This article has been accepted for publication in IEEE Open Journal of the Communications Society.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

Received XX Month, XXXX; revised .

Multi-RIS-assisted 3D Localization and

Synchronization via Deep Learning
Alireza Fadakar1 , Maryam Sabbaghian12 (Member, IEEE) , Henk Wymeersch3 (Fellow, IEEE)

1
A. Fadakar and M. Sabbaghian are with the School of Electrical and Computer Engineering, University of Tehran, Iran. Email:
[email protected], [email protected].
2
Corresponding author: Maryam Sabbaghian
3
H. Wymeersch is with the Department of Electrical Engineering, Chalmers University of Technology, Sweden. Email: [email protected].
The work of Henk Wymeersch has been supported by the SNS JU project 6G-DISAC under the EU’s Horizon Europe research and
innovation Program under Grant Agreement No 101139130 as well as the Swedish Research Council under project No. 2022-03007.

ABSTRACT Reconfigurable intelligent surfaces (RISs) have received considerable attention in applications
related to localization. However, operation in multi-path scenarios is challenging from both complexity
and performance perspectives. This study presents a two-stage low complexity method for joint three-
dimensional (3D) localization and synchronization using multiple RISs. Firstly, the received signals are
preprocessed, and an efficient deep learning architecture is proposed to initially estimate the angles of
departure (AODs) of the virtual line of sight paths from the RISs to the user. Then, a hybrid asynchronous
AOD time-of-arrival-based approach is proposed in the first stage to estimate an initial guess of the position
of the user equipment (UE). Finally, in the second stage, an optimization problem is formulated to refine the
position of the UE by effectively utilizing the estimated delays and the clock offset. Our comparative study
reveals that the proposed method outperforms the existing methods in terms of accuracy and complexity.
Notably, the proposed method showcases enhanced robustness against multipath effects when compared
to the state-of-the-art approaches.

INDEX TERMS 3D localization, deep learning, mmWave, reconfigurable intelligent surface, synchroniza-
tion.

I. Introduction In 5G, accurate radio localization is achievable through the

The application of localization technologies has witnessed utilization of a multitude of antennas and a large radio band-
extensive adoption in various domains, including wireless width, combining AOA, AOD, and TDOA. Looking ahead
networks, disaster rescue, augmented reality, and the Internet to 6G, radio localization is anticipated to become even more
of Things (IoT) [1], [2]. Consequently, there has been ubiquitous, with the integration of reconfigurable intelligent
a growing interest in advancing localization systems by surfaces (RISs) [13]. RISs have attracted enormous interest,
leveraging innovative technologies over the past few decades mainly for their ability to overcome line-of-sight (LoS)
[3]–[5]. In the presence of noise-corrupted sensor mea- blockages in millimeter wave (mmWave) communications.
surements, various approaches are utilized to estimate the They are composed of multiple nearly passive metamaterials
position of the source, encompassing time-of-arrival (TOA) and are poised to revolutionize the performance of wireless
[6], time-difference-of-arrival (TDOA) [7], angle-of-arrival applications. This includes enhancing capabilities in localiza-
(AOA) [3], [4], angle-of-departure (AOD), received-signal- tion and sensing through dynamic manipulation of the signal
strength (RSS) [8], [9], and their hybrid counterparts [10]. propagation environment [12], [14]. RISs have emerged
Time-based measurements rely on synchronization [4], [6], as a pivotal tool in propelling cost-effective and energy-
while RSS-based measurements are subject to propagation efficient approaches for optimizing wireless communication.
effects [8]. Finally, AOA or AOD-based localization has been By modifying the characteristics of electromagnetic waves
a focal point in recent research [4], [11], [12], as it eliminate in the environment, RISs enable significant improvements in
the need for synchronization expanding its applicability to a wireless communication performance [11], [12], [15], [16].
broader range of scenarios [3], [4]. RIS-aided localization approaches leverage RISs to ac-
tively manipulate multipath propagation, which enhances the

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/

VOLUME , 1

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

precision and robustness of localization systems, extending A critical sub-problem in RIS-aided localization is AOD
the concept of multipath-aided localization to RIS [17]–[19]. estimation, which is necessarily performed in beamspace.
In [12], a closed-form least squares (LS) based localization Although deep learning (DL) has been applied for channel
method is proposed for three-dimensional (3D) localization. estimation or sensing in RIS-assisted systems [23]–[25], to
This method utilizes partially connected receiving RISs and the best of our knowledge, our work is among the first to
employs the popular subspace-based root multiple signal utilize deep neural networks for 2D-AOD estimation for RIS-
classification (R-MUSIC) method for angle estimation. How- assisted localization and synchronization. Recent advance-
ever, it is worth noting that the R-MUSIC method requires a ments in AOA estimation have introduced DL methods, in
high computation. In [19], a single RIS is utilized for local- order to deal with various imperfections, low-SNR condi-
ization, with the base station (BS) receiving signals reflected tions, or multipath, or to provide lower-complexity solutions
by the RIS for AOA and TOA estimation. The proposed than model-based counterparts. In [26] a hierarchical deep
method involves an algorithm conducting two 1D searches neural network (DNN) framework for 1D-AOA estimation
over the TOAs and one two-dimensional (2D) search over is proposed, utilizing a multitask autoencoder and parallel
the AODs, resulting in high computational complexity. Ad- multilayer perceptron (MLP) classifiers for denoising and
ditionally, due to the propagation path of the probing signal AOA classification across spatial subregions. However, limi-
traversing through the BS-RIS-user equipment (UE)-RIS- tations in low-SNR conditions and a lack of generalization to
BS, significant signal attenuation is incurred. Consequently, 2D scenarios are observed in previous studies [27], [28]. A
the method demonstrates limited accuracy, particularly over subsequent study in [27] presents a deeper DNN architecture
longer ranges. The study in [20] achieves environment map- with three residual blocks, relying heavily on computation
ping and user localization through array signal processing for training and output spectrum calculation. Both methods
and atomic norm denoising, utilizing two RISs for channel employ the upper triangular elements of the correlation
parameter estimation. More recently, [21] demonstrates that matrix as input vectors, overlooking the 2D matrix structure
the absolute positions of UEs can be estimated through the and resulting in suboptimal AOA performance due to lost
assistance of at least two RISs and sidelink communication features. Later in [29], a two-stage method is introduced
between the UEs, even in the absence of BSs. In recent for 2D-AOA estimation. Firstly, using a similar technique
years, various maximum likelihood (ML) based approaches to [26], a deep convolutional network (DCN) classifies
have been proposed for localization using RISs [11], [17], 2D angles into discrete subregions of the arrival plane.
[22]. ML-based methods, while asymptotically optimal, de- Subsequently, the orthogonal matching pursuit algorithm
pend on a known signal model at the receiver to achieve estimates 2D angles within each subregion. In [30], the
optimal performance. Thus in the presence of multipath, due authors introduce a similar approach using MLP classifiers
to the unknown number of non-line-of-sight (NLoS) paths, for 1D-AOA estimation. Initially, they employ a full-row
ML methods have an inconsistent performance. Moreover, Toeplitz matrices reconstruction (FTMR) algorithm to utilize
the nonconvex nature of ML estimators necessitates the all rows of the sample covariance matrix (SCM), followed
use of iterative solvers, leading to significant computational by computing the sum of squares of these matrices. Utilizing
demands at each iteration. In [22], a semi-passive RIS is the forward/backward spatial smoothing (FBSS) technique,
employed which utilizes the MUSIC algorithm for angle similar to the MUSIC algorithm, they then perform eigen-
of arrival (AOA) estimation. Subsequently, the received value decomposition (EVD) to identify the noise subspace.
sensor data is transformed into frequency-domain sequences The resulting polynomial coefficients are used as features for
using discrete Fourier transform (DFT), enabling the ML the MLP classifiers. Although this method reduces input size
estimator to jointly estimate the TOA and the UE related compared to previous DL-based approaches, the integration
coefficient. Among ML methods, [17] also addresses the of FTMR, FBSS, and EVD significantly increases com-
synchronization using a single BS and a single RIS in 2D putational demands. Moreover, [28] notes its performance
scenarios. In particular, first, it proposes a relaxed ML-based limitations in low-SNR scenarios. In [31], a convolutional
method (RML), which searches the uncertainty environment neural network (CNN) is proposed for 1D-AOA estimation,
to obtain an initial guess for both the location of the UE and utilizing the 2D covariance matrix as input. The CNN com-
the clock offset for synchronization. Thus, it depends on the prises four convolutional layers and three fully-connected
dimensions of the environment, which requires significant layers. However, it exhibits high computational complexity
computation, especially in large 3D environments. Next, it due to its complex architecture. Recently, [32] proposed
proposes a joint ML based method (JML) that aims to opti- a deep residual network (ResNet) utilizing raw in-phase
mize the ML estimator accurately. This is achieved by uti- (I) and quadrature (Q) components of the received signal
lizing the derivative-free Nelder-Mead algorithm. However, as input. However, its input is dependent on the number
when generalizing this method to 3D muti-RIS scenarios, of snapshots, limiting its generalizability. Additionally, the
the algorithm is computationally demanding and requires a increased depth of ResNets leads to higher complexity during
high number of iterations to converge. testing. Furthermore, [28] improves 2D-AOA estimation by
integrating attention mechanisms into CNNs, while [33]

2 VOLUME ,

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

investigates multi-source 2D-AOA estimation using uniform multipath effect, ensuring its robustness against this
circular arrays (UCAs) under receiver mobility. In all of phenomenon. For simplicity, in the rest of the paper,
the aforementioned DL-based methods for AOA estimation, this network is referred to as VNet.
scenarios involving multipath and RISs are not considered. • A low-complexity 3D localization routine: A novel
This paper addresses this issue by employing two techniques: low-complexity two-stage method is proposed to esti-
effective dataset generation taking into account the multipath mate the location and the clock offset of the UE in
and the utilization of a less common loss function to mitigate the presence of a single BS and multiple RISs. In the
the impact of multipath on AOA (or AOD) estimation first stage (SG1), a hybrid asynchronous AOD/TOA-
performance. based method is proposed to estimate the delays of
The presence of multipath imposes requirements on reso- the received signals and an initial guess for the UE’s
lution. Resolution can be provided in the delay domain (by position, using the proposed VNet and beamforming
the use of larger bandwidths) or in the angle domain (by the techniques. Moreover, the clock offset is also estimated
use of larger arrays). Various array structures can be used in this stage. In the second stage (SG2), a hybrid
to provide 2D angle resolution, including uniform planar AOD/TOA-based optimization problem is formulated
arrays (UPA), circular arrays, conformal arrays, and L- to refine the estimated position in SG1. An efficient
shaped arrays. While UPA structures have been more widely iterative approach is proposed for solving this problem
considered in RIS-aided localization [11], [12], [14], [19], using the estimated position from SG1 as the starting
[21], the advantages of other array topologies are given less point of the algorithm.
attention in previous studies. Linear arrays for RISs and BSs • Comprehensive complexity and performance analy-
have been recently considered in [34] for UE localization. sis: Our comparative study validates that the proposed
Recently, [35] introduced a novel circular RIS architecture method shows more robustness against the multipath
for precise environmental information acquisition and unique effect compared to the existing methods. Moreover, our
decoupling of channel parameters for user localization, not complexity analysis demonstrates the superiority and
feasible with traditional UPA RIS topologies. Furthermore, efficiency of the proposed method.
in [36], the authors suggest employing conformal metasur-
faces on vehicles’ bodies to alleviate blockage effects by The rest of this paper is organized as follows. Sec. II
generating artificial reflections. This compensates for the presents the system model and problem formulation. Sec. III
non-flat shape of the vehicle’s body using appropriate phase simplifies the signal model and defines some key notations
patterns. and metrics. Moreover, it details the design of beamforming
In this paper, we introduce RISs with L-shaped topolo- at the BS and RIS phase profiles. Sec. IV introduces and
gies which have several advantages over other topologies investigates the details of the signal pre-processing and
including UPAs. First of all, for a given aperture (and thus the proposed VNet. Sec. V, presents the details of the
angular resolution) L-shaped arrays exhibit a significantly proposed two-stage method. Sec. VI carries out extensive
reduced number of elements compared to UPAs [37]–[40]. simulations to demonstrate the effectiveness and efficiency of
Secondly, by employing L-shaped arrays, we can decompose the proposed method. Finally, Sec. VII concludes the paper.
RIS elements into two symmetric ULAs. Thus, the 2D-
AOD estimation problem can be easily decomposed into
Notations
two simpler 1D-AOD estimation problems, which results in
a considerable reduction in computational complexity [38]– Matrices are denoted by capital bold letters, such as X , while
[41], compared to UPAs. For a given number of elements, vectors are represented by bold lowercase letters, such as
L-shaped arrays, thanks to their higher subarray widths, have x. The submatrix Xu,v is defined as the selection of rows
better AOD estimation performance compared to UPAs, so indexed by u and columns indexed by v from the matrix X .
there is no loss in terms of SNR. If : is used instead of u or v , it indicates the selection of
In this paper, we consider the multi-RIS-aided localization all rows or columns, respectively, from X . The superscripts
problem in a 3D complex multipath environment using L- (.)T , (.)H and (.)−1 denote vector or matrix transpose,
shaped RIS. We present a novel two-stage hybrid AOD/TOA hermitian, and inverse, respectively. [x1 , . . . , xn ] shows the
localization method aided by deep learning for joint user horizontal concatenation of the vectors x1 , . . . , xn . The no-
localization and synchronization. The main contributions of tation diag(x) represents a diagonal matrix constructed using
this work are as follows: the vector x as its diagonal. Similarly, diag(X) refers to a
vector composed of the diagonal elements of the matrix X .
• A novel architecture for deep learning-based AOD For a matrix X whose elements are in dB, Xlin represents
estimation: An efficient low-complexity deep learning- the linear values of X , where [Xlin ]i,j = 10[X]i,j /10 . The
based architecture is proposed to initially estimate the n × n identity matrix is shown by In , and 1M denotes an
AODs of the virtual line of sight (VLoS) paths between M × 1 all ones vector. ∥X∥F and ∥x∥ show the Frobenius
the RISs and the UE. In contrast to existing DL- and l2 norms, respectively, and finally, A ⊗ B and A ⊙ B
based methods, our approach effectively addresses the are the Kronecker and Hadamard productions, respectively.

VOLUME , 3

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

𝒛
by √
yt [n] = P hTt [n]ft st [n] + vt [n], (2)
𝒛
𝒙
𝒔𝒑𝟏 𝒑𝑏
𝒙
𝒛
𝒙
where st [n] ∈ C is the n-th component of the OFDM symbol
St ∈ CNs ×1 in the t-th transmission, ft ∈ CMb ×1 represents
the beamforming vector at the BS in the t-th transmission.
LoS Blocker 𝒔𝒑𝟐 RIS-UE path
BS-RIS path
Moreover, P denotes the total transmitted power, vt [n] is the
𝒔𝒑𝟐 BS-SP-UE path zero-mean complex Gaussian noise with variance σv2 , and
𝒔𝒑𝟏
𝑧 𝑧
RIS-SP-UE path
ht [n] ∈ CMb ×1 signify the channel between the BS and UE
Base station Reconfigurable intelligent surface User equipment
BS-UE scatter point
for the n-th subcarrier and t-th transmission. This channel
𝑥 𝑥 RIS-UE scatter point
can be expressed as follows, assuming the LoS path between
FIGURE 1: System model where the LoS path between the UE and BS is the UE and BS is blocked
blocked. The G RISs provide VLoS paths, useful for localization of the UE. G
X
hTt [n] = hTb,u [n] + hTrg ,u [n]Ωg,t Hb,rg [n], (3)
II. System Model g=1
As illustrated in Fig. 1, we consider a BS, G RISs, and a where hb,u [n] ∈ C Mb ×1
denotes the channel from the BS to
UE in an indoor or outdoor environment. We aim to estimate the UE, hrg ,u [n] ∈ CMg ×1 is the channel from the g -th RIS
the UE location and clock offset in the presence of NLoS to the UE, and Ωg,t ∈ CMg ×Mg represents the phase control
paths using the downlink signal received from the BS via coefficients of the g -th RIS during the t-th transmission for
established VLoS paths with RISs. which we adopt a commonly used model, as seen in prior
literature [17], [21], [34],
A. Geometry Model 1 Mg

Both the BS and each of the G RISs are equipped with Ωg,t = diag(ejωg,t , . . . , ejωg,t ), (4)
an L-shaped array, comprising two orthogonal subarrays. where m
ωg,t∈ [0, 2π) is the amount of phase change by the
Specifically, one subarray is oriented in the elevation direc- m-th element of the g -th RIS at the t-th transmission. In
tion (perpendicular to the ground), while the other is in the addition, Hb,rg ∈ CMg ×Mb denotes the channel from the
azimuth direction (parallel to the ground). Let Mbθ and Mgθ BS to the g -th RIS.
denote the number of elements in the elevation subarray of
the BS and the g -th RIS, respectively. Similarly, Mbψ and C. Channel Model
Mgψ present the number of elements in azimuth subarrays. The tandem channel ht [n] comprises three parts: the channel
Thus, the total of Mb = Mbθ + Mbψ and Mg = Mgθ + Mgψ from BS to UE hb,u [n], the channel from BS to RIS Hb,rg [n]
elements are utilized at the BS and the g -th RIS, respectively. and the channel from RIS to UE hrg ,u [n]. These are detailed
Moreover, pb ∈ R3×1 , prg ∈ R3×1 , and pu ∈ R3×1 present next.
the position of the centers of the BS, the g -th RIS, and the
UE, respectively, where g = 1, . . . , G. We assume that the
LoS path between the BS and UE is blocked. In addition, 1) BS-UE Channel
scatter points (SPs) are utilized to model the NLoS paths for Based on the geometric channel model [43], since the LoS
the multipath effect as shown in Fig. 1. channel between the BS and UE is blocked, the first tandem
In this work, it is assumed that the UE resides within the channel hb,u [n] is defined as follows assuming the presence
(i) Ib,u
far-field (FF) region relative to each RIS. According to [42], of Ib,u SPs located at {pb,u }i=1 affecting this channel:
the FF condition is met when the distance d between the UE Ib,u
and any RIS surpasses X (i) (i) (i)
hb,u [n] = αb,u e−j2πτb,u n∆f ab (Φdb,u ), (5)
DF = 2D2 /λ, (1) i=1
(i)
(i) (i)
with λ indicating the carrier wavelength and D specifying where αb,u = ρb,u ejφb,u is the complex gain between BS
the RIS’s aperture size, defined as the greatest distance (i) (i)
and UE through the i-th SP, with ρb,u and φb,u denoting
between any pair of elements on the RIS. This assumption (i)
its modulus and phase components. τb,u represents the delay
will be verified later when presenting the results.
from BS to UE via the i-th SP up to a clock offset δ , which
can be obtained as follows
B. Signal Model (i) (i)
The BS communicates with the UE via G RISs by trans- (i) ∥pb − pb,u ∥ + ∥pb,u − pu ∥
τb,u = + δ, (6)
mitting orthogonal frequency division multiplexing (OFDM) c
pilots over Ns subcarriers across Nt transmissions. The where c is the speed of light. In (5), ab (.) ∈ CMb ×1 denotes
subcarrier spacing, denoted by ∆f , can be determined by the BS array steering vector. With the assumption that the
∆f = B/Ns , where B represents the bandwidth. In partic- element located at the origin of the array belongs to the
ular, the t-th transmission over the n-th subcarrier is given elevation subarray, this vector is defined as follows for the

4 VOLUME ,

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

𝒛 Here, Φab,rg = [θb,r

a a
, ψb,r ]T and Φdb,rg = [θb,r d d
, ψb,r ]T , as
g g g g
illustrated in Fig. 2, denote the AOAs and AODs from the
a d
𝒛 𝜽𝒅𝒃,𝒓𝒈 (𝒊)
BS to the g -th RIS, respectively. θb,r , θb,r are the elevation
𝜽𝒅𝒃,𝒖 a d
g g

𝒙 AODs, and ψb,rg , ψb,rg correspond to the spatial angles with

𝜽𝒂𝒃,𝒓𝒈 respect to the x-axis of the array, which are defined similar
𝒑𝑏 𝒙
𝝍𝒂𝒃,𝒓𝒈 (𝒊)
𝝍𝒅𝒃,𝒖 to (8) as
𝜽𝒅𝒓𝒈,𝒖 𝝍𝒅𝒃,𝒓𝒈    
𝒔𝒑𝒊 a [pb;rg ]3 d [prg ;b ]3
(𝒋)
𝜽𝒅𝒓𝒈,𝒖 𝝍𝒅𝒓𝒈,𝒖 θb,r = arccos , θ b,rg = arccos ,
g
∥pb;rg ∥ ∥prg ;b ∥
𝒅(𝒋)
𝝍𝒓𝒈,𝒖 𝒑𝑢 (11)
   
𝒔𝒑𝒋 a [p ]
b;rg 1 d [p ]
rg ;b 1
ψb,r = arccos , ψb,r = arccos ,
g
∥pb;rg ∥ g
∥prg ;b ∥
FIGURE 2: Definitions of AODs from RISs and from the BS and AOAs at
(12)
the RISs.
where prg ;b = RbT (prg − pb ) and pb;rg = RgT (pb − prg ),
elevation (θ) and spatial angle (ψ ) pair Φ = [θ, ψ]T : with Rg ∈ R3×3 being the rotation matrix corresponding to
θ the orientation of the g -th RIS.
ab (Φ) =[1, ejkd cos(θ) , . . . , ejk(Mb −1)d cos(θ) ,
ψ 3) RIS-UE Channel
ejkd cos(ψ) , . . . , ejkMb d cos(ψ) ]T , (7) Assuming the presence of Irg ,u SPs affecting the channel
where k = 2π/λ denotes the wave number with λ = c/fc from the g -th RIS to UE, the third tandem channel hrg ,u [n]
being the wavelength and fc the carrier frequency, and d is in (3) is defined as
the distance between any two adjacent elements. The spatial hrg ,u [n] = αrg ,u e−j2πτrg ,u n∆f arg (Φdrg ,u )
(i)
d(i) d(i) T d(i)
angle is split into Φdb,u = [θb,u , ψb,u ] , where θb,u is
Irg ,u
d(i)
the elevation AOD, and ψb,u denotes the spatial AOD with X −j2πτr(i) n∆f (i)
+ αr(i)
g ,u
e g ,u arg (Φdrg ,u ), (13)
respect to the x-axis of the array, which are shown in Fig. 2.
i=1
It is important to highlight that the latter differs from the
actual azimuth angle. Specifically, the relationship between in which Φdrg ,u = [θrdg ,u , ψrdg ,u ]T denotes the elevation and
the spatial angle ψ and the elevation and azimuth pair spatial (with respect to x-axis) AODs of the LoS path from
(θ, ϕ) is given by cos(ψ) = sin(θ) cos(ϕ). The geometric the g -th RIS to the UE, defined similarly to (11). In addition,
d(i) d(i) αrg ,u = ρrg ,u ejφrg ,u is the complex channel gain between
relationships of the angles θb,u and ψb,u with the positions (i) (i) jφ(i)
of the BS, UE, and SPs are presented as follows: the g -th RIS and the UE. The notation αrg ,u = ρrg ,u e rg ,u
(i)
! (i)
! is defined similarly, which corresponds to the i-th SP with
[pb,u;b ]3 [pb,u;b ]1 (i) (i) (i) (i)
d(i)
θb,u = arccos d(i)
, ψb,u = arccos , position prg ,u ∈ R3×1 . Φdrg ,u = [θrdg ,u , ψrdg ,u ]T is defined as
(i) (i) (i)
∥pb,u;b ∥ ∥pb,u;b ∥ the AODs from the g -th RIS to prg ,u , obtained in a similar
(8) manner to (11). In (13), τrg ,u is the delay between the g -th
(i) (i)
where pb,u;b = RbT (pb,u − pb ) denotes the coordinates of RIS and the UE up to a clock offset δ ,
the i-th SP with respect to the local coordinate system of the
∥pu − prg ∥
BS. Here, Rb ∈ R3×3 depicts the rotation matrix defining τrg ,u = + δ. (14)
the orientation of the BS. c
(i)
Similarly, τrg ,u denotes the NLoS delay between the g -th
RIS and the UE through the i-th SP,
(i) (i)
2) BS-RIS Channel ∥prg ,u − prg ∥ + ∥pu − prg ,u ∥
Assuming LoS path between BS and RISs, the second τr(i)
g ,u
= + δ. (15)
c
tandem channel Hb,rg [n] is defined as It is noteworthy that in (14) and (15), the inclusion of
Hb,rg [n] = αb,rg e−j2πτb,rg n∆f arg (Φab,rg )aTb (Φdb,rg ), (9) the clock offset δ , addresses synchronization discrepancies
specifically between the BS and the UE. Importantly, since
where αb,rg = ρb,rg ejφb,rg is the complex gain between BS the RISs operate passively, they inherently do not introduce
and the g -th RIS, with ρb,rg and φb,rg being its modulus any clock offset.
and phase components, respectively. Here, τb,rg expresses Moreover, it is important to note that while the detection
the delay between the g -th RIS and the BS as τb,rg = and estimation of NLoS paths (i.e., SP locations) have the
∥pb − prg ∥/c. The g -th RIS steering vector arg (.) ∈ CMg ×1 potential to enhance system performance, the realized im-
is defined similarly to (7) as provement is limited by the considerable path loss associated
θ
arg (Φ) =[1, ejkd cos(θ) , . . . , ejk(Mg −1)d cos(θ) , with these paths and the fact that the number of such paths
ψ is a priori unknown [21]. In this study, similar to [17], [21],
ejkd cos(ψ) , . . . , ejkMg d cos(ψ) ]T . (10) [34], we interpret these NLoS paths as interference. Unlike

VOLUME , 5

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

conventional methodologies, our proposed method neither (b,i)

and XNLoS denotes the i-th NLoS signal from the BS to UE
ignores nor attempts to estimate or detect these paths and through the i-th SP. Moreover, V ∈ CNs ×Nt denotes the
thus, we do not estimate the numbers Ib,u or {Irg ,u }G g=1 .
(g)
noise matrix with elements [V ]n,t = vt [n]. XVLoS can be
Instead, as shown in Sec. IV-B, it leverages machine learning obtained as
techniques to intelligently mitigate the impact of NLoS (g)
√
paths. XVLoS = P αrg d(τrg )bTrg (Φdrg ,u , Φab,rg )Wg
 
× diag aTb (Φdb,rg )F , (22)
III. Preliminaries: Metrics and Phase Profile Design
In this section, the signal model and beamforming models with Wg = [ωg,1 , . . . , ωg,Nt ], F = [f1 , . . . , fNt ]. In (21)
(g,i) (g)
are further elaborated and specialized, in order to simplify XNLoS is defined similarly to XVLoS in (22), corresponding
(b,i)
the processing. to the i-th SP from the g -th RIS to UE. Finally, XNLoS can
be represented as
(b,i)
√ (i) (i) (i)
A. Signal Simplification XNLoS = P αb,u d(τb,u )aTb (Φdb,u )F . (23)
In this section, we simplify the signal model and derive The signal-to-noise ratio (SNR) and the line-of-sight to
equations that will be used in the following sections. By multipath ratio (LMR) metrics are defined as follows:
substituting (5), (9) and (13) into (3), the expression is
simplified to the following compact form: ∥XVLoS ∥2F ∥XVLoS ∥2F
SNR = , LMR = . (24)
Ib,u ∥V ∥2F ∥XNLoS ∥2F
X (i) (i) (i)
hTt [n] = αb,u [d(τb,u )]n aTb (Φdb,u ) (16) These terms will be used in Sec. VI to evaluate the perfor-
i=1 mance of the proposed method.
G 
X B. Beamforming and RIS Phase Profile Design
+ αrg [d(τrg )]n aTrg (Φdrg ,u )Ωg,t arg (Φab,rg )aTb (Φdb,rg ) In order to cover the environment around the RISs, based
g=1
on [44]–[46], an optimized choice of codebooks for phase
(17) profiles {W }G in terms of improving the received signal
g g=1
Irg ,u  power at the UE, is the DFT matrix W , whose (i, j)-th
X (i) g
+ αr(i)
g
[d(τr(i)
g
)]n aTrg (Φdrg ,u )Ωg,t arg (Φab,rg )aTb (Φdb,rg ) , elements is given by [W g ]i,j = φg(j−1)(i−1) , 1 ≤ ∀i, j ≤
i=1 Mg , and φg = e−j2π/Mg for g = 1, . . . , G.
where αrg = αrg ,u αb,rg and τrg = τrg ,u + τb,rg denote the Regarding BS beamforming design, we recall that the LoS
total complex gain and the total delay between the BS and path between the BS and UE is assumed to be blocked
UE through the g -th RIS, respectively. The delay steering (see Fig. 1). So in this paper we use VLoS paths between
vector d(.) ∈ CNs ×1 is defined as BS and the UE established by RISs for joint localization
d(τ ) = [1, e−jξ1 τ , . . . , e−jξNs −1 τ ]T , (18) and synchronization. Since the locations of the BS and
(i) RISs pb and {prg }G g=1 are known beforehand, the angles
where ξn = 2πn∆f . Similarly, the notations αrg = {Φa , Φd }G are also known. Thus, we design the BS
(i) (i) (i) (i) (i) b,r g b,r g g=1
αrg ,u αb,rg and τrg = τrg ,u + τb,rg are defined for the i- beamforming matrix F using the optimal directional beam-
th SP between the g -th RIS and UE. We can express forming to concentrate the power P towards individual RISs,
aTrg (Φdrg ,u )Ωg,t arg (Φab,rg ) = bTrg (Φdrg ,u , Φab,rg )ωg,t , (19) which maximizes the SNR at the RISs [17], [34]. Hence,
the codebook F ∈ CMb ×G for designing the beamforming
where ωg,t = diag (Ωg,t ) and the notation brg (., .) ∈ CMg ×1 matrix F is chosen as
is defined as
F = a∗b Φdb,r1 , . . . , a∗b Φdb,rG .



(25)
brg (Φ1 , Φ2 ) = arg (Φ1 ) ⊙ arg (Φ2 ), (20)
Hence, using the codebooks W g and F , the BS beam-
for Φ1 = [θ1 , ψ1 ]T and Φ2 = [θ2 , ψ2 ]T . For simplicity, in forming matrix F and RIS phase profiles {W }G are
g g=1
the rest of the paper, we assume that st [n] = 1 for all t = designed as follows:
1, . . . , Nt and n = 1, . . . , Ns . After simplifying (16) using
1
[F ]:,1 ⊗ 11×M1 , . . . , [F ]:,G ⊗ 11×Mg ,
 
(19) and substituting the simplified equation into (2), and F =√
stacking all Ns Nt received signals over all Nt transmissions M b Nt
and Ns subcarriers, we obtain the matrix Y ∈ C Ns ×Nt
as (26)
(
G Ib,u G IX
rg ,u [W g ]:,(t−Lg−1 ) , Lg−1 + 1 ≤ t ≤ Lg
X (g)
X (b,i)
X (g,i) [Wg ]:,t = ,
Y = XVLoS + XNLoS + XNLoS +V , (21) 0 (absorption mode), otherwise
g=1 i=1 g=1 i=1 Pg
| {z } | {z } where Lg = g ′ =1 Mg for g = 1, . . . , G with L0 = 0.
XVLoS XNLoS
Hence for each subcarrier, PG the total number of transmissions
(g) (g,i)
where XVLoS and XNLoS are the noise-free received VLoS is equal to Nt = g=1 Mg . In (26),
√
each beamforming
and i-th SP NLoS signals through the g -th RIS, respectively, vector is divided by a factor of Mb to normalize the

6 VOLUME ,

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

vectors and ensure that the total transmitted power remains that the SPs between the BS and UE approximately do not
constant throughout
√ the entire period. Moreover, we divide interfere with the VLoS paths (though they will be included
each vector by Nt to satisfy the condition tr(F F H ) = 1. in the simulations). On the contrary, since the phase profiles
Only Mg beams with indices Lg−1 + 1 ≤ t ≤ Lg are of RISs are designed to illuminate the space in front of them
concentrated towards the g -th RIS. Thus, almost no power is uniformly, the SPs between RISs and UE are more likely to
received from other samples at the UE from the BS through interfere with the VLoS paths at the receiver. Our proposed
the g -th RIS. Thus, for t ≤ Lg−1 or Lg +1 ≤ t, the g -th RIS deep learning-based AOD estimator explained in Sec. IV-B2,
is configured in absorption mode to prevent signal scattering. meticulously considers these SPs to mitigate their impact.
In conclusion to this section, we emphasize that the focus Next, the pLS estimates of the channels are obtained by
of this paper is on a downlink scenario. In this framework, multiplying Mg W −1 g
1
by the right side of (28) as
all methodologies proposed in the subsequent sections are q
specifically designed for autonomous operation at the UE X̂g = P Mg Mb /Nt αrg d(τrg )bTrg (Φdrg ,u , Φab,rg )
level, for instance within a mobile device. The beamform- Irg ,u
Xq (i)
ing designs in (26) ensures compatibility with scenarios + P Mg Mb /Nt αr(i) d(τr(i) )bTrg (Φdrg ,u , Φab,rg )
g g
involving multiple UEs. As a result, our proposed approach i=1
supports independent implementation at each UE, thereby + Ib,u,g + Vg ,
′
(30)
enabling personalized 3D localization and synchronization
based on the signals each device receives. where Ib,u,g is the (small) p
interference term stemming from
′
hb,u [n]. Moreover, Vg = Mg Vg W −1 g denotes the noise
IV. Proposed VNet AOD Estimation
matrix after LS estimation with zero-mean and the same
For clarity and systematic presentation, the proposed method covariance matrix σv2 IMg as Vg . This LS estimation has a
is divided into sequential stages. In the first stage, the re- complexity of O(Ns Mg2 ) for the g -th RIS. Subsequently,
ceived signal Y , as defined in (21), undergoes preprocessing since the AOAs Φab,rg are known by the UE, we multiply
 
to yield G matrices, which are provided to VNet, to yield (30) by the matrix diag a∗rg (Φab,rg ) to eliminate the depen-
AOD estimates. The subsequent positioning and refinement dency of X̂g on Φab,rg , which results the following calibrated
stages are deferred to Sec. V. signal:
q
A. Generating the VNet Input X̂gc = P Mg Mb /Nt αrg d(τrg )aTrg (Φdrg ,u )
First, Y from (21) is partitioned into G submatrices Yg ∈ Irg ,u
Xq
(g) (i)
CNs ×M as + P Mg Mb /Nt αr(i)
g
d(τr(i)
g
)aTrg (Φdrg ,u )
i=1
Yg = [Y ]:,(Lg−1 +1):Lg , g = 1, . . . , G. (27) c ′′
+ Ib,u,g + Vg , (31)
According to the designed RIS phase profiles in (26), the
g -th RIS only reflects in Mg transmissions with indices t = which is done with complexity
 O(Ns Mg ) for the g -th RIS.
′′ ′
Lg−1 + 1, . . . , Lg . Thus, Yg can be obtained as: In (31), Vg = Vg diag a∗rg (Φab,rg ) denotes the calibrated
√ noise matrix with zero-mean and the covariance matrix of
Yg = P αrg d(τrg )bTrg (Φdrg ,u , Φab,rg )W g diag(ug )
σv2 IMg .
Irg ,u
X√ (i)
+ P αr(i)
g
d(τr(i)
g
)bTrg (Φrdg ,u , Φab,rg )W g diag(ug ) B. The Proposed VNet
i=1
Ib,u
Although (31) represents a conventional model for an-
X√ (i) (i) gle/delay estimation, this paper diverges from traditional
+ P αb,u d(τb,u )u(i)
g + Vg , (28)
methods by introducing a deep learning-based approach de-
i=1
signed to mitigate interference impact with lower complexity.
where Vg = [V ]:,(Lg−1 +1):Lg , and Specifically designed to excel in challenging scenarios char-
(i)
ug = aTb (Φdb,rg )Fg , u(i) T d acterized by low-SNR, low-LMR conditions, and proximity
g = ab (Φb,u )Fg (29)
1×Mg
of SPs to the UE, the proposed approach addresses these
where Fg = [F ]:,(Lg−1 +1):Lg = [F ]:,g ⊗ 1 . Hence, complexities effectively, as will be verified later in Sec. VI
according to BS precoder design in (25) and (26), we
√deduce in comparison to existing methods.
p (i)
that diag(ug ) = Mb /Nt IMb and ug = βi 1TMg / Mb Nt , This section proposes a low-complexity deep learning
(i)
where βi = aTb (Φdb,u )a∗b (Φdb,rg ) denotes the gain of the BS structure to obtain an initial estimate of the VLoS AODs
array at the direction of the i-th SP. Observe that βi depends {Φdrg ,u }G
g=1 which will be refined in the subsequent sections.
on the number of BS array elements Mb and the position of
the i-th SP. We assume that the SPs are not located on the W g is a DFT matrix, W −1
1 Since
g = M1 W H g . For simplicity, the
g
LoS paths between the BS and RISs, so that |βi | becomes
p
scalar Mg is also multiplied to keep the variance of the noise, unchanged
negligible, especially for larger Mb . Hence, we conclude after the LS estimation.

VOLUME , 7

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

1) VNet Input (1)

f1
(1)
f2 l(1) d(1) o(1)
Conv2D Conv2D Flatten FC FC
Recall from (10), that the first Mgθ elements of arg (Φdrg ,u )
solely depend on the elevation AOD θrdg ,u and the rest R
(2)
f1
(2)
f2 l(2) d(2) o(2)
g
Mgψ elements only depend on the spatial AOD ψrdg ,u . Conv2D Conv2D Flatten FC FC

Hence X̂gc can be partitioned into two submatrices X̂gθ = ..

.

[X̂gc ]1:Mgθ ,: , X̂gψ = [X̂gc ](M θ +1):(M θ +Mgψ ),: . To eliminate (Q)
f1
Conv2d
(Q)
f2
Flatten
l(Q)
FC
d(Q)
FC
o(Q)
g g Conv2d
the dependency of these two submatrices on unknown vari-
(i) (i)
ables {αrg , τrg , αrg , τrg }, we compute the sample covari- FIGURE 3: Proposed architecture for the VNet, including Q parallel
ance matrix (SCM) as follows: branches, each covering a specific range of AODs.

Rb gθ = 1 X̂gθT X̂gθ∗
Ns
= Pg arg ,θ (θrdg ,u )aH d θ bθ
rg ,θ (θrg ,u ) + Cg,I + Rg,v , (32) works instead of a single network is to enable generalization
where arg ,θ (.) = [arg (.)]1:Mgθ ,: , and the first term, dom- and scalability with low complexity. By employing this
inated by power, depends on the elevation AOD θrdg ,u . technique, which is also used in [30], [47], we can seamlessly
θ expand the angle scope by incorporating a new subnetwork,
Additionally, R b g,v is the SCM of the noise covariance matrix
θ θ without the necessity of retraining existing networks. Con-
θ 2
Rg,v = σv IMgθ , and Cg,I θ
∈ CMg ×Mg denotes all remain- versely, for a more targeted angle scope, we can efficiently
ing interference terms. In addition, Pg = P Mg Mb ρ2rg /Nt focus on the corresponding subnetworks.
represents the overall gain of the g -th VLoS path through
the g -th RIS. Similar notations, such as R b gψ , ar ,ψ (.) =
ψ ψ ψ
g
• Input: The input is shared across all Q subnetworks.
[arg (.)]((Mgθ +1):Mg ),: , Cg,I , Rg,v and R b g,v are defined ac- (g)
The input is denoted by R and is either Rθ or Rψ .
(g)
cordingly. Since the neurons in neural networks only take • Layers: Given the input in (33), the output of the q -th
real values, we decompose the complex-valued SCMs R b gθ
subnetwork is given by
ψ
and R b g into their real and imaginary parts, yielding tensors
(q) (q)
θ θ
Mgψ ×Mgψ ×2 g (q) = o(q) (d(q) (l(q) (f2 (f1 (R))))). (34)
Rg ∈ RMg ×Mg ×2 and Rψ
θ
g ∈ R that can be
(q) (q)
represented as two-channel images. Moreover, Rθg and Rψ g The functions f1 (.) and f2 (.) denote the outputs of
are normalized to reduce the input variability and sensitivity convolutional layers followed by a batch normalization
of the network as follows: layer and a rectified linear unit (ReLU) layer used as
[ℜ{R b gθ }; ℑ{R b gθ }] [ℜ{R b gψ }; ℑ{R b gψ }] an activation function. The function l(q) (.) is a flatten
Rθg = , Rψg = . layer, and the functions d(q) (.) and o(q) (.) present the
∥R b gθ ∥F ∥R b gψ ∥F
fully connected layers (or dense layer), which construct
(33)
the classification part of the network.
The resulting normalized tensors are then fed into the
• Output: The output g (q) is of length K and relates to
proposed network. The overall complexity order of the (q) (q)
the desired output angle scope (denoted by [φmin , φmax ],
derivations for equations (32) and (33) can be expressed as
where φ presents either elevation (θ) or spatial (ψ )
O(((Mgθ )2 + (Mgψ )2 )Ns ). (q) (q)
AODs) and resolution r, so that K = (φmax −φmin )/(Q·
r).

2) VNet Architecture To obtain the overall output of the VNet, the outputs
As presented in (33), the input of the proposed VNet
depends on the number of elements in each subarray of RISs. h all Q subnetworks
of iT are concatenated as follows: g =
(1)T (Q)T QK×1
g ,...,g ∈ R+ .
Therefore, a separate neural network must be trained for
subarrays with a different number of elements. To simplify
the exposition, in the rest of the paper, we assume that all
subarrays in the RISs, whether horizontal or vertical, have
the same number of elements, i.e., M = Mgθ = Mgψ for 3) VNet Loss Function
g = 1, . . . , G. This assumption is made without loss of We use one-hot binary vectors for the labels of the training
generality and can be relaxed in practice by defining separate dataset. These vectors have exactly one non-zero element,
networks for different subarray configurations. corresponding to the true AOD of the VLoS path. To reduce
The proposed architecture of VNet is illustrated in Fig. 3. the impact of NLoS paths and thus to down-weight the con-
It consists of Q distinct subnetworks with identical struc- tribution of the zero classes in the loss function, we employ
tures, each responsible for predicting the AODs for a differ- an asymmetric loss function (ASL) [48] which generalizes
ent region of the output scope (i.e., non-overlapping ranges the commonly used binary cross-entropy (BCE). For the
of angles). The primary reason for considering Q subnet- output g ∈ RQK with the ground truth label y ∈ RQK ,

8 VOLUME ,

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

the total classification loss is defined as and I ∈ I be the number of NLoS paths in this data
QK
X sample.
LASL (g, y) = L(gl , yl ), where (35) • Based on the definitions of SNR and LMR in (24),
l=1 the total SNR of the I NLoS paths can be obtained
L(g, y) = −y(1 − g)γ+ log(g) − (1 − y)gϵγ− log(1 − gϵ ) as ΓNLoS = Γ − χ (in dB). The SNR of each of these
I NLoS paths might differ. To handle this, we define
where gϵ = max(0, g − ϵ), for a small ϵ > 0, and
Γdiff , and select the SNR of NLoS paths randomly in
the parameters γ+ and γ− serve as positive and negative
the interval [Γ − Γdiff /2, Γ + Γdiff /2] such that their
focusing parameters, respectively. The dynamic adjustment
sum equals ΓNLoS .
in the loss function via the parameters ϵ, γ+ , γ− allows
• let τmin and τmax denote the minimum and maximum
for reduced emphasis on simple classes.2 Note that when
delay between the BS and the UE through RISs.
γ+ = γ− = ϵ = 0, ASL reverst to BCE.
The proposed training dataset generation for VNet is
4) VNet Dataset Generation Strategy detailed in Algorithm 1. The notation a(φ) refers to the
The proposed VNet outputs the desired angle scope steering vector of a generic subarray with M elements:
[φmin , φmax ] for either the elevation AOD (θ) or spatial AOD a(φ) = [1, ejkd cos(φ) , . . . , ejk(M −1)d cos(φ) ]T . (36)
(ψ ), where φ denotes the AOD. This interval is evenly
divided into Z = QK grid points to obtain the vector V. Positioning and Refinement
φ = [φ1 , . . . , φZ ]T , as explained in the previous section. From the 2D-AODs from each RIS, the position can be
For simplicity, a single network is trained to estimate both recovered. With the estimated position, the clock offset can
the elevation (θ) and spatial (ψ ) AODs due to their symmetry. be estimated and all parameters can be refined. We proceed
Hence, we assume that φmin and φmax are determined based in two stages.
on the minimum and maximum values of these AODs.
Each of the signals X̂gθ and X̂gψ , defined in Sec. IV-B1 A. Stage 1 – Coarse Estimation
(obtained from (31)), represents a single-input multiple- 1) Position Estimation
output (SIMO) system with multiple subcarriers and a linear
Let {Φ̌drg ,u }Gg=1 denote the estimated elevation and spatial
array comprising M elements at the receiver, operating in
AODs by the VNet. For g = 1, . . . , G, define ℓg to represent
the presence of multipath and noise. Hence, to generate
the line passing through the center prg of the g -th RIS with
training data, as detailed in Algorithm 1, we propose this
unit direction ug ∈ R3×1 specified by the estimated AODs
equivalent interpretation to generate a LoS SIMO signal with
Φ̌drg ,u . For an arbitrary point v ∈ R3×1 , the sum of squared
a random delay and an AOD φz (line 7), in the presence
distances from the lines ℓ1 , . . . , ℓG is obtained as:
of multipath (line 13) and noise (line 15) with the correct
G G
statistical properties. We outline our proposed approach as X X
S(v) = ∥v − prg ∥22 − (v − prg )T ug uTg (v − prg )
follows:
g=1 g=1
• [Γmin , Γmax ] is defined to denote the desired SNR range G
X
(in dB). This range is evenly divided into NΓ values to = (v − prg )T (I3 − ug uTg )(v − prg ). (37)
obtain the vector Γ = [Γ1 , . . . , ΓNΓ ]T , with Γ1 = Γmin g=1
and ΓNΓ = Γmax . After setting the gradient of S(v) to zero, the point p̌u that
• [χmin , χmax ] presents the desired LMR range, and it is minimizes S(v) can be obtained with complexity O(G) as:
divided evenly into Nχ values to obtain the vector χ = G  G

†X
X
[χ1 , . . . , χNχ ]T , where χ1 = χmin and χNχ = χmax . p̌u = I3 − ug uTg I3 − ug uTg prg . (38)


• Let I = [I1 , . . . , INI ]T be a set of possible values g=1 g=1
for the number of SPs between one of RISs and UE
denoted by I ∈ {Ir1 ,u , . . . , IrG ,u }. It is important to
note that for a fixed LMR, higher values of I correspond 2) Clock Offset Estimation
to smaller power levels for each NLoS path. As a result, Next, all the delays are estimated. Using (38), the delay
fewer NLoS paths (I ) leads to more intricate scenarios. between the BS and the UE without considering the clock
• Let Γ ∈ Γ and χ ∈ χ be the selected SNR and LMR offset, can be obtained as:
for a specific data sample during dataset generation, ∥pb − prg ∥ + ∥p̌u − prg ∥
τ̂rg ,s = , (39)
2 This formulation enables the network to mitigate the impact of easily c
classified zero classes in two ways. Initially, adjusting parameters such as where the subscript s emphasizes not taking clock offset into
γ+ ≥ 0 and γ− > 0 softens the impact of easy classes (g ≪ 0.5) in the account. Based on (39), the uncertainty regions {Ig }G g=1 for
loss function. Subsequently, the introduction of the shift parameter ϵ ≥ 0 the delays {τrg }G g=1 can be defined as I g = [τ̂ r g ,s + δmin −
imposes a stricter threshold, discarding very easy classes (g ≤ ϵ) as outlined τϵ , τ̂rg ,s +δmax +τϵ ], where δmin and δmax are lower and upper
in [48]. bounds of the clock offset δ . Moreover, τϵ denotes a small

VOLUME , 9

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

Algorithm 1 Proposed Dataset Generation for Training VNet as follows:

G 
∥prg − pb ∥ + ∥prg − p̂u ∥

Inputs: φ, Γ, Γdiff , χ, I , τmin , τmax , Ns 1 X
Output: Dataset D with corresponding labels L δ̂ = τ̂rg − . (42)
G g=1 c
1: Define empty lists D, L. 3) Fine grid 2D AOD Estimation For Improved Coarse
2: for nΓ = 1 : NΓ do Positioning
3: for z = 1 : Z do Since the proposed VNet uses a relatively coarse grid for
4: Choose I and χ from I and χ, randomly. computational complexity reasons, the positioning perfor-
mance in (37) will be limited by this coarse grid. To refine
5: Generate I + 1 random unit modulus complex
the AOD estimation and the corresponding position estimate,
numbers c0 , . . . , cI (|ci | = 1).
we proceed as follows.
6: Choose a random delay τ0 ∈ [τmin , τmax ]. First, from (31), the LS estimate of the array steering
vector arg (Φdrg ,u ) is obtained as
XLoS ← ΓnΓ ,lin c0 a(φz )dT (τ0 )
p
7:
T

8: ΓNLoS ← ΓnΓ − χ (in dB) ârg (Φdrg ,u ) = X̂gc d∗ (τ̂rg ), (43)

with complexity O(GNs M ), which is then used to refine the
9: Choose I random values Γ(1) , . . . , Γ(I) from the
estimations of the elevation and azimuth AODs {Φdrg ,u }G
interval [ΓNLoS − Γdiff Γdiff
2 , ΓNLoS + 2 ], with the
g=1

sum ΓNLoS , and store them in ΓNLoS . θ̂rdg ,u = argmax aH (θ)[ârg (Φdrg ,u )]1:M ,
θ
10: Choose I random values τ1 , . . . , τI uniformly
ϕ̂drg ,u = argmax aH (ψ)[ârg (Φdrg ,u )](M +1):2M , (44)
from [τmin , τmax ], and store them in τNLoS . ψ

11: Choose I random values φ(1) , . . . , φ(I) from which are again 1D searches with overall complexity
[φmin , φmax ], and store them in φNLoS . O(G(NΘ + NΨ )M ), considering NΘ and NΨ grid points
for θ and ψ , respectively.
12: A ← [c1 a(φ(1) , . . . , cI a(φ(I) )], (based on (36)) Next, the estimated UE location p̌u is updated by em-
D ← [d(τ1 ), . . . , d(τI )], (based on (18)) ploying (38) once again, this time using the refined AODs
{ϕ̂drg ,u }G
g=1 . This yields the updated location estimate p̂u .
p
13: XNLoS ← A diag(ΓNLoS,lin )D T
14: Generate a standard complex Gaussian noise B. Stage 2 – Refined Position Estimation
matrix V ∈ CM ×Ns . The first stage, although providing a consistent estimation,
√
15: V ← M Ns V /∥V ∥F does not fully exploit the location information embedded
in the estimated delays {τ̂rg }G g=1 and the clock offset δ̂ .
16: Y ← XLoS + XNLoS + V To address this issue, first, we estimate the distance Rg
17: Obtain R ∈ RM ×M ×2 based on (32) and (33). between the center of the g -th RIS and the UE as R̂g =
c(τ̂rg − δ̂) − ∥prg − pb ∥, where τ̂rg and δ̂ are the estimated
18: Define the one-hot label y ∈ {0, 1}Z such that values for the delay through the g -the RIS and the clock-
[y]z = 1 and [y]i = 0 for i ̸= z . offset, respectively. Similar to Sec. V-A, the lines {ℓg }Gg=1
19: Append R and y to D, L, respectively. are defined using the refined AODs {ϕ̂drg ,u }G g=1 . Moreover,
let Sg denote the sphere with the center prg and radius
R̂g for g = 1, . . . , G. In the absence of noise, all the
lines {ℓg }G G
g=1 and spheres {Sg }g=1 pass through the UE’s
position. Motivated by this fact, we exploit this geometric
constant to determine the margin of error. Subsequently, we model by using the estimated AODs, delays, and the clock
obtain estimates of the delay steering vectors {d(τrg )}G offset to refine the UE’s position. To this end, we find a
g=1
using (31), point whose sum of squared distances from the lines and
ˆ r ) = X̂ c a∗ (Φ̌d ),
d(τ (40) the nearest point on the surface of spheres is minimum. The
g g rg rg ,u
optimization problem can be written as follows:
which has a complexity O(GNs M ). Next, using the beam- X G
forming technique, the delays {τrg }G
g=1 are estimated as: p̃u = argmin (pu − prg )T (I3 − ug uTg )(pu − prg )
ˆ r ), p u ∈R3
τ̂rg = argmax dH (τ )d(τ g
(41) g=1
τ ∈Ig G
X 
2
which is solved by a simple line search with complexity + (∥pu − prg ∥ − R̂g ) , (45)
O(GNIg Ns ), considering NIg grid points. Using the esti- g=1

mated UE position p̂u and the estimated delays {τ̂rg }Gg=1 in where the first term is accounted for the distances from the
the first stage, based on (14), we estimate the clock-offset δ lines, explained in Sec. V-A, and the term |∥pu − prg ∥ − R̂g |

10 VOLUME ,

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

TABLE 1: System parameters TABLE 2: Proposed method parameters

Default System Parameters Symbol Value Parameter Value Parameter Value
Carrier frequency fc 28 GHz Q 3 φmin 45◦
Environment Dimensions x0 × y0 × z0 100 × 100 × 10 [m] r 0.5◦ φmax 135◦
UE position pu [48, 51, 1]T [m] Γ [−18, −17, . . . , −5]T Γdiff 5 dB
BS position pb [0, 50, 7.5]T [m] χ [0, 1, . . . , 10]T τmin 0 ns
BS subarray length Mbθ , Mbψ 50, 50 I [1, . . . , 10]T τmax 1000 ns
RIS 1 position pr1 [50, 0, 10]T [m] Ns 200 Ng 500
RIS 2 position pr2 [50, 100, 5]T [m] δmin 0 ns τϵ 5 ns
RIS subarrays length M 15 δmax 30 ns θ0 0.01◦
Clock Offset δ 10 ns θϵ 1◦ ψ0 0.01◦
ψϵ 1◦ γ− , γ+ , ϵ 0, 3, 0.4

denotes the minimum distance between the point pu and the

VNet
surface of the sphere Sg . In general, the problem (45) is non- 101
R-MUSIC

RMSE [deg]
convex, and thus it may have multiple optimal solutions, and RSNet
the specific solution obtained depends on the initial guess of VNet (using BCE)

the optimization algorithm. To address this issue, we use the 100

previously estimated UE location p̂u in the first stage as the

initial point for gradient descent, with 10−1
G −18 −16 −14 −12 −10 −8 −6 −4 −2 0 2
SNR [dB]
X
∇pu f (pu ) =2 (I3 − ug uTg )(pu − p rg )
g=1 (a) LMR = 3 dB.
G  
X R̂g 101
+2 1− (pu − prg ), (46) VNet
∥pu − prg ∥ R-MUSIC
RMSE [deg]

g=1
RSNet
where f (.) denotes the objective function in (45). The 100 VNet (using BCE)
resulting order complexity is O(GNSG2 ), where NSG2 is the
number of gradient descent iterations until convergence.
10−1
0 2 4 6 8 10
LMR [dB]
VI. Simulation Results and Discussion
In this section, we evaluate the performance of the proposed (b) SNR = −5 dB.
joint 3D localization and synchronization scheme for RIS-
FIGURE 4: RMSE of the VNet and the state-of-the-art methods versus
assisted mmWave systems through numerical simulations.
SNR and LMR.

A. Simulation Setup
The considered RIS-assisted mmWave system consists of B. Baselines and Metrics
one BS, two RISs, and a single UE with default system We evaluate the 2D-AOD estimation performance in terms of
parameters presented in Table 1. Some of these parameters the root mean squared error (RMSE) of the proposed VNet
may change in different simulations. Moreover, to ensure against several baselines, namely:
that the FF condition is met, according to (1) and Table 1,
and the fact that the maximum distance between two el- • Classical R-MUSIC [49]: This method utilizes the
ementspin the considered L-shaped RIS configuration is roots of the characteristic equation derived from the
D = M 2 + (M − 1)2 λ/2 we have DF ≈ 2.25 m. The eigenvalues of the SCM to estimate the AODs.
UE position is selected to ensure a minimum distance of • Roots-spectrum network (RSNet) [30]: As explained in
DF from the RIS centers, satisfying this condition. Sec. I, this method utilizes FTMR followed by FBSS
The parameters for training the proposed VNet (Algo- technique, and then performs EVD to extract noise
rithm 1) and the constants in the two-stage proposed local- subspace features for its proposed network’s input,
ization method are presented in Table 2. Despite the selected which are used to estimate the MUSIC-like spatial
SNR and LMR range specified in Table 2 and a fixed value spectrum.
of Ns = 200 for training VNet, it will be observed that the • VNet using the classical BCE loss function: To demon-
VNet generalizes for higher SNRs and LMRs, as well as a strate the effectiveness of ASL, we employ the same
wide range of Ns . VNet structure with BCE loss function.

VOLUME , 11

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

We evaluate the localization performance of the SG1 esti- 1.2 True θ = 70◦ VNet

Pseudospectra
mator (from Sec. V-A) and SG2 estimator (from Sec. V-B) RSNet
◦
VNet (BCE)
θ̂VNet = 70.5 θ̂RSNet = 129.5◦
in terms of the RMSE against the following baselines: 0.8
θ̂VNet(BCE) = 71◦ θ̂R-MUSIC = 127.416◦

• The approach from [11], [12], which relies on R- 0.4

MUSIC AOD estimation and LS position estimation.
0
• The convergent iterative method (CIM) from [4], which
40 50 60 70 80 90 100 110 120 130
only performs positioning from AOD estimations. θ [deg]
Hence, for a fair comparison, we combine it with the
proposed VNet. (a) (SNR, LMR) = (−16, 0) dB.

• RML and JML proposed in [17], which are direct 1.2 True θ = 60◦ VNet

Pseudospectra
positioning methods as they compute the position esti- RSNet VNet (BCE)
mate directly from the observed waveforms. RML is a 0.8 θ̂VNet = 60◦ θ̂RSNet = 58.5◦
grid-based approach, while JML is based on a gridless θ̂VNet(BCE) = 59.5◦ θ̂R-MUSIC = 59.137◦
0.4
Nelder-Mead algorithm.
0
The system models used in these baselines may differ
50 60 70 80 90 100 110 120 130
from our considered system model. Therefore, for a fair
θ [deg]
comparison, certain modifications and generalizations have
been made to adapt these methods to our system. (b) (SNR, LMR) = (−14, 3) dB.

C. Results and Discussion FIGURE 5: Spectrum plots for two test data samples for different SNR
and LMR values.
1) Results of the Proposed VNet
Fig. 4a compares the RMSE results for a fixed LMR = 3 dB,
L-shaped: Mt = 16 UPA: Mt = 16
and Fig. 4b, shows the results in terms of LMR for a fixed L-shaped: Mt = 36 UPA: Mt = 36
RMSE [deg]

SNR = −5 dB. For Fig. 4a, at each SNR the number of 100 100
data samples are generated for every angle θ ∈ [45◦ , 135◦ ]
(with step 0.5◦ ), resulting the total of 18000 Monte Carlo
10−1
samples for each SNR. A similar process is done for Fig. 4b.
It can be seen that the proposed VNet outperforms the state-
of-the-art methods such as R-MUSIC, especially in low-SNR −14 −12 −10 −8 −6 −4 −2 0 2
and low-LMR regimes. The results also highlight the benefit SNR [dB]

of the proposed loss function over BCE. This superiority FIGURE 6: Comparing AOD estimation performance of L-shaped arrays
stems from the VNet’s ability to mitigate interference effec- with UPAs with the same number of array elements Mt .
tively.
The pseudospectra of the proposed VNet, RSNet, and
VNet trained with BCE are depicted in Fig. 5 for varying
SNR and LMR values, with the number of subcarriers azimuth AOD pair to be (θ, ϕ) = (99.416◦ , 120.528◦ ). For
set to Ns = 200. Due to the ASL function (defined in each scenario, we employ 1D-MUSIC twice for an L-shaped
(35)), the proposed method exhibits nonzero outputs at other array with M = Mt /2 for 2D-AOD estimation. For UPAs,
AODs, especially in low-SNR and low-LMR scenarios. This we utilize 2D-MUSIC for direct 2D-AOD estimation. Fig. 6
behavior, driven by the ASL, enables VNet to focus more illustrates the RMSE results across SNR ∈ [−14, dB, 2, dB]
on the dominant VLoS path, resulting in sharper peaks at for a fixed LMR = 10, dB using 1000 data samples at each
the true AODs compared to when trained using BCE loss. SNR. As anticipated, the larger aperture of L-shaped arrays
Notably, R-MUSIC and RSNet underperform in low-SNR results in more precise estimations compared to UPAs. This
and low-LMR regimes. highlights the effectiveness of L-shaped arrays in decom-
posing 2D-AOD estimation into two simpler 1D-AOD tasks
2) Comparison with UPA
while achieving superior accuracy. Consequently, L-shaped
To demonstrate the superiority of L-shaped arrays over
arrays offer enhanced AOD estimation performance, albeit
UPAs, we compare their RMSE results in this section. For
requiring larger subarray widths.
2D-AOD estimation with UPAs, we utilize 2D-MUSIC. To
this end it is sufficient to utilize X̂gc from (31) to derive
the SCM, following (32), for input generation. Subsequently,
EVD is performed to obtain MUSIC spectra. Let Mt repre- 3) Results of the Proposed Two-stage Method
sent the total number of elements in the corresponding RIS. In order to show that the proposed method can attain optimal
In this experiment, we consider two scenarios with Mt = 16 performance, we will first consider results without NLoS
and Mt = 36, respectively. We assume the true elevation and paths. This allows us to compare it to the CRLB. Then, we

12 VOLUME ,

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

JML RML LMR = 5 dB Proposed: SG1

Proposed: SG1 Proposed: SG2 LMR = 10 dB Proposed: SG2
101 R-MUSIC+LS VNet+CIM R-MUSIC+LS
101
RMSE [m]

CRLB VNet+CIM

RMSE [m]
JML
100

100
10−1

−18 −16 −14 −12 −10 −8 −6 −4 −2 0 10−1

SNR [dB]
FIGURE 7: Localization RMSE versus SNR without NLoS interference. −18 −16 −14 −12 −10 −8 −6 −4 −2 0
SNR [dB]
FIGURE 8: Localization RMSE results in the presence of NLoS paths.
will add multipath to show the robustness of the proposed
method.
Fig. 7 shows localization RMSE values of the UE com-
pared to the state-of-the-art methods. In this experiment, performance is accompanied by a significant reduction in
we assume that there are no NLoS paths (i.e., Irg ,u = computational complexity, particularly when contrasted with
0, Ib,u = 0). Note that since the CIM algorithm assumes ML-based methodologies.
that the estimates of AODs are available at the receiver,
we use the AODs obtained from the proposed VNet. We 4) Proposed Method Scalability
can make several observations. First of all, SG2 operates In this section, we conduct simulations to demonstrate the
close to the CRLB for SNRs greater than −14 dB, and as scalability of the proposed method in different scenarios.
the SNR increases, the gap between the RMSE and the First, we analyze the performance with respect to the number
CRLB decreases. In the absence of multipath effect, JML of subcarriers Ns and the number of elements M per RIS
eventually reaches the theoretical bound CRLB at SNRs subarray. It is noteworthy that the same VNet is used to
exceeding −10 dB. Consistent with its nature as a grid- obtain the result across Ns . However, since the input depends
based approach, the RML method saturates at −12 dB. In the on M , we train a separate network for each value of M
low-SNR regime, SG1 demonstrates superior performance to obtain the results. The training process utilizes the same
compared to the LS and CIM methods. parameters as outlined in Table 2.
In the next experiment, we compare the performance Fig. 9 illustrates the localization RMSE results across
of the methods in the presence of multipath in Fig. 8. different numbers of subcarriers Ns ∈ [50, 400] and Fig. 10
Consider a challenging scenario with Ir1 ,u = Ir2 ,u = 4 SPs for different number of elements M ∈ [10, 30] for different
(1) (1) SNR values. The results across Ns are obtained using a
between RISs and UE with same3 positions pr1 ,u = pr2 ,u =
(2) (2) (3) (3) fixed number of antennas M = 15, while the results for
[40, 50, 0]T , pr1 ,u = pr2 ,u = [60, 50, 0]T , pr1 ,u = pr2 ,u =
T (4) (4) T
[60, 45, 0] , pr1 ,u = pr2 ,u = [50, 50, 5] , selected close to different values of M are obtained using a fixed value of
the true position of the UE. Moreover, we assume Ib,u = 3 Ns = 200. 5000 Monte Carlo trials are used for each Ns
(1)
SPs between BS and UE with positions pb,u = [35, 55, 2]T , and M . The position of the UE is fixed at pu = [48, 51, 1]T ,
(2) (3) and we provide the CRLB values, assuming that there are
pb,u = [30, 40, 0]T and pb,u = [25, 50, 1]T . We observe that no NLoS paths. From Fig. 9, it is evident that the proposed
across two LMR values (LMR ∈ 5, 10, dB) the proposed VNet, although trained only for Ns = 200, demonstrates
SG2 outperforms existing methods, especially in low-SNR excellent generalization across different values of Ns . In
scenarios. In the presence of NLoS paths, the performance addition, both stages SG1 and SG2 in the proposed method
of the ML-based JML method degrades with lower LMR showcase improvement with higher Ns and SNR. The results
values, but as LMR increases, its results converge toward the of SG2 approach the CRLB values, and the gap between
scenario without multipath. SG1 consistently outperforms LS them diminishes for higher values of Ns or SNR. Similar
and CIM methods across the entire SNR range. observations can be drawn from Fig. 10. The RMSE value
As shown in Figs. 7 and 8, the proposed approach also improves with a higher number of elements deployed in
outperforms existing methods under the challenging condi- each subarray of the RISs, demonstrating robustness across
tions of low-SNR and low-LMR regimes, achieving accu- different values of Ns and M .
racy within a minimal gap from the theoretically optimal
solution. Importantly, as detailed in Sec. VI-D, this superior
D. Complexity Analysis
3 In
general, SPs may have different locations for each RIS, and some In this section, we provide a complexity analysis of the
of them may be common among multiple RISs. Here, for simplicity, we proposed methods and compare them to the state-of-the-art
assume the same set of SPs for both RISs. methods. Our focus will be on inference only, not training.

VOLUME , 13

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

SNR = −10 dB Proposed: SG1

SNR = 0 dB Proposed: SG2 TABLE 4: Complexity analysis of the proposed method
SNR = 10 dB CRLB
100 Component Complexity
Channel Estimation O(GNs M 2 )
RMSE [m]

Calibration O(GNs M )
SG1 O(GNs M ) + O(F M 2 ) + O(G) + O(GNs M )
2
10−1
+ O(GNIg Ns ) + O(G(NΘ + NΨ )M )
SG2 O(GNSG2 )

10−2
derived from the network architecture.4 Hence, by focusing
50 100 150 200 250 300 350 400
Ns on the dominant part, the overall complexity of the proposed
method is given by O(GNs M 2 ), which refers to the LS
FIGURE 9: Localization RMSE for different numbers of subcarriers.
channel estimation and the computations of the SCMs in
(32). In all simulations, the number of iterations NSG2
SNR = −10 dB
required was between 3 and 11 iterations with an average
Proposed: SG1
SNR = 0 dB Proposed: SG2 of 6.35.
100 SNR = 10 dB CRLB

3) Overall Complexity Comparison

RMSE [m]

Table 5 compares the order of complexities of the proposed

10−1 method with the state-of-the-art ones and also the average
amount of time required for each method. It should be noted
that only the proposed method and the ML-based RML and
10−2 JML can estimate the clock offset. Unlike RML and JML, the
proposed method is environment-independent. Since RML
10 15 20 25 30 is grid-based, it requires a 3D search in the uncertainty
M environment. In all simulations in this paper, we considered
FIGURE 10: Localization RMSE for different numbers of RIS elements. an environment where the horizontal coordinates (x, y) are
between 35 and 65 and the vertical (z) are between 0 and
6 for these two methods. We divide each dimension with
1) Complexity of AOD Estimators
grid length of 1.5 meter resulting a grid mesh of dimensions
(Nx , Ny , Nz ) = (21, 21, 5) for RML. We use NFFT = 512
Table 3 summarizes the complexities of the proposed VNet
for the RML method’s FFT computation, as in [17]. In our
and existing methods. It includes the order of complexity for
3D scenario, NNM , representing the Nelder-Mead algorithm’s
signal pre-processing (column 2), exact FLOPs calculated
iterations in JML, averages 235.65 iterations, peaking at 287.
(column 3), and average total time for angle estimation on
This contrasts with the simpler 2D scenario in [17], where
the central processing unit (CPU) (column 4). The results
only up to 30 iterations were reported.
demonstrate the computational advantage of our VNet.
VII. Conclusion
In this study, we investigated the problem of 3D localization
2) Detailed Complexity Analysis of the Proposed Two-stage
and synchronization of a UE in the presence of a single BS
Method
and multiple L-shaped RISs, in a multipath environment.
A novel 2D-AOD estimator was proposed, leveraging the
The order complexities of various components of our method
RISs’ high angular resolution. The proposed method relies
are detailed in Table 4, with corresponding explanations
on deep learning and can effectively operate in multipath
provided in relevant sections. The term O(F M 2 ) denotes
environments while being agnostic of the geometric config-
the simplified complexity of the VNet, where F is a constant
uration of RISs, UE, and BS. In addition, a novel hybrid
4 By solely considering the number of multiplications, a general approx-
imation for the count
 of floating-point operations (FLOPs) [50] can be
P c (i−1) (i) (i)2 P Nf
derived as F = O QM0 N i=1 nc n c q + Q j=1 Lj−1 Lj .
TABLE 3: Complexity comparison of AOD estimators which applies to an arbitrary network comprising Q identical parallel sub-
Preprocessing Flops Avg. Est. time (CPU) [ms] networks, each consisting of Nc convolution layers and Nf fully-connected
2
VNet O(M 2 Ns ) 645330 0.471 layers. Here, M0 denotes the length of the input and L0 = M0 q (Nc ) is
RSNet O(M 2 Ns ) + O(M 4 ) 161400 1.38 the length of the output of the flatten layer. Also, Lj represents the number
R-MUSIC O(M 2 Ns ) - 1.48
of cells in the j-th fully connected layer.

14 VOLUME ,

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

[12] J. He, A. Fakhreddine, C. Vanwynsberghe, H. Wymeersch, and G. C.

TABLE 5: Overall computational complexity comparison Alexandropoulos, “3D localization with a single partially-connected
Method Order of Complexity Description Avg. Est. [ms]
receiving RIS: Positioning error analysis and algorithmic design,”
Proposed SG1 O(GNs M 2 ) - 4.13
IEEE Transactions on Vehicular Technology, vol. 72, no. 10, pp.
Proposed SG2 O(GNSG2 ) NSG2 ≈ 6.35 2.45
13 190–13 202, 2023.
Overall Proposed O(GNs M 2 ) - 6.58
[13] K. Keykhosravi, B. Denis, G. C. Alexandropoulos, Z. S. He, A. Al-
RML O(GNx Ny Nz Ns M 3 ) Nx = 21, Ny = 21, 1164.3
banese, V. Sciancalepore, and H. Wymeersch, “Leveraging RIS-
+ O(NFFT log NFFT ) Nz = 5, NFFT = 512 enabled smart signal propagation for solving infeasible localization
JML O(NNM Ns M (M 2 + M (b) )) NNM ≈ 235.65 964.7 problems: Scenarios, key research directions, and open challenges,”
R-MUSIC+LS O(GM 3 ) + O(GNs M 2 ) - 6.1 IEEE Vehicular Technology Magazine, vol. 18, no. 2, pp. 20–28, 2023.
R-MUSIC+CIM O(GM 3 ) + O(GNs M 2 ) NCIM ≈ 311.76 18.6 [14] R. Ghazalian, H. Chen, G. C. Alexandropoulos, G. Seco-Granados,
+ O(GNCIM ) H. Wymeersch, and R. Jäntti, “Joint user localization and location
calibration of a hybrid reconfigurable intelligent surface,” IEEE Trans-
actions on Vehicular Technology, vol. 73, no. 1, pp. 1435–1440, 2024.
[15] C. Huang, A. Zappone, G. C. Alexandropoulos, M. Debbah, and
AOD/TOA-based two-stage method was proposed, which C. Yuen, “Reconfigurable intelligent surfaces for energy efficiency in
effectively utilizes the 2D-AOD estimates and combines wireless communication,” IEEE Transactions on Wireless Communi-
cations, vol. 18, no. 8, pp. 4157–4170, 2019.
them to efficiently obtain TOA estimates as well as a [16] Q. Wu and R. Zhang, “Beamforming optimization for wireless network
UE position estimate. Extensive simulations showed that aided by intelligent reflecting surface with discrete phase shifts,” IEEE
the proposed method significantly outperforms the state-of- Transactions on Communications, vol. 68, no. 3, pp. 1838–1851, 2020.
[17] A. Fascista, M. F. Keskin, A. Coluccia, H. Wymeersch, and G. Seco-
the-art methods, and closely approaches the CRLB values Granados, “RIS-aided joint localization and synchronization with
in various scenarios. Moreover, the results revealed that a single-antenna receiver: Beamforming design and low-complexity
the proposed method demonstrates more robustness against estimation,” IEEE Journal of Selected Topics in Signal Processing,
vol. 16, no. 5, pp. 1141–1156, 2022.
NLoS paths. Furthermore, the complexity analysis reveals [18] H. Zhang, H. Zhang, B. Di, K. Bian, Z. Han, and L. Song, “To-
that the proposed method necessitates considerably fewer wards ubiquitous positioning by leveraging reconfigurable intelligent
computations compared to previous studies. surface,” IEEE Communications Letters, vol. 25, no. 1, pp. 284–288,
2021.
[19] K. Keykhosravi, M. F. Keskin, G. Seco-Granados, and H. Wymeersch,
“Siso RIS-enabled joint 3D downlink localization and synchroniza-
REFERENCES tion,” in IEEE International Conference on Communications, 2021,
[1] J. A. del Peral-Rosado, R. Raulefs, J. A. López-Salcedo, and G. Seco- pp. 1–6.
Granados, “Survey of cellular mobile radio localization methods: From [20] Y. Lin, S. Jin, M. Matthaiou, and X. You, “Channel estimation
1g to 5g,” IEEE Communications Surveys & Tutorials, vol. 20, no. 2, and user localization for IRS-assisted MIMO-OFDM systems,” IEEE
pp. 1124–1148, 2018. Transactions on Wireless Communications, vol. 21, no. 4, pp. 2320–
[2] A. Bochem and H. Zhang, “Robustness enhanced sensor assisted 2335, 2022.
monte carlo localization for wireless sensor networks and the internet [21] H. Chen, P. Zheng, M. F. Keskin, T. Al-Naffouri, and H. Wymeersch,
of things,” IEEE Access, vol. 10, pp. 33 408–33 420, 2022. “Multi-RIS-enabled 3D sidelink positioning,” IEEE Transactions on
[3] Y. Wang and K. C. Ho, “An asymptotically efficient estimator in Wireless Communications, pp. 1–1, 2024.
closed-form for 3-D AOA localization using a sensor network,” IEEE [22] M. Hua, Q. Wu, W. Chen, Z. Fei, H. C. So, and C. Yuen, “Intelligent
Transactions on Wireless Communications, vol. 14, no. 12, pp. 6524– reflecting surface-assisted localization: Performance analysis and algo-
6535, 2015. rithm design,” IEEE Wireless Communications Letters, vol. 13, no. 1,
[4] Y. Zou, L. Wu, J. Fan, and H. Liu, “A convergent iteration method for pp. 84–88, 2024.
3-D AOA localization,” IEEE Transactions on Vehicular Technology, [23] N. K. Kundu and M. R. McKay, “Channel estimation for reconfig-
vol. 72, no. 6, pp. 8267–8271, 2023. urable intelligent surface aided MISO communications: From LMMSE
[5] Y. Zheng, M. Sheng, J. Liu, and J. Li, “Exploiting AoA estimation to deep learning solutions,” IEEE Open Journal of the Communications
accuracy for indoor localization: A weighted AoA-based approach,” Society, vol. 2, pp. 471–487, 2021.
IEEE Wireless Communications Letters, vol. 8, no. 1, pp. 65–68, 2019. [24] Y. Liu, I. Al-Nahhal, O. A. Dobre, and F. Wang, “Deep-learning chan-
[6] T. Jia, H. Liu, K. C. Ho, and H. Wang, “Mitigating sensor motion effect nel estimation for IRS-assisted integrated sensing and communication
for AOA and AOA-TOA localizations in underwater environments,” system,” IEEE Transactions on Vehicular Technology, vol. 72, no. 5,
IEEE Transactions on Wireless Communications, vol. 22, no. 9, pp. pp. 6181–6193, 2023.
6124–6139, 2023. [25] S. Liu, Z. Gao, J. Zhang, M. D. Renzo, and M.-S. Alouini, “Deep
[7] T. Wang, H. Xiong, H. Ding, and L. Zheng, “TDOA-based joint denoising neural network assisted compressive channel estimation
synchronization and localization algorithm for asynchronous wireless for mmwave intelligent reflecting surfaces,” IEEE Transactions on
sensor networks,” IEEE Transactions on Communications, vol. 68, Vehicular Technology, vol. 69, no. 8, pp. 9223–9228, 2020.
no. 5, pp. 3107–3124, 2020. [26] Z.-M. Liu, C. Zhang, and S. Y. Philip, “Direction-of-arrival estimation
[8] K. You, W. Guo, T. Peng, Y. Liu, P. Zuo, and W. Wang, “Parametric based on deep neural networks with robustness to array imperfections,”
sparse Bayesian dictionary learning for multiple sources localization IEEE Transactions on Antennas and Propagation, vol. 66, no. 12, pp.
with propagation parameters uncertainty,” IEEE Transactions on Sig- 7315–7327, 2018.
nal Processing, vol. 68, pp. 4194–4209, 2020. [27] W. Liu, “Super resolution DOA estimation based on deep neural
[9] Y. Duan, K.-Y. Lam, V. C. S. Lee, W. Nie, K. Liu, H. Li, and C. J. network,” Scientific Reports, vol. 10, no. 1, pp. 1–9, 2020.
Xue, “Data rate fingerprinting: A WLAN-based indoor positioning [28] A. Fadakar, A. Jafari, P. Tavana, R. Jahani, and S. Akhavan, “Deep
technique for passive localization,” IEEE Sensors Journal, vol. 19, learning based 2D-DOA estimation using l-shaped arrays,” Journal of
no. 15, pp. 6517–6529, 2019. the Franklin Institute, vol. 361, no. 6, p. 106743, 2024.
[10] Y. Liu, Y. Wang, Y. Shen, and X. Shi, “Hybrid TOA-AOA WLS [29] J. Ma, M. Wang, Y. Chen, and H. Wang, “Deep convolutional
estimator for aircraft network decentralized cooperative localization,” network-assisted multiple direction-of-arrival estimation,” IEEE Signal
IEEE Transactions on Vehicular Technology, vol. 72, no. 7, pp. 9670– Processing Letters, vol. 31, pp. 576–580, 2024.
9675, 2023. [30] K. Lee et al., “Deep learning-aided coherent direction-of-arrival es-
[11] G. C. Alexandropoulos, I. Vinieratou, and H. Wymeersch, “Local- timation with the FTMR algorithm,” IEEE Transactions on Signal
ization via multiple reconfigurable intelligent surfaces equipped with Processing, vol. 70, pp. 1118–1130, 2022.
single receive RF chains,” IEEE Wireless Communications Letters, [31] G. K. Papageorgiou, M. Sellathurai, and Y. C. Eldar, “Deep networks
vol. 11, no. 5, pp. 1072–1076, 2022. for direction-of-arrival estimation in low SNR,” IEEE Transactions on

VOLUME , 15

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in IEEE Open Journal of the Communications Society. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/OJCOMS.2024.3399605

Signal Processing, vol. 69, pp. 3714–3729, 2021. Alireza Fadakar received the B.Sc. and M.Sc.
[32] S. Zheng, Z. Yang, W. Shen, L. Zhang, J. Zhu, Z. Zhao, and degrees in electrical engineering from University
X. Yang, “Deep learning-based DOA estimation,” IEEE Transactions of Tehran, Tehran, Iran, in 2020 and 2023, re-
on Cognitive Communications and Networking, pp. 1–1, 2024. spectively. Currently, he is a research assistant
[33] A. Fadakar, A. Mansourian, and S. Akhavan, “Deep learning aided at University of Tehran, specializing in wireless
multi-source passive 3D AOA wireless positioning using a moving communication. His research interests lie in the
receiver: A low complexity approach,” Ad Hoc Networks, vol. 154, p. application of machine learning for wireless com-
103382, 2024. munications, channel estimation, MIMO networks,
[34] W. Wang and W. Zhang, “Joint beam training and positioning for in- RIS-assisted communications, signal processing,
telligent reflecting surfaces assisted millimeter wave communications,” digital twin and optimization.
IEEE Transactions on Wireless Communications, vol. 20, no. 10, pp.
6282–6297, 2021.
[35] Y. Lin, S. Jin, M. Matthaiou, and X. Yi, “Circular RIS-enabled Maryam Sabbaghian (S’01–M’08) received the B.Sc. and M.Sc. degrees
channel estimation and localization for multi-user ISAC systems,” from the Sharif University of Technology, Tehran, Iran, in 1999 and 200,
IEEE Transactions on Wireless Communications, pp. 1–1, 2024. respectively, and the Ph.D. degree from Carleton University, Ottawa, ON,
[36] M. Mizmizi, R. A. Ayoubi, D. Tagliaferri, K. Dong, G. G. Gentili, and
U. Spagnolini, “Conformal metasurfaces: A novel solution for vehicu- Canada, in 2007, all in electrical engineering. From 2008 to 2010, she
lar communications,” IEEE Transactions on Wireless Communications, was a Post-Doctoral Fellow with the School of Engineering and Applied
vol. 22, no. 4, pp. 2804–2817, 2023. Sciences, Harvard University. In 2010, she joined the School of Electrical
[37] J.-F. Gu, W.-P. Zhu, and M. Swamy, “Joint 2-D DOA estimation
via sparse l-shaped array,” IEEE Transactions on Signal Processing, and Computer Engineering, University of Tehran. Her current research
vol. 63, no. 5, pp. 1171–1182, 2015. interests include 5G and beyond technologies, neural communications, and
[38] X. Wu and W.-P. Zhu, “On efficient gridless methods for 2-D DOA es- applications of information theory in interdiciplinary studies. She is a
timation with uniform and sparse l-shaped arrays,” Signal Processing,
vol. 191, p. 108351, 2022. recipient of the Natural Sciences and Engineering Council NSERC-PDF
[39] J.-F. Gu and P. Wei, “Joint svd of two cross-correlation matrices to of Canada, Ontario Graduate Scholarship ,and Nortel Scholarship. She was
achieve automatic pairing in 2-D angle estimation problems,” IEEE a Technical Program Committee for several IEEE Telecomm. conferences.
Antennas and Wireless Propagation Letters, vol. 6, pp. 553–556, 2007.
[40] N. Tayem and H. M. Kwon, “L-shape 2-dimensional arrival angle
estimation with propagator method,” IEEE transactions on antennas
Henk Wymeersch(S’01, M’05, SM’19, F’24)
and propagation, vol. 53, no. 5, pp. 1622–1630, 2005.
obtained the Ph.D. degree in Electrical Engineer-
[41] S. Kikuchi, H. Tsuji, and A. Sano, “Pair-matching method for esti-
ing/Applied Sciences in 2005 from Ghent Uni-
mating 2-D angle of arrival with a cross-correlation matrix,” IEEE
versity, Belgium. He is currently a Professor of
Antennas and Wireless Propagation Letters, vol. 5, pp. 35–40, 2006.
Communication Systems with the Department of
[42] C. Ozturk, M. F. Keskin, V. Sciancalepore, H. Wymeersch, and
Electrical Engineering at Chalmers University of
S. Gezici, “RIS-aided localization under pixel failures,” IEEE Trans-
Technology, Sweden. Prior to joining Chalmers,
actions on Wireless Communications, pp. 1–1, 2024.
he was a postdoctoral researcher from 2005 until
[43] B. Zheng, C. You, W. Mei, and R. Zhang, “A survey on channel
2009 with the Laboratory for Information and
estimation and practical passive beamforming design for intelligent
Decision Systems at the Massachusetts Institute of
reflecting surface aided wireless communications,” IEEE Communica-
Technology. Prof. Wymeersch served as Associate
tions Surveys & Tutorials, vol. 24, no. 2, pp. 1035–1071, 2022.
Editor for IEEE Communication Letters (2009-2013), IEEE Transactions
[44] Y. Yang, B. Zheng, S. Zhang, and R. Zhang, “Intelligent reflecting
on Wireless Communications (since 2013), and IEEE Transactions on
surface meets OFDM: Protocol design and rate maximization,” IEEE
Communications (2016-2018) and is currently Senior Member of the IEEE
Transactions on Communications, vol. 68, no. 7, pp. 4522–4535, 2020.
Signal Processing Magazine Editorial Board. During 2019-2021, he was an
[45] T. L. Jensen and E. De Carvalho, “An optimal channel estimation
IEEE Distinguished Lecturer with the Vehicular Technology Society. His
scheme for intelligent reflecting surfaces based on a minimum variance
current research interests include the convergence of communication and
unbiased estimator,” in IEEE International Conference on Acoustics,
sensing, in a 5G and Beyond 5G context.
Speech and Signal Processing (ICASSP), 2020, pp. 5000–5004.
[46] C. Liu, X. Liu, D. W. K. Ng, and J. Yuan, “Deep residual learning for
channel estimation in intelligent reflecting surface-assisted multi-user
communications,” IEEE Transactions on Wireless Communications,
vol. 21, no. 2, pp. 898–912, 2022.
[47] A. M. Elbir, “DeepMUSIC: Multiple signal classification via deep
learning,” IEEE Sensors Letters, vol. 4, no. 4, pp. 1–4, 2020.
[48] T. Ridnik, E. Ben-Baruch, N. Zamir, A. Noy, I. Friedman, M. Protter,
and L. Zelnik-Manor, “Asymmetric loss for multi-label classification,”
in IEEE/CVF International Conference on Computer Vision, 2021, pp.
82–91.
[49] A. Barabell, “Improving the resolution performance of eigenstructure-
based direction-finding algorithms,” in ICASSP ’83. IEEE Interna-
tional Conference on Acoustics, Speech, and Signal Processing, vol. 8,
1983, pp. 336–339.
[50] R. Hunger, Floating point operations in matrix-vector calculus. Inst.
Circuit Theory Signal Process., Munich Univ. Technol., München,
Germany, Rep. 2007, 2005.

16 VOLUME ,

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/