Channel Estimation for RIS-Aided mmWave MIMO

Channels
Jiguang He† , Markus Leinonen† , Henk Wymeersch? , Markku Juntti†
† Centre for Wireless Communications, FI-90014, University of Oulu, Finland
? Department of Electrical Engineering, Chalmers University of Technology, Gothenburg, Sweden

Abstract—A reconfigurable intelligent surface (RIS) can shape CE for RIS-aided MIMO systems has been recently studied
the radio propagation by passively changing the directions of in [11], [12]. In [11], CE is performed using CS and deep
impinging electromagnetic waves. The optimal control of the RIS learning methods in a setup with a few active elements at the
arXiv:2002.06453v1 [[Link]] 15 Feb 2020

requires perfect channel state information (CSI) of all the links

connecting the base station (BS) and the mobile station (MS) RIS. In [12], sparse matrix factorization and matrix completion
via the RIS. Thereby the channel (parameter) estimation at the are exploited in a sequential manner to perform iterative CE.
BS/MS and the related message feedback mechanism are needed. In this work, full advantage of the RIS is not achieved due
In this paper, we adopt a two-stage channel estimation scheme to the on/off state applied to the RIS elements. Instead of
for the RIS-aided millimeter wave (mmWave) MIMO channels estimating the MIMO channels, a multi-level codebook based
using an iterative reweighted method to sequentially estimate the
channel parameters. We evaluate the average spectrum efficiency scheme was leveraged to design the phase control matrix at
(SE) and the RIS beamforming gain of the proposed scheme and the RIS and the combining vector at the MS jointly [13].
demonstrate that it achieves high-resolution estimation with the In the paper, we study the CE problem of the passive RIS-
average SE comparable to that with perfect CSI. aided mmWave MIMO system. We divide the CE problem into
Index Terms—Channel estimation, compressive sensing, mil- two subproblems and apply an iterative reweighted method to
limeter wave MIMO, reconfigurable intelligent surface.
find the estimates of the channel parameters sequentially. In
I. I NTRODUCTION the first stage, we estimate the angle of departure (AoD) for
the BS-RIS link, the angle of arrival (AoA) of the RIS-MS
Large unused spectrum is available in the millimeter wave link, and the corresponding effective propagation path gains.
(mmWave) bands. In order to compensate for the high free In the second stage, we estimate the product of the propagation
space path loss, large antenna arrays are needed both at the path gains of the BS-RIS link and the RIS-MS link and the
transmitters and receivers [1]–[3]. This makes the channel difference of directional sine of the AoA of the BS-RIS and the
estimation (CE) more difficult than that in conventional sub-6 AoD of RIS-MS link. Besides evaluating the mean square error
GHz multiple-input multiple-output (MIMO) systems having (MSE) of the estimated channel parameters, we design the RIS
less transmit and receive antennas. The mmWave MIMO phase control matrix, the BS beamforming (BF) vector, and the
channel is inherently sparse due to the limited number of dis- MS combining vector based on the estimates and evaluate the
tinguishable paths in the angular domain. Thus, compressive average SE and RIS BF gain. The performance of the proposed
sensing (CS) techniques, which take advantage of the sparsity, scheme is compared to that of an orthogonal matching pursuit
have been widely applied in the channel (parameter) estimation (OMP) approach. Simulation results demonstrate that average
of mmWave MIMO channels, e.g., in [4], [5]. SE achieved by the proposed method approaches that with
In order to further improve the spectrum efficiency (SE) perfect channel state information (CSI), even in the low signal-
and to guarantee a wide communication coverage, the concept to-noise ratio (SNR) regime with limited training overhead.
of a reconfigurable intelligent surface (RIS) has been recently Notation: A bold lowercase letter a denotes the column
proposed to smartly shape the propagation of electromagnetic vector, a bold capital letter A denotes the matrix, (·)H ,
waves [6]–[9]. The RIS also has great potential to offer higher (·)T , and (·)∗ denote the Hermitian transpose, transpose, and
accuracy of positioning and localization, both for indoor and conjugate, respectively, diag(a) denotes a square diagonal
outdoor, compared to the system without RISs [9], [10]. The matrix with entries of a on its diagonal, a ◦ b denotes the
RIS can be made of an array of phase shifters, which can Hadamard product of a and b, [a]i denotes the ith element of
passively steer the beams towards the dedicated user(s) by vector a, [A]ij denotes the (i, j)th element of A, and k · kF
controlling the phase of each RIS unit. This RIS architecture is is the Frobenius norm.
called the discrete RIS. Another type of RIS is the contiguous
RIS which can be seen as an active transceiver [6]. II. C HANNEL M ODEL
This work is supported by Horizon 2020, European Union’s Framework We consider the RIS-aided mmWave MIMO system that
Programme for Research and Innovation, under grant agreement no. 871464 comprises one multi-antenna BS, one multi-antenna MS, and
(ARIADNE). This work is also partially supported by the Academy of Finland one multi-element RIS, as depicted in Fig. 1. The numbers
6Genesis Flagship (grant 318927) and Swedish Research Council (grant no.
2018-03701). The work of M. Leinonen has been financially supported in part of antenna elements are defined as NBS , NMS , and NRIS ,
by Infotech Oulu and the Academy of Finland (grant 319485 and 323698). respectively. The antenna array is assumed to be an uniform
 RIS
CE DT

…
MS
Fig. 2: The sounding and CE procedure. Each CE subinterval contains
BS multiple blocks where Ωt varies over the blocks.

Fig. 1: The considered RIS-aided mmWave MIMO system with one

BS, one MS, and one RIS.
ignore the NLoS paths in the BS-RIS and RIS-MS links, and
approximate the composite channel in (3) as
linear array (ULA); an extension to an uniform planar array
H ≈ α(φR,M,0 )ρR,M,0 αH (θR,M,0 )Ωα(φB,R,0 )ρB,R,0 αH (θB,R,0 )
(UPA) is feasible. We assume that the direct link between the
BS and the MS are obstructed, which renders the potential = gα(φR,M,0 )αH (θB,R,0 ), (4)
usage of a RIS for maintaining the connectivity.
where g ∈ C is the effective propagation path gain, defined as
We assume geometric channel model. The channel between
the BS and the RIS HB,R ∈ CNRIS ×NBS is g = ρB,R,0 ρR,M,0 αH (θR,M,0 )Ωα(φB,R,0 ). (5)
LB,R
X According to (4) and (5), the composite channel is virtually a
HB,R = ρB,R,l α(φB,R,l )αH (θB,R,l ), (1)
point-to-point MIMO channel with one path via the RIS. The
l=0
rank of H is 1  {NBS , NMS }. Thus, the channel encompasses
where θB,R,l and φB,R,l denote the lth AoD and AoA of the BS- a sparse structure in the angular domain – characteristic to an
RIS link, respectively, LB,R denotes the number of resolvable mmWave channel – which can efficiently be leveraged by CS.
paths, which is usually on the order of 3–5 in mmWave
bands [14], and ρB,R,l denotes the ith propagation path III. S OUNDING P ROCEDURE
gain. Index l = 0 refers to the line-of-sight (LoS) path, and
l = 1, · · · , LB,R refer to the non-line-of-sight (NLoS) paths, We assume that the channels suffer from block fading. For
e.g., single-bounce or multi-bounce reflection paths. Usually, the sounding process, one coherence time interval is divided
|ρB,R,0 |2  |ρB,R,l |2 for l = 1, · · · , LB,R , and the difference into two subintervals, the first one for CE and the second
is easily more than 20 dB [15]. Finally, α(θB,R,l ) ∈ CNBS ×1 for data transmission (DT), as depicted in Fig. 2. The CE
and α(φB,R,l ) ∈ CNRIS ×1 are the array response
vectors with subinterval is further divided into multiple blocks.
[α(θB,R,l )]k = exp j2π λd (k − 1) sin(θB,R,l ) for k =
In each block t = 1, · · · , T , the BS sends a (random) train-
1, · · · , NBS and [α(φB,R,l )]k = exp j2π λd (k − 1) sin(φB,R,l ) ing matrix Xt ∈ CNBS ×NX which, after reflected from the RIS
for k = 1, · · · , NRIS , where d is the antenna element spacing, having a (random) phase control matrix Ωt , is received at the
4 √ MS through a (random) combining matrix Wt ∈ CNMS ×NW .
λ is the wavelength of the carrier frequency, and j = −1.
Thus, the received signal at the MS is given as
Similarly, the channel between the RIS and the MS, denoted
as HR,M ∈ CNMS ×NRIS , is Yt = WtH H(Ωt )Xt + WtH Nt , t = 1, · · · , T, (6)
LR,M
X where we write H explicitly as a function of Ωt . Further, let
HR,M = ρR,M,l α(φR,M,l )αH (θR,M,l ). (2)
l=0 gt = ρR,M,0 ρB,R,0 αH (θR,M,0 )Ωt α(φB,R,0 ). (7)
Using (1) and (2), the composite channel H ∈ CNMS ×NBS
The following assumptions are made for the CE subinterval:
between the BS and MS becomes
• All the channel parameters stay unchanged within the
H = HR,M ΩHB,R , (3) coherence time.
• Ωt is varying with block index t, and it is constructed
where Ω ∈ CNRIS ×NRIS is the phase control matrix at the RIS.
The matrix Ω is a diagonal matrix with unit-modulus elements with randomly generated phases ω ∈ [−π π].
• Wt and Xt are varying with block index t.
on the diagonal, i.e., [Ω]kk = exp(jω) with ω ∈ R. In practice,
the reflection may not be perfect so that reflection coefficient
IV. T WO -S TAGE CE A PPROACH
a ∈ [0, 1] as [Ω]kk = a exp(jω) describes the amplitude
scaling and power loss [16]; however, we assume that a = 1. We consider a two-stage approach to simplify the CE
Since the LoS path is typically much stronger than the problem. The solution process is summarized in Algorithm 1.
NLoS paths according to the field measurements in [15], we The details of the algorithm will be provided in the sequel.
 where
Algorithm 1: Two-stage CE Approach  1 1 T 
G(i) = diag (i)
,··· , (i)
. First stage |ĝ1 |2 +  |ĝT |2 + 
Input: Received signals Y1 , . . . , YT , combining matrices (i)
W1 , . . . , WT , training matrices X1 , . . . , XT , phase con- and ĝt is an estimate of gt at the ith iteration. Setting the
trol matrices Ω1 , . . . , ΩT , and threshold value τth . first-order partial derivative of S̃(g, θ, φ) in (9) with respect
Output: θ̂, φ̂, and ĝ. to gt to zero yields
1: Initial θ (i) and φ(i) for i = 0 by the SVD based approach.
!−1 N
X
(i+1) 1 2
X
2: repeat ĝt = (i)
+ kA t X t k F xH H
t,k At yt,k ,
(i+1) ξ(|ĝt |2 + )
3: Compute ĝt by (10). k=1
4: Construct the objective function S̃(ĝ(i+1) , θ, φ) (10)
by (11). where xt,k and yt,k is the kth column of Xt and Yt ,
5: Use gradient descent to minimize S̃(ĝ(i+1) , θ, φ). respectively, and At = WtH α(φ)αH (θ). Thus, for given
(i+1) (i+1) T
6: until The maximum number of iterations reached or ĝ(i+1) = [ĝ1 , · · · , ĝT ] , (9) can be written as
kĝ(i+1) − ĝ(i) k2 < τth . T X T N
1X 2 ∗ 1 XX
. Second stage S̃(ĝ(i+1) , θ, φ) = βt zt zt (i) + H
yt,k yt,k
Input: ĝ and threshold value τ̃th . ξ t=1 |ĝt |2 +  t=1 k=1
ˆ T n
Output: q̂ and θ̃. X (i+1) ∗

(i+1)
∗
(i+1) 2
o
7: Initial θ̃ (i)
for i = 0. + −ĝt zt − ĝt zt + |ĝt | kAt Xt k2F
t=1
8: repeat !
T NX
9: Compute q̂ (i+1) . X X
˜ (i+1) , θ̃) by (17). = −βt zt∗ zt + H
yt,k yt,k , (11)
10: Construct the objective function J(q̂
˜ (i+1) , θ̃). t=1 k=1
11: Use gradient descent to minimize J(q̂
N X
12: until The maximum number of iterations reached or
xH H
P
where we defined the quantities zt = t,k At yt,k and
|q̂ (i+1) − q̂ (i) | < τ̃th . k=1
 −1
1 2
βt = (i) + kA t X t kF .
ξ(|ĝt |2 +)
A. First Stage The above equations give rise to an iterative algorithm
In the first stage, we estimate the AoD of the BS-RIS link aiming at minimizing the objective function in (11). To this
θB,R,0 in (4), the AoA of the RIS-MS link φR,M,0 in (4), end, we propose Algorithm 1 where at each iteration i, we use
and the effective propagation path gain of the reflection link a gradient descent algorithm to find estimates for the angles
gt in (7) based on the received signals {Yt }Tt=1 . Typically, θ and φ for a given ĝ(i+1) in (10). The calculation of the
the CS methods like OMP have been proposed for CE. One required first-order partial derivatives associated with θ and φ
drawback is that they only can recover angular estimates that is presented in Appendix A. The initial values for θ and φ,
lie on a pre-defined grid of discrete angles. This “on-the-grid” defined as θ(0) and φ(0) , can be determined by the singular
problem, which inevitably deteriorates the CE performance, value decomposition (SVD) based approach [8]. After a certain
can be counteracted by the iterative reweighted method [8], stopping criterion is met, we proceed to the second stage.
used herein as well. B. Second Stage
Using the iterative reweighted method, the CE problem in
In the second stage, we estimate the remaining channel
the first stage is formulated as
parameters based on the final estimate of g obtained in the
T
X first stage, denoted as ĝ = [ĝ1 , · · · , ĝT ]T . Obtaining a separate
min S(g, θ, φ) = ln(|gt |2 + ) + ξkYt − WtH H(Ωt )Xt k2F , estimate of the AoA for the BS-RIS link φB,R,0 and the AoD
g,θ,φ
t=1 for the RIS-MS link θR,M,0 seems infeasible; the same holds
(8)
for the propagation path gains in the BS-RIS and RIS-MS links
where g = [g1 , · · · , gT ]T , parameter  > 0 ensures that the
ρR,M,0 and ρB,R,0 . Thus, instead, we will estimate the product
argument of ln(·) is positive, ξ > 0 controls the tradeoff
of the propagation path gains ρR,M,0 ρB,R,0 and the difference
between the sparsity of g and data fitting. Here, we replaced
of directional sine, defined as θdiff = sin(φR,M,0 ) − sin(θB,R,0 ).
φR,M,0 and θB,R,0 with φ and θ to simplify the notations. The
According to (7), we rewrite gt as
sparsity-inducing ln(·) term accounts for the fact that due to
∗
gt = ρR,M,0 ρB,R,0 ω T


the random generation of Ωt , t = 1, · · · , T , some elements in t α (θR,M,0 ) ◦ α(φB,R,0 ) , (12)
g may be much smaller than others.
where ω t ∈ CNRIS ×1 is the vector of phase control matrix val-
The problem in (8) can be further formulated as [17]
ues as Ωt = diag(ω t ). By stacking the T different ω t ’s row-
T
X wise as Ω̃ = [ω 1 , · · · , ω T ]T and introducing q = ρR,M,0 ρB,R,0
min S̃(g, θ, φ) = gH G(i) g + ξ kYt − WtH H(Ωt )Xt k2F , and θ̃ = asin(θdiff ), we have
g,θ,φ
t=1
g = Ω̃ρR,M,0 ρB,R,0 α∗ (θR,M,0 )◦α(φB,R,0 ) = q Ω̃α(θ̃). (13)


(9)
 Recall that the the angular sparsity is one. Regarding this 10-1
fact, we aim at finding the sparsest representation while
minimizing the MSE of the data fitting, which fits in the 10-2
, Proposed, T = 10

MSE
problem of line spectral estimation [17]. Further, we assume , Proposed, T = 10
, Two-stage OMP, T = 10
that , Two-stage OMP, T = 10
10-3
ĝ = g + n = q Ω̃α(θ̃) + n, (14) , Proposed, T = 16
, Proposed, T = 16
, Two-stage OMP, T = 16
where the estimation error n from the first stage is modelled 10-4
, Two-stage OMP, T = 16
-25 -20 -15 -10 -5 0
as additive Gaussian noise, independent of q and θ̃. Thus, the SNR
objective function of the second stage estimation problem is Fig. 3: Angular parameter estimate in the first stage.
formulated as
min J(q, θ̃) = ln(|q|2 + ) + µkĝ − Ω̃hk22 , (15) 10-1

where h = qα(θ̃) and µ is a parameter that controls the

10-2
tradeoff between the sparsity and data fitting. The objective

MSE
function can be further reformulated as
˜ θ̃) = qq ∗ 10-3 Two-stage OMP, T = 10
min J(q, + µkĝ − Ω̃hk22 . (16) Proposed, T = 10
|q̂ (i) |2 +  Two-stage OMP, T = 16
Proposed, T = 16
10-4
We also follow the iterative reweighted method to find the -25 -20 -15
SNR
-10 -5 0

high-resolution estimate of θ̃. At the beginning, we find an

Fig. 4: Angular difference estimate in the second stage.
initial estimate of θ̃ on the grid as θ̃(0) = arg maxθ̃∈Υ |ĝH Ω̃h|,
where Υ is a quantized set of angles within [−π π]. Setting
the derivative of the objective function J(q, ˜ θ̃) in (16) with
H is uses in the second stage. The dictionary for the
(i+1)
respect to q to zero yields q̂ = γα(θ̃)H Ω̃ ĝ, where γ = first stage is constructed by quantizing the angles as
H
−1 NBS,q NMS,q

1/(µ(|q̂ (i) |2 + )) + α(θ̃)H Ω̃ Ω̃α(θ̃) . For fixed q̂ (i+1) , Ψ1 = A∗ ({θ̄k }k=1 ) ⊗ A({φ̄k }k=1 ) with NBS,q = 2NBS
the objective function can be further written as and NMS,q = 2NMS ; the dictionary in the second stage is
NRIS,q
NMS,q

˜ (i+1) , θ̃) = −γτ + ĝH ĝ.
J(q̂ (17) Ψ2 = A {θ̄k }k=1 with NRIS,q = 2NRIS . A {φ̄k }k=1 is
NMS,q
 
H
defined as A {φ̄k }k=1 = α(φ̄1 ), · · · , α(φ̄NMS,q ) , and the
where τ = α(θ̃)H Ω̃ ĝĝH Ω̃α(θ̃). Based on the first-order same principle is applied to A({θ̄k }k=1
NBS,q
) and A({θ̄k }k=1
NRIS,q
).
derivative in Appendix B, gradient descent is applied to find a The simulation results are shown in Figs. 3–6. As shown in
(suboptimal) θ̃ that minimizes the objective function in (17). Fig. 3, the proposed scheme outperforms the two-stage OMP
V. P ERFORMANCE E VALUATION scheme in terms of the MSE of the AoA and AoD estimates
obtained in the first stage. A super-resolution estimate can be
In this section, we evaluate the MSE performance of the
achieved, e.g., the average MSE is at the level of 10−3 at
angular parameter estimation in the first stage, the average
the SNR of −5 dB when T = 16. This offers a near-optimal
SE, and the RIS BF gain based on the estimated parameters.
design of the BF and combining vectors at the BS and the
The simulation parameters are set as follows: NBS = NBS =
MS.
32, NRIS = 64, NX = NW = 10, T = {10, 16},  = 1,
Fig. 4 shows the average MSE of the angular difference
ξ = µ = 1000, and d = λ2 . The number of RF chains is
estimate in the second stage. The increase of the time slots
defined as NRF = 10. The number of required time slots for
improves performance. This in turn brings higher RIS BF gain
CE is T NXNN W
= {100, 256}. We assume that the product
RF and average SE, as seen in Figs. 5 and 6. It is worth noting
of the propagation path gains, q, follows CN (0, 1) and each
that for T = 16, the average SE of the proposed scheme is
element of Nt follows CN (0, σ 2 ). The SNR is defined as σ12 ,
close to that with full CSI, even in the low SNR regime.
and 2000 realizations are considered for averaging.
The average SE is defined as VI. C ONCLUSION
R = E log2 (1 + |wH H(Ω̂)f |2 /σ 2 ) bits/s/Hz,
 
(18) We studied the CE problem for the RIS-aided mmWave
p p MIMO systems. We proposed a two-stage iterative reweighted
where w = 1/NMS α(φ̂R,M,0 ), f = 1/NBS α(θ̂B,R,0 ), and method that finds estimates of the channel parameters in a
the estimate based optimal phase control matrix is [Ω̂]kk = sequential optimization loop. Simulation results confirmed the
ˆ
exp(−j2π λd (k − 1) sin(θ̃)). The RIS BF gain is defined as advantages of the proposed scheme compared to the two-stage
GBF = |αH (θR,M,0 )Ω̂α(φB,R,0 )|/NRIS . OMP approach in terms of CE and SE performance. Since the
As a benchmark, we consider a two-stage OMP al- used gradient descent method cannot guarantee the optimality
gorithm that has the following one-to-one correspon- of the solutions and its convergence may be slow, it is of
dence with our proposed scheme: the first stage ap- interest to study more computational efficient algorithms in
plies the simultaneous OMP whereas the standard OMP future work.
 1
0.9
B. Derivatives in the Second Stage
0.8 ˜ (i+1) , θ̃) in (17) with respect to θ̃ can
The derivative of J(q̂
0.7
RIS BF Gain be written as
0.6
˜ (i+1) , θ̃)
∂ J(q̂ ∂γ ∂τ
0.5 =− τ −γ ,
Proposed, T = 10
Two-stage OMP, T = 10
∂ θ̃ ∂ θ̃ ∂ θ̃
0.4
Proposed, T = 16 ∂γ ∂τ
Two-stage OMP, T = 16
where ∂ θ̃ and ∂ θ̃ are in the form of
-25 -20 -15 -10 -5 H H
∂γ/∂ θ̃ = −γ 2 α(θ̃)H DH Ω̃ Ω̃α(θ̃) + α(θ̃)H Ω̃ Ω̃Dθ̃ α(θ̃) ,


SNR θ̃
H H
Fig. 5: RIS BF gain vs. SNR. ∂τ /∂ θ̃ = α(θ̃)H DH
θ̃
Ω̃ ĝĝH Ω̃α(θ̃) + α(θ̃)H Ω̃ ĝĝH Ω̃Dθ̃ α(θ̃),
and Dθ̃ = diag [0, j2π λd cos( d

θ̃), · · · , j2π λ (i−1) cos(θ̃), · · · ,
25
d T
j2π λ (NRIS − 1) cos(θ̃)] .
Average SE (bits/s/Hz)

20
R EFERENCES
15
[1] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, “Channel
10 Proposed, T = 10
estimation and hybrid precoding for millimeter wave cellular systems,”
Two-stage OMP, T = 10 IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831–846, Oct
5
Perfect CSI 2014.
Proposed, T = 16
Two-stage OMP, T = 16 [2] J. He, T. Kim, H. Ghauch, K. Liu, and G. Wang, “Millimeter wave
0 MIMO channel tracking systems,” in proc. of IEEE Globecom Work-
-25 -20 -15 -10 -5 0 shops (GC Wkshps), Dec 2014, pp. 416–421.
SNR [3] R. W. Heath, N. Gonzlez-Prelcic, S. Rangan, W. Roh, and A. M.
Fig. 6: Average SE vs. SNR. Sayeed, “An overview of signal processing techniques for millimeter
wave MIMO systems,” IEEE J. Sel. Topics Signal Process., vol. 10,
no. 3, pp. 436–453, April 2016.
[4] Z. Marzi, D. Ramasamy, and U. Madhow, “Compressive channel esti-
A PPENDIX mation and tracking for large arrays in mm-Wave picocells,” IEEE J.
Sel. Topics Signal Process., vol. 10, no. 3, pp. 514–527, April 2016.
[5] J. Lee, G. Gil, and Y. H. Lee, “Channel estimation via orthogonal
A. Derivatives in the First Stage matching pursuit for hybrid MIMO systems in millimeter wave com-
munications,” IEEE Trans. Commun., vol. 64, no. 6, pp. 2370–2386,
The first-order partial derivative of S̃(ĝ(i+1) , θ, φ) in (11) June 2016.
with respect to θ is written as [6] S. Hu, F. Rusek, and O. Edfors, “Beyond massive MIMO: The potential
of positioning with large intelligent surfaces,” IEEE Trans. Signal
Process., vol. 66, no. 7, pp. 1761–1774, Apr 2018.
T
∂ S̃(ĝ(i+1) , θ, φ) X ∂βt ∗ ∂z ∗ ∂zt [7] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and
= − zt zt − βt t zt − βt zt∗ , C. Yuen, “Large intelligent surfaces for energy efficiency in wireless
∂θ t=1
∂θ ∂θ ∂θ communication,” arXiv, vol. abs/1810.06934, 2018.
[8] C. Hu, L. Dai, T. Mir, Z. Gao, and J. Fang, “Super-resolution channel
∗
∂kAt Xt k2F
estimation for mmwave massive MIMO with hybrid precoding,” IEEE
∂βt ∂zt ∂zt
where ∂θ , ∂θ , ∂θ , and ∂θ are given by Trans. Veh. Technol., vol. 67, no. 9, pp. 8954–8958, Sep. 2018.
[9] J. He, H. Wymeersch, L. Kong, O. Silvén, and M. Juntti, “Large
N intelligent surface for positioning in millimeter wave MIMO systems,”
∂βt X
X ∂AHt H ∂At accepted by IEEE VTC Spring, 2020.
= −βt2 xH
t,k At xt,k + xH
t,k At xt,k , [10] H. Wymeersch, J. He, B. Denis, A. Clemente, and M. Juntti, “Radio
∂θ ∂θ ∂θ localization and mapping with reconfigurable intelligent surfaces,” 2019.
k=1
N N [11] A. Taha, M. Alrabeiah, and A. Alkhateeb, “Enabling large intelligent
∂zt∗ X
X
H ∂At ∂zt X
X ∂AHt H surfaces with compressive sensing and deep learning,” arXiv, 2019.
= yt,k xt,k , = xt,k y ,
∂θ ∂θ ∂θ ∂θ t,k [12] Z. He and X. Yuan, “Cascaded channel estimation for large intelligent
k=1 k=1 metasurface assisted massive MIMO,” IEEE Wireless Commun. Lett.,
NX vol. 9, no. 2, pp. 210–214, Feb 2020.
∂kAt Xt k2F X ∂AH
t ∂At [13] J. He, H. Wymeersch, T. Sanguanpuak, O. Silvén, and M. Juntti,
= xH
t,k At xt,k + xH H
t,k At At xt,k ,
∂θ ∂θ ∂θ “Adaptive beamforming design for mmwave RIS-aided joint localization
k=1 and communication,” accepted by IEEE WCNC, 2020.
[14] T. S. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N.
and ∂A d
∂θ = At Dθ , Dθ = diag [0, −j2π λ cos(θ), · · · ,

t Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave
d d mobile communications for 5G cellular: It will work!” IEEE Access,
− j2π λ (i − 1) cos(θ), · · · , −j2π λ (NBS − 1) cos(θ)]T . vol. 1, pp. 335–349, 2013.
Similarly, we can write the first-order partial derivative of [15] M. R. Akdeniz, Y. Liu, M. K. Samimi, S. Sun, S. Rangan, T. S.
S̃(ĝ(i+1) , θ, φ) in (11) with respect to φ as Rappaport, and E. Erkip, “Millimeter wave channel modeling and
cellular capacity evaluation,” IEEE J. Sel. Areas Commun., vol. 32, no. 6,
pp. 1164–1179, Jun 2014.
T
∂ S̃(ĝ(i+1) , θ, φ) X ∂βt ∗ ∂z ∗ ∂zt [16] Q. Wu and R. Zhang, “Beamforming optimization for wireless network
= − zt zt − βt t zt − βt zt∗ , aided by intelligent reflecting surface with discrete phase shifts,” IEEE
∂φ t=1
∂φ ∂φ ∂φ Trans. Commun., pp. 1–1, 2019.
[17] J. Fang, F. Wang, Y. Shen, H. Li, and R. S. Blum, “Super-resolution
compressed sensing for line spectral estimation: An iterative reweighted
and ∂A t H H
∂φ = Wt Dφ α(φ)α (θ), Dφ = diag [0, j2π λ cos(φ),
d
approach,” IEEE Trans. Signal Process., vol. 64, no. 18, pp. 4649–4662,
· · · , j2π λd (i − 1) cos(φ), · · · , j2π λd (NMS − 1) cos(φ)]T .


Sep. 2016.