Γ∗G ΓH ,
is M T × M KN and involves φt instead of φ̃t since we there is a scaling ambiguity; in particular, γ H and γ G
are assuming that no direct UE-BS channel is present, yield the same composite channel as αγ H and γ G /α∗ ,
which also implies that hc = vec(Hc ) = vec(G∗ H). respectively. To obtain an identifiable parameterization η
Using various properties of the Kronecker and Khatri– for the composite channel in (51), one could, for example,
Rao products, the composite channel can be decomposed set χ RH,1 = [0 0]T so that aR (χ RH,1 ) is a vector of ones
using either a parametric approach or via overcomplete and set γ H,1 = 1. With these assumptions, the set of unique
dictionaries as follows: parameters η that describe the composite channel matrix
is given in Table 1. The total number of parameters is thus
Hc = (A∗U (χ U G ) ⊗ AB (χ BH )) ΓGH dim(η) = 5dH + 5dG − 4, which for typical values of dH
and dG is much smaller than the number of parameters
M K×dH dG dH dG ×dH dG
2M KN that must be estimated for a nonparametric Hc ,
× ATR (χ RG ) AH
R (χ RH ) (51a) which could be in the thousands.

dH dG ×N 1) Channel Estimation for the General Case: Using (51),
the composite channel is given by
= (A∗U D ⊗ ABD ) QGH ATRD AH (51b)

RD

M K×NU D NBD NU D NBD ×N 2
RD 2 ×N
NRD hc = A (χ) γ GH (55a)
= AD qGH (55b)
where
where γ GH = γ ∗G ⊗ γ H , qGH = vec(QGH ),

ΓGH = Γ∗G ⊗ ΓH (52a)
T
QGH = QU G Γ∗G QTRG ⊗ QBH ΓH QTRH (52b) A (χ) = ATR (χ RG ) AH
R (χ RH )

NU D ×NRD NBD ×NRD
(A∗U (χ U G ) ⊗ AB (χ BH )) (56a)
= (QU G ⊗ QBH ) ΓGH (QRG ⊗ QRH )T . (52c) T

NU D NBD ×dH dG 2
dH dG ×NRD AD = ATRD AH
RD ⊗ (A∗U D ⊗ ABD ) . (56b)
and χ T = [χ TBH χ TRH χ TRG χ TU G ]. Similarly, there are estimate χ , setting χ RH,1 = [0 0]T to make the model
two forms of the general data model in (49) identifiable
√
y = P ZA (χ) γ GH + n (57a) χ̂ = arg min yH P⊥
ZA(ω ) y (62)
√ ω
= P ZAD qGH + n. (57b)
where P⊥ ZA(ω ) is the projection orthogonal to the columns
At first glance, (57a) has the standard form assumed in of the effective array response ZA(χ). This would require
AoA estimation problems, where A(χ ) = ZA(χ), but there a nonconvex optimization over the 3(dG + dH ) − 2 spatial
are some caveats. First, γ GH is not arbitrary but is instead frequencies in χ .
a nonlinear function of γ H and γ G . Second, note that the One special case worth mentioning occurs when dH =
kth column of A(χ) can be expressed as dG = 1 or when the UE-RIS and RIS-BS channels are LoS.
In this case, (57a) simplifies to hc = γ a(χ), where

[A(χ)]:k = CT (χ RG , χ RH ) ⊗ a∗U (ωU G, ) ⊗ aB (ωBH,p )
:k
a(χ) = aR (χ RG ) ⊗ a∗U (ωU G ) ⊗ aB (ωBH ) (63)
(58)
with four angle parameters of interest: ωBH , ωU G , and

where = k/dH , p = moddH (k), and
χ RG . This results in a standard single-snapshot AoA
estimation problem, and the vector χ can be deter-
C (χ RG , χ RH ) = ATR (χ RG ) AH
R (χ RH ). (59) mined either by maximizing the beamforming criterion
|yH Za(χ)|2 or minimizing the DML criterion√ yH P⊥
Za(ω ) y,
†
Each column of CT (χ RG , χ RH ) only depends on a pair of and in either case setting γ̂ = (Za(χ̂)) y/ P . Another way
2-D angles, one each from χ RG and χ RH [21] that beamforming can be applied in the general case with
arbitrary dH and dG is to ignore the interdependence of the
T columns of A(χ) on different combinations of the elements
CT (χ RG , χ RH ) = ATR (χ RG ) AH
R (χ RH ) (60a)
:k of χ and just treat the angle parameters of each column as

k:
T
= aTR (χ RG, ) aH (60b) if they were independent variables. This reparameterizes
R (χ RH,p )
the model in (57a) and (61) as
= aR (χ RG, − χ RH,p ) . (60c)
√ $ %
y = P ZA χ γ + n (64a)
Thus, from (58), we see that the kth composite steering $ % $ % $ %
A χ = a χ 1 · · · a χ dH dG (64b)
vector in (57a) is parameterized by different pairs of $ % $ % ∗$ % $ %
entries from χ RG and χ RH and from χ BH and χ U G a χ k = aR χ k,1 ⊗ aU χ k,2 ⊗ aB χ k,3 (64c)
[A(χ)]:k = aR (χ RG, − χ RH,p ) ⊗ a∗U (ωU G, ) ⊗ aB (ωBH,p ). where γ is an arbitrary dH dG vector and χ has four
(61) elements, one each for χ k,2 and χ k,3 since they are
1-D spatial frequencies and two for χ k,1 since it is 2-D.
While this increases the number of angle parameters that
The fact that A(χ) only depends on the differences must be estimated to 4dH dG , the beamforming criterion
between the elements of χ RG and χ RH is a direct conse- yH Za(χ ) 2 can be applied since the columns of A(χ )
quence of the fact that they are not separately identifiable, are identically parameterized. This would require a search
as mentioned above. An AoA estimation algorithm that for dH dG local maxima in a 4-D space. In this case,
takes the special structure of γ GH into account is difficult √
γ̂ = (ZA(χ̂ ))† y/ P .
to formulate, and 1-D methods, such as beamforming and CSI estimation for the dictionary-based model in (57b)
MUSIC, cannot exploit the inherent relationship between also requires an unconventional approach, due to the
the columns of A(χ). However, assuming that A(χ) is full sparsity pattern in qGH [15], [21], [23]. The first issue
rank4 for all possible χ , one could ignore the structure is again due to the ambiguity in specifying the spatial
of γ GH and use the deterministic ML (DML) criterion to frequencies of the RIS AoAs and AoDs; any circular shift
4 The matrix A(ω) will generically be full rank as long as the BS, in the columns of ARD or the rows of QRG and QRH will
RIS, and UE arrays have unambiguous array manifolds (e.g., elements leave AD unchanged in (56b). As before, this ambiguity
spaced no more than λ/2 apart for a ULA or UPA). However, there can be rectified by forcing the first column of QRH to
are pathological cases where A(ω) can drop rank. This could occur,
for example, if any RIS angle differences are repeated, i.e., ωRG, −
have its nonzero element in the position corresponding to
ωRH,p = ωRG, − ωRH,p for = or p = p , or if the BS χ RH = [0 0]T . The second issue is most easily understood
and UE arrays have identical manifolds and share an angle between the via (52b). The matrix Q1 = (QU G Γ∗G QTRG ) has dG nonzero
vectors ωBH and ωU G . Although such cases occur with probability
zero, A(ω) will become ill-conditioned if they are approximately true
elements at the row/column coordinates corresponding
and numerical problems would ensue. to the AoA/AoD pairs of G, while Q2 = (QBH ΓH QTRH )
has dH nonzero elements at positions corresponding to We see immediately that (67b) and (67c) are in the form
the AoA/AoD pairs of H. The Kronecker product of Q1 of a standard AoA estimation problem as in (39), and as
with Q2 then repeats the sparse structure of Q2 at the long as M ≥ max{2dH − T1 , dH + 1} [43], χ BH could
block rows/columns in QGH corresponding to the nonzero be estimated from Y1 using an AoA estimation algorithm.5
elements of Q2 . In addition, the nonzero values in these Similarly, the estimation of χ U G can also be performed
repeating blocks are scaled versions of one another; the separately based on the following equation, provided that
dH nonzero elements in Q2 correspond to γ H , and each K ≥ dG + 1 since presumably M > dG :
time they are repeated in QGH , they are multiplied by a
different element of γ G . This structure in QGH in turn X1 YH X1 N H
creates a corresponding repeated pattern in qGH . Thus, √1 = AU (χ U G ) ΞH AH
B (χ BH ) + √1 (69a)
T1 P T1 P
qGH is dH dG -sparse, but it depends only on dH + dG − 1 1
= AU (χ U G ) S + √ X1 NH 1 . (69b)
unique complex values. While cumbersome, imposing the T1 P
required structured sparsity constraint on qGH is possible
[15], [23]. For the second stage, the UE repeatedly sends T2 /dG
A bigger issue with the CS approach in the general blocks of the pilot signal matrix X2 ∈ CK×dG during
case is the size of the dictionary in (57b), which has T2 time slots, while the reflection shift Φb varies from
2
NBD NU D NRD elements. A smaller dimensional problem block to block. The observed signal in the bth pilot block
can be formulated based on the parameterization in (64) (b = 1, . . . , T2 /dG ) is given by
such that
√
√ Yb = P HΦb GH X2 + N2,b . (70)
y= P ZAD q +n (65)
With sufficiently accurate estimates of χ BH and χ U G from
where q is dH dG -sparse and unstructured, and the dictio- Stage 1, we have
nary ZAD is composed of NRD NU D NBD entries defined
by 1
YB,b = √ A†B (χ̂ BH ) Yb B†
P
∗
ZAD = Z (ARD ⊗ A∗U D ⊗ ABD ) . (66) Γ H AH
R (χ RH ) Φb AR (χ RG ) ΓG + NB,b (71)
√
While the dictionary is now a factor of NRD smaller, where NB,b = (1/ P )A†B (χ̂ BH )N2,b (AH U (χ̂ U G )X2 ) and
the resulting problem is likely still intractable. Assuming B = AH U (χ U G ) X 2 . The b = 1, . . . , T 2 /d G blocks defined
LoS propagation only serves to reduce the sparsity level, in (71) are gathered together to form
without reducing the dimension of the dictionary. Conse-
quently, a more tractable approach is needed, as described " #T
next. YB = vec (YB,1 ) , . . . , vec YB, T2 (72a)
dG
T
= Ψ∗ ATR (χ RG ) AH
R (χ RH ) ΓGH + NB (72b)

2) Simpler Decoupled Approach: In this approach, the CT (ω RG ,ω RH )
channel estimation is decoupled into two stages. In the T
∗
=Ψ ATRD AH
RD (QRG ⊗ QRH ) ΓGH + NB (72c)
first stage, the BS and UE components of the channel
are determined from an initial set of pilot data and then
removed from the composite channel. In the second stage, where ΨH = [φ 1 , . . . , φ(T2 /dG ) ], C(χ RG , χ RH ) is defined
additional pilot data are used to estimate the remainder of in (59) and NB is defined similar to YB .
the channel [22], [25]. In particular, Stage 1 assumes that A sparse estimation problem could be set up for χ RG
the UE transmits a K × T1 matrix of orthogonal (T1 ≥ K ) and χ RH based on vectorizing (72c), and while the result-
pilot data X1 , while the RIS holds a fixed reflection pattern ing dictionary is significantly smaller than in (57b), it still
2
φ, which results in the following received signal at the BS: has NRD elements. Instead, a much simpler solution can be
found by noting the structure of C(χ RG , χ RH ) described
√
Y1 = P HΦ̄GH X1 + N1 (67a) in (60), which indicates that each of the dG dH columns
√ of (72b) depends only on a pair of 2-D angles, one from
= P AB (χ BH ) ΞAH
U (χ U G ) X1 + N1 (67b)
√
= P ABD QBH ΞQU G AH
T
U D X1 + N 1 (67c)
5 Note that in some work, it is assumed that the BS-RIS channel
where changes slowly since the BS and RIS are in fixed locations. In such cases,
ωRH can be estimated infrequently and thus may already be known,
and hence, the above estimation step may not always be necessary [5],
∗
Ξ = ΓH AH
R (χ RH ) ΦAR (χ RG ) ΓG . (68) [8], [17], [55].
χ RG and one from χ RH . In particular, for the kth column Ω = [χ(0)T · · · χ T (L − 1)]T , and where χ(k), γ GH (k),
and qGH (k) are the spatial frequencies, path gains, and
yB,k γk Ψ∗ aR (χ RG, − χ RH,p ) + nB,k (73) sparse vectors associated with the channel for tap k,
respectively. Note that we have assumed the general case
where the AoAs/AoDs are potentially different for each
where γk is the kth element of γ ∗G ⊗ γ H and k = ( − tap. Consequently, while we see from (76) that the full-
1)dH + p for = 1, . . . , dH and p = 1, . . . , dG . Each of scale parameterization of the problem is similar to that
the dH dG columns of YB is thus approximately equivalent in (57), the dictionary size and parameter dimensions are
to a single snapshot from a T2 -element “array” with a all a factor of L larger.
single 2-D spatial frequency. As before, the gains and 2-D The decoupled approach described in Section IV-C2 can
frequencies of the columns are interrelated, but if one be exploited to significantly reduce the required complex-
ignores this fact, the unknown part of the channel in (70) ity. As before, we ignore the direct channel and assume
can be reconstructed in a suboptimal way by solving a an initial training period of T1 = To1 Nc samples from To1
series of dH dG 1-D AoA estimation problems on each of OFDM symbols where the RIS reflection state is fixed at φ.
the columns of YB . In particular, if cTk represents the kth Using (24), after removal of the cyclic prefix, we can collect
row of ΓGH C(χ RG , χ RH ), then cTk is estimated using the all Nc samples from OFDM symbol t in the matrix Y1,t as
estimates of γk and χ k = χ RG, − χ RH,p obtained for the follows:
kth column of YB
√
Y1,t = P H (0) ΦGH (0) · · · H (L − 1) ΦGH (L − 1)
ĉk = γ̂k aR (χ̂ k ). (74)
×X1,t + N1,t (78a)
√
For LoS channels with dH = dG = 1, only a single sparse = P AB (ΩBH ) ΠX1,t + N1,t (78b)
estimation problem needs to be solved.
where X1,t is block circulant with first block row
D. Wideband Single-User MIMO
[xt,1 · · · xt,Nc ], ΩBH = [χ BH (0)T · · · χ TBH (L − 1)]T , and
As in Section III-B, for geometric channel models
with OFDM signals, either time- or frequency-domain
AB (ΩBH ) = [AB (χ BH (0) · · · AB (χ BH (L − 1)] (79a)
approaches can be employed. We will focus on the time-
domain approach here as it is a bit more straightforward
L−1
Π = blkdiag Ξ (k) AH
U (χ U G (k)) (79b)
(see [17] for methods that focus on estimating the indi- k=0
vidual subchannel responses). Our starting point is (26),

which differs from (57) in that the wideband channel where Ξ(k) is the matrix corresponding to (68) for
parameter vector hc is composed of L single-tap terms tap k. Concatenating data from the To1 OFDM symbols
like hc stacked together. Partition the columns of Z into L yields
blocks, each block corresponding to one of the taps hc (k),
k = 1, . . . , L, in hc √
Y1 = P [Y1,1 · · · Y1,To1 ] (80a)
√
= P AB (ΩBH ) ΠX1 +N1 (80b)
Z = [Z(0) Z(1) · · · Z(L − 1)] . (75)
M ×LdH LdH ×T1
Using (55), we can write

where X1 and N1 are defined like Y1 . Assuming that M >
LdH , the BS AoAs for each tap of the impulse response can
γ GH (0)
√
..
be estimated using standard approaches. The AoDs at the
y = P A (Ω) . +n (76a) UE can be found by noting that
γ GH (L − 1)
D(0)
qGH (0)
√ (Y1 X†1 )H (N 1 X † )H
=
..
= P AD √ + √ 1
.. +n (81)
. (76b) P T1 . P T1
qGH (L − 1) D(L − 1)
where where D(k) = AU (χ U G (k))Ξ(k)AH B (χ BH (k)). Assum-

ing that K > dG , the UE angles are found by
solving AoA estimation problems on the L blocks
A (Ω) = [Z (0) A (χ (0)) · · · Z (L − 1) A (χ (L − 1))]
in (81).
(77a) In Stage 2, To2 additional OFDM training symbols are
AD = [Z (0) AD · · · Z (L − 1) AD ] (77b) transmitted, for a total of T2 = To2 Nc samples. As
in (70), the received signal in the bth symbol block (b = E. Single-Antenna Scenarios
1, . . . , To2 /LdG ) is given by
1) Single-Antenna UE: When the UE has only a single
√
L−1 antenna, χ U G = ∅ is the empty set and AU (χ U G ) = 1TdG .
Yb,s = P H (k) Φb,s GH (k) Xs−k + Nt,s The matrix A(χ) in (57a) still has dH dG columns, now
k=0 given by
= AB (ΩBH ) blkdiag {Vk }L−1 H
k=0 Bs + Nt,s (82a)
∗
Vk = ΓH (k) AH
R (χ RH (k)) Φb,s AR (χ RG (k)) ΓG (k) [A(χ)]:k = aR (χ RG, − χ RH,p ) ⊗ aB (ωBH,p ) (86)
√
(82b)
Bs = P X H
2,s AU (χ U G (0)) where = k/dH and p = moddH (k). We assume without

· · · XH (χ U G (L − 1)) loss of generality that xt = 1, ∀ t, so (50) simplifies to
2,s−L+1 AU (82c)
T
φ1 ⊗ IM
where Xs−k is K × LdG . Replacing χ U G (k) and χ BH (k)
.
with their estimates from Stage 1 and assuming M ≥ LdH Z= .. . (87)
and K ≥ dG , we multiply each Yb,s on the left by the φTT ⊗ IM
estimate of the pseudoinverse of AB (ΩBH ) and on the
right by the estimate of the pseudoinverse of BH
s , and we Ignoring the structure of γ GH in (57a), the DML criterion
stack the vectorized Vk together as in (62) can be applied to estimate the 3dh + 2dG − 2
spatial frequencies in χ , which represents only a slight
vec (V0 ) savings compared with the multiantenna UE case. The

.. simpler beamforming criterion can be used if one ignores
yB,b,s = . + nB,b,s
the relationship between the columns of A(χ) as in (64),
vec (VL−1 )
treating them as independent vectors that are a function
ΓGH (0) C (0) φb,s of three frequency variables: one for the BS and two for

..
= . + nB,b,s (83) the RIS. This results in a search for dH dG local maxima
ΓGH (L − 1) C (L − 1) φb,s−L+1 in a 3-D space. In the LoS case, the problem is solved by
optimizing a function of three frequency variables in either
the DML or beamforming approach.
where C(k) = C(χ RG (k), χ RH (k)). We stack the Nc vec-
The CS-based model in (57b) for single-antenna UEs can
tors defined in (83) together and take the transpose as
also be approached in two ways. The first retains the full
in (72)
geometric structure of the channel, with

T
YB,b

= YB,b,0 · · · YB,b,L−1 (84a) AD = ATRD AH
RD ⊗ ABD (88a)

Nc ×LdH dG Nc ×dH dG QGH = γ H
G QRG ⊗ QBH ΓH QRH .
T T
(88b)
YB,b,k Ψ∗b,k C (k) ΓGH (k) +NB,b,k
T
(84b)
1×NRD NBD ×NRD
Nc ×N

N×dH dG
Ψ∗b = Ψ∗b,0 · · · Ψ∗b,L−1 (84c) In this case, the dH dG -sparse vector qGH = vec(QGH )
has a similar structure as before, and the dictionary has
2
where Ψ∗b is block circulant with first set of columns NBD NRD elements. The second approach ignores the
defined by ΨTb,0 = [Ψb,1 ··· Ψb,Nc ]. Stacking the sparse structure as in (65), except in this case that the
result from all To2 /LdG training symbol blocks YTB = dictionary ZAD = Z(ARD ⊗ ABD ) has only NRD NBD
[YTB,1 · · · YTB,To2 /LdG ] and partitioning them into L blocks terms.
of dH dG columns each, we have Further complexity reduction is possible using the
decoupled approach described in Section IV-C2. In
∗ Stage 1, χ BH is estimated from Y1 as before, but
Ψ1,k

. estimation of χ U G is not required. Stage 2 proceeds as
ȲB,k .. CT (k)ΓGH (k) + NB,k (85)
before, but the solution is obtained in a different way. In
Ψ∗Nc ,k particular, in this case, we define
√
where ȲB,k holds columns k + 1 through k + L of YB . yt = P AB (χ BH ) γ H
G ⊗ ΓH C (χ RG , χ RH ) φt + nt
1 † T
This equation is equivalent in form to (72b), and thus, the
methods discussed previously can be used to solve for the YB = √ AB (χ̂ BH ) Y1 (89a)
P
remaining channel parameters for path k. The process is
Ψ∗ CT (χ RG , χ RH ) (γ ∗G ⊗ ΓH ) + NB (89b)
then repeated for all L paths, k = 0, . . . , L − 1.
and note that the dH columns of YB are now linear where the superscripts indicate the user index. UE u is
combinations of RIS array response vectors assumed to have Ku antennas and the corresponding
!
channel has dGu propagation paths so that dG = u dGu .

dG The composite channel and its vectorized form are
yB,k γH,k Ψ∗ γ ∗G,n aR (χ RG,n − χ RH,k ) + nB,k (90a) given by
n=1
γH,K Ψ∗ diag (AR (χ RG ) γ ∗G ) aR (−χ RH,k ) + nB,k . 1
G H
(90b)
.
Hc = .. (94a)
G H
U
This special structure allows for a simpler solution than 1
that required in the multiantenna UE case considered in hc

.
Section IV-C2 [21]. To see this, note that because of our hc = .. = Jvec (Hc ) (94b)
identifiability conditions γH,1 = 1 and χ RH,1 = 0, the first hUc
column of YB is given by
where the permutation matrix J is implicitly defined to

yB,1 Ψ∗ AR (χ RG ) γ ∗G + nB,1 . (91) group the blocks of hc by user rather than by RIS ele-
ments. Compact expressions for the multiuser case can
It is clear that the parameters χ RG and γ G can be be found for the general data model that parallel those
estimated from yB,1 using an AoA estimation algorithm in (57)
such as beamforming or DML or using a dG -sparse CS $ 1%
approach. Once estimated, these parameters can be A χ γ G1 H
√
substituted into (90), and yB,k can be used to estimate y = P ZJ .. +n (95a)
γH,k and χ RH,k for k = 2, . . . , dH . Thus, for Stage 2, $ U %.
instead of solving dH dG 1-sparse AoA estimation problems A χ γ GU H
that ignore the underlying structure of the data as in
Section IV-C2, for single-antenna UEs, we can estimate √ $ % γ G.1 H
= P Z J1 A χ 1 · · · JU A χ U .. +n (95b)
the channel with one dG -sparse estimation, followed by
γ
dH − 1 1-sparse problems, that when combined provide GU H
estimates of χ RH , χ RG , γ H , and γ G .
A(ω )

γ GH
2) Single-Antenna UE and BS: Here, there is no need for

qG1 H
a first stage, as there are no angles to estimate at the BS or √
.
= P Z [J1 AD · · · JU AD ] .. +n (95c)
UE. Instead, we simply collect T observations at the BS and
stack them together in a T × 1 vector y = [y1 · · · yT ]T , effective dictionary qGU H
which yields
qGH
y = Ψ∗ CT (χ RG , χ RH ) (γ ∗G ⊗ γ H ) + n (92a) where {χ u , γ Gu H , qGu H } are the angle parameters, chan-

∗

dG dH
∗ nel gains, and sparse vectors for UE u as in (55), respec-
=Ψ γG,k γH,n aR (χ RG,k − χ RH,n ) + n. (92b)
tively. The block column Ju of the permutation matrix
k=1 n=1
J = [J1 · · · JU ] is of dimension M KN × M Ku N . General
methods can be developed for simultaneous estimation of
This leads to an equivalent single-snapshot AoA estimation
all the channel parameters based on these equations and
problem, although with a nonlinear dependence on the
the approaches discussed in Section IV-C, but as before, the
gains via the Kronecker product. Ignoring this as before,
dimension of the resulting optimization problems is likely
the spatial frequencies can be determined using a 2dH +
prohibitive except for certain simple cases.
2dG − 2 dimensional DML search or a single dH dG -sparse
The decoupled approach of Section IV-C2 can be applied
CS problem with or without application of the special row
to reduce the estimation complexity. Provided that M ≥
structure induced by the model.
dH dG , the BS AoAs χ BH are estimated as before, while the
F. Multiuser Scenarios equation for estimating the UE AoDs is slightly different
Multiple users in the geometric model can be accounted than (69)
for by redefining G as follows:
AU (χ 1U G )
1 X1 YH
.. X1 NH
AU (χ 1U G )ΓG1 AH 1
R (χ RG ) G √1 = . S+ √1

.
.
T1 P T1 P
G= .. ≡ .. (93) AU (χ U
U G)
AU (χ U H U
GU (96)
U G )ΓGU AR (χ RG )
where S = ΞH AH
B (χ BH ) and For geometric channel models, this approach could be
implemented as follows [21].
H 1
AR (χ RG ) 1) Choose one antenna from one of the UEs, and trans-

. ∗ mit training data while the other antennas are silent
Ξ H
= ΓG .. Φ AR (χ RH )Γ∗H . (97)
to estimate the channel Hdiag(g1 ).
AH U
R (χ RG )
2) Transmit training data from the remaining UE anten-
nas, and estimate the UE AoDs χ kU G as in (69) if there
Assuming that Ku > dGu , UE AoD estimates can be is only one user with multiple antennas or as in (96)
found separately for each user by processing different if there are multiple multiantennas users.
block rows of (96). Once estimates of χ BH and χ uU G 3) Set the RIS reflection vector to a fixed value φ̄ and
are determined, we collect additional training data and send at least dG additional training symbols xt to
proceed as in (71) where obtain
U

B = blkdiag AH u
U (χ U G )u=1 X2 . (98) Y = HΦ̄AR (χ RG ) ΓG AH
U (χ U G ) X + N. (101)
†
U (χ̂ U G )X) . Since
Stacking the vectorized block columns of (71) together, we Then, multiply on the right by (AH
have we have an estimate of H, the resulting matrix is
$ $ 1 % 1,∗ % approximately in the standard form for AoA esti-

vec ΓH AH R (χ RH ) Φb AR χ RG ΓG mation, using either steering vectors drawn from
.. ĤΦ̄AR (χ RG ) or an overcomplete dictionary ĤΦ̄ARD .
yB,b = . + nB,b
$ U % U,∗ Use of the geometric model offers a further opportunity
vec ΓH AR (χ RH ) Φb AR χ RG ΓG
H
T$ 1 % H for dramatic reductions in both computational complexity

AR χ RG AR (χ RH ) and training overhead. The AoAs and AoDs in the geo-

= ΓGH ..
. φb + nB,b . (99) metric model change very slowly and can be considered
$ % to be stationary over multiple coherence blocks. Only the
AR χ RG AR (χ RH )
T U H
complex gains γ G and γ H change substantially from block
C(ω 1
RG
,...,ω U
RG
,ω RH ) to block. This suggests that, once estimated, the spatial
frequencies for subsequent blocks can be considered con-
Collecting the vectors from (99) together as columns of a stant, and only the gains need to be reestimated [21].
matrix as in (72) leads to For CS-based methods, this means that the support of the
sparse solution is already known. Consequently, with at
least dH dG training samples, the channel gains in (57),
YB Ψ∗ C χ 1RG , . . . , χ U
RG , χ RH ΓGH + NB . (100)
(76), or (95) can be estimated using LS, e.g., for (57a),
we have
As before, each column of (100) involves only the differ-
ence between one spatial frequency from {χ 1RG , . . . , χ U
RG }
1
γ̂ GH = √ (ZA(χ̂))† y. (102)
and one from χ RH . There are dH dG such combinations, P
and thus, the remainder of the channel parameters can be
found by solving dH dG single spatial frequency estimation V. U N S T R U C T U R E D V E R S U S
problems. Note also that when the UEs have only a sin- GEOMETRIC MODELS:
gle antenna, the simplification discussed in Section IV-E1 THE TRADEOFFS
holds, where only one dG -dimensional AoA estimation fol- The choice of channel model and estimation algorithm
lowed by dH − 1 1-D AoA estimation problems is necessary. for RIS-based systems depends on many factors, including
estimation accuracy, training overhead, algorithm com-
plexity, and robustness to the modeling assumptions. We
G. Reducing the Complexity and will examine these factors in this section and then focus
Training Overhead on estimation performance for some specific scenarios in
The methods discussed in Section III-E can be used to Section VI. The discussion below in this section will assume
further reduce the training required even for geometric the case of either a single K -antenna UE or K single-
channel models. The availability of prior knowledge of antenna UEs.
low-rank channel covariance matrices is useful for geo-
metric models, although less for reducing the amount of
training than for designing the pilot symbols xt and RIS A. Estimation Accuracy
training φt to improve the received SNR. Also, the common Rather than study the performance of individual algo-
channel H associated with multiple UE antennas can be rithms, of which there are too many examples to consider
exploited as before to reduce the algorithm complexity. comprehensively in this article, we can get a sense for
what is achievable by examining performance bounds see [19] for a detailed derivation for the special case
based just on the models themselves. The most common of a single-antenna BS and UE). When one forms an
T
example is the CRB of the composite channel h̆c = estimate of the composite channel hc from the estimate
[Re(hc )T Im(hc )T ], where Re(hc ) and Im(hc ), respectively, of the parameters in η , the resulting CRB for the channel
represent the real and imaginary parts of hc . is given by
In the unstructured case, the parameters to be estimated
are the 2M KN elements of the channel itself: η = h̆c . T
∂ h̆c ∂ h̆c
Assuming temporally and spatially white Gaussian noise CRBs h̆c = CRB(η) (109)
∂η ∂η
with variance σ 2 as in (1) and (5), the log likelihood is
given by
where the subscript s is for “structured.” Simplification
1 √ 2 of (109) for the general RIS-based model in (57a) is still
fu (h̆c ) = −M T ln(πσ 2 ) − y− P Zhc (103) an open problem, but some general observations based
σ2
on studies from traditional AoA estimation can be made
where the subscript u denotes “unstructured.” The CRB is (e.g., see [56]). In particular, like the unstructured case,
defined in terms of the Fisher information matrix (FIM) CRB(η) and, hence, CRBs (h̆c ) are inversely proportional
to T and the SNR P/σ 2 . However, unlike the unstructured
case, whose performance cannot be improved by increas-
CRBu h̆c = FIM−1
u h̆c (104)
ing the number of antennas (except in the sense that
& T * T must increase with N to maintain identifiability), the
'
( ∂fu h̆c ∂fu h̆c '
+
FIMu h̆c = E (105)
geometric CRB will improve as either M or N increases.
' ∂ h̆c
) ∂ h̆c '
, For example, the CRB for estimation of the AoAs at the
BS decreases proportionally with M 3 . In Section VI, an
example will be given to show the decrease with N .
where E(·) denotes expectation with respect to the noise The examples in Section VI all show that the geomet-
distribution. It is straightforward to show that the CRB for ric model provides a much smaller channel estimation
the unstructured model is given by error for all of the scenarios considered. However, as
discussed in Section V-C, some of this performance advan-
σ 2 T −1 tage will inevitably be lost in practice due to modeling
CRBu (h̆c ) = (Z̆ Z̆) (106)
2P errors.
For large arrays such as those considered in RIS-based
where systems, the CRB is achievable by standard AoA estimation
algorithms under the ideal model, even for small (nonas-
Re(Z) −Im(Z) ymptotic) values for T [44]. Such algorithms estimate
Z̆ = (107)
Im(Z) Re(Z) the spatial frequencies as continuous variables, unlike
CS-based methods that employ a dictionary on a finite grid.
and Z is defined in (5). If Z is designed to be orthogonal as Since the density of the grid cannot be made arbitrarily
in Section III-A1, then the CRB simplifies to fine due to the resulting computational complexity and
ill-conditioning, for CS-based AoA estimation algorithms,
the grid spacing becomes the limiting factor rather than
σ2
CRBu (h̆c ) = I2M K(N+1) (108) the noise, and thus, the CRB is an optimistic performance
2P T
indicator. The examples in Section VI will demonstrate
signifying that the lower bound is identical for every that even with the loss due to the use of an AoA grid,
element of h̆c (note that we have assumed the RIS element the geometric model still provides a sizable performance
gains are β = 1). An important observation is that the advantage.
CRB in the unstructured case is independent of the number
of BS antennas, RIS elements, and UEs (although T ≥ B. Training Overhead
K(N + 1) must hold for the model to be identifiable and, When considering the amount of training required for a
hence, for the FIM to be invertible). given model, there are two aspects to consider: the mini-
The CRB for the geometric channel model depends on mum number of samples required for unique identification
the parameter vector η defined by the angles and gains of the channel model and the number of samples needed
listed in Table 1 and is of much smaller dimension than for to obtain some desired level of performance. In many
the unstructured model when the channels are dominated cases, these two quantities may be quite different. Some
by a small number of propagation paths. The CRB for performance-related issues were discussed above and will
the structured η is more difficult to compute, requiring a be illustrated in Section VI, so here, we focus on the
large number of tedious derivative calculations which we uniqueness of the models. A more comprehensive analysis
do not include here (for an example of such derivations, of the impact of the training would account for its impact
on the system throughput, both in terms of time lost for C. Robustness

data transmission and the impact of the resulting channel While geometric methods typically enjoy an advantage
estimation errors on the achievable rate (see [57] for a in terms of training overhead, their reliance on a more
recent study that addresses this question). detailed parameterized model makes them subject to per-
We know from our discussion in Section III-A1 that formance degradation when the actual system deviates
identifiability of the unstructured model requires T ≥ from the assumed model. In particular, the geometric
K(N + 1) pilot symbols. Establishing a similar condi- model requires that one determine the number of prop-
tion for the general geometric model in (57) is difficult agation paths in each link and relies on prior knowledge
since it amounts to a single-snapshot multidimensional of the array response or calibration in order to determine
frequency estimation problem, with a complicated inter- the AoAs and AoDs. Errors in determining the model order
relationship between the frequency dimensions. A simple or inevitable deviations of the array response from the
dimensionality analysis gives the necessary condition that nominal model will lead to a corresponding performance
the number of observations M T must exceed the number loss. To get a sense for the magnitude of this loss compared
of parameters to be estimated, which is 5dH + 5dG − 4. For to that due to background noise, consider the following
typical values of dH and dG assumed for sparse channels simple model involving a single source with spatial fre-
and a large BS array, this condition could be satisfied quency ω that transmits T pilot symbols:
for T as small as 1 or 2. However, the complexity of
directly using (57) for estimating the channel parameters is √
Y= P (a(ω) + ã) x + N (110)
prohibitive due to the nonconvex optimization required for
χ based on (57a) using classical AoA estimation methods
or the huge dictionary needed by CS methods for the where Y and N are M × T matrices containing the received
model in (57b). signal and noise, respectively, x is a 1 × T vector of pilot
Instead, consider the simpler decoupled approach dis- data, and ã is a vector that represents the deviation of
cussed in Section IV-C2, which assumes the more stringent the M × 1 array response from the nominal model a(ω).
case for identifiability with a single K -antenna UE where A conventional matched filter beamformer steered to angle
the channel depends on the dG AoDs at the UE. This ω produces the following statistic:
approach assumes K ≥ dG + 1 and a training interval
√
divided into two parts: T = T1 + T2 . For large enough 1 H √ P H 1 H
a (ω)YxH = P + H
M , the first stage requires at least T1 ≥ K pilot symbols to MT M a (ω)ã + M T a (ω)Nx
estimate the BS AoAs and the UE AoDs, while the second desired
ζa ζn
stage requires that T2 /dG ≥ 4 in order to uniquely estimate (111)
the 2-D angle and gain for each of the dH dG columns
defined in (73) [58]. Together, this implies that a minimum
assuming unit amplitude symbols and antenna gains. For
of T = 5K = 5dG + 1 pilots are required to identify
simplicity, assume that the elements of ã are uncorrelated
the composite channel. Note that for the case where there
random perturbations with variance σa2 . Then, the variance
are K single-antenna UEs rather than one K -antenna UE,
of the calibration error term ζa is σa2 P/M , while that of
the condition K ≥ dG + 1 can be eliminated since the
the noise term ζn is σ 2 /M T . Thus, the impact of the
estimation of χ U G is no longer required. In this case, the
array calibration error boils down to the relative size of
minimum number of training samples is T = K + 4dG . In
σa2 compared with σ 2 /P T . For example, if we assume that
either case, the training overhead is considerably less than
the antennas are calibrated to within a 5% accuracy (σa =
for the unstructured model.
0.05), then the array calibration errors would dominate
In addition to the identifiability conditions that must
the effects of the noise when T > 400 for an SNR of 0
be satisfied for unique AoA/AoD estimation, CS-based
dB and when T > 125 for an SNR of 5 dB. This simple
methods must also satisfy conditions that will guarantee
analysis suggests that array calibration errors, while not
with high probability the existence of the desired sparse
negligible, are only an important source of error at higher
factorization. Unfortunately, such conditions are not possi-
SNRs not typical of scenarios envisioned for RIS-based
ble to specify precisely. For the decoupled approach with a
systems (especially those at millimeter wave and higher
single K -antenna UE, sparse estimation of QBH in (67c)
frequencies). However, further study is needed to confirm
requires M ≥ O(dH log(NBD /dH )), while estimation of
this result for the more complicated channel conditions
QU G in (67c) and (69) requires K ≥ O(dG log(NU D /dG )),
when RISs are involved.
neither with a specific requirement for T1 . Stage 2, on the
other hand, requires T2 ≥ dG O(log(NRD )) to find a sparse
factorization for the 1-sparse problems in (69), which is D. Algorithm Complexity
still likely much smaller than for the unstructured model. Another important factor to consider in choosing the
As before, the condition K ≥ O(dG log(NU D /dG )) can type of channel model is the complexity required to com-
be dropped when the scenario involves K single-antenna pute the channel estimate. For the LS approach described
users. in (13)–(15) and implemented with orthogonal X and Ψ,
the required number of operations is O(M (K + N )T ) ≈

O(M N T ), where the approximation is due to the fact
that typically, N K . Since T = O(KN ) is needed for
identifiability, the complexity order is O(M KN 2 ), which is
quadratic in the number of RIS elements. If nonorthogonal
RIS phase sequences are employed, the computation of
the pseudoinverse of ΨH requires O(N 3 ) operations in
addition to the O(M KN 2 ) needed to calculate the channel
estimate, and hence, the relative values of M K and N
determine the overall computational load.
For the geometric case, tackling the estimation problem
implied by the general data model in (57a) via classical
AoA estimation methods is only practical for very small
values of dH and dG . The sparsity-based model in (57b)
is difficult to implement even for small dH and dG since
2
the size of dictionary is NBD NU D NRD , which is usually
Fig. 2. Channel estimate NMSE and CRB for various channel
of order O(M KN 2 ) if we assume that the resolution of
models versus SNR. Channel parameters are
the grid points is inversely proportional to the number M 30, N 30, T 31, K 1, dF 5, dG 5, and dH 2.
of antennas on each side of the link. Since the overall
complexity order would involve the product of the dictio-
nary size with the sparsity level dH dG and the number of
measurements M T , it is clear that this direct approach is points for the RIS. The performance metric adopted in
not feasible. all numerical examples is the average of the diagonal
Instead, we again consider the decoupled approach of elements of the CRB over a large number of different
Section IV-C2 as a practical alternative. We will focus on (geometric) channel realizations with random AoAs/AoDs
the complexity of CS-based approaches since they have and path gains.
been most frequently considered for RIS-based channel The first case considers a single-UE with K = 1 antenna,
estimation. For a single K -antenna user, the steps involved orthogonal pilot signals, and uniformly distributed RIS
in the decoupled approach are a dH -sparse estimation with phases during training. The AoAs/AoDs were generated
an M ×NBD dictionary, followed by a dG -sparse estimation using a uniform distribution over [−90◦ , 90◦ ] for azimuth
with a K × NU D dictionary, and concluding with dH dG and [0◦ , 90◦ ] for elevation. The path gains were generated
1-sparse problems with an N × NRD dictionary. The total as unit-variance Rayleigh random variables. Since the path
complexity is thus of order gains are independent of the number of paths, a channel
with more paths will generally make a larger contribution
O (dH M NBD ) + O (dG KNU D ) + O (dH dG N NRD ) to the overall SNR. The path gain distribution is the same
$ % $ % $ % for all three channels Hd , H, G, which corresponds to a case
≈ O dH M 2 + O dG K 2 + O dH dG N 2 (112)
where the BS, RIS, and UE are approximately located at
the vertices of an equilateral triangle with similar propaga-
where our approximation again assumes grid resolution tion characteristics. Also, this assumption means that even
on the order of the reciprocal of the number of anten- though during channel estimation the RIS phases Φt are
nas. We see that the decoupled CS-based approach will not chosen to maximize the coherent gain offered by the
have a lower complexity, provided that dH KN 2 /M , RIS, the RIS provides sufficient combining gain such that
dG M N 2 /K , and dH dG M K , all of which would HΦt G Hd . The SNR in the following examples is
be considered reasonable for large RIS and BS arrays. The defined as P/σ 2 , but the effective SNR will be a function
computational load is lower for the case where there are of the channel gain Hd + HΦt G , which will change as
K single-antenna users, based on the approach described a function of the number of antennas, RIS elements, and
in Section IV-E1. path gains.
Figs. 2–4 assume a BS with M = 30 antennas and an RIS
VI. N U M E R I C A L E X A M P L E S with N = 30 elements. Fig. 2 plots the performance as a
The examples in the following compare CRBu (h̆c ) and function of SNR for a case with dH = 2, dF = dG = 5, and
CRBs (h̆c ) for several different scenarios. The first set T = 31 training samples, which is the minimum required
of examples also includes the performance of the LS for the case of a K = 1 antenna UE. The CRB for the
algorithm and an implementation of the decoupled esti- unstructured model and the geometric model with and
mator for the geometric channel model described in without the direct BS-UE channel Hd is included, along
Section IV-C2, referred to as “CS-Geom” in the plots and with the performance of the LS and CS-Geom algorithms.
discussion next. The CS optimizations were implemented The CRB curves are plotted with solid lines, while the
using OMP on a grid with M points for the BS and N performance of the algorithms is plotted with dashed lines.
Fig. 3. Channel estimate NMSE and CRB for various channel Fig. 5. Channel estimate NMSE and CRB for various channel
models versus number of training samples T. Channel parameters models versus number of RIS elements N, where M N 70. Channel
are M 30, N 30, SNR 5 dB, K 1, dF 5, dG 5, and dH 2. parameters are SNR 5 dB, K 2, dF 5, dG
5, and dH 2. For the
unstructured model and the blue curves (geometric model), we have
T KN. For the red curves (geometric model), T 30.
We see that the performance of the LS algorithm exactly

matches the unstructured CRB, while for CS-Geom, there
is a significant gap between the bound and the accuracy channel is present; this is due to the observation above
of the algorithm. This is due to a number of factors, the that for the assumed model, we have HΦt G Hd . In
most important being the rather coarse grid that was used effect, when Hd is present, we are faced with the task of
to estimate the AoAs and AoDs. Despite this gap, the CS- estimating additional parameters that are observed more
Geom approach still provides a 10-dB improvement in weakly in the data.
channel estimation accuracy across all SNRs. It is impor- Fig. 4 plots the performance as a function of the number
tant to note that this gap can be closed considerably using of paths present in the geometric channel model, again
a technique that estimates the AoAs/AoDs as continuous assuming M = N = 30 and the minimum number of
variables. T = K(N + 1) = 31 training samples. The unstructured
The advantage of the geometric model in terms of model is unaffected by these changes, and both the CRB
training data is apparent from Fig. 3 for the same scenario and LS algorithm show constant accuracy versus the num-
with an SNR of 5 dB; we see that the same performance ber of paths. While the presence of more paths increases
can be achieved with more than order of magnitude fewer the power of the received signals in this example, this
training samples. Note that the plots show degradation in gain is offset by the fact that more parameters must be
geometric channel estimation performance when the direct estimated, and hence, the geometric CRB and normalized
mean squared error (NMSE) of CS-Geom increase with the
number of paths. It is clear from this example that the
geometric model can include a relatively large number of
paths before its performance degrades to the level of the
unstructured case.
The last example for the first scenario is shown in Fig. 5,
which shows the channel estimation performance as a
function of N assuming that M + N = 70, with the other
parameters set as before (K = 2, dF = 5, dG = 5, dH = 2,
and SNR = 5 dB). The curve for the unstructured model
assumes that T is increasing with N according to T = KN ,
which explains why the performance improves as the size
of the RIS increases. The blue curves for the geometric
model assume the same increasing values for T , while the
red lines assume a fixed value of T = 30. For fixed T , the
best performance is achieved when M N , while larger
values of T favor the use of a larger RIS and a smaller
Fig. 4. Channel estimate NMSE and CRB for various models
number of BS antennas.
versus number of propagation paths. Channel parameters are The second scenario is different from the first in a
M 30, N 30, T 31, K 1, and SNR 5 dB. number of ways, but the conclusions are essentially the
Fig. 6. CRB for various channel models versus SNR. Channel

parameters are M 6, N 64, T 260, U 2, K 4, dF 2, dG 3,
and dH 2. Fig. 8. CRB for various models versus number of propagation
paths dG . Channel parameters are M 6, N 64, T 260, U 2,
K 4, SNR 0 dB, dF 2, and dH 2.
same, so for these plots, we include only the CRB results.

Unlike the previous case, there are U = 2 users with 2
antennas each (K = 4), the RIS training sequences are bound for unstructured model (independent of Hd );
not random but rather chosen such that the diagonal CRB (4) “Geom. Hd only”—bound for Hd assuming geometric
in (108) holds, the BS has M = 6 antennas, and there are model when no RIS is present; and 5) “Geometric no
dH = dF = 2 paths in the BS-RIS and BS-UE channels. Hd ”—bound for Hc assuming geometric model without Hd
Except for Fig. 8, dG = 3. In addition, the Rayleigh present. In all cases, the results for “Geometric Hd ” and
distributed path gains in this example have a variance “Geom. Hd only” are essentially identical, indicating that
equal to the reciprocal of the number of paths. Figs. 6–8 the presence or absence of the RIS should not impact the
assume a larger RIS with N = 64 elements, and thus, the quality of the estimate of Hd .
LS channel estimator requires at least T = K(N +1) = 260 Fig. 6 shows the CRB versus the SNR, and as in the
training samples in order to obtain a unique estimate. This previous case, we see an estimation gain of about 15 dB
value for T will be assumed, except for the case where the for the geometric model in the ideal case. Achievable
performance is plotted versus T . performance versus T is shown in Fig. 7 for 0-dB SNR.
Figs. 6–9 each show the CRB for five different cases: The curve for the unstructured case is only shown for T ≥
1) “Geometric Hd ”—bound for estimate of Hd assum- 260 where the channel is identifiable, and as before, there
ing geometric model and with the RIS channel present; is at least an order of magnitude reduction in training data
2) “Geometric Hc ”—bound for estimate of Hc using required for the geometric model. Fig. 8 shows the results
geometric model with Hd included; 3) “Unstructured”— versus dG for 0-dB SNR, and we still see a considerable gap
between the achievable performance of the unstructured
and geometric models even as the number of propagation
paths grows larger. The CRB versus the number of RIS
elements N = Nx2 for 0-dB SNR is shown in Fig. 9 with
T = K(N + 1) = 788 in order for the unstructured model
to be identifiable for the case of the largest RIS.
VII. E M E R G I N G T O P I C S
A. Alternatives to Passive Reflection-Only RIS
The methods discussed above have all assumed a
passive reflecting RIS with elements that only apply a
tunable phase shift to the impinging signal, together
with a phase- and frequency-dependent attenuation fac-
tor. Recently, several new RIS devices and architectures
have been introduced, which include active receivers, local
baseband processing, or the ability to perform passive
Fig. 7. CRB for various channel models versus number of training “transmission” through the RIS (akin to refraction) rather
samples T. Channel parameters are M 6, N 64, SNR 0 dB, U 2, than just reflection. Proponents argue that the resulting
K 4, dF 2, dG 3, and dH 2. benefits of such implementations outweigh the resulting
and receive (STAR)-RIS is another recently proposed RIS

architecture in which each element can split the received
energy into reflected or refracted (“transmitted”) signals
in a controllable way. LS-based channel estimators for an
STAR-RIS system are presented in [67] for two cases, one
where the elements time-switch between reflection and
refraction (a simple extension of existing methods) and
one where the received energy is split between reflection
and refraction. We also mention here the related work
of [68], which, instead of active RIS elements, proposes
deploying active anchor nodes in known locations near the
UEs to reduce the CSI estimation complexity as discussed
in Section IV-G and resolve the ambiguity of the composite
channel.
Fig. 9. CRB for various channel models versus number of columns

B. Double RIS Systems
Nx in a square RIS with N N2
x elements. Channel parameters are
M 6, T 788, SNR 0 dB, U 2, K 4, dF 2, dG 3, and dH 2. We have only considered communication links between
a BS and UEs that include a single RIS, but channel
estimation for scenarios with two RIS has been studied
increase in hardware complexity and power consumption
for certain special cases. In [69], a double RIS scenario
at the RIS. The work focusing on active RIS elements has
is considered where a BS communicates with a single UE
assumed only active receivers and not active transmission,
over an indirect channel that passes from the BS to RIS
and hence, the additional required power is still signifi-
1 to RIS 2 to the UE, assuming that all other channels
cantly less than a similarly sized active relay. Instead of the
are blocked. They propose an unstructured LS channel
lower SNR data at the BS that has propagated from the UEs
estimator that requires at least N1 N2 training samples,
and is reflected by the RIS, data collected by an active RIS
assuming that the two RISs have N1 and N2 elements.
element will have higher SNR due to lower path loss and
They also consider the special case of an LoS channel
could lead to higher fidelity channel estimates. In addition,
between the two RIS, which reduces the training overhead
data from the active receivers would allow the channels
to N1 + N2 . A more general scenario is considered in [70],
H and G to be separately identified, which as we have
in which both RIS have unblocked channels with the BS
seen above is useful in cases where one component varies
and UEs, in addition to the direct link between the two
less rapidly than the other and need not be estimated as
RIS. A two-stage unstructured LS approach is proposed
frequently.
where, in the first stage, the phase shifts of RIS 1 are held
The most common channel estimation scenario con-
fixed, while those of RIS 2 are varied, and the composite
sidered thus far involves an RIS with a few dedicated
single-bounce (UE-IRS 2-BS) and double-bounce (UE-IRS
active receivers whose collected data are used to estimate
2-IRS 1-BS) channels are estimated. In Stage 2, the phase
the gains and angles of the geometric channel model.
of both RIS is varied in order to estimate the corresponding
These estimates are then used in turn to infer the full H
components of RIS 1, which are superimposed on the
and G channel matrices using the known RIS geometry
composite channels estimated in Stage 1. Assuming K
[59]–[61] or a deep neural network (DNN) [62]. This
single-antenna UEs, the required training overhead for this
matrix completion approach is similar to methods pro-
approach is of the following order:
posed for channel estimation in hybrid digital/analog sys-
tems. A tensor completion approach is proposed in [63]
KN1 N2
and also extended to the wideband OFDM case. The active T =O + K (N1 + N2 ) (113)
M
elements can either be embedded in fixed positions with
a separate RF chain for each as in the methods above
which can be large when the number of BS antennas M
or different linear combinations of the RIS outputs can
is small. An earlier approach to the double-RIS problem
be combined together through a single RF chain over the
in [71] assumes LoS propagation between two RIS that
training interval to achieve a similar result [64].
both possess active receivers, which allowed the individual
A more general hybrid architecture is proposed in [65]
channel components to be determined. It is unclear if a
and [66], where each RIS element is designed to act as
geometric model could be used for a purely passive double-
both a passive reflector and an active receiver. In such
RIS scenario to achieve an identifiable parameterization.
designs, not only is the reflected phase a tunable parameter
but also the fraction of the power absorbed at each element
and applied to the receiver. The development of channel C. Learning-Based Methods
estimation methods that exploit this additional flexibility The application of machine learning to RIS systems and
is an interesting open problem. Simultaneous transmit channel estimation has also been growing recently. For
Table 2 Summary of Topics Covered by References Related to Channel Estimation
example, the approach in [72] designs a DNN that directly described above, such as LS [35], [74] or CS in the geo-
takes the received training data and produces the channel metric model [75]. Multistage DNN denoising networks
estimate. The DNN is trained using synthetic uplink data are proposed in [76] and [77], where separate DNNs
generated, assuming that the direct and composite chan- are used to first denoise the direct channel (if present)
nels undergo Rayleigh fading. As has been noted in other with the RIS switched OFF, then to denoise a reduced-
work, the performance improves if not only the real and dimension version of the composite channel with only a
imaginary parts of the input data are provided but also a subset of the RIS elements active, and finally to map the
third component related to the data, in this case its magni- sampled composite channel to its full dimensions. A similar
tude. Other work has employed the phase of the data as the technique is proposed in [62] that uses a DNN to map
third component [73]. However, Elbir and Coleri [73] used the channel estimated with a few active RIS elements to
a federated learning approach in the downlink, in which the full RIS array. The approach in [60] also assumes
the UEs generate local channel estimates using a learning- the availability of active RIS receivers, but instead of
based optimization and the gradients of the local networks attempting to estimate the channel with the full RIS, the
are fed back to the BS to update the global model. sampled channels from the active RIS subarray are used as
A common learning-based approach for RIS systems “environment descriptors” during training to find the best
involves denoising an initial channel estimate with a DNN. codebook of RIS phases that yield the highest sum rate.
The initial estimate can be found using the methods Finally, we mention here the approach described in [78],
which uses deep unfolding of the LS fit of the channel to we have also briefly discussed on-going research related
the data with a nuclear norm regularization to promote a to RIS with active elements, double-RIS systems, and
low-rank estimate. machine learning algorithms. In Table 2, we summarize
The primary advantage of the above learning-based the references related to channel estimation that we have
methods is that, once the neural networks are trained, discussed in this work, and to aid those seeking more
the channel estimates are obtained with relatively little details, we indicate whether these papers consider channel
computation. However, since the training is performed estimation for geometric channel models, wideband chan-
using synthetic data, the ability of the network to handle nel models, multiple UEs, multiantenna BS, multiantenna
nonidealities not present in the simulated training data is UEs, active receivers, double RIS, and learning-based
still an open question. algorithms.
In addition to these emerging topics discussed in
VIII. C O N C L U S I O N A N D F U T U R E W O R K Section VII, many open research problems related to
Wireless channel estimation for RIS-based systems pro- RIS remain. For example, while we have considered
vides a rich source of interesting research problems. We only relatively simple scenarios involving a single RIS,
have highlighted some of the solutions to these problems a challenging question to answer is how to efficiently
for two general classes of channel models: unstructured coordinate the estimation of channels for multiple RIS
models that make no assumptions about the propagation when they are visible to multiple BSs and their reflections
environment or the RIS or array geometries, and struc- interact. We have considered only simple models for RIS
tured or geometric models that rely on the assumption of element behavior, but future work should investigate the
sparse propagation paths and calibrated antenna arrays implications of more realistic models that account for
and RIS element responses. Algorithms for estimating the coupled dependence of the RIS element gains and
unstructured channels are conceptually simple and robust, phases, their variation in frequency for wideband sce-
but for RIS-based systems, they require a large training narios, mutual coupling, nonlinear properties, and so on.
overhead, they have a higher computational complexity, More work is also needed to further reduce the training
and their achievable accuracy is limited due to the large overhead and computational complexity associated with
number of channel coefficients that must be estimated. RIS channel estimation. Possible topics include reducing
On the other hand, geometric channel models lead to the the complexity of CS-based approaches that normally
estimation of many fewer parameters and hence have a require large dictionaries for RIS-based CSI estimation,
much smaller training burden, lower overall complexity, exploiting available side information such as knowledge
and can achieve dramatically better performance. How- of the local propagation environment or known channel
ever, this improvement comes at the cost of determining statistics for optimal design of pilots and RIS phases dur-
the model order (i.e., number of propagation paths), and ing training, and advancing the application of learning-
some of the performance gains will be lost in practice due based methods. Finally, a systematic study of the impact of
to the use of grid-based estimators and inevitable mod- modeling errors on RIS-based systems is needed to more
eling errors. We have highlighted these issues using both rigorously establish the potential benefits of geometric
theoretical derivations and numerical CRB examples, and models.
REFERENCES
[1] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, progressive refinement,” IEEE J. Sel. Areas no. 6, pp. 4144–4157, Mar. 2021.
M. Alouini, and R. Zhang, “Wireless Commun., vol. 38, no. 11, pp. 2604–2620, [12] Y. Wei, M.-M. Zhao, M.-J. Zhao, and Y. Cai,
communications through reconfigurable intelligent Nov. 2020. “Channel estimation for IRS-aided multiuser
surfaces,” IEEE Access, vol. 7, pp. 116753–116773, [7] Q.-U.-A. Nadeem, H. Alwazani, A. Kammoun, communications with reduced error propagation,”
2019. A. Chaaban, M. Debbah, and M.-S. Alouini, IEEE Trans. Wireless Commun., vol. 21, no. 4,
[2] M. Di Renzo et al., “Smart radio environments “Intelligent reflecting surface-assisted multi-user pp. 2725–2741, Apr. 2022.
empowered by reconfigurable intelligent surfaces: MISO communication: Channel estimation and [13] B. Zheng, C. You, and R. Zhang, “Fast channel
How it works, state of research, and the road beamforming design,” IEEE Open J. Commun. Soc., estimation for IRS-assisted OFDM,” IEEE Wireless
ahead,” IEEE J. Sel. Areas Commun., vol. 38, no. 11, vol. 1, pp. 661–680, 2020. Commun. Lett., vol. 10, no. 3, pp. 580–584,
pp. 2450–2525, Nov. 2020. [8] C. Hu, L. Dai, S. Han, and X. Wang, “Two-timescale Mar. 2021.
[3] T. L. Jensen and E. De Carvalho, “An optimal channel estimation for reconfigurable intelligent [14] L. Wei, C. Huang, G. C. Alexandropoulos, Z. Yang,
channel estimation scheme for intelligent reflecting surface aided wireless communications,” IEEE C. Yuen, and Z. Zhang, “Joint channel estimation
surfaces based on a minimum variance unbiased Trans. Commun., vol. 69, no. 11, pp. 7736–7747, and signal recovery in RIS-assisted multi-user MISO
estimator,” in Proc. IEEE ICASSP, Barcelona, Spain, Nov. 2021. communications,” in Proc. IEEE Wireless Commun.
May 2020, pp. 5000–5004. [9] Z. Wang, L. Liu, and S. Cui, “Channel estimation for Netw. Conf. (WCNC), Nanjing, China, Mar. 2021,
[4] B. Zheng and R. Zhang, “Intelligent reflecting intelligent reflecting surface assisted multiuser pp. 1–6.
surface-enhanced OFDM: Channel estimation and communications: Framework, algorithms, and [15] J. Chen, Y.-C. Liang, H. V. Cheng, and W. Yu,
reflection optimization,” IEEE Wireless Commun. analysis,” IEEE Trans. Wireless Commun., vol. 19, “Channel estimation for reconfigurable intelligent
Lett., vol. 9, no. 4, pp. 518–522, Apr. 2020. no. 10, pp. 6607–6620, Oct. 2020. surface aided multi-user MIMO systems,”
[5] H. Liu, X. Yuan, and Y.-J.-A. Zhang, [10] B. Zheng, C. You, and R. Zhang, “Intelligent Dec. 2019, arXiv:1912.03619.
“Matrix-calibration-based cascaded channel reflecting surface assisted multi-user OFDMA: [16] G. T. de Araujo, A. L. F. de Almeida, and R. Boyer,
estimation for reconfigurable intelligent surface Channel estimation and training design,” IEEE “Channel estimation for intelligent reflecting
assisted multiuser MIMO,” IEEE J. Sel. Areas Trans. Wireless Commun., vol. 19, no. 12, surface assisted MIMO systems: A tensor modeling
Commun., vol. 38, no. 11, pp. 2621–2636, pp. 8315–8329, Dec. 2020. approach,” IEEE J. Sel. Topics Signal Process.,
Nov. 2020. [11] L. Wei, C. Huang, G. C. Alexandropoulos, C. Yuen, vol. 15, no. 3, pp. 789–802, Apr. 2021.
[6] C. You, B. Zheng, and R. Zhang, “Channel Z. Zhang, and M. Debbah, “Channel estimation for [17] Z. Wan, Z. Gao, and M.-S. Alouini, “Broadband
estimation and passive beamforming for intelligent RIS-empowered multi-user MISO wireless channel estimation for intelligent reflecting surface
reflecting surface: Discrete phase shift and communications,” IEEE Trans. Commun., vol. 69, aided mmWave massive MIMO systems,” in Proc.
IEEE Int. Conf. Commun. (ICC), Dublin, Ireland, solutions,” IEEE Open J. Commun. Soc., vol. 2, vol. 9, no. 11, pp. 1845–1849, Nov. 2020.
Jun. 2020, pp. 1–6. pp. 471–487, 2021. [56] P. Stoica and N. Arye, “MUSIC, maximum
[18] P. Wang, J. Fang, H. Duan, and H. Li, “Compressed [36] S. Sumanthiran, D. Kudathanthirige, likelihood, and Cramer–Rao bound,” IEEE Trans.
channel estimation for intelligent reflecting K. T. Hemachandra, T. Samarasinghe, and Acoust., Speech Signal Process., vol. 37, no. 5,
surface-assisted millimeter wave systems,” IEEE G. A. Baduge, “Rank-1 matrix approximation-based pp. 720–741, May 1989.
Signal Process. Lett., vol. 27, pp. 905–909, 2020. channel estimation for intelligent reflecting [57] A. Zappone, M. Di Renzo, F. Shams, X. Qian, and M.
[19] S. Noh, H. Yu, and Y. Sung, “Training signal design surface-aided multi-user MISO communications,” Debbah, “Overhead-aware design of reconfigurable
for sparse channel estimation in intelligent IEEE Commun. Lett., vol. 25, no. 8, pp. 2589–2593, intelligent surfaces in smart radio environments,”
reflecting surface-assisted millimeter-wave Aug. 2021. IEEE Trans. Wireless Commun., vol. 20, no. 1,
communication,” IEEE Trans. Wireless Commun., [37] E. Björnson and L. Sanguinetti, “Rayleigh fading pp. 126–141, Jan. 2021.
vol. 21, no. 4, pp. 2399–2413, Apr. 2022. modeling and channel hardening for reconfigurable [58] X. Liu and N. D. Sidiropoulos, “Almost sure
[20] X. Ma et al., “Joint channel estimation and data rate intelligent surfaces,” IEEE Wireless Commun. Lett., identifiability of constant modulus
maximization for intelligent reflecting surface vol. 10, no. 4, pp. 830–834, Apr. 2021. multidimensional harmonic retrieval,” IEEE Trans.
assisted terahertz MIMO communication systems,” [38] W. Cai, H. Li, M. Li, and Q. Liu, “Practical modeling Signal Process., vol. 50, no. 9, pp. 2366–2368,
IEEE Access, vol. 8, pp. 99565–99581, 2020. and beamforming for intelligent reflecting surface Sep. 2002.
[21] G. Zhou, C. Pan, H. Ren, P. Popovski, and aided wideband systems,” IEEE Commun. Lett., [59] X. Hu, R. Zhang, and C. Zhong, “Semi-passive
A. L. Swindlehurst, “Channel estimation for vol. 24, no. 7, pp. 1568–1571, Jul. 2020. elements assisted channel estimation for intelligent
RIS-aided multiuser millimeter-wave systems,” [39] W. Yang, H. Li, M. Li, Y. Liu, and Q. Liu, “Channel reflecting surface-aided communications,” IEEE
IEEE Trans. Signal Process., vol. 70, pp. 1478–1492, estimation for practical IRS-assisted OFDM Trans. Wireless Commun., vol. 21, no. 2,
2022. systems,” in Proc. IEEE Wireless Commun. Netw. pp. 1132–1142, Feb. 2022.
[22] J. He, H. Wymeersch, and M. Juntti, “Channel Conf. Workshops (WCNCW), Nanjing, China, [60] A. Taha, M. Alrabeiah, and A. Alkhateeb, “Enabling
estimation for RIS-aided mmWave MIMO systems Mar. 2021, pp. 1–6. large intelligent surfaces with compressive sensing
via atomic norm minimization,” IEEE Trans. [40] R. O. Schmidt, “Multiple emitter location and signal and deep learning,” IEEE Access, vol. 9,
Wireless Commun., vol. 20, no. 9, pp. 5786–5797, parameter estimation,” IEEE Trans. Antennas pp. 44304–44321, 2021.
Sep. 2021. Propag., vol. AP-34, no. 3, pp. 276–280, Mar. 1986. [61] X. Chen, J. Shi, Z. Yang, and L. Wu,
[23] X. Wei, D. Shen, and L. Dai, “Channel estimation [41] R. Roy, A. Paulraj, and T. Kailath, “ESPRIT—A “Low-complexity channel estimation for intelligent
for RIS assisted wireless communications—Part II: subspace rotation approach to estimation of reflecting surface-enhanced massive MIMO,” IEEE
An improved solution based on double-structured parameters of cisoids in noise,” IEEE Trans. Acoust., Wireless Commun. Lett., vol. 10, no. 5,
sparsity,” IEEE Commun. Lett., vol. 25, no. 5, Speech, Signal Process., vol. ASSP-34, no. 5, pp. 996–1000, May 2021.
pp. 1403–1407, May 2021. pp. 1340–1342, Oct. 1986. [62] Y. Jin, J. Zhang, X. Zhang, H. Xiao, B. Ai, and
[24] W. Zhang, J. Xu, W. Xu, D. W. K. Ng, and H. Sun, [42] P. Stoica and K. C. Sharman, “Maximum likelihood D. W. K. Ng, “Channel estimation for semi-passive
“Cascaded channel estimation for IRS-assisted methods for direction-of-arrival estimation,” IEEE reconfigurable intelligent surfaces with enhanced
mmWave multi-antenna with quantized Trans. Acoust., Speech, Signal Process., vol. 38, deep residual networks,” IEEE Trans. Veh. Technol.,
beamforming,” IEEE Commun. Lett., vol. 25, no. 2, no. 7, pp. 1132–1143, Jul. 1990. vol. 70, no. 10, pp. 11083–11088, Oct. 2021.
pp. 593–597, Feb. 2021. [43] M. Wax and I. Ziskind, “On unique localization of [63] Y. Lin, S. Jin, M. Matthaiou, and X. You,
[25] K. Ardah, S. Gherekhloo, A. L. F. de Almeida, and multiple sources by passive sensor arrays,” IEEE “Tensor-based algebraic channel estimation for
M. Haardt, “TRICE: A channel estimation Trans. Acoust., Speech, Signal Process., vol. 37, hybrid IRS-assisted MIMO-OFDM,” IEEE Trans.
framework for RIS-aided millimeter-wave MIMO no. 7, pp. 996–1000, Jul. 1989. Wireless Commun., vol. 20, no. 6, pp. 3770–3784,
systems,” IEEE Signal Process. Lett., vol. 28, [44] M. Viberg, B. Ottersten, and A. Nehorai, Jun. 2021.
pp. 513–517, 2021. “Performance analysis of direction finding with [64] G. C. Alexandropoulos and E. Vlachos, “A hardware
[26] S. Ma, W. Shen, J. An, and L. Hanzo, “Wideband large arrays and finite data,” IEEE Trans. Signal architecture for reconfigurable intelligent surfaces
channel estimation for IRS-aided systems in the Process., vol. 43, no. 2, pp. 469–477, Feb. 1995. with minimal active elements for explicit channel
face of beam squint,” IEEE Trans. Wireless Commun., [45] G. Tang and A. Nehorai, “Performance analysis for estimation,” in Proc. IEEE ICASSP, Barcelona,
vol. 20, no. 10, pp. 6240–6253, Oct. 2021. sparse support recovery,” IEEE Trans. Inf. Theory, Spain, May 2020, pp. 9175–9179.
[27] Z. Chen, J. Tang, X. Y. Zhang, D. K. C. So, S. Jin, vol. 56, no. 3, pp. 1383–1399, Mar. 2010. [65] G. C. Alexandropoulos, N. Shlezinger,
and K.-K. Wong, “Hybrid evolutionary-based sparse [46] I. Bilik, “Spatial compressive sensing for I. Alamzadeh, M. F. Imani, H. Zhang, and
channel estimation for IRS-assisted mmWave direction-of-arrival estimation of multiple sources Y. C. Eldar, “Hybrid reconfigurable intelligent
MIMO systems,” IEEE Trans. Wireless Commun., using dynamic sensor arrays,” IEEE Trans. Aerosp. metasurfaces: Enabling simultaneous tunable
vol. 21, no. 3, pp. 1586–1601, Mar. 2022. Electron. Syst., vol. 47, no. 3, pp. 1754–1769, reflections and sensing for 6G wireless
[28] Y. Liu, S. Zhang, F. Gao, J. Tang, and O. A. Dobre, Jul. 2011. communications,” Apr. 2021, arXiv:2104.04690.
“Cascaded channel estimation for RIS assisted [47] A. C. Gurbuz, V. Cevher, and J. H. McClellan, [66] I. Alamzadeh, G. C. Alexandropoulos,
mmWave MIMO transmissions,” IEEE Wireless “Bearing estimation via spatial sparsity using N. Shlezinger, and M. Imani, “A reconfigurable
Commun. Lett., vol. 10, no. 9, pp. 2065–2069, compressive sensing,” IEEE Trans. Aerosp. Electron. intelligent surface with integrated sensing
Sep. 2021. Syst., vol. 48, no. 2, pp. 1358–1369, Apr. 2012. capability,” Sci. Rep., vol. 11, Oct. 2021,
[29] H. Liu, J. Zhang, Q. Wu, H. Xiao, and B. Ai, “ADMM [48] Q. Shen, W. Liu, W. Cui, and S. Wu, Art. no. 20737.
based channel estimation for RISs aided millimeter “Underdetermined DOA estimation under the [67] C. Wu, C. You, Y. Liu, X. Gu, and Y. Cai, “Channel
wave communications,” IEEE Commun. Lett., compressive sensing framework: A review,” IEEE estimation for STAR-RIS-aided wireless
vol. 25, no. 9, pp. 2894–2898, Sep. 2021. Access, vol. 4, pp. 8865–8878, 2016. communication,” IEEE Commun. Lett., vol. 26,
[30] S. Kim, H. Lee, J. Cha, S.-J. Kim, J. Park, and [49] R. E. Carrillo and K. E. Barner, “Lorentzian iterative no. 3, pp. 652–656, Mar. 2022.
J. Choi, “Practical channel estimation and phase hard thresholding: Robust compressed sensing with [68] X. Guan, Q. Wu, and R. Zhang, “Anchor-assisted
shift design for intelligent reflecting surface prior information,” IEEE Trans. Signal Process., channel estimation for intelligent reflecting surface
empowered MIMO systems,” IEEE Trans. Wireless vol. 61, no. 19, pp. 4822–4833, Oct. 2013. aided multiuser communication,” IEEE Trans.
Commun., early access, Feb. 8, 2022, doi: [50] J. A. Tropp and A. C. Gilbert, “Signal recovery from Wireless Commun., early access, Nov. 13, 2021, doi:
10.1109/TWC.2022.3147825. random measurements via orthogonal matching 10.1109/TWC.2021.3123674.
[31] E. Björnson, “Optimizing a binary intelligent pursuit,” IEEE Trans. Inf. Theory, vol. 53, no. 12, [69] C. You, B. Zheng, and R. Zhang, “Wireless
reflecting surface for OFDM communications under pp. 4655–4666, Dec. 2007. communication via double IRS: Channel estimation
mutual coupling,” in Proc. 25th Int. ITG Workshop [51] K. Yu, M. Shen, R. Wang, and Y. He, “Joint nuclear and passive beamforming designs,” IEEE Wireless
Smart Antennas (WSA), Nov. 2021. norm and l1−2 -regularization sparse channel Commun. Lett., vol. 10, no. 2, pp. 431–435,
[32] Z.-Q. He, H. Liu, X. Yuan, Y.-J. A. Zhang, and estimation for mmWave massive MIMO systems,” Feb. 2021.
Y.-C. Liang, “Semi-blind cascaded channel IEEE Access, vol. 8, pp. 155409–155416, 2020. [70] B. Zheng, C. You, and R. Zhang, “Efficient channel
estimation for reconfigurable intelligent surface [52] J. P. Vila and P. Schniter, “Expectation-maximization estimation for double-IRS aided multi-user MIMO
aided massive MIMO,” Feb. 2021, Gaussian-mixture approximate message passing,” system,” IEEE Trans. Commun., vol. 69, no. 6,
arXiv:2101.07315. IEEE Trans. Signal Process., vol. 61, no. 19, pp. 3818–3832, Jun. 2021.
[33] Z.-Q. He and X. Yuan, “Cascaded channel pp. 4658–4672, Oct. 2013. [71] Y. Han, S. Zhang, L. Duan, and R. Zhang,
estimation for large intelligent metasurface assisted [53] J. T. Parker, P. Schniter, and V. Cevher, “Bilinear “Cooperative double-IRS aided communication:
massive MIMO,” IEEE Wireless Commun. Lett., vol. generalized approximate message passing—Part I: Beamforming design and power scaling,” IEEE
9, no. 2, pp. 210–214, Feb. 2020. Derivation,” IEEE Trans. Signal Process., vol. 62, Wireless Commun. Lett., vol. 9, no. 8,
[34] J. Mirza and B. Ali, “Channel estimation method no. 22, pp. 5839–5853, Nov. 2014. pp. 1206–1210, Aug. 2020.
and phase shift design for reconfigurable intelligent [54] Z. Yang and L. Xie, “On gridless sparse methods for [72] A. M. Elbir, A. Papazafeiropoulos, P. Kourtessis, and
surface assisted MIMO networks,” IEEE Trans. line spectral estimation from complete and S. Chatzinotas, “Deep channel learning for large
Cogn. Commun. Netw., vol. 7, no. 2, pp. 441–451, incomplete data,” IEEE Trans. Signal Process., intelligent surfaces aided mm-wave massive MIMO
Jun. 2021. vol. 63, no. 12, pp. 3139–3153, Jun. 2015. systems,” IEEE Wireless Commun. Lett., vol. 9, no. 9,
[35] N. K. Kundu and M. R. McKay, “Channel estimation [55] C. You, B. Zheng, and R. Zhang, “Fast beam pp. 1447–1451, Sep. 2020.
for reconfigurable intelligent surface aided MISO training for IRS-assisted multiuser [73] A. M. Elbir and S. Coleri, “Federated learning for
communications: From LMMSE to deep learning communications,” IEEE Wireless Commun. Lett., channel estimation in conventional and
RIS-assisted massive MIMO,” Aug. 2020, assisted compressive channel estimation for multi-stage CSI acquisition for reconfigurable
arXiv:2008.10846. mmWave intelligent reflecting surfaces,” IEEE intelligent surface aided MIMO systems,” IEEE
[74] C. Liu, X. Liu, D. W. K. Ng, and J. Yuan, “Deep Trans. Veh. Technol., vol. 69, no. 8, pp. 9223–9228, Commun. Lett., vol. 25, no. 6, pp. 2024–2028,
residual learning for channel estimation in Aug. 2020. Jun. 2021.
intelligent reflecting surface-assisted multi-user [76] Y. Wang, H. Lu, and H. Sun, “Channel estimation in [78] J. He, H. Wymeersch, M. Di Renzo, and M. Juntti,
communications,” IEEE Trans. Wireless Commun., IRS-enhanced mmWave system with “Learning to estimate RIS-aided mmWave
vol. 21, no. 2, pp. 898–912, Feb. 2022. super-resolution network,” IEEE Commun. Lett., channels,” IEEE Wireless Commun. Lett., vol. 11,
[75] S. Liu, Z. Gao, J. Zhang, M. D. Renzo, and vol. 25, no. 8, pp. 2599–2603, Aug. 2021. no. 4, pp. 841–845, Apr. 2022.
M.-S. Alouini, “Deep denoising neural network [77] S. Gao, P. Dong, Z. Pan, and G. Y. Li, “Deep
ABOUT THE AUTHORS

A. Lee Swindlehurst (Fellow, IEEE) Cunhua Pan (Member, IEEE) received the
received the B.S. degree in electrical B.S. and Ph.D. degrees from the School
engineering from Brigham Young University of Information Science and Engineering,
(BYU) Provo, UT, USA, in 1985, the M.S. Southeast University, Nanjing, China, in
degree in electrical engineering from BYU, 2010 and 2015, respectively.
in 1986, and the Ph.D. degree in electrical From 2015 to 2016, he was a Research
engineering from Stanford University, Associate with the University of Kent,
Stanford, CA, USA, in 1991. Canterbury, U.K. He held a postdoctoral
He was with the Department of Electrical position at the Queen Mary University of
and Computer Engineering, BYU, from 1990 to 2007. From 1996 London, London, U.K., from 2016 and 2019, where he was a
to 1997, he held a joint appointment as a Visiting Scholar at Lecturer from 2019 to 2021. Since 2021, he has been a Full
Uppsala University, Uppsala, Sweden, and the Royal Institute of Professor with Southeast University. He has coauthored over
Technology, Stockholm, Sweden. From 2006 to 2007, he was on 120 IEEE journal articles. His research interests mainly include
leave working as the Vice President of Research at ArrayComm reconfigurable intelligent surfaces (RIS), intelligent reflection
LLC, San Jose, CA, USA. Since 2007, he has been a Professor with surface (IRS), ultrareliable low-latency communication, machine
the Department of Electrical Engineering and Computer Science, learning, unmanned aerial vehicles, the Internet of Things, and
University of California at Irvine, Irvine, CA, USA. From 2014 to mobile edge computing.
2017, he was a Hans Fischer Senior Fellow with the Institute Dr. Pan has served as a TPC Member for numerous con-
for Advanced Studies, Technical University of Munich, Munich, ferences, such as IEEE International Conference on Commu-
Germany. His research focuses on array signal processing for radar, nications (ICC) and IEEE Global Communications Conference
wireless communications, and biomedical applications, and he has (GLOBECOM). He was the Student Travel Grant Chair for ICC
over 375 publications in these areas. in 2019. He received the IEEE ComSoc Leonard G. Abra-
Dr. Swindlehurst was elected as a Foreign Member of ham Prize in 2022. He is the Workshop Chair of IEEE WCNC
the Royal Swedish Academy of Engineering Sciences (IVA) 2024 and the TPC Co-Chair of IEEE ICCT 2022. He currently
in 2016. He was the inaugural Editor-in-Chief of the serves as an Editor for IEEE WIRELESS COMMUNICATION LETTERS,
IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING. He recei- IEEE COMMUNICATIONS LETTERS, and IEEE ACCESS. He has served
ved the 2000 IEEE W. R. G. Baker Prize Paper Award, the 2006 as a Guest Editor or Lead Guest Editor for special issues
IEEE Communications Society Stephen O. Rice Prize in the Field of in the IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS,
Communication Theory, the 2006, 2010, and 2022 IEEE Signal IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING, IEEE
Processing Society’s Best Paper Awards, the 2017 IEEE Signal Vehicular Technology Magazine, IEEE OPEN JOURNAL OF VEHICULAR
Processing Society Donald G. Fink Overview Paper Award, and TECHNOLOGY, and IEEE ACCESS. He has organized Workshops on
the Best Paper Award at the 2020 IEEE International Conference on RIS in IEEE ICCC 2021 and IEEE GLOBECOM 2021. He is also the
Communications. Workshops and Symposia Officer for the Reconfigurable Intelligent
Surfaces Emerging Technology Initiative.
Gui Zhou (Graduate Student Member, IEEE)
received the B.S. and M.E. degrees from
the School of Information and Electron-
ics, Beijing Institute of Technology, Beijing, Ming Li (Senior Member, IEEE) received the
China, in 2015 and 2019, respectively. She M.S. and Ph.D. degrees in electrical engi-
is currently working toward the Ph.D. degree neering from The State University of New
at the School of Electronic Engineering and York at Buffalo (SUNY-Buffalo), Buffalo, NY,
Computer Science, Queen Mary University USA, in 2005 and 2010, respectively.
of London, London, U.K. From January 2011 to August 2013, he
Her major research interests include intelligent reflection sur- was a Postdoctoral Research Associate with
faces and signal processing. the Department of Electrical Engineering,
SUNY-Buffalo. From August 2013 to June
Rang Liu (Graduate Student Member, IEEE) 2014, he joined Qualcomm Technologies, Inc., Buffalo, as a Senior
received the B.S. degree in electronics infor- Engineer. Since June 2014, he has been with the School of Informa-
mation engineering from the Dalian Univer- tion and Communication Engineering, Dalian University of Technol-
sity of Technology, Dalian, China, in 2018, ogy, Dalian, China, where he is currently a Professor. His current
where she is currently working toward the research interests are in the general areas of communication the-
Ph.D. degree at the School of Information ory and signal processing with applications to integrated sensing
and Communication Engineering. and communication, reconfigurable intelligent surfaces, millimeter-
Her current research interests are focused wave communications, massive multi-input–multioutput (MIMO)
on signal processing, massive multi-input– systems, and secure wireless communications.
multi-output (MIMO) systems, reconfigurable intelligent surfaces, Dr. Li was the Co-Chair of the IEEE INFOCOM 2019 Work-
and integrated sensing and communication. shop on Wireless Communications and Networking in Extreme
Ms. Liu was a recipient of the National Scholarship in 2020. Environments.

Swindlehurst Channel Estimation With Reconfigurable Intelligent SurfacesA General Framework

Uploaded by

Copyright:

Available Formats

You might also like

Swindlehurst Channel Estimation With Reconfigurable Intelligent SurfacesA General Framework

Uploaded by

Document Information

Original Description:

Original Title

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

Swindlehurst Channel Estimation With Reconfigurable Intelligent SurfacesA General Framework

Uploaded by

Copyright:

Available Formats

Channel Estimation With

1312 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1313

represented as IN , and N × 1 vectors composed of all

1314 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

Provided that T ≥ K(N + 1) and Z is full rank, there are T 

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1315

1316 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

B. Wideband Single-User MIMO

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1317

symbol t as where the composite channel is now Hc = gH H =

1318 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

ZT = UV (30) where we drop the first column of Hd to create H̃d and we

√ " # which for large M is significantly less than the value

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1319

a(χ) = ax (ωx ) ⊗ ay (ωy ) (35)

where the vertical array response component is similar

1320 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

for d peaks in the spectrum representation

pB (χ) = aH (χ)RY a(χ) (40) AR (χ RH ) = ARD QRH (44)

Alternatively, one can employ higher resolution algorithms

where (·)† represents the pseudoinverse, although esti-

1 in the position corresponding to ωBH,k , assuming that

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1321

where γ GH = γ ∗G ⊗ γ H , qGH = vec(QGH ),

1322 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

with four angle parameters of interest: ωBH , ωU G , and

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1323

1324 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

vidual subchannel responses). Our starting point is (26),

Using (55), we can write

where where D(k) = AU (χ U G (k))Ξ(k)AH B (χ BH (k)). Assum-

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1325

1326 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

where the permutation matrix J is implicitly defined to

2) Single-Antenna UE and BS: Here, there is no need for 

y = Ψ∗ CT (χ RG , χ RH ) (γ ∗G ⊗ γ H ) + n (92a) where {χ u , γ Gu H , qGu H } are the angle parameters, chan-

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1327

 T$ 1 % H for dramatic reductions in both computational complexity

1328 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1329

on the system throughput, both in terms of time lost for C. Robustness

1330 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

the required number of operations is O(M (K + N )T ) ≈

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1331

We see that the performance of the LS algorithm exactly

1332 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

Fig. 6. CRB for various channel models versus SNR. Channel

same, so for these plots, we include only the CRB results.

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1333

and receive (STAR)-RIS is another recently proposed RIS

Fig. 9. CRB for various channel models versus number of columns

1334 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

Table 2 Summary of Topics Covered by References Related to Channel Estimation

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1335

1336 P ROCEEDINGS OF THE IEEE | Vol. 110, No. 9, September 2022

Vol. 110, No. 9, September 2022 | P ROCEEDINGS OF THE IEEE 1337

Provided that T ≥ K(N + 1) and Z is full rank, there are T

2) Single-Antenna UE and BS: Here, there is no need for

T$ 1 % H for dramatic reductions in both computational complexity