Localization Algorithm Design for

RIS-Aided Wireless Localization System

by
Tuo Wu

Doctor of Philosophy

School of Electronic Engineering and Computer Science

Queen Mary University of London
United Kingdom

August 2024
 Abstract
Reconfigurable intelligent surfaces (RISs) offer several advantages over traditional base

stations (BSs) for localization systems, such as lower hardware costs and energy-efficient,

high-precision estimation capabilities due to its passive panel nature. Additionally, the

large RIS size allows for highly accurate radio localization parameter estimation. Conse-

quently, developing localization algorithms for RIS-assisted wireless localization systems

holds great significance. However, prevailing studies tend to assume an ideal and simpli-

fied communication environment when designing the localization algorithm for RIS-aided

wireless localization systems. Addressing this oversight, this thesis investigates the algo-

rithms design for RIS-aided wireless localization systems, specifically tackling practical

conditions, such as non-Gaussian angle estimation errors (AEE), multiple users, faulty

reflecting elements, and near-field.

First, this thesis presents a comprehensive framework for jointly analyzing the AEE

and designing a three-dimensional (3D) localization algorithm for a multiple RISs aided

localization systems. The azimuth and elevation AoAs at the RISs are estimated by

applying the two-dimensional discrete Fourier transform (2D-DFT) algorithm. The AEE

is then analyzed in terms of probability density functions (PDF), revealing that the AEE

is non-Gaussian. Then, the closed-form expressions of the variances are formulated using

generalized hypergeometric series. Finally, the two-stage weighted least square (TSWLS)

algorithm is employed to estimate the 3D position of the mobile user (MU) using the

estimated AoAs and the obtained non-Gaussian variance.

Second, this thesis designs a specific localization algorithms for accommodating the non-

Gaussian nature of AoAs errors and the Gaussian character of TDoA error. Following the

classical two-step 3D localization, the AoAs and TDoAs at the RISs are estimated using

different methods, resulting in non-Gaussian and Gaussian errors, respectively. Then,

i
 ii

a multiple weighted least squares (mWLS) algorithm is designed to accurately localize

MU. Besides, this thesis presents a unique bias analysis for evaluating the performance

of the proposed localization algorithm under both Gaussian and non-Gaussian errors.

Third, this thesis investigates the localization algorithm designs for multiple MUs local-

ization problem in the multiple-input multiple-output (MIMO) mmWave systems aided

by the RIS. A novel type of fingerprint, space-time channel response vector (STCRV), is

designed for fingerprint based localization algorithm design. Then, this thesis proposes

a novel residual convolution network regression (RCNR) learning algorithm to output

the estimated 3D position of the MU with higher accuracy.

Fourth, this thesis explores the algorithm design addressing a practical scenario where

RIS contains some unknown (number and places) faulty elements that cannot receive

signals. Transfer learning is employed to design a two-phase transfer learning (TPTL)

algorithm for accurate detection of faulty elements. Then, this thesis proposes a transfer-

enhanced dual-stage (TEDS) algorithm to regain the information lost from the faulty

elements and reconstruct the complete high-dimensional RIS information for localization.

Fifth and last, this thesis investigates a near-field mobile tracking system assisted by

extremely large-scale RIS (XL-RIS). An XL-RIS information reconstruction (XL-RIS-IR)

algorithm is designed to reconstruct the high-dimensional RIS information from the low-

dimensional BS received signal. Then, this thesis proposes a comprehensive framework

for mobile tracking, consisting of a Feature Extraction Module and a Mobile Tracking

Module. The Feature Extraction Module is designed for extracting a comprehensive from

the reconstructed RIS information. A time-varying sequence formed by the extracted

feature vector is fed into the Mobile Tracking Module, which employs an Auto-encoder

(AE) with a stacked bidirectional long short-term memory (Bi-LSTM) encoder and a

standard LSTM decoder to predict MUs’ positions in the upcoming time slot.
 iii

Acknowledgments
Firstly, I extend my sincere gratitude to my supervisor, Prof. Maged Elkashlan, for

his immense support over the past three years. His suggestions, encouragement, and

constructive feedback have been invaluable. I consider myself fortunate to have had

the opportunity to learn from him. I would also like to thank Prof. Mona Jaber and

Prof. Akram Alomainy for their roles as my progression panel members and for their

constructive advice.

Secondly, I would also like to sincerely thank Prof. Cunhua Pan for offering me a chance

to pursue my PhD at the QMUL, and for his support of my research endeavors. I would

also like to thank Prof. Kai-Kit Wong, Prof. Robert Schober, Prof. Jiangzhou Wang,

Prof. Cheng-Xiang Wang, and Prof. Xiaohu You for providing constructive comments

on my research. I would also like to thank Prof. Yijin Pan and Dr. Kangda Zhi gave

me comprehensive guidance, led my progress on my research.

Thirdly, I extend my heartfelt gratitude to Mrs Melissa Yeo for her unwavering and

kindness support throughout the time in QMUL. My thanks also go to my lab mates

Zhendong Peng, Lanting Zha, Dr. Ruikang Zhong, and Na Xue for their invaluable

help and encouragement. I am especially grateful to my friends Dr. Gui Zhou, Dr.

Kangda Zhi, Na Yan, Bingjie Zhu, Guanyu Hu, Chunfeng Xie, and Zhixiong Chen for the

memorable lunches in 115&CS321 and for the countless joyful moments and adventures

we shared. Additionally, I would like to thank Dr. Xiazhi Lai, Dr. Jianchao Zheng, Dr.

Junteng Yao, and Dr. Lifeng Mai for being there through the highs and lows, sharing in

our Wechat group, and collaborating on research projects. Finally, I would like to thanks

all the younger schoolmates in Southeast University for the warmest family feeling.

Lastly, I owe a deep sense of gratitude to my wife, my brother and his wife, and my

parents for their boundless love and unwavering support.

Table of Contents

Abstract i

Acknowledgments iii

Table of Contents iv

List of Figures x

1 Introduction 2

1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2.1 RIS-aided localization . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2.2 DL-aided localization . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Research Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

1.5 Author’s Publication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

1.6 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Fundamental Concepts 11

2.1 Reconfigurable Intelligent Surface . . . . . . . . . . . . . . . . . . . . . . . 11

2.2 RIS-aided Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.3 Localization Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.3.1 Two-Step Localization Scheme . . . . . . . . . . . . . . . . . . . . 14

iv
Table of Contents v

2.3.2 Fingerprint-Based Localization Scheme . . . . . . . . . . . . . . . 14

2.4 Localization Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.4.1 ToA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.4.2 TDoA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.4.3 AoA and AoD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.5 Weighted Least Squares (WLS) Algorithm . . . . . . . . . . . . . . . . . . 17

2.6 Convolutional Neural Networks (CNNs) . . . . . . . . . . . . . . . . . . . 18

2.6.1 Convolutional Layers . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.6.2 Pooling Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.6.3 Fully Connected Layers . . . . . . . . . . . . . . . . . . . . . . . . 19

2.7 Transfer Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 21

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.2.1 Geometry Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.2.2 Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3.2.3 Received Signal Model . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.2.4 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.3 Angle Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.3.1 Initial Angle Estimation . . . . . . . . . . . . . . . . . . . . . . . . 29

3.3.2 Fine Angle Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 30

3.4 Derivation of the PDF of Angle Estimation Error . . . . . . . . . . . . . . 32

3.4.1 PDF of Φ̃1,i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.4.2 PDF of Θ̃1,i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

3.5 Variance of Angle Estimation Error . . . . . . . . . . . . . . . . . . . . . . 43

3.5.1 Variance of Φ̃1,i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.5.2 Variance of Θ̃1,i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.6 3D Position Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

Table of Contents vi

3.7 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

4 RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 55

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

4.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

4.2.1 Wireless Channel Model . . . . . . . . . . . . . . . . . . . . . . . . 58

4.2.2 Received Signal Model . . . . . . . . . . . . . . . . . . . . . . . . . 61

4.3 Two Step 3D Positioning Scheme . . . . . . . . . . . . . . . . . . . . . . . 62

4.3.1 Step I : Channel Parameters Estimation . . . . . . . . . . . . . . . 62

4.3.2 Step II: Proposed mWLS Algorithm . . . . . . . . . . . . . . . . . 66

4.4 Bias Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

4.4.1 Decomposing Estimation Error Expression . . . . . . . . . . . . . 76

4.4.2 Bias of Preliminary Estimation . . . . . . . . . . . . . . . . . . . . 79

4.4.3 Bias of Refining Estimation . . . . . . . . . . . . . . . . . . . . . . 81

4.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

5 RIS-Aided Localization Algorithm Design: Employing Fingerprint 90

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

5.2 System Model and Problem Formulation . . . . . . . . . . . . . . . . . . . 92

5.3 Residual Convolution Network Regression Learning Algorithm . . . . . . 96

5.3.1 Regression-oriented Positioning . . . . . . . . . . . . . . . . . . . . 96

5.3.2 The structure of the RCNR learning algorithm . . . . . . . . . . . 96

5.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

6 RIS-Aided Localization Algorithm Design: Overcoming Faulty Ele-

ments 104

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

Table of Contents vii

6.2 System Model and Problem Formulation . . . . . . . . . . . . . . . . . . . 107

6.2.1 Geometry and Channel Model . . . . . . . . . . . . . . . . . . . . 107

6.2.2 Faulty Element Model . . . . . . . . . . . . . . . . . . . . . . . . . 109

6.2.3 Signal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

6.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

6.3.1 Faulty Element Detection Problem . . . . . . . . . . . . . . . . . . 111

6.3.2 Localization Problem Under Faulty Elements . . . . . . . . . . . . 114

6.4 Faulty Element Detection Algorithm Design . . . . . . . . . . . . . . . . 115

6.4.1 Adoption of Transfer Learning . . . . . . . . . . . . . . . . . . . . 115

6.4.2 The Architecture of DenseNet 121 . . . . . . . . . . . . . . . . . . 116

6.4.3 Phase I: Transfer Learning of DenseNet 121 . . . . . . . . . . . . . 117

6.4.4 Phase II: Transfer Learning of ‘Phase I’ Network . . . . . . . . . . 121

6.4.5 Trained TPTL Algorithm Usage Guide . . . . . . . . . . . . . . . 124

6.5 RIS Signal Reconstruction and Localization . . . . . . . . . . . . . . . . . 125

6.5.1 Stage I: RIS Signal Reconstruction . . . . . . . . . . . . . . . . . . 125

6.5.2 Stage II: Enhanced Localization using Reconstructed Signal ȳr . . 129

6.6 Transfer-Enhanced Direct Fingerprint Algorithm . . . . . . . . . . . . . . 132

6.6.1 Transfer Learning for Localization . . . . . . . . . . . . . . . . . . 132

6.6.2 Training Using BS Received Signal . . . . . . . . . . . . . . . . . . 132

6.6.3 Localization with Well-trained TEDF Algorithm . . . . . . . . . . 133

6.7 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

6.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142

7 RIS-aided Localization Algorithm Design: Tracking in Near-field 144

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

7.2 System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

7.2.1 Geometry and Channel Model . . . . . . . . . . . . . . . . . . . . 148

7.2.2 Signal Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150

7.3 XL-RIS Information Reconstruction Algorithm . . . . . . . . . . . . . . . 151

Table of Contents viii

7.3.1 Data Processing Module . . . . . . . . . . . . . . . . . . . . . . . . 152

7.3.2 DenseNet Module . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

7.3.3 Output Module . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

7.3.4 Training of XL-RIS-IR Algorithm . . . . . . . . . . . . . . . . . . . 155

7.4 Employing XL-RIS Information For Tracking . . . . . . . . . . . . . . . . 155

7.4.1 Tracking Problem Formulation . . . . . . . . . . . . . . . . . . . . 155

7.4.2 Tracking Framework Assisted by XL-RIS Information . . . . . . . 156

7.5 XL-RIS Information Based Feature Extraction . . . . . . . . . . . . . . . 157

7.5.1 CNN Extractor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158

7.5.2 T& F Feature Extractor . . . . . . . . . . . . . . . . . . . . . . . 160

7.5.3 AoAs Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . 163

7.5.4 Final Feature Vector Construction . . . . . . . . . . . . . . . . . . 166

7.6 XL-RIS Information Based Tracking Algorithm . . . . . . . . . . . . . . . 166

7.6.1 LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

7.6.2 Stacked Bi-LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

7.6.3 Encoder Construction for Stacked Bi-LSTM AE . . . . . . . . . . 170

7.6.4 Decoder Construction for Stacked Bi-LSTM AE . . . . . . . . . . 171

7.7 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

7.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

8 Conclusion 181

8.1 Summary of Contributions and Insights . . . . . . . . . . . . . . . . . . . 181

8.2 Future Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

8.2.1 Passive Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183

8.2.2 Active RIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

8.2.3 Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . 184

8.2.4 Fundamental Limitation of RIS Information . . . . . . . . . . . . . 185

8.2.5 Extending Tracking Beyond Predefined Areas . . . . . . . . . . . . 185

8.2.6 Optimization of RIS Phase Shift Matrix . . . . . . . . . . . . . . . 186

Table of Contents ix

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

Appendix A Derivations in Chapter 3 194

A.1 Derivation of D1,i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195

A.2 Derivation of D2,i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198

′
A.3 Derivation of D1,i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
′
A.4 Derivation of D2,i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

Appendix B Proofs and Derivations in Chapter 4 203

B.1 Derivation of E[ũ] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

B.2 Derivation of W1 and W̃1 . . . . . . . . . . . . . . . . . . . . . . . . . . 205

B.3 Proof of Proposition 4. 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 205

B.4 Derivations of E[W̃1 P2 B1 ũ] . . . . . . . . . . . . . . . . . . . . . . . . . . 206

B.5 Derivation of the Covariance Matrix of q̂ . . . . . . . . . . . . . . . . . . . 208

List of Figures

2.1 An example of the RIS-aided communications. . . . . . . . . . . . . . . . 12

3.1 RISs aided localization system model. . . . . . . . . . . . . . . . . . . . . 23

3.2 Communication link from RIS to BS. . . . . . . . . . . . . . . . . . . . . . 25

3.3 PDF of Θ̃1,i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3.4 PDF of Φ̃1,i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

3.5 PDF of Θ̃1,i in the cases of Gaussian and non-Gaussian. . . . . . . . . . . 50

3.6 Variance of Θ̃1,i versus RIS size. . . . . . . . . . . . . . . . . . . . . . . . 50

3.7 Variance of Φ̃1,i versus RIS size. . . . . . . . . . . . . . . . . . . . . . . . . 51

3.8 Variances of Φ̃1,i and Θ̃1,i versus SNR (dB). . . . . . . . . . . . . . . . . . 51

3.9 Comparison of the proposed framework and geometry algorithm. . . . . . 52

3.10 MSE of the proposed framework versus RISs number. . . . . . . . . . . . 52

3.11 The MSE comparison of the proposed framework. . . . . . . . . . . . . . . 53

4.1 The system model of multiple RISs aided localization systems. . . . . . . 59

4.2 Standard deviation of ni and ωi versus SNR (dB). . . . . . . . . . . . . . 85

4.3 Standard deviation of ni and ωi versus Ny (Nz ). . . . . . . . . . . . . . . . 85

4.4 The MSE comparison of Gaussian and non-Gaussian. . . . . . . . . . . . . 86

4.5 The MSE comparison with other algorithms. . . . . . . . . . . . . . . . . 87

4.6 The MSE comparison versus RIS size. . . . . . . . . . . . . . . . . . . . . 88

4.7 The bias comparison of Gaussian and non-Gaussian. . . . . . . . . . . . . 89

4.8 The bias comparison versus RIS size. . . . . . . . . . . . . . . . . . . . . . 89

x
List of Figures xi

5.1 RIS-aided multiuser localization system model. . . . . . . . . . . . . . . . 92

5.2 The structure of the RCNR learning algorithm. . . . . . . . . . . . . . . . 93

5.3 Two blocks used for the proposed algorithm. . . . . . . . . . . . . . . . . 97

5.4 Two kinds of convolution block used for the proposed algorithm. . . . . . 98

5.5 The CDF of estimation error with different fingerprints. . . . . . . . . . . 101

5.6 The CDF of estimation error with different algorithms. . . . . . . . . . . . 101

5.7 The CDF of estimation error with different kinds of CSI error. . . . . . . 102

5.8 The CDF of estimation error with different phase shift matrices. . . . . . 103

6.1 Network structure of TPTL algorithm. . . . . . . . . . . . . . . . . . . . . 111

6.2 Network structure of DenseNet 121. . . . . . . . . . . . . . . . . . . . . . 115

6.3 Network structure of a Dense Block. . . . . . . . . . . . . . . . . . . . . . 117

6.4 Network structure of TEDS algorithm. . . . . . . . . . . . . . . . . . . . . 125

6.5 Network structure of TEDF algorithm. . . . . . . . . . . . . . . . . . . . . 130

6.6 Train and test loss of the Phase I of TPTL algorithm. . . . . . . . . . . . 133

6.7 Train and test loss of the Phase II of TPTL algorithm. . . . . . . . . . . . 133

6.8 The CDF of detection accuracy with different kinds of algorithm. . . . . . 134

6.9 Detection accuracy versus SNR (dB) across different Nfau . . . . . . . . . . 134

6.10 Detection accuracy versus SNR (dB) across different positions and Nfau . . 136

6.11 Test loss and train loss of TEDF algorithm. . . . . . . . . . . . . . . . . . 137

6.12 Test loss and train loss of Stage I of TEDS algorithm. . . . . . . . . . . . 137

6.13 Test loss and train loss of Stage II of TEDS algorithm. . . . . . . . . . . . 137

6.14 Comparison when Nfau = 10, 15. . . . . . . . . . . . . . . . . . . . . . . . 140

6.15 Comparison when when SNR = 10, 30 dB. . . . . . . . . . . . . . . . . . . 140

6.16 NMSE Comparison between far field and near field. . . . . . . . . . . . . . 141

6.17 NMSE Comparison between TEDS algorithm and RCNR algorithm. . . 142

6.18 NMSE Comparison between different methods. . . . . . . . . . . . . . . . 142

7.1 The structure of the XL-RIS-IR algorithm. . . . . . . . . . . . . . . . . . 151

7.2 Network structure of DenseNet. . . . . . . . . . . . . . . . . . . . . . . . . 153

List of Figures xii

7.3 Network structure of a Dense Block. . . . . . . . . . . . . . . . . . . . . . 154

7.4 3D random walk example. . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

7.5 3D wave example. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

7.6 3D spiral example. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172

7.7 Test and train losses of random walk trajectory. . . . . . . . . . . . . . . . 173

7.8 Test and train losses loss of spiral trajectory. . . . . . . . . . . . . . . . . 173

7.9 Test and train losses loss of wave trajectory. . . . . . . . . . . . . . . . . . 173

7.10 MSE comparison of different inputs. . . . . . . . . . . . . . . . . . . . . . 176

7.11 MSE comparison of different trajectory. . . . . . . . . . . . . . . . . . . . 177

7.12 MSE comparison of different number of reflecting elements. . . . . . . . . 178

7.13 MSE comparison with and without feature extraction. . . . . . . . . . . . 179

List of Abbreviations

RIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . reconfigurable intelligent surface

AoA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . angle of arrival

TDoA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . time difference of arrival

MIMO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . multiple-input multiple-output

BS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . base station

6G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . sixth generation

IoT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . internet of things

ToA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . time of arrival

AE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .autoencoder

LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . long short-term memory

DL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .deep learning

LoS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . line-of-sight

CNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . convolution neural network

TSWLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . two-stage weighted least squares

xiii
3D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . three-dimensional

SISO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . single-input single-output

MUs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .mobile users

DNN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . deep neural network

STCRV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . space-time channel response vector

AEE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . angle estimation errors

2D-DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . two-dimensional discrete Fourier transform

PDF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . probability density functions

RSSI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . received signal strength information

DoF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . degree of freedom

TPTL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . two-phase transfer learning

TEDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . transfer-enhanced dual-stage

XL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . extremely large-scale

XL-RIS . . . . . . . . . . . . . . . . . . . . . . . extremely large-scale reconfigurable intelligent surface

XL-RIS-IR . . . . . . . . . . . . . . . . . . . . extremely large-scale RIS information reconstruction

mWLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . multiple weighted least squares

RCNR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .residual convolution network regression

CDF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . cumulative distribution function

1
Chapter 1

Introduction

1.1 Background

The upcoming shift towards the sixth generation (6G) Internet of Things (IoT) wire-

less network marks a pivotal moment in our technological journey [1–3]. Enhanced

localization accuracy becomes central in this new era. With expanding boundaries in

applications like smart cities, autonomous vehicles, and precision agriculture, knowing

the exact real-time location of objects and devices is essential [4]. This heightened

demand for accuracy challenges us to not only innovate but to fundamentally rethink

our approach to localization [5–8]. The localization algorithms specifically tailored to

diverse challenges and scenarios are required. To cater to such demands, the design of

localization algorithms that customized for specific scenarios becomes imperative [9–11].

Traditionally, the localization algorithms have been categorized into two main types:

the two-step method and the fingerprint-based method. The two-step method

estimates parameters like angle of arrival (AoA) and time difference of arrival (TDoA)

to deduce the location by utilizing the geometric relationships. The fingerprint-based

method , on the other hand, relies on a database of pre-recorded signal characteristics,

such as received signal strength (RSS), and compares real-time signals to the database

2
Chapter 1. Introduction 3

to pinpoint a location.

While both methods have been dependable, they heavily rely on energy-consuming

base stations (BS). Given the growing emphasis on sustainability and efficiency, this

reliance may not be ideal for the advanced needs of 6G localization. In this context,

reconfigurable intelligent surfaces (RIS) have been introduced as a revolutionary tech-

nology for both wireless communication and localization [12–16]. Beyond its ability

to reconstruct obstructed line-of-sight (LoS) communication links, RIS provides a cost-

effective alternative to traditional BSs, ensuring consistent connectivity even in chal-

lenging environments [17]. Its slim and adaptable design makes it seamlessly integrate

into urban structures, while its energy-efficient nature emphasizes its suitability for 6G

networks [18]. Furthermore, compared with the BS with fewer antennas, the large num-

ber of reflecting elements within the RIS naturally store richer signal information which

could be exploited in localization parameter estimation, providing the potential for high-

accuracy localization [19]. Given these advantages, RIS-aided localization stands out as

a promising technique for 6G era.

Building upon the foundation set by the traditional localization methods and the

emerging potential of RIS, the challenges of the complex 6G communication environment

come into focus. For the two-step method, there is an increased number of localiza-

tion parameters to estimate because of the unparalleled proliferation of devices and the

heterogeneity of connections within 6G networks [20]. Additionally, the error of esti-

mating localization parameters can be large due to the multi-path propagation. Hence,

deep learning (DL) [21–25] can be employed to estimate a large number of localization

parameters with high accuracy, leading to required localization performance. For the

fingerprint-based method, the fingerprint database can become exceedingly extensive

while the features of the fingerprint can be extremely greater [26]. Therefore, DL algo-

rithms (e.g. convolution neural network (CNN) [27]) are introduced as a more advanced

solution to deal with the intricate database and extract more features. Given the diverse

challenges posed by the 6G environment, DL holds the potential to significantly improve

Chapter 1. Introduction 4

the localization performance within 6G networks.

1.2 Literature Review

1.2.1 RIS-aided localization

Given the advantages mentioned, RIS-aided localization algorithms have garnered sig-

nificant research interest. Wu et al. [17] proposed a two-stage weighted least squares

(TSWLS) algorithm for the three-dimensional (3D) localization in mmWave systems

utilizing RISs. Addressing synchronization challenges, Kamran et al. [28] introduced

a streamlined estimation algorithm for 3D localization and synchronization in a single-

input single-output (SISO) multi-carrier system incorporating an RIS. To tackle multi-

user passive localization issues, Čišija et al. [29] presented an efficient estimator com-

bined with an innovative RIS phase profile design to counteract inter-path interference.

Recognizing user mobility in the system, authors of [30] developed a simplified localiza-

tion estimator for MUs in RIS-enhanced 6G systems. Integrating sidelink communica-

tion technology, Chen et al. [19] unveiled a two-stage 3D sidelink localization algorithm

using RISs for intelligent transportation systems (ITSs), facilitating precise user equip-

ment (UE) localization without relying on BSs. Taking into account the near-field effect,

various studies [18, 31–35] have crafted corresponding RIS-aided localization algorithms.

1.2.2 DL-aided localization

Given the aforementioned advantages, DL-aided localization algorithms have become a

focal point of contemporary research endeavors. For the two-step method, the researchers

in [20, 36–38] estimated the channel parameters with the DL-based algorithm, which can

be used to determine the location of the MU. Hao et al. in [36] proposed a DL-based

technique for channel estimation and signal detection within OFDM systems. Hu et al.

in [20] presented theoretical analysis on DL-based channel estimation for single-input

multiple-output (SIMO) systems, while Huang et al. [37] introduced a framework inte-

grating massive multiple-input multiple-output (MIMO) systems with deep learning.

Chapter 1. Introduction 5

Furthermore, Yang et al. [38] put forth an online DL-based channel estimation algo-

rithm tailored for doubly selective fading channels. For the fingerprint-based method,

Bhattacherjee et al. in [39] explored the use of a deep neural network (DNN) for wireless

localization across mmWave and sub-6 GHz frequencies. The authors of [26] proposed a

CSI-based 3D localization method for the MIMO-OFDM system using a 3DCNN. Addi-

tionally, employing the space-time channel response vector (STCRV) as a new type of

fingerprint, Wu et al. [40] developed a DL-aided algorithm for multi-MU localization.

1.3 Research Motivation

Research in the design of localization algorithms for RIS-aided systems has seen consider-

able advancements. However, algorithms tailored for ideal communication environments

may fall short in real-world settings. It is critical to develop localization algorithms

that are attuned to practical scenarios. Challenges such as non-Gaussian angle esti-

mation errors (AEE), support for multiple users, limited number of antennas at the

BS, faulty reflecting elements, and near-field conditions are still emerging areas with

numerous unanswered questions. Addressing these issues is essential for the evolution of

localization systems in practical applications.

First, many studies fail to incorporate AEE into their algorithm designs or assume

AEE follows a Gaussian distribution. However, in practice, the distribution of the AEE

depends on the practical estimation methods (e.g., 2D-DFT algorithm [41]), which may

not follow the Gaussian distribution. Therefore, existing Gaussian-based localization

algorithms and performance analysis are not applicable when considering practical AoA

estimation methods. However, it is a new challenge to model the AEE in terms of

distribution when considering the practical angle estimation methods. Consequently, the

computation of AEEs’ variances should align with their specific non-Gaussian probability

density functions (PDF). Therefore, designing algorithms that effectively address non-

Gaussian AEE is crucial for accurately reflecting real-world conditions. Furthermore, it

is another non-trivial difficulty to analyze the localization performance. To be specific,

Chapter 1. Introduction 6

traditionally deriving the CRLB in the context of non-Gaussian AEE can be complex,

making it difficult to gain insights through comparisons between the CRLB and the

RMSE.

Second, most of the existing researches investigating RIS-aided localization algo-

rithm on two-step algorithm. However, localizing MUs with the two-step localization

algorithms is very challenging due to the large number of channel parameters to be esti-

mated. Fortunately, a promising and important technique, fingerprint based localization

algorithms can directly predict the position by using the fingerprint (e.g., RSSI) with-

out estimating the channel parameters. However, the extraction of features from the

cascaded channel, including the RIS, still requires further investigation.

Third, the passive nature of the RIS has led to the design of these algorithms primarily

based on the received signal at the BS, i.e., BS information. However, due to the high

cost and power consumption of radio-frequency chains may result in a limited number of

antennas at the BS and therefore the dimension of received signal at the BS is low. This

limitation could hinder the capacity to distinguish locations. Fortunately, the received

signal at the RIS, i.e., RIS information, actually contains much richer feature information

about the user location than the BS signal which can be exploited to realize higher

localization accuracy. Hence, RIS information can be used for a type of fingerprint

to directly predict the position of the mobile users. However, it remains a non-trivial

challenge to reconstruct the RIS information due to the passive nature of the RIS.

Fourth, considering a more general scenario where some of the RIS elements may

have been damaged (refer to as faulty elements in this thesis), localization algorithms

designed under the assumption of a perfect RIS without faulty elements will no longer

be reliable and may result in low localization accuracy. Hence, it is crucial to design an

efficient localization algorithm to gain the lost accuracy from the faulty elements.

Fifth but not least, to overcome the high path-loss attenuation and fully unleash

the gain of RIS, the RIS is expected to be constructed on an extremely large scale and
Chapter 1. Introduction 7

equipped with large numbers of passive and low-cost reflecting elements. However, with a

large physical dimension, the communication between an RIS and users will be located in

the near field. Besides, another important challenge is about mobile localization aiming

to accommodate the movement of MU. However, it remains a non-trivial challenge to

design a mobile tracking algorithm within the near-field condition.

1.4 Contributions

Motivated by the above factors, this thesis aims at investigating the localization algo-

rithm design for RIS-aided localization system subject to several practical scenarios. To

this end, this thesis propose a complete framework to jointly analyze the AEE in terms of

PDF and design the 3D localization algorithm. Besides, this thesis propose a novel deep

learning based algorithm to localize multiple users. Furthermore, this thesis proposes an

effective localization algorithm based on the RIS information in the presence of faulty

elements. Finally, propose a comprehensive data-driven framework for near-field mobile

tracking. The specific contributions can be summarised as follows.

• This thesis provides a comprehensive RIS-aided localization framework for ana-

lyzing AEE and designing a 3D localization algorithm. The PDF of the AEE is

investigated given the properties of the 2D-DFT angle estimation algorithm, which

reveals that the resulting AEE is non-Gaussian. This thesis also provides the com-

plex expression for the variance based on the obtained PDF. By employing the

estimated AoAs and the obtained non-Gaussian variance, the 3D position of the

MU is estimated the TSWLS algorithm.

• This thesis introduces an effective 3D localization algorithm that combines AoA

and TDoA estimated at the RIS, along with a comprehensive theoretical analysis

focusing on bias. This thesis conducts a joint estimation of AoAs and TDoAs at

the RIS using different methods, which lead to non-Gaussian and Gaussian errors,

respectively. To address the non-Gaussian distribution of AEEs and the Gaus-

Chapter 1. Introduction 8

sian nature of TDoA errors, the thesis develops a multiple weighted least squares

(mWLS) algorithm for precise localization of MU. Furthermore, it provides a novel

bias analysis to assess the performance of the proposed localization algorithm in

the presence of both Gaussian and non-Gaussian errors.

• In this thesis, a fingerprint-based algorithm is employed to localize MUs. It intro-

duces the concept of the STCRV as a novel form of fingerprint for localization. To

effectively process the STCRV, the thesis presents a new learning algorithm known

as residual convolution network regression (RCNR). This algorithm is designed to

predict the 3D position of MUs with enhanced accuracy, leveraging the unique

properties of the STCRV fingerprint for precise localization.

• This thesis explores the localization algorithm design based on the received signal

at the RIS, e.g., RIS information. Compared with BS information, RIS information

offers higher dimension and richer feature set, thereby significantly improving the

ability to extract positions of the MUs. By capitalizing on this high-dimensional

RIS information, the thesis introduces a novel degree of freedom (DoF) in the design

of algorithms for RIS-assisted localization, which has the potential to markedly

increase localization accuracy.

• This thesis address a practical scenario where RIS contains some unknown (num-

ber and places) faulty elements that cannot receive signals. Transfer learning is

employed to design a two-phase transfer learning (TPTL) algorithm for accurate

detection of faulty elements. A transfer-enhanced dual-stage (TEDS) algorithm is

proposed to regain the information lost from the faulty elements and reconstruct

the complete high-dimensional RIS information for localization. The CNN and

variational autoencoder (VAE) is integrated in Stage I to obtain the RIS informa-

tion. In Stage II, transferred DenseNet 121 is employed to estimate the location of

the MU.

• This thesis provides a novel mobile tracking framework leveraging the high-dimensional
Chapter 1. Introduction 9

extremely large-scale (XL) RIS information, named as XL-RIS information. An

XL-RIS information reconstruction (XL-RIS-IR) algorithm to reconstruct the high-

dimensional XL-RIS information from the low-dimensional BS information. A com-

prehensive framework for mobile tracking is proposed in this thesis, consisting of

a Feature Extraction Module and a Mobile Tracking Module.

1.5 Author’s Publication

Journal Articles:

1. T. Wu, C. Pan, Y. Pan, S. Hong, H. Ren, M. Elkashlan, F. Shu and J. Z. Wang

“Joint Angle Estimation Error Analysis and 3D Positioning Algorithm Design for

mmWave Positioning System” IEEE Internet of Things Journal, Early Access.

=⇒ Corresponding to Chapter 3

2. T. Wu, C. Pan, Y. Pan, H. Ren, M. Elkashlan and C. -X. Wang, “Fingerprint-

Based mmWave Positioning System Aided by Reconfigurable Intelligent Surface,”

IEEE Wireless Communications Letters, vol. 12, no. 8, pp. 1379-1383, Aug. 2023.
=⇒ Corresponding to Chapter 5

3. T. Wu, C. Pan, K. Zhi, H. Ren, M. Elkashlan, C. X. Wang, R. Schober, X. You,

“Exploit High-Dimensional RIS Information to Localization: What Is the Impact

of Faulty Element?,” Accepted by IEEE Journal on Selected Areas in Communica-

tions.
=⇒ Corresponding to Chapter 6

4. T. Wu, C. Pan, K. Zhi, H. Ren, M. Elkashlan, and J. Z. Wang, “Employing High-

Dimensional RIS Information for RIS-aided Localization Systems,” Accepted by

IEEE Communication Letters.

=⇒ Corresponding to Chapter 6 and Chapter 7

Under Review Articles:

1. T. Wu, H. Ren, C. Pan, Y. Pan, S. Hong, M. Elkashlan, F. Shu and J. Z. Wang,

Chapter 1. Introduction 10

“RIS-Aided Localization Algorithm and Analysis: Tackling Non-Gaussian Angle

Estimation Errors,” Minor revision in IEEE Internet of Things Journal .

=⇒ Corresponding to Chapter 4

2. T. Wu, C. Pan, K. Zhi, H. Ren, M. Elkashlan, C. X. Wang, R. Schober, X. You,

“Near-Field Mobile Tracking: A Framework of Using XL-RIS Information,” Under

Review in EEE Journal on Selected Areas in Communications.

=⇒ Corresponding to Chapter 7

1.6 Thesis Organization

The remainder of this thesis is organized as follows. Chapter 2 introduces some funda-

mental concepts for RIS-aided localization systems. Chapters 3 and 4 focus on far-field,

single-user localization scenarios without considering direct LoS paths, utilizing multi-

ple RISs. Chapter 3 explores the development of an RIS-aided localization algorithm

designed to address non-Gaussian AEEs. Chapter 4 advances this exploration by devel-

oping a comprehensive RIS-aided localization algorithm that integrates the AoAs and

TDoAs estimated at the RIS, including an in-depth bias analysis. Chapters 5 through 7

discuss scenarios involving a single RIS, focusing on algorithms designed to handle mul-

tiple users, without incorporating direct LoS paths. Chapters 5 and Chapters 6 begin

in the far-field context and gradually moves towards the near-field in the Chapters 7.

Chapter 5 investigates the challenges of designing a fingerprint-based localization algo-

rithm for multiple users in the far-field. Chapter 6 considers practical challenges such as

unknown faulty elements in the RIS panel and presents an efficient algorithm tailored

for these complications in a multi-user environment. Chapter 7 shifts the focus to the

near-field, detailing a tracking algorithm that addresses the unique challenges posed by

near-field conditions in RIS-aided systems. Chapter 8 presents the conclusion and some

thoughts for future work.

Chapter 2

Fundamental Concepts

In this Chapter, some basic concepts used in the study of RIS-aided localization systems

are introduced, which serves as a solid foundation for the following technical chapters.

2.1 Reconfigurable Intelligent Surface

As an emerging candidate for next-generation communication systems, RIS, also called

intelligent reflecting surfaces (IRS), have attracted significant interest from both academia

and industry [13]. RIS is a reconfigurable engineered surface that does not require active

RF chains, power amplifiers, and digital signal processing units, and is usually made of

a large number of low-cost and passive scattering elements that are coupled with simple

low-power electronic circuits. By intelligently tuning the phase shifts of the impinging

waves with the aid of a controller, an RIS can constructively strengthen the desired signal

or can deconstructively weaken the interference signals, which results in an appealing

nearly-passive beamforming gain. Meanwhile, as a thin surface, RIS can be flexibly

deployed on the facade of the builds in the urban area, and help effectively eliminate the

communication dead zone.

The reconfigurability of the RIS is ensured by low-power tunable electronic circuits,

such as the positive intrinsic negative (PIN) diodes or varactors [42]. The phase shifts

11
Chapter 2. Fundamental Concepts 12

RI
S

BS Obs
tac
le Us
er

Figure 2.1: An example of the RIS-aided communications.

of the impinging signal can be configured based on the adjustment of the on/off state of

the PIN diodes or the bias voltage of the varactors. As shown in Fig. 2.1, the RIS can

reflect the impinging signal from the BS to the user, with the adjustment of the phase

shifts of the signal. Assume that the phase shift matrix of the RIS is Φ, the cascaded

channel from the BS to the user via the RIS will be

hcas = hH
rd Φhsr . (2.1)

Due to the unit modulus constraints of the phase shift, the matrix Φ is subject to

condition |[Φ](n,n) | = 1. Clearly, by properly designing the phase shift matrix, the

RIS could intelligently tailor the wireless propagation environment, and then achieve

promising performance gains.

2.2 RIS-aided Localization

In wireless communication, particularly in environments lacking GPS, radio localization

emerges as a crucial alternative for determining user locations. This technique relies on

radio signals to infer the positions of mobile devices or vehicles (agent nodes) with the

help of fixed-position nodes like BSs or access points (APs), known as anchor nodes. Ini-

tially, the process involves extracting distance or angle measurements from the received

signal, such as time of arrival (ToA), TDoA, RSS, AoA, and angle of departure (AoD),

followed by estimating the agent node’s location based on these measurements. In 4G

networks, the typical localization accuracy requirement is within 50 meters in urban

Chapter 2. Fundamental Concepts 13

areas and 150 meters in rural areas for emergency services, while 5G networks aim to

significantly improve this, targeting 3-10 meters for general outdoor environments and

as precise as 1 meter or less for indoor scenarios. For high-precision applications like

autonomous vehicles and industrial automation, 5G seeks to achieve centimeter-level or

even millimeter-level accuracy.

The localization accuracy has progressively improved from 3G systems, offering tens

of meters accuracy using TDoA, to 4G, and notably to 5G mmWave communication

systems achieving centimeter-level accuracy. With the advent of 5G and the forthcom-

ing 6G networks, which support high-frequency mmWave and THz bands, the demand

for high precision and reliable localization is escalating, especially for applications like

smart factories, autonomous driving, and augmented reality. However, these systems

face challenges such as susceptibility to blockages that disrupt LoS, crucial for precise

location estimations.

RISs present a solution by offering reliable, precise, and energy-efficient position esti-

mates at a low cost. RISs enhance localization by creating virtual LoS links to counteract

blockages, acting as quasi-passive anchor nodes that reduce hardware costs and energy

consumption. Additionally, their large physical aperture provides higher angular reso-

lution, and their capability to adjust phase shifts offers significant beamforming gains,

making RISs a valuable addition to next-generation wireless localization systems.

2.3 Localization Scheme

Traditionally, the localization schemes can be categorized into two main types: the two-

step scheme and the fingerprint-based scheme. The two-step scheme estimates

parameters like AoAs to deduce the location by utilizing the geometric relationships.

The fingerprint-based scheme, on the other hand, relies on a database of pre-recorded

signal characteristics, such as RSS, and compares real-time signals to the database to

pinpoint a location.
Chapter 2. Fundamental Concepts 14

2.3.1 Two-Step Localization Scheme

The position and orientation of the MU can be estimated based on the observed signal

channel gains, and time delays are estimated using some existing methods. For example,

the channel gains can be estimated with the LS approach, while the angles can be

estimated by using array signal processing algorithms in the spatial domain, such as

MUSIC and ESPRIT. The estimated time delays can be extracted from the pilot signal.

In the second step, the position of the MU can be obtained by using multi-angulation

(or triangulation) from the angle-related measurements, or by using multi-lateration (or

trilateration) methods from distance-related measurements, or by a combination of both.

2.3.2 Fingerprint-Based Localization Scheme

The fingerprint-based localization scheme operates by creating and utilizing a detailed

database of signal characteristics specific to various known locations within the coverage

area. This database, often referred to as a “fingerprint map”, encompasses a wide array of

signal attributes such as RSS, ToA, and even AoA collected in a pre-survey phase. Each

unique set of signal characteristics, or “fingerprint,” corresponds to a specific physical

location.

In the operational phase, real-time measurements of the signal characteristics are

captured by the MU or the receiving device. These measurements are then matched

against the pre-established fingerprint map using algorithms designed to recognize pat-

terns or similarities. The matching process can be accomplished through various tech-

niques such as nearest neighbor algorithms, machine learning models, or sophisticated

pattern recognition methods that account for fluctuations in signal strength and envi-

ronmental changes.

The primary advantage of the fingerprint-based scheme lies in its ability to accom-

modate complex indoor environments and areas where direct line-of-sight methods may

be less effective due to multipath propagation or non-line-of-sight conditions. However,

this method requires an extensive and sometimes labor-intensive initial mapping phase
Chapter 2. Fundamental Concepts 15

to create a reliable and comprehensive fingerprint database. Additionally, the database

may need regular updates to maintain accuracy as environmental conditions change over

time.

2.4 Localization Measurements

Denote the position of the MU as sMU = [xu , yu , zu ]T , the position of the BS as sBS =

[xs , ys , zs ]T , the position of the k-th RIS as sRISk = [xr,k , yr,k , zr,k ]T . The localization

measurements are closely related with the coordinates and rotation angle of MU, which

means that there exists a mapping from the MU’s location to the measurements, which

can be described as follows.

2.4.1 ToA

The propagation delay t̂0 from the BS to the MU, factoring in the MU’s coordinates

sMU , is calculated as:

∥sBS − sMU ∥
t̂0 = + nt0 , (2.2)
c

where c represents the speed of light, and nt0 is the ToA measurement error. The ToA

t̂k0 from the MU to the k-th RIS can be obtained by subtracting the propagation delay

of the k-th RIS-BS path from the propagation delay of the path from BS to MU via the

k-th RIS, which are expressed as

∥sMU − sRISk ∥ ∥sBS − sRISk ∥

t̂k0 = + ntk = δ̂k − + ntk , (2.3)
c c

where ntk denotes the ToA error for the k-th RIS, and δ̂k denotes the ToA from the MU

to the BS.

The ToA measurements are associated with the path lengths, which in turn correlate

with the position of the MU. Specifically, the terms τ̂0 and τ̂k0 represent the estimated

distance of the direct path and the estimated distance from the MU to the k-th RIS,
Chapter 2. Fundamental Concepts 16

respectively. These distances can be mathematically expressed as follows

dˆ0 = ct̂0 = ∥sBS − sMU ∥ + nd0 ,

dˆk = ct̂k0 = ∥sMU − sRISk ∥ + ndk , (2.4)

where nd0 and ndk represent the distance estimation errors.

2.4.2 TDoA

The position of the MU can also be ascertained through the TDoA measurement. TDoA

utilizes the distance differences between different pairs of the direct MU-BS path and

the k-th MU-RIS paths.

Specifically, using the direct path from the BS to the MU as the reference path, TDoA

is defined by the following equation:

t̂TD,k = t̂k0 − t̂0 . (2.5)

Accordingly, the distance difference referring to the TDoA is formulated as

dˆd,k = ct̂TD,k = (dˆk − dˆ0 ) = ∥sMU − sRISk ∥ − ∥sBS − sMU ∥ + nTDk , (2.6)

where nTDk is the estimation error of dd,k .

2.4.3 AoA and AoD

The AoAs and AoDs at the RIS are estimated with angle estimation algorithms, e.g.,

MUSIC. Let φ̂ak , φ̂ek denote the estimated azimuth, elevation AoA at the k-th RIS,

respectively. Besides, θ̂ka , θ̂ke denote the estimated azimuth, elevation AoD at the k-th

RIS, respectively. They are also have mathematical formulation with the MU position,
Chapter 2. Fundamental Concepts 17

which can be formulated as

 
yu − yr,k
φ̂ak = arcsin  q  + na,ak , (2.7)
2 2
(xu − xr,k ) + (yu − yr,k )
 
e −zr,k
φ̂k = arccos + na,ek , (2.8)
∥sMU − sRISk ∥2
 
xu − xr,k
θ̂ka = arctan + nd,ak , (2.9)
yu − yr,k
!
zu − zr,k
θ̂ke = arctan + nd,ek . (2.10)
a (x − x ) + cos θ̂ a (y − y )
sin θ̂in,k u r,k in,k u r,k

where na,ak , na,ek , nd,ak , and nd,ek denote the estimation error associated to the AoA

and AoD, respectively.

2.5 Weighted Least Squares (WLS) Algorithm

The weighted least squares (WLS) algorithm is an advanced statistical method used

to estimate unknown parameters in a model by minimizing the sum of the squared

differences between observed and predicted values, with each difference being weighted

to reflect the confidence or reliability of its observation. This approach is particularly

useful in scenarios where measurements vary in precision or quality, allowing for a more

accurate and reliable estimation by giving more importance to higher quality data.

The core principle of the WLS algorithm revolves around the concept of weighted

minimization. Unlike the ordinary least squares (OLS) method, which treats all dis-

crepancies between model predictions and actual observations equally, WLS assigns a

weight to each discrepancy. These weights are inversely proportional to the variance

of the observations, meaning that measurements with lower variance (and hence higher

precision) have a greater influence on the outcome of the parameter estimation process.

Mathematically, the objective of the WLS algorithm is to find the parameter vector
Chapter 2. Fundamental Concepts 18

θ that minimizes the weighted sum of squared residuals:

n
X
min wi (yi − f (xi , θ))2 ,
θ
i=1

where yi represents the observed values, f (xi , θ) denotes the predicted values determined

by the model, wi are the weights assigned to each observation, and n is the number of

observations.

In the context of localization, the WLS algorithm can be particularly effective for

refining position and orientation estimates. For instance, in estimating the location of a

mobile device using signals from multiple sources (e.g., satellites in GPS, base stations in

cellular networks, or access points in Wi-Fi networks), the WLS algorithm can account

for the varying signal qualities and interference effects by assigning appropriate weights

to each signal. This ensures that signals with higher reliability have a more significant

impact on the final estimation, enhancing the accuracy of the localization process.

2.6 Convolutional Neural Networks (CNNs)

CNNs are a class of deep neural networks, highly effective in processing data with a

grid-like topology, such as images. CNNs have revolutionized the field of computer

vision, enabling breakthroughs in image recognition, image classification, object detec-

tion, and many other areas where automated understanding of visual data is required.

Furthermore, CNNs can be applied to fingerprint-based localization algorithms, where

the fingerprint is treated as an image. This approach allows for the leveraging of CNNs’

robust feature extraction capabilities to enhance the accuracy and efficiency of localiza-

tion systems.

The architecture of a CNN is specifically designed to automatically and adaptively

learn spatial hierarchies of features from input images. This is achieved through the use

of multiple layers of convolutional layers, pooling layers, and fully connected layers. The

key components that distinguish CNNs from other neural network models are given as
Chapter 2. Fundamental Concepts 19

follows.

2.6.1 Convolutional Layers

These layers perform a convolution operation, applying filters to the input to create a

feature map that summarizes the presence of detected features in the input. Convolu-

tional layers can capture spatial relationships between pixels in an image by considering

small chunks of the input image at a time.

2.6.2 Pooling Layers

Following convolutional layers, pooling (or subsampling or down-sampling) layers reduce

the dimensionality of each feature map but retain the most essential information. Pooling

helps in making the detection of features invariant to scale and orientation changes.

2.6.3 Fully Connected Layers

At the end of the network, one or more fully connected layers perform classification

based on the features extracted and down-sampled by the convolutional and pooling

layers. The last fully connected layer (often followed by a softmax activation function)

produces the final output, such as the probabilities of different classes.

2.7 Transfer Learning

Transfer learning [21] is particularly advantageous for tasks like detecting faulty elements

within the RIS. One of the primary reasons behind its suitability is the analogous nature

of received signals to images. Received signal at the BS can be viewed as an image,

also with features to be extracted and recognized. This similarity allows us to harness

the power of models pre-trained on image recognition tasks and adapt them for faulty

elements detection. Transfer learning can efficiently use data from image domains to

facilitate not only the detection of faulty elements but also to improve the precision of

localization algorithms. By leveraging pre-trained models, transfer learning introduces

Chapter 2. Fundamental Concepts 20

significant improvements in localization performance, especially in environments where

direct training data may be scarce or noisy. Transfer learning can efficiently use data

from image domains to facilitate faulty elements detection. Mathematically, transfer

learning can be given by

min LT (Ttarget |θT , Dtarget ) + λL(Tsource |θT , Dsource ). (2.11)

θT

In the above equation, LT is the loss function and λ is a regularization parameter ensuring

the model’s adaptability to y while preserving knowledge from image domains. The

variable θT denotes the model’s parameters, fine-tuned in transfer learning. Ttarget and

Tsource refer to the target and source tasks. The datasets for these tasks are denoted by

Dtarget and Dsource , respectively.

Chapter 3

RIS-Aided Localization
Algorithm Design: Tackling
Non-Gaussian AEE

3.1 Introduction

For the mmWave RIS-aided localization systems, there are some researchers conduct the

research on CRLB analysis and optimizing RIS phase shifts. However, a detailed 3D

localization algorithm capable of acquiring the closed-form expression of the localization

of the MU has not been proposed. Such a localization algorithm is crucial for practical

system deployment and serves as a foundation for further error analysis.

Besides, for those who designed the classical localization algorithm without consider-

ing RIS, they predominantly assumes Gaussian-distributed localization parameter errors

for tractable algorithmic design. However, the distribution of the estimation error of

these parameters, may not follow the Gaussian distribution. In practice, the distribu-

tion of estimation error of these parameters depends on the practical estimation methods

(e.g., 2D-DFT algorithm [41]), which shows the non-Gaussian property. Hence, it is of

21
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 22

great interest and important for us to investigate the distribution of the estimation error

of these parameters. As a result, the impact of these non-Gaussian distribution on the

localization algorithm can be further studied.

Against the above background, this chapter propose a complete framework to jointly

analyze the angle estimation and design the 3D localization algorithm for the mmWave

RIS-aided localization system. To model the angle estimation error, this chapter adopt

the 2D-DFT estimation method as an example and leverage the geometric relationship

between the AoAs and their triangle functions. Furthermore, to derive the closed-form

solution of the MU’s position, this chapter derive the variance of the angle estimation

error and apply the TSWLS algorithm. Our main contributions are summarized as

follows:

1) A comprehensive framework is proposed to jointly analyze the angle estimation

and design the 3D localization algorithm for the mmWave RIS-aided localization

system.

2) This chapter obtain closed-form expressions for the estimated azimuth and eleva-

tion AoAs using the 2D-DFT method. Our analysis reveals that the angle estima-

tion error at the RISs depends on the search grid, RIS panel size, and the number

of reflecting elements.

3) The angle estimation error, including both azimuth and elevation, is characterized

by a non-Gaussian PDF. The PDF is derived by utilizing the geometric relationship

between the AoAs and the corresponding triangle functions. To simplify the com-

plex geometric expression, this chapter employ a first-order linear approximation

of the triangle function. Specifically, this chapter develop an algorithm to derive

and approximate the PDF expression for the azimuth angle estimation error.

4) Using the derived PDF of the estimation error, this chapter analytically derive the

variance of the angle estimation error. For the azimuth estimation error, which has

three distinct non-zero intervals in its PDF, this chapter calculate the integral sep-
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 23

Figure 3.1: RISs aided localization system model.

arately for each interval to obtain the variance. This variance is then incorporated

into the WLS algorithm to obtain the closed-form 3D position of the MU.

5) Simulation results verify the accuracy of the derived results and demonstrate the

superiority of the proposed framework. It is observed that the variance decreases

with the number of elements, which means that increasing the number of RISs’

elements improves the estimation accuracy.

3.2 System Model

3.2.1 Geometry Model

Consider a RISs-aided localization system, where a MU sends pilot signals to the BS to

locate the MU with the assistance of an RIS. The BS is equipped with a ULA with Nb

antennas, and the MU is equipped with a single antenna. Moreover, this chapter assume

that there are M RISs, and each RIS is a UPA. Furthermore, this chapter assume that

the direct link between the MU and the BS has been blocked by the obstacles.

The BS is placed parallel to the x-axis with the center located at p = [xp , yp , zp ]T .

The ith RIS is placed parallel to the y-o-z plane with their centers located at si =
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 24

[xi , yi , zi ]T ,i = 1, 2, · · · , M . The UPA-based RIS has Ny,z = Ny × Nz reflecting elements,

where Ny and Nz denote the numbers of reflecting elements along the y-axis and z-

axis, respectively. The real location of the MU is q = [xq , yq , zq ]T and is assumed to

be placed parallel to the x-o-y plane. The estimated location of the MU is denoted as

q̂ = [x̂q , ŷq , ẑq ]T . Generally, once the RISs and BS have been deployed, the coordinates

si and p are known and invariant. In order to locate the MU, this chapter need to obtain

the estimated q̂.

3.2.2 Channel Model

Assuming that the number of propagation paths between the MU and the ith RIS is

Ni , the AoA of the nth path from the MU to the ith RIS can be decomposed into the

elevation angle 0 ≤ Θn,i ≤ π in the vertical direction, and the azimuth angle 0 ≤ Φn,i ≤ π

in the horizontal direction, respectively. As a result, the array response vector at the ith

RIS of the nth path can be expressed as

a(Θn,i , Φn,i ) = ae (Φn,i ) ⊗ aa (Θn,i , Φn,i ), (3.1)

where ⊗ denotes the Kronecker product. Moreover, this chapter have

−j2πdr sin Φn,i −j2π(Nx −1)dr sin Φn,i

ae (Φn,i ) = [1, e λc , ..., e λc ]T , (3.2)

and

−j2πdr cos Θn,i cos Φn,i −j2π(Nz −1)dr cos Θn,i cos Φn,i
aa (Θn,i , Φn,i ) = [1, e λc ,··· ,e λc ]T , (3.3)

where dr and λc denote the distance between the elements of the RISs and the carrier

wavelength, respectively. Then, the channel between the MU and the ith RIS, denoted
 Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 25

as hi , can be modeled as

Ni
X
hi = αn,i a(Θn,i , Φn,i )
n=1
Ni
X
= α1,i a(Θ1,i , Φ1,i ) + αn,i a(Θn,i , Φn,i ), (3.4)
| {z }
n=2
LoS | {z }
H N LoS
h
i,zi]
T

where αn,i denotes the complex channel gain of the nth path and the ith RIS. Moreover,
y
i α1,i , Θ1,i , and Φ1,i denote the complex channel gain, the elevation AoA, and the azimuth

AoA of the LoS path, respectively. As shown from (3.4), channel components of hi can

be categorized into two types, namely LoS and NLoS. LoS path component is the direct
MU
xq,yq,zq]T path between the RISs and the MU, NLoS path component consists of the paths between

the RISs and the MU reflected by scatters, e.g., walls, human bodies, and etc. Moreover,

according to [43], the complex channel gain of LoS is given by

λc e−j2πd1,i
α1,i = , (3.5)
4πd1,i

where d1,i is the distance between the ith RIS and the MU.

MR,i

Figure 3.2: Communication link from RIS to BS.

The RISs are assumed to be placed in the vicinity of the BS [44]. Then, the ray-

tracing LoS channel model can be employed according to [45, 46]. Fig. 3.2 illustrates

the transmission from the RIS to the BS. The inter-element spacing is assumed to be d1
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 26

while the inter-antenna spacing is assumed to be d2 . For the sake of analysis, in Fig. 3.2,

the first antenna of the BS is assumed to be placed at the original point. R represents

the distance from the first antenna of the BS to the first reflecting element of the RIS,

while ΦBR and ΘBR are the angles of the local spherical coordinate system at the RIS

and the BS. For notation brevity, let ny,z = (ny , nz ) denote the reflecting element of

the RIS at column ny ∈ {0, 1, · · · , Ny − 1} and row nz ∈ {0, 1, · · · , Nz − 1}. Thus, the

distance between BS antenna nb ∈ {0, 1, · · · , Nb − 1} and reflecting element ny,z at the

RIS can be calculated as

rnb ,ny,z
q
= (R sin ΦBR cos ΘBR + nb d2 )2 + (R cos ΦBR + ny d1 )2 + (R sin ΦBR sin ΘBR − nz d1 )2 .

(3.6)

The array vector from reflecting element ny,z to the Nb antennas at the BS is written as

2π 2π
hny,z = [ej r
λ 0,ny,z , ..., ej r
λ Nb −1,ny,z ]T . (3.7)

Thus, the channel matrix[45] from the RIS to the BS is given by [45]

√
H= ρ[h0 , h1 , ..., hNy,z −1 ], (3.8)

where ρ is the common power attenuation.

3.2.3 Received Signal Model

Let et ∈ CNy,z ×1 denote the phase shift vector of the RIS at time slot t, which satisfies

|[et ]ny,z | = 1 for 1 ≤ ny,z ≤ Ny,z . Accordingly, the received signal from the MU via the

RIS to the BS at time slot t could be expressed as

√
yi (t) = HDiag(et )hi ps(t) + n(t), (3.9)
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 27

where s(t) denotes the transmitted pilot signal of the MU and n(t) ∈ CNb ×1 is additive

white Gaussian noise (AWGN) following the distribution of CN (0, δ 2 I). Moreover, p is

the transmit power of the MU.

Since the positions of the BS and the RISs are fixed and known at the BS, the channel

matrix H can be directly calculated based on (3.8).This chapter considers the massive

mmWave MIMO system, where the number of antennas at the BS is larger than the

number of reflecting elements at the RIS. This assumption is primarily made to sim-

plify the mathematical derivations and theoretical analysis in the subsequent chapters.

While acknowledging the cost implications of such a setup, more practical scenarios with

fewer BS antennas will be explored in Chapters 6 and 7. These chapters will discuss

how utilizing RIS information can address the limitations imposed by fewer antennas,

thereby providing a more cost-effective yet efficient system configuration. As a result, it

is obtained that the pseudo-inverse matrix of the channel matrix H, which is given by

Hp = (HH H)−1 HH . By multiplying both sides of (3.12) by Hp , one obtains

√
ỹi (t) = Diag(et )hi ps(t) + ñ(t), (3.10)

where ỹi (t) = Hp yi (t), and ñ(t) = Hp n(t) that follows the distribution of CN (0, δ 2 (HH H)−1 ).

Assuming that the phase shifts of the reflecting elements are available at the BS, by mul-

tiplying both sides of (3.10) by the inverse matrix of Diag(et ), with the RIS phase set

to all ones, it is written as

√
ȳi (t) = hi ps(t) + n̄(t), (3.11)

where ȳi (t) = (Diag(et ))−1 ỹi (t) and n̄(t) = (Diag(et ))−1 ñ(t). The influence of the

optimized phase matrix of the RIS on localization would be an interesting topic for future

work. Here, n̄(t) follows the distribution of CN (0, δ 2 ((HDiag(et ))H HDiag(et ))−1 ).

Besides, since the NLoS path component varies fast and its contribution to the channel

is marginal, especially in the mmWave band, this chapter are more interested in the LoS
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 28

path component [47–49]. Thus, this chapter consider the NLoS path component as

a noise component and intend to estimate the LoS path component in this chapter.

Therefore, the received signal can be further written as

√
ȳi (t) = ĥi ps(t) + n̂(t), (3.12)

PNi √
where ĥi = α1,i a(Θ1,i , Φ1,i ) and n̂(t) = n=2 αn,i a(Θn,i , Φn,i ) ps(t) + n(t).

3.2.4 Framework

In the existing works concerning the localization algorithm design, for tractability, the

estimation error of channel parameters is assumed to be additive zero-mean complex

Gaussian noise [50]. However, in practice, the distribution of these channel parame-

ters depends on the practical estimation methods, which may not follow the Gaussian

distribution. To investigate the impact of the practical estimation error of the channel

parameters, this chapter propose to design a comprehensive framework to jointly analyze

the angle estimation error and estimate the 3D position of the MU.

Based on the 2D-DFT angle estimation technique [51], the angle estimation error

analysis and 3D localization algorithm design is investigated in this chapter. First, this

chapter apply the 2D-DFT algorithm to estimate (Θ1,i , Φ1,i ) by using the received signal

ȳi (t) in (3.12). Then, based on the estimation of the AoAs, this chapter first derive the

PDF of the angle estimation error, based on which the closed-form expression of the

variance of the angle estimation error is derived. Finally, this chapter apply the TSWLS

algorithm to estimate the 3D position of the MU by using the estimated AoAs and the

derived variance.

The details of the proposed framework are summarized in Algorithm 1. The descrip-

tions of each step of the proposed framework will be introduced in the following sections.
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 29

Algorithm 1 Joint Angle Estimation Error Analysis and 3D Positioning Framework

1: Estimate (Θ1,i , Φ1,i ) by using the 2D-DFT estimation algorithm;
2: Derive the PDF of the angle estimation error;
3: Derive the closed-form expression of the variance of the angle estimation error;
4: Estimate the 3D position of the MU by using the variance of the angle estimation
error.

3.3 Angle Estimation

The first step of the proposed framework is to estimate (Θ1,i , Φ1,i ) by using the received

signal ȳi (t) in (3.12). Hence, in this section, this chapter apply the 2D-DFT algorithm

[51] to estimate the AoA at the RISs.

3.3.1 Initial Angle Estimation

According to the expression of a(Θ1,i , Φ1,i ) in (3.1), a(Θ1,i , Φ1,i ) can be further derived

as

a(Θ1,i , Φ1,i ) = vec{A(Θ1,i , Φ1,i )} = vec{aa (Θ1,i , Φ1,i )aTe (Φ1,i )}. (3.13)

To estimate the AoA at the ith RIS, this chapter define two DFT matrices FNy and
2π
−j N bi b′i
FNz , elements of which are written as [FNy ]bi b′i = e y (bi , b′i = 0, 1, · · · , Ny − 1)
2π ′
and [FNz ]qi qi′ = e−j Nz qi qi (qi , qi′ = 0, 1, · · · , Nz − 1), respectively. Meanwhile, define
2πdr cos Θ1,i cos Φ1,i 2πdr sin Φ1,i
u1,i = λc and v1,i = λc . Then, this chapter define the normalized

2D-DFT of the matrix A(Θ1,i , Φ1,i ) in (3.13) as ADF T (Θ1,i , Φ1,i ) = FNy A(Θ1,i , Φ1,i )FNz ,

whose (bi , qi )th element is calculated as

[ADF T (Θ1,i , Φ1,i )]bi qi

Ny −1 Nz −1 bi ny qi nz
X X −j2π( + )
= [A(Θ1,i , Φ1,i )]bi qi e Ny Nz

ny =0 nz =0
Ny u1,i Nz v1,i
j
Ny −1 2πb
(u1,i − N i ) j Nz2−1 (v1,i −
2πqi
) sin(πbi − 2 ) sin(πqi − 2 )
=e 2 y e Nz × Ny u1,i
· Nz v1,i
.
sin((πbi − 2 )/Ny ) sin((πqi − 2 )/Nz )

(3.14)
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 30

When the number of reflecting elements becomes infinite, i.e., Ny → ∞, Nz → ∞,

Ny u1,i Nz v1,i
there always exist some integers bni = 2π , qni = 2π such that [ADF T (Θ1,i , Φ1,i )]bni qni

= 1, while the other elements are all zero. Therefore, all power is concentrated on the

(bni , qni )th element and ADF T (Θ1,i , Φ1,i ) is a sparse matrix. However, the RIS size
Ny u1,i Nz v1,i
could not be infinitely large, thus 2π and 2π may not be integers in general, which

leads to the channel power leakage from the (bni , qni )th element to its adjacent element.

However, ADF T (Θ1,i , Φ1,i ) can still be approximated as a sparse matrix with the most

power concentrated around the (bni , qni )th element. Therefore, the peak power position

of ADF T (Θ1,i , Φ1,i ) is still useful for estimating the AoAs at the achor. Then, the initial

estimation is derived as follows:

λc bni
cos Θ̂ini ini
1,i cos Φ̂1,i = ,
Ny d r
λc qni
sin Φ̂ini
1,i = , (3.15)
Nz dr

where Θ̂ini ini

1,i and Φ̂1,i denote the initial estimated angles at the RIS.

3.3.2 Fine Angle Estimation

The resolution of the estimated sin Φ̂ini ini ini

1,i and cos Θ̂1,i cos Φ̂1,i is limited by the half of the

DFT interval. In order to improve the estimation accuracy, angle rotation is provided

to solve the mismatch issue in this subsection [41].

Define the angle rotation matrix of A(Θ1,i , Φ1,i ) as Aro (Θ1,i , Φ1,i ), expressed as

Aro (Θ1,i , Φ1,i ) = UNy (ϖ̃1,i )A(Θ1,i , Φ1,i )UNz (ϖ̃2,i ), (3.16)

where the diagonal matrices UNy (ϖ̃1,i ) and UNz (ϖ̃2,i ) are given by

UNy (ϖ̃1,i ) = Diag{1, ej ϖ̃1,i , ..., ej(Ny −1)ϖ̃1,i },

UNz (ϖ̃2,i ) = Diag{1, ej ϖ̃2,i , ..., ej(Nz −1)ϖ̃2,i }, (3.17)

Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 31

with ϖ̃1,i ∈ [−π/Ny , π/Ny ] and ϖ̃2,i ∈ [−π/Nz , π/Nz ] being the angle rotation parame-

ters. By using the angle rotation operation, the (bi , qi )th element of the 2D-DFT of the

rotated matrix Aro

DF T (Θ1,i , Φ1,i ) is calculated as

[Aro
DF T (Θ1,i , Φ1,i )]bi qi
Ny −1 Nz −1 b i ny q i nz
X X −j2π( + ) j2π( ϖ̃1,i ny + ϖ̃2,i nz )
= [A(Θ1,i , Φ1,i )]bi qi e Ny Nz
e 2π 2π

ny =0 nz =0
Ny −1 2πb Nz −1 2πq
j (u1,i +ϖ̃1,i − N i ) (v1,i +ϖ̃2,i − N i )
=e 2 y ej 2 z

Ny u1,i N ϖ̃ Nz v1,i N ϖ̃
sin(πbi − 2 − y 2 1,i ) sin(πqi − 2 − z 2 2,i )
× Ny u1,i N ϖ̃
· Nz v1,i N ϖ̃
, (3.18)
sin((πbi − 2 − y 2 1,i )/Ny ) sin((πqi − 2 − z 2 2,i )/Nz )

where ϖ̃1,i and ϖ̃2,i could be optimized with the one-dimensional search. Then, this

chapter could obtain the estimated results as

λc bni λc ϖ̃1,i
cos Θ̂1,i cos Φ̂1,i = − ,
Ny dr 2πdr
λc qni λc ϖ̃2,i
sin Φ̂1,i = − , (3.19)
Nz dr 2πdr

where Θ̂1,i and Φ̂1,i denote the final estimated azimuth and elevation AoAs after angle

rotation, respectively. Furthermore, Θ̂1,i and Φ̂1,i could be expressed as

 
 λc bni λc ϖ̃2,i 
Φ̂1,i = arcsin  − ,
Nz dr 2πdr

,s
λc bni λc ϖ̃1,i λc qni λc ϖ̃2,i 2
Θ̂1,i = arccos(( − ) (1 − ( − ) )). (3.20)
Ny dr 2πdr Nz dr 2πdr

Furthermore, the estimation of the AoAs (the elevation angle Θn,i and the azimuth angle

Φn,i ) at the ith RIS can be also formulated as

Θ̂1,i = Θ1,i + Θ̃1,i

Φ̂1,i = Φ1,i + Φ̃1,i , (3.21)

Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 32

where Θ1,i and Φ1,i denote the true AoAs at the ith RIS, and Θ̃1,i and Φ̃1,i denote the

estimation error at the ith RIS.

3.4 Derivation of the PDF of Angle Estimation Error

In this section, as the second step of the proposed framework, the PDF of the angle

estimation error is derived, which will be used for deriving the variance of the angle

estimation error. To obtain the PDF of the angle estimation error, the first step is to

derive the PDF of Φ̃1,i based on the estimated angle Φ̂1,i and the property of the 2D-

DFT. Then, the second step is to design an algorithm by deriving the PDF of Θ̃1,i based

on the estimated angle Θ̂1,i and the property of the 2D-DFT. The details are given as

follows:

3.4.1 PDF of Φ̃1,i

Based on the above section, it can be observed that ϖ̃2,i could be obtained by the one-

dimensional search in the interval of [− Nπz , Nπz ]. It is assumed that there are S2,i grids

points in the interval [− Nπz , Nπz ] and s2,i ∈ {1, · · · , S2,i } is the optimal point. Therefore,
2πs2,i
the optimal solution for the one-dimensional search is ϖ̃2,i = Nz S2,i , and the estimation

of sin Φ1,i is thus written as

λc qni λc ϖ̃2,i λc qni λc s2,i

sin Φ̂1,i = − = − . (3.22)
Nz dr 2πdr Nz dr Nz dr S2,i

For notation simplicity, define Ŷi = sin Φ̂1,i and Yi = sin Φ1,i . Then, the estimation

of Φ1,i could be expressed as

 
λc qni λc s2,i
Φ̂1,i = arcsin Ŷi = arcsin − . (3.23)
Nz dr Nz dr S2,i

According to (3.22) and the property of the one-dimensional search method, the value
h i
of Yi follows the uniform distribution within the region of Ŷi − ai , Ŷi + ai , where ai =
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 33

λc
2Nz dr S2,i , which is given by


1


2ai , Ŷi − ai ≤ y ≤ Ŷi + ai
fYi (y) = (3.24)
0, others.



By denoting the estimation error of Yi as Ỹi , it is formulated as Ŷi = Yi + Ỹi . Then, the

distribution of Ỹi is given by


1


2ai , −ai ≤ ỹ ≤ ai
fY˜i (ỹ) = (3.25)
0, others.



Since Φ1,i = arcsin Yi , the CDF of Φ1,i is derived as

FΦ1,i (ϕ1,i ) = Pr(Φ1,i ≤ ϕ1,i )

= Pr(arcsin Yi ≤ ϕ1,i ) = Pr(Yi ≤ sin ϕ1,i ). (3.26)

By assuming that Φ1,i ∈ (− π2 , π2 ), the PDF of Φ1,i could be derived as

∂FΦ1,i (ϕ1,i )
fΦ1,i (ϕ1,i ) = = cos ϕ1,i fYi (sin ϕ1,i )
∂ϕ1,i

 cos ϕ1,i ,
 arcsin(Ŷi − ai ) ≤ ϕ1,i ≤ arcsin(Ŷi + ai )
2ai
= (3.27)
0, others.



As Φ̂1,i = Φ1,i + Φ̃1,i , the CDF of the estimation error Φ̃1,i is calculated as

     
FΦ̃1,i ϕ̃1,i = Pr Φ̃1,i ≤ ϕ̃1,i = Pr Φ̂1,i − Φ1,i ≤ ϕ̃1,i
   
= Pr Φ1,i ≥ Φ̂1,i − ϕ̃1,i = 1 − Pr Φ1,i ≤ Φ̂1,i − ϕ̃1,i . (3.28)

   
Define a1,i = arcsin Ŷi + ai and a2,i = arcsin Ŷi − ai . Based on (3.28), the PDF of
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 34

Φ̃1,i is written as

 
  ∂FΦ̃1,i ϕ̃1,i  
fΦ̃1,i ϕ̃1,i = = fΦ1,i Φ̂1,i − ϕ̃1,i
∂ ϕ̃1,i

cos(Φ̂1,i −ϕ̃1,i )


2ai , Φ̂1,i − a1,i ≤ ϕ̃1,i ≤ Φ̂1,i − a2,i
= (3.29)
0, others.



Moreover, as the estimation techniques are relatively mature, it is assumed that

 
the estimation error is very small. Therefore, it is formulated as sin ϕ̃1,i ≈ ϕ̃1,i and
 
cos ϕ̃1,i ≈ 1. Consequently, it is formulated as the following approximation:

 
cos Φ̂1,i − ϕ̃1,i = cos Φ̂1,i cos ϕ̃1,i + sin Φ̂1,i sin ϕ̃1,i ≈ cos Φ̂1,i + ϕ̃1,i sin Φ̂1,i . (3.30)

To simplify the variance calculation of Φ̃1,i in the subsequent section, the PDF of Φ̃1,i

can be approximated as

cos Φ̂1,i +ϕ̃1,i sin Φ̂1,i
  

2ai , Φ̂1,i − a1,i ≤ ϕ̃1,i ≤ Φ̂1,i − a2,i
fΦ̃1,i ϕ̃1,i ≈ (3.31)
0, others.



3.4.2 PDF of Θ̃1,i

In this subsection, the PDF of the estimation error Θ̃1,i can be derived. As the derivations

are complicated, the main procedures are summarized in Algorithm 2.

Algorithm 2 Algorithm of deriving the PDF of Φ̃1,i

1: Derive the PDF of cos Φ1,i by using the PDF of Ỹi in (3.25);
2: Derive the PDF of cos Θ1,i for two cases by using the PDF of cos Φ1,i in (3.38) and
the PDF of cos Φ1,i cos Θ1,i in (3.40);
3: Derive the PDF of Θ̃1,i based on the two cases of the PDF of cos Θ1,i in (3.49) and
(3.50).
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 35

[Link] PDF of cos Φ1,i

First of all, the PDF of cos Φ1,i should be derived, so that the PDF of cos Θ1,i could be

calculated.

Let Xi = cos Φ1,i , which can be expressed as a function of Ỹi as follows:

q q
Xi = cos Φ1,i = 1 − sin2 Φ1,i = 1 − Yi2
q
= 1 − (Ŷi − Ỹi )2 . (3.32)

By defining the estimation of Xi as X̂i = cos Φ̂1,i = Xi + X̃i , the estimation error X̃i

could be calculated as:

q
X̃i = X̂i − 1 − (Ŷi − Ỹi )2 . (3.33)

However, the expression (3.33) is complicated and thus challenging to derive a compact

form of the PDF of X̃i . Fortunately, since the value of Ỹi is relatively small, X̃i in (3.33)

can be approximated by using the Taylor expansion, which is given by

 
q q
Ỹi · Ŷi Ŷi
q
2 2 2 1
X̃i = 1 − Ŷi − 1 − (Ŷi − Ỹi )2 1 − Ŷi − (1 − Ŷi ) 2 +  = − Ỹi .
 
2 1
(1 − Ŷi ) 2 X̂i

(3.34)

Utilizing (3.34), the CDF of X̃i could be calculated as

     
 Ŷi X̂i  X̂i 
FX̃i (x̃i ) = Pr(X̃i ≤ x̃i ) ≈ Pr − Ỹi ≤ x̃i  = PrỸi ≥ − x̃i  = 1 − PrỸi ≤ − x̃i .
  
X̂i Ŷi Ŷi

(3.35)
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 36

Then, by using (3.35), the PDF of X̃i could be calculated as

    
X̂i ˆ Yˆi
 X̂i   X̂i 
 , − Yi ai ≤ x̃i ≤ a
X̂i i

2ai Yˆi X̂i
fX̃i (x̃i ) = − − fY˜i  − x̃i  = (3.36)
Ŷi Ŷi 
 0, others.

Furthermore, by using X̂i = Xi + X̃i , the CDF of Xi can be calculated as

FXi (xi ) = Pr(Xi ≤ xi ) = Pr(X̂i − X̃i ≤ xi ) = Pr(X̃i ≥ X̂i − xi ) = 1 − Pr(X̃i ≤ X̂i − xi ).

(3.37)

By using (3.36) and (3.37), the PDF of Xi could be written as


X̂i Yˆi Yˆi
∂FXi (xi )
 , X̂i − a ≤ xi ≤ X̂i − a
X̂i i X̂i i

2ai Yˆi
fXi (xi ) = = fX̃i (X̂i − xi ) =
∂xi
0, others.



(3.38)

[Link] PDF of cos Θ1,i

By using the PDF of cos Φ1,i and cos Φ1,i cos Θ1,i , the PDF of cos Θ1,i can be derived as

follows.

Firstly, from the above section, it is known that ϖ̃1,i could be obtained by using the

one-dimensional search in the interval of [− Nπy , Nπy ]. Similarly, it is assumed that there

are S1 grids points in the interval [− Nπy , Nπy ] and s1,i ∈ {1, · · · , S1,i } is the optimal point.
2πs1,i
Therefore, the optimal solution for the one-dimensional search is ϖ̃1,i = Ny S1,i , and the

estimated cos Φ1,i cos Θ1,i is given by

λc bni λc ϖ̃1,i λc bni λc s1,i

Ẑ = cos Θ̂1,i cos Φ̂1,i = − = − . (3.39)
Ny d1 2πdr Ny dr Ny dr S1,i

By using (3.39) and the nature of the one-dimensional search method, the real value of
h i
Zi = cos Θ1,i cos Φ1,i follows the uniform distribution within the region of Ẑi − bi , Ẑi + bi ,
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 37

λc
where bi = 2Ny d1 S1,i . The PDF of Zi is thus given by


1


2bi , Ẑi − bi ≤ zi ≤ Ẑi + bi
fZi (zi ) = (3.40)
0, others.



By denoting the estimation error of Zi as Z̃i , it is noted that Ẑi = Zi + Z̃i . Then, the

distribution of Z̃i is given by


1


2bi , −bi ≤ z˜i ≤ bi
fZ̃i (z˜i ) = (3.41)
0, others.



Next, for simplicity, denote Ui = cos Θ1,i . Then, by utilizing the definition of Zi

below (3.39) and Xi above (3.32), it is derived that Zi = Ui Xi . By combining the PDF

of Xi in (3.38) with the PDF of Zi in (3.40), the PDF of Ui can be derived as follows

Yˆi
Z X̂i + ai
X̂i
fUi (ui ) = xfXi (x)fZi (ui x)dx. (3.42)
Yˆi
X̂i − ai
X̂i

According to (3.40), it is observed that fZi (ui x) is non-zero when Ẑi − bi ≤ ui x ≤

Ẑi +bi , which determines the PDF of Ui . Therefore, it is necessary to discuss the different

conditions according to the non-zero intervals of fZi (ui x) and fXi (x) in the following.
Yˆi Yˆi
For notation brevity, it is denoted that α1,i = X̂i − a, α2,i = X̂i + a, β1,i = Ẑi − bi
X̂i i X̂i i

and β2,i = Ẑi + bi .

β β
Condition 1: If α1,i < u1,i i
≤ α2,i < u2,i
i
, the integral interval of x in (3.42) can be
h i
β
recast as u1,i
i
, α2,i , thereby yielding the PDF of Ui as

Z α2,i Z α2,i  2 !
1 X̂i X̂i X̂i 2 β1,i
fUi (ui ) = · · xdx = xdx = α2,i − .
β1,i
2bi 2ai Ŷi 4ai bi Ŷi β1,i
8ai bi Ŷi ui
ui ui

(3.43)
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 38

Furthermore, the interval of ui is given by

   
β1,i β1,i β2,i
, ∩ −∞, . (3.44)
α2,i α1,i α2,i

β1,i β2,i
Based on (3.44), it is necessary to compare α1,i with α2,i ,
so that the interval of ui can
h 
β2,i β1,i β β1,i
be further determined. If α2,i > α1,i holds, the interval can be rewritten as α1,i ,
2,i α1,i
.
h 
β1,i β2,i
Otherwise, the interval is α2,i , α2,i .

β β
Condition 2: If u1,i ≤ α1,i < u2,i ≤ α2,i , the integral interval of x in (3.42) can be
h i i i
β
recast as α1,i , u2,i
i
, thus the PDF of Ui is given by

Z β2,i Z β2,i  2 !
ui 1 X̂i X̂i ui X̂i β2,i
fUi (ui ) = · · xdx = xdx = − α1,i 2 .
α1,i 2bi 2ai Ŷi 4ai bi Ŷi α1,i 8ai bi Ŷi ui

(3.45)

The interval of ui is given by

   
β2,i β2,i β1,i
, ∩ , +∞ . (3.46)
α2,i α1,i α1,i

h 
β2,i β1,i β2,i β2,i
As a result, if α2,i >
holds, the interval can be further recast as
α1,i α2,i , α1,i . Other-
h 
β β2,i
wise, the interval can be derived as α1,i ,
1,i α1,i
.

β1,i β2,i
Condition 3: If ui ≤ α1,i < α2,i < ui , the integral interval of x in (3.42) can be

derived as [α1,i , α2,i ]. Thus, the PDF of Ui is

fUi (ui )
Z α2,i Z α2,i
1 X̂i X̂i
= · · xdx = xdx
α1,i 2bi 2ai Ŷi 4ai bi Ŷi α1,i
 !2 !2 
X̂i  Ŷi Ŷi X̂i
= X̂i + ai − X̂i − ai  = . (3.47)
8ai bi Ŷi X̂i X̂i 2bi

h 
β1,i β2,i
The interval of ui is accordingly given by α1,i , α2,i .
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 39

β β
Condition 4: If α1,i < u1,i i
< u2,i
i
≤ α2,i , the integral interval of x in (3.42) can be
h i
β β
derived as u1,i
i
, U2,ii . Thus, the PDF of Ui is

fUi (ui )
Z β2,i Z β2,i
ui X̂i X̂i ui
= β1,i
· xdx = xdx
ui
4ai bi Ŷi 4ai bi Ŷi βu1,i
i
 !
β2,i 2 β1,i 2
  
X̂i X̂i Ẑi
= − = . (3.48)
8ai bi Ŷi ui ui 2ai Ŷi u2i

Additionally, after some mathematical manipulations, the interval of ui can be derived

h 
β β1,i
as α2,i ,
2,i α1,i
.

β2,i β1,i
Based on the above discussions, by comparing α2,i with α1,i , the PDF of Ui can be

simplified as the following two cases:

β β
Case 1: If α2,i > α1,i holds, the intervals of u in Condition 1 and Condition 2 are given
h  2,i h 1,i 
β β1,i β2,i β2,i
by α1,i ,
2,i α1,i
and α2,i α1,i . Additionally, Condition 3 is valid, whereas Condition 4
,

is invalid. Therefore, the PDF of Ui is written as

   2 
X̂i 2 β1,i β1,i β1,i
α2,i − , ≤ ui <


8ai bi Yˆi ui α2,i α1,i





 X̂i β1,i β2,i


2bi , α1,i ≤ ui < α2,i
fUi (ui ) =  2  (3.49)
X̂i β2,i 2 β2,i β2,i
− α1,i ≤ ui ≤

 ,
8ai bi Yˆi ui α2,i α1,i






0, others.



β β
Case 2: If α2,i2,i
< α1,i holds, the intervals of ui in Condition 1 and Condition 2 are
h  1,i h 
β β2,i β1,i β2,i
given by α1,i ,
2,i α2,i
and α1,i α1,i , respectively. Moreover, Condition 4 is valid, while
,
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 40

Condition 3 is invalid. Accordingly, the PDF of Ui can be derived as

   2 
X̂i 2 β1,i β1,i β2,i
α2,i − , ≤ ui <


8ai bi Yˆi ui α2,i α2,i





 X̂i Ẑi β2,i β1,i

 , ≤ ui <
2ai Yˆi u2i α2,i α1,i
fUi (ui ) =  2  (3.50)
X̂i β2,i 2 β1,i β2,i
− α1,i ≤ ui ≤

 ,
8ai bi Yˆi ui α1,i α1,i






0, others.



[Link] PDF of Θ1,i

By using the PDF of cos Θ1,i , the PDF of Θ1,i can be derived as follows.

First of all, the CDF of Θ1,i can be derived as follows

FΘ1,i (θ1,i ) = Pr(Θ1,i ≤ θ1,i ) = Pr(arccos Ui ≤ θ1,i )

= 1 − Pr(Ui ≤ cos θ1,i ). (3.51)

Then, the PDF of Θ1,i is obtained as follows

∂FΘ1,i (θ1,i )
fΘ1,i (θ1,i ) = = sin θ1,i fUi (cos θ1,i ). (3.52)
∂θ1,i

Furthermore, the indoor localization system is considered in this chapter, where the

RISs are supposed to be mounted on the wall. Hence it is formulated as Θ1,i ∈ (0, π).

Thus, Θ1,i decreases monotonically with Ui . According to the PDF of Ui in the afore-

mentioned two cases, the PDF of Θ1,i is derived in the following.

β2,i β1,i
Case 1: When α2,i > α1,i , according to (3.49), the PDF of Θ1,i can be derived, which

is given by (3.53).

β2,i β1,i
Case 2: When α2,i < α1,i , the PDF of Θ1,i is derived based on (3.50), which is given

by (3.54).
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 41

  2 
X̂i sin θ1,i β2,i 2 β β

 − α1,i , arccos α2,i ≤ θ1,i ≤ arccos α2,i


 8ai bi Yˆi cos θ1,i 1,i 2,i

 X̂i sin θ1,i β β
fΘ1,i (θ1,i ) = 2bi , arccos α2,i
2,i
< θ1,i ≤ arccos α1,i
1,i (3.53)
h i
X̂i sin θ1,i β1,i 2 β1,i β1,i

α2,i 2 − ( cos θ1,i ) , arccos < θ1,i ≤ arccos


8ai bi Yˆi α1,i α2,i



0, others.


[Link] PDF of Θ̃1,i

By using the CDF and PDF of Θ1,i in (3.51), (3.53) and (3.54), the CDF and PDF of

Θ̃1,i can be calculated as follows.

Firstly, based on Θ̂1,i = Θ1,i + Θ̃1,i , the CDF of Θ̃1,i can be derived as follows

FΘ̃1,i (θ̃1,i ) = Pr(Θ̃1,i ≤ θ̃1,i ) = Pr(Θ̂1,i − Θ1,i ≤ θ̃1,i )

= Pr(Θ1,i ≥ Θ̂1,i − θ̃1,i )

= 1 − Pr(Θ1,i ≤ Θ̂1,i − θ̃1,i ). (3.55)

Then, based on the above two cases of fΘ1,i (θ1,i ) in (3.53) and (3.54), the PDF of

Θ̃1,i can be derived accordingly by using (3.55).

β2,i β1,i
Case 1: When α2,i > α1,i , by using (3.53), the PDF of Θ̃1,i is calculated as (3.56).
β2,i β1,i
Case 2: When α2,i < α1,i , the PDF of Θ̃1,i can be derived by using (3.54), which is given

by (3.57).

To facilitate the error analysis, this chapter aims to derive the approximation of

  2 
X̂i sin θ1,i β2,i 2 β β
 − α1,i , arccos α2,i ≤ θ1,i ≤ arccos α1,i
8ai bi Yˆi
 cos θ1,i 1,i 1,i




 X̂ Ẑ sin θ
i i 1,i β β
fΘ1,i (θ1,i ) = , arccos α1,i ≤ θ1,i < arccos α2,i (3.54)
h 2ai Yˆi (cos θ1,i )i2 1,i 2,i
X̂i sin θ1,i β β2,i β1,i

α2,i 2 − ( cos1,i 2
θ1,i ) , arccos ≤ θ1,i < arccos


8ai bi Yˆi α2,i α2,i




0, others.
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 42

∂FΘ̃1,i (θ̃1,i )
fΘ̃1,i (θ̃1,i ) = = fΘ1,i (Θ̂1,i − θ̃1,i )
∂ θ̃1,i
 h i
X̂i sin(Θ̂1,i −θ̃1,i ) β1,i β β
 α2,i 2 − ( )2 , Θ̂1,i − arccos α1,i ≤ θ̃1,i < Θ̂1,i − arccos α1,i
8ai bi Yˆi cos(Θ̂1,i −θ̃1,i )

 2,i 1,i

X̂i sin(Θ̂1,i −θ̃1,i ) β β

, Θ̂1,i − arccos α1,i ≤ θ̃1,i < Θ̂1,i − arccos α2,i


2bi 1,i 2,i
=  2 
X̂i sin(Θ̂1,i −θ̃1,i ) β2,i β β
− α1,i 2 , Θ̂1,i − arccos α2,i ≤ θ̃1,i ≤ Θ̂1,i − arccos α2,i


8a b Yˆ


 i i i cos(Θ̂ −θ̃ ) 1,i 1,i 2,i 1,i

0, others.

(3.56)

∂Fθ̃1,i (θ̃1,i )
fΘ̃1,i (θ̃1,i ) = = fΘ1,i (Θ̂1,i − θ̃1,i )
∂ θ̃1,i
   2 
X̂i sin(Θ̂1,i −θ̃1,i ) β1,i β β

 α2,i 2 − , Θ̂1,i − arccos α1,i ≤ θ̃1,i < Θ̂1,i − arccos α2,i


 8ai bi Yˆi cos(Θ̂1,i −θ̃1,i ) 2,i 2,i

 X̂i Ẑi sin(Θ̂1,i −θ̃1,i ) β β
, Θ̂1,i − arccos α2,i ≤ θ̃1,i < Θ̂1,i − arccos α1,i

=  2ai Yˆi [cos(Θ̂1,i −θ̃1,i )]2 2,i 1,i
2
X̂i sin(Θ̂1,i −θ̃1,i ) β2,i β β

− α1,i 2 , Θ̂1,i − arccos α1,i ≤ θ̃1,i ≤ Θ̂1,i − arccos α2,i


8ai bi Yˆi cos(Θ̂ −θ̃ )
 1,i 1,i
1,i 1,i



0, others.

(3.57)

fΘ̃1,i (θ̃1,i ). As it is formulated as assumed that the estimation error is very small, it is

formulated as sin θ̃1,i ≈ θ̃1,i and cos θ̃1,i ≈ 1, leading to

sin(Θ̂1,i − θ̃1,i ) ≈ sin Θ̂1,i − θ̃1,i cos Θ̂1,i ,

cos(Θ̂1,i − θ̃1,i ) ≈ cos Θ̂1,i + θ̃1,i sin Θ̂1,i . (3.58)

Using the approximations in (3.58), the approximation of fΘ̃1,i (θ̃1,i ) can be derived

according to the above two cases. Furthermore, by denoting sin Θ̂1,i = V̂i , cos Θ̂1,i = Ûi ,
β β β
B1,i = Θ̂1,i − arccos α1,i
2,i
, B2,i = Θ̂1,i − arccos α1,i
1,i
, B3,i = Θ̂1,i − arccos α2,i
2,i
, B4,i =
β
Θ̂1,i − arccos α2,i
1,i
, the expression could be further simplified in the following.

β2,i β1,i
Case 1: When α2,i > α1,i , based on (3.56), the approximation of fΘ̃1,i (θ̃1,i ) can be

derived, which is written as (3.59).

Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 43

β2,i β1,i
Case 2: When α2,i < α1,i , as it is formulated as (3.57), the approximation of fΘ̃1,i (θ̃1,i )

can be derived as (3.60).

3.5 Variance of Angle Estimation Error

This section aims to calculate the variance of Φ̃1,i and Θ̃1,i by using the PDF in the

above section, which will be used for the 3D position estimation in the next section.

   
X̂i (Vˆi −θ̃1,i Ûi )

β1,i 2

 α2,i −2 , B1,i ≤ θ̃1,i < B2,i
8ai bi Yˆi Ûi +θ̃1,i Vˆi



X̂i (Vˆi −θ̃1,i Ûi )


, B2,i ≤ θ̃1,i < B3,i

fΘ̃1,i (θ̃1,i ) ≈  2
2bi  (3.59)
 X̂i (Vˆi −θ̃1,i Ûi ) β2,i
− α1,i 2 , B3,i ≤ θ̃1,i ≤ B4,i


8ai bi Yˆi Ûi +θ̃1,i Vˆ

i



0, others.


  2 
X̂i (Vˆi −θ̃1,i Ûi )

2 β1,i

 α2,i − , B1,i ≤ θ̃1,i < B3,i


 8ai bi Yˆi Ûi +θ̃1,i Vˆi
X̂i Ẑi (Vˆi −θ̃1,i Ûi )


, B3,i ≤ θ̃1,i < B2,i

fθ̃1,i (θ̃1,i ) ≈  2aYˆi (Ûi +θ̃1,i Vˆi )2 (3.60)
 X̂i (Vˆi −θ̃1,i Ûi ) β2,i
 2
− α1,i 2 , B2,i ≤ θ̃1,i ≤ B4,i


8a b Yˆ Û +θ̃ Vˆ

i i i i 1,i i



0, others.

Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 44

3.5.1 Variance of Φ̃1,i

This subsection provides the variance expression of Φ̃1,i . Based on the PDF of Φ̃1,i in

(3.31), the variance of Φ̃1,i can be calculated as

D(ϕ̃1,i )

= E{ϕ̃21,i } − (E{ϕ̃1,i })2

! Φ̂1,i −a2,i
1 ϕ̃31,i ϕ̃41,i
= cos Φ̂1,i + sin Φ̂1,i
2ai 3 4
Φ̂1,i −a1,i
 2
! Φ̂1,i −a2,i
1  ϕ̃2 ϕ̃3
 1,i cos Φ̂1,i + 1,i sin Φ̂1,i

−  , (3.61)
4a2i  2 3 
Φ̂1,i −a1,i

x1
where f (x) = f (x1 ) − f (x2 ) and E{ϕ̃1,i } denotes the expectation of ϕ̃1,i .
x2

3.5.2 Variance of Θ̃1,i

This subsection aims to derive the variance of Θ̃1,i . Different from the PDF of Φ̃1,i , the

PDF of Θ̃1,i is more complicated. Therefore, analyze the variance should be analyzed

according to the aforementioned two cases as follows.

β2,i β1,i
Case 1: When α2,i > α1,i , based on the PDF of θ̃1,i in (3.59), the variance of Θ̃1,i is

derived as
 2
Z Z
2
D(θ̃1,i ) = θ̃1,i fΘ̃1,i (θ̃1,i )dθ̃1,i − θ̃1,i fΘ̃1,i (θ̃1,i )dθ̃1,i  . (3.62)
 
| {z } | {z }
D1,i D2,i

For D1,i in (3.62), it is divided into three different non-zero intervals that can be

expressed as

D1,i = D11,i + D12,i + D13,i , (3.63)

Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 45

where D11,i , D12,i and D13,i are the integral expressions in the intervals of [B1,i , B2,i ),

[B2,i , B3,i ) and [B3,i , B4,i ], respectively. The expressions of D11,i , D12,i and D13,i are

given in Appendix A.1.

For D2,i in (3.62), it is the expectation of θ̃1,i , which is denoted by E(θ̃1,i ). It can be

divided into three different non-zero intervals that are given by

D2,i = D21,i + D22,i + D23,i , (3.64)

where D21,i , D22,i and D23,i are the integral expressions in the intervals of [B1,i , B2,i ),

[B2,i , B3,i ) and [B3,i , B4,i ], respectively. The expressions of D21,i , D22,i and D23,i are

given in Appendix A.2.

β2,i β1,i
Case 2: When α2,i < α1,i , according to the PDF of Θ̃1,i in (3.60), the variance of Θ̃1,i

can be calculated as
 2
Z Z
D′ (θ̃1,i ) = 2
θ̃1,i fΘ̃1,i (θ̃1,i )dθ̃1,i − θ̃1,i fΘ̃1,i (θ̃1,i )dθ̃1,i  . (3.65)
 
| {z } | {z }
′
D1,i ′
D2,i

′ , it is divided into three different non-zero intervals that are given by

For D1,i

′ ′ ′ ′
D1,i = D11,i + D12,i + D13,i , (3.66)

′ , D′
where D11,i ′
12,i and D13,i are the integral expressions in the intervals of [B1,i , B3,i ),
′ , D′
[B3,i , B2,i ) and [B2,i , B4,i ]. The expressions of D11,i ′
12,i and D13,i are given in Appendix

A.3.

′ in (3.65), it is the expectation of θ̃ , which is denoted by E(θ̃ ). It is also

For D2,i 1,i 1,i

divided into three different non-zero intervals that are given by

′ ′ ′ ′
D2,i = D21,i + D22,i + D23,i , (3.67)
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 46

′ , D′
where D21,i ′
22,i and D23,i are the integral expressions in the intervals of [B1,i , B3,i ),
′ , D′
[B3,i , B2,i ) and [B2,i , B4,i ]. The expressions of D21,i ′
22,i and D23,i are given in Appendix

A.4.

3.6 3D Position Estimation

Using the estimated AoA in Section 3.3 and the variance of the estimation error in

Section 3.5, the expression of the estimation of the MU’s 3D position is derived. First,

it is formulated as the AoAs at the ith RIS Θ̂1,i and Φ̂1,i given in (3.21). For the sake of

illustration, all the estimated AoAs are collected in the following vectors:

Θ̂ = Θ + Θ̃,

Φ̂ = Φ + Φ̃, (3.68)

where

Θ̂ = [Θ̂L,1 , · · · , Θ̂L,I ], Φ̂ = [Φ̂L,1 , · · · , Φ̂L,I ],

Θ = [ΘL,1 , · · · , ΘL,I ], Φ = [ΦL,1 , · · · , ΦL,I ],

Θ̃ = [Θ̃L,1 , · · · , Θ̃L,I ], Φ̃ = [Φ̃L,1 , · · · , Φ̃L,I ]. (3.69)

In the existing works, for tractability, Θ̃ and Φ̃ are assumed to be the additive

complex Gaussian noise with zero mean. However, according to the angle estimation

error analysis in the previous section, the PDF of the estimation error Φ̃ should be

modeled as (3.31), while the PDF of the estimation error Θ̃ should be modeled as (3.59)

or (3.60). Furthermore, the covariance matrices can be written as

2 2
QΘ = diag[σΘ L,1
, · · · , σΘ L,I
],
2 2
QΦ = diag[σΦ L,1
, · · · , σΦ L,I
], (3.70)

2
where σΘ 2
and σΦ 2
denote the variance of Θ̃1,i and Φ̃i , respectively. σΦ can be
1,i 1,i 1,i
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 47

2
calculated according to (3.61), and σΘ can be calculated according to (3.62) or (3.65).
1,i

Accordingly, it is proposed to derive the closed-form expression of the MU’s estimated

3D position. First, the pseudolinear equations can be can derived [52] based on the

estimated AoAs as follows:

T T
ĝΘ s − ĝΘ
i i i
q ≈ −Θ̃1,i d1,i cos Φ̂1,i ,
T T
ĝΦ s − ĝΦ
i i i
q ≈ −Θ̃1,i d1,i , (3.71)

where d1,i denotes the distance between the ith RIS and the MU, and it is formulated

as

ĝΘi = [− cos Θ̂1,i , sin Θ̂1,i , 0]T ,

ĝΦi = [− sin Φ̂1,i sin Θ̂1,i , sin Φ̂1,i cos Θ̂1,i , − cos Φ̂1,i ]T . (3.72)

As a result, the following compact form of equations can be derived as

ĥ − Ĝq = Bz (3.73)

where

ĥ = [1T (ĜΘ ⊙ S)T , 1T (ĜΦ ⊙ S)T ]T , Ĝ = [ĜTΘ , ĜTΦ ]T , B = [B1 , B2 ]T , z = [Θ̃T , Φ̃T ]T ,

ĜΘ = [ĝΘ1 , · · · , ĝΘI ]T , ĜΦ = [ĝΦ1 , · · · , ĝΦI ]T , S = [s1 , · · · , sI ]T , B1 = [BΘ , O]T ,

B2 = [O, BΦ ]T , BΘ = −diag[dL,1 cos Φ̂L,1 , · · · , dL,I cos Φ̂L,I ]T , BΦ = −diag[dL,1 , · · · , dL,I ]T .

(3.74)

Based on (3.73), the TSWLS algorithm [52] is applied to derive the closed-form expression

of the MU’s estimated position, which is written as:

q̂ = (ĜT WĜ)−1 ĜT Wĥ, (3.75)

Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 48

where W = BQBT and Q = diag[QΘ , QΦ ]. The details of the derivation can be found

in [52].

3.7 Simulation Results

This section presents simulation results to validate the accuracy of the derivations and

approximations. In the simulation, a TDD mmWave MISO channel from the MU to the

RISs is considered. Moreover, the MU, the RISs are assumed to be placed in a 3D area.

The number of BS antennas is set as 400. The locations of four RISs are s1 = [2, 20, 3]T ,

s2 = [−12, 15, 50]T , s3 = [−10, −6, −8]T and s4 = [10, 6, −20]T 1 . It is assumed that the

inter-antenna spacing of UPA at the RISs is dr = λc /2. The RIS phase is set to all ones,

and the MU’s location is within a reasonable 3D range of [0, 30] meters in the x-axis,

[0, 50] meters in the y-axis, and [-20, 20] meters in the z-axis to ensure the effectiveness

of the Monte Carlo simulations. There are 10 paths between the MU and the RISs.

The following results are obtained by averaging over 10,000 random estimation error

realizations. Unless otherwise stated, it is assumed S = S1 × S2 = 64 × 64 for rotation

angle search grids, and the SNR is assumed to be 10 dB. The experiments are conducted

in the mmWave frequency band of 92 GHz. The localization accuracy is assessed in terms

of the MSE and the bias.

Fig. 3.3 and Fig. 3.4 illustrate the PDF of estimation errors. It is assumed that the

RIS size is Ny = Nz = 16. It can be observed from Fig. 3.3 and Fig. 3.4 that the derived

results match well with the simulation results, which verify the accuracy of the derived

results and confirm that the angle estimation error is non-Gaussian.

Fig. 3.5 presents the PDF of Θ̃1,i in cases of Gaussian and non-Gaussian. As depicted

in Fig. 3.5, the simulation results demonstrate that as the RIS size increases, the PDF of

error based on the Gaussian distribution (referred to as ‘Gaussian’) gradually approaches

the PDF of actual errors (referred to as ‘Simulation’). In contrast, the PDF derived from
1
In this chapter, the positions of the RISs are fixed for the proposed IoT mmWave localization system.
However, exploring the optimal geometric configuration of the RISs for the IoT mmWave localization
system would be an intriguing avenue for future work.
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 49

200
simulation
theory
180

160

140

Probability Density
120

100

80

60

40

20

0
-8 -6 -4 -2 0 2 4 6 8
10 -3

Figure 3.3: PDF of Θ̃1,i .

700
simulation
theory

600

500
Probability Density

400

300

200

100

0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
10 -3

Figure 3.4: PDF of Φ̃1,i .

the theoretical analysis (referred to as ‘Theory’) consistently approximates the true error

PDF across all scenarios. This finding indicates that the derived PDF offers a broader

applicability compared to the Gaussian distribution-based error model.

Fig. 3.6 and Fig. 3.7 display the variance σθ21,i of estimation error Θ̃1,i and the vari-

ance σϕ2 1,i of estimation error Φ̃1,i as the functions of RIS size Ny (Nz ), respectively. Fig.

3.6 and Fig. 3.7 display the variance σθ21,i , which show that the theoretical results coin-

cide with the simulation results, which validates the correctness of the derived results.

Moreover, it is observed that the variances of Θ̃1,i and Φ̃1,i decrease with the RIS size,

which means that increasing the number of antennas could improve the estimation accu-
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 50

1400
Simulation, N y =N z =100
Simulation, N y =N z =10
1200
Simulation, N y =N z =50
Gaussian, Ny =N z =100
1000 Gaussian, Ny =N z =10
Gaussian, Ny =N z =50
Theory, Ny =N z =100
800
Theory, Ny =N z =10
Theory, Ny =N z =50
600

400

200

0
-8 -6 -4 -2 0 2 4 6 8
10-3

Figure 3.5: PDF of Θ̃1,i in the cases of Gaussian and non-Gaussian.

racy.

10 -1
theory
simulation

10 -2
L,i
2

10 -3

-4
10
0 5 10 15 20
RIS size N y(N z)

Figure 3.6: Variance of Θ̃1,i versus RIS size.

Fig. 3.8 clearly demonstrates that the variances in both azimuth and elevation angles

decrease gradually with increasing SNR. However, in low SNR conditions, the impact

of noise on angle estimation is more pronounced. As the SNR increases, the two curves

approach a plateau, indicating that the influence of SNR on localization performance

diminishes progressively. This observation highlights the crucial role of SNR in angle

estimation accuracy, particularly in challenging environments characterized by low SNR

levels. In such conditions, the presence of noise can significantly affect the accuracy

of angle estimation. However, as the SNR improves, the impact of noise becomes less
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 51

theory
simulation

-2
10

L,i
2
10 -3

10 -4
0 2 4 6 8 10 12 14 16 18 20
RIS size N y(N z)

Figure 3.7: Variance of Φ̃1,i versus RIS size.

significant, leading to more stable and reliable angle estimates.

2
-10 L,i
2
L,i

-20
10*log 10( 2)

-30

-40

-50

-60

-70
-20 -15 -10 -5 0 5 10 15 20 25 30
SNR(dB)

Figure 3.8: Variances of Φ̃1,i and Θ̃1,i versus SNR (dB).

Fig. 3.9 compares the MSE of the proposed framework aided by 2 RISs with that

aided by 3 RISs when the size of RISs increases from Ny = Nz = 1 to Ny = Nz = 20. As

shown in the figure, the MSE decreases with the RIS size as expected. Furthermore, it

is shown that the MSE of 2 RISs is larger than that of 3 RISs. It implies that increasing

the number of RISs can significantly improve the localization accuracy of the proposed

framework. Additionally, the figure illustrates that incorporating Gaussian AEE into

the TSWLS algorithm results in a smaller MSE, suggesting that the actual AEE are not

Gaussian and exhibit some discrepancies.

Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 52

0
TSWLS, 3 RISs
-5 Gaussian AEE, 2RISs
TSWLS, 2 RISs
-10 Geometry
-15

10log 10 (MSE)
-20

-25

-30

-35

-40

-45

-50
0 5 10 15 20
RIS size Ny (Nz)

Figure 3.9: Comparison of the proposed framework and geometry algorithm.

The results in Fig. 3.10 show a significant gap between the curves for Ny = Nz = 1

and Ny = Nz = 2. However, the improvement from Ny = Nz = 4 to Ny = Nz = 8 is rel-

atively small. This suggests that the impact of RIS size diminishes when Ny (Nz ) exceeds

4, providing insights for optimizing localization performance. By carefully balancing the

dimensions of RIS antennas and the number of RISs, localization can be enhanced while

reducing costs and energy consumption, contributing to greener IoT deployments. The

-15
Ny = N z = 1
-20 Ny = N z = 2
Ny = N z = 4
-25 Ny = N z = 8

-30
10log (MSE)

-35
10

-40

-45

-50

-55

-60
2 2.5 3 3.5 4
Number of RISs

Figure 3.10: MSE of the proposed framework versus RISs number.

simulation results in Fig. 3.11 and Fig. 4.7 illustrate the MSE and the bias performance

of three algorithms under varying SNR conditions. As the SNR increases, all algorithms

exhibit improved localization performance, particularly in low SNR scenarios. However,

beyond a certain threshold, typically around 20dB to 30dB, the impact of noise becomes
Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 53

10
Geometry
TSWLS
0 WLS

-10

10log 10(MSE)
-20

-30

-40

-50

-60
-20 -15 -10 -5 0 5 10 15 20 25 30
SNR(dB)

Figure 3.11: The MSE comparison of the proposed framework with other algo-
rithms versus SNR (dB).

less significant, resulting in negligible changes in the MSE and the bias for all three algo-

rithms. Notably, the geometry algorithm, used as a benchmark, demonstrates relatively

consistent performance throughout.

As shown in Fig. 3.11, when comparing the performance in terms of MSE, the geom-

etry algorithm, which relies on geometric relationships for localization estimates, shows

approximately 10 times higher MSE compared to the WLS and TSWLS algorithms. On

the other hand, the TSWLS algorithm consistently outperforms the WLS algorithm,

achieving approximately 5 times lower MSE. These results strongly demonstrate the

superior performance of the proposed TSWLS algorithm within the framework. It out-

performs both the geometry algorithm and the WLS algorithm, providing more accurate

and reliable localization estimations, particularly in challenging SNR conditions. These

findings reinforce the effectiveness and superiority of the proposed TSWLS algorithm

within the framework. It significantly outperforms both the geometry algorithm and the

WLS algorithm, providing more accurate and reliable localization estimations, especially

in challenging SNR conditions.

Chapter 3. RIS-Aided Localization Algorithm Design: Tackling Non-Gaussian AEE 54

3.8 Summary

In this Chapter, a comprehensive framework is designed to analyze the angle estimation

error and design the 3D localization algorithm for the mmWave system. Firstly, the AoAs

at the RISs are estimated by applying the 2D-DFT algorithm. Based on the property

of the 2D-DFT algorithm, the angle estimation error was analyzed in terms of PDF.

Then, the intricate geometric expression of the error PDF is simplified by employing the

first-order linear approximation of triangle functions. Besides, the variance expression

of the error is given by using the error PDF. Finally, the TSWLS algorithm is applied

to estimate the 3D position of the MU using the estimated AoAs and the obtained

non-Gaussian variance.
Chapter 4

RIS-Aided Localization
Algorithm Design: Leveraging
AoA and TDoA

4.1 Introduction

Research in RIS-aided localization systems has significantly advanced. However, most

of the existing researches investigating RIS-aided localization mainly based on only one

channel parameter, such as AoAs. However, designing the corresponding localization

algorithm based on AoA alone will actually waste the good localization performance

of RIS [53], and the corresponding localization accuracy will not be high enough [54].

Therefore, it is essential to consider integrating AoA and TDoA to design the localization

algorithm.

However, according to the classical localization literature like [55], TDoA errors often

follow a Gaussian distribution due to the nature of time measurement techniques, which

typically involve averaging multiple signal observations to reduce randomness, leading

to a central limit theorem effect. Hence, it remains another non-trivial challenge to

55
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 56

design the localization algorithm. To be specific, incorporating non-Gaussian AEE and

Gaussian TDoA estimation error into algorithm design significantly complicates the pro-

cess due to complex mathematical transformations and heightened optimization require-

ments. Furthermore, it is another non-trivial difficulty to analyze the localization per-

formance and analyze the performance. To be specific, traditionally deriving the CRLB

in the context of non-Gaussian AEE can be complex, making it difficult to gain insights

through comparisons between the CRLB and RMSE.

The WLS method [56] is an effective choice for handling both Gaussian and non-

Gaussian errors simultaneously due to its ability to assign different weights to different

estimation, thus accurately accommodating the varying nature and complexities of these

error distributions. However, it often suffers from limitations such as sensitivity to initial

value selection and potential convergence to local minima. Therefore, it is still a challenge

to design an advanced WLS based algorithm that both utilizes the AoA and the TDoA,

harnessing their different error properties to enhance localization accuracy.

For an effective performance analysis of the proposed algorithm, bias is adopted as

a new evaluation metric in this chapter. This choice is driven by the limitations of

CRLB in scenarios involving non-Gaussian AoA and Gaussian TDoA errors. The CRLB

assumes a likelihood function that is often inappropriate for non-Gaussian distributions,

rendering it inapplicable in such cases. Moreover, the sophisticated mathematical oper-

ations required for CRLB, like the computation and inversion of the Fisher Information,

are less practical for the analysis. In contrast, bias analysis focuses on calculating the

expectation of the estimator and comparing it with the true parameter value, thereby

simplifying the evaluation process. It also offers greater flexibility in handling vari-

ous distribution types and nonlinear effects, making it a more intuitive and adaptable

method for evaluating estimator performance. However, analyzing localization perfor-

mance through bias analysis remains challenging due to the inherently complex nature

of WLS, which often involves non-linear relationships and varying error variances across

data points. This complexity introduces difficulties in accurately modeling and quanti-
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 57

fying the bias, especially since traditional linear assumptions are not applicable in such

scenarios.

To address the above-mentioned challenges, this chapter designs a mWLS localization

algorithm to enhance localization accuracy, aiming at tackling the non-Gaussian AEE.

This algorithm jointly utilizes the AoAs with their associated Non-Gaussian estimation

errors and the TDoAs with Gaussian estimation errors, enabling the derivation of the

closed-form expression for the 3D position of the MU. Additionally, the expression of the

expectation of the error is decomposed to simplify the calculation and conduct the bias

analysis to assess the performance of the proposed system. The main contributions can

be summarized as:

1) This chapter employ a two-step localization scheme for IoT mmWave systems

aided by multiple RISs. In the initial step, this chapter estimate the AoAs and

TDoA, highlighting that the AoA errors exhibit non-Gaussian properties while

TDoA errors are Gaussian. This chapter also define the geometric relationship of

the MU’s position. Then, the uniquely designed mWLS algorithm leverages both

types of errors, capitalizing on their distinct properties, to accurately determine

the MU’s position based on the estimated parameters and their distinct covariance.

2) To rigorously evaluate the proposed localization algorithm, this chapter undertake

a thorough analysis to deduce the theoretical estimation bias, uniquely factoring in

both Gaussian and non-Gaussian errors. By decomposing matrices that encompass

AoA and TDoA estimation into true values and estimation errors, this chapter

ascertain the preliminary estimation error and its theoretical bias. Using this

information, this chapter further determine the refining estimation’s bias. This

nuanced approach, more streamlined than traditional CRLB, clearly highlights the

accuracy of the localization algorithm.

3) Simulation results are provided to evaluate the effectiveness of the proposed local-

ization algorithm for non-Gaussian AEE. Besides, these results not only demon-
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 58

strate the superior performance of the proposed algorithm, but also validate the

applicability of the employed bias analysis for non-Gaussian AEE. Finally, the

derived bias results are in good agreement with the simulation results, indicat-

ing that the performance of the proposed algorithm in the real-world applications

aligns with theoretical expectations.

4.2 System Model

As shown in Fig. 4.1, this chapter considers an RIS-aided mmWave localization system,

where an MU sends pilot signals to the BS to locate the MU with the assistance of

multiple RISs. RISs could support high data rate while maintaining low costs and

energy consumption. Besides, RISs can constructively reflect the signal from the BS to

users. The BS is equipped with a ULA of Nb antennas, and the MU is equipped with a

single antenna. Moreover, there are M RISs, and each RIS is a UPA 1 .

The BS is placed parallel to the x-axis with the center located at p = [xp , yp , zp ]T . The

i-th RIS is placed parallel to the y-o-z plane with its center located at si = [xi , yi , zi ]T ,

i = 1, 2, · · · , M . The UPA-based RIS has Ny,z = Ny × Nz reflecting elements, where Ny

and Nz denote the numbers of reflecting elements along the y-axis and z-axis, respectively.

The true position of the MU is q = [xq , yq , zq ]T and is assumed to be placed parallel to

the x-o-y plane. The estimated location of the MU is q̂ = [x̂q , ŷq , ẑq ]T . Generally, once

the RISs and BS have been deployed, the coordinates si and p are known and invariant.

4.2.1 Wireless Channel Model

This subsection presents the channel model of the system under consideration. It is

important to note that the model includes two types of communication links character-

ized by distinct channel parameters: the direct link from the MU to the BS, and the

reflecting links via the RISs.

1
By installing GPS receivers on devices such as BS and RIS, microsecond time accuracy can be
achieved, thus ensuring network-wide synchronization.
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 59

th RIS

...
th RIS

BS

MU

Figure 4.1: The system model of multiple RISs aided localization systems.

s =[x
For the direct link from the MU to the BS, it is assumed that ]T number of propa-
,y ,zthe i i i q

gation paths between the BS and the MU is M , the azimuth AoA of the m-th path from

the MU to the BS is 0 ≤ θU a,m ≤ π. The array response vector at the BS of the m-th

path can be expressed as

q =[xq,yq,zq]T

−j2π(Nb −1)db sin θU a,m

aU a (θBa,m ) = [1, · · · , e λc ]T , (4.1)

where db and λc denote the distance between the antennas of the BS and the carrier

wavelength, respectively. The channel response of the direct link is given by

M
X
hBU = δm aU a (θU a,m ), (4.2)
m=1

where δm denotes the complex channel gain of the m-th path.

For the reflecting links between the MU and the BS via RISs, they are decomposed
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 60

into two sub-links: the MU-RISs link and the RISs-BS link.

For the MU-RISs link, it is assumed that the number of propagation paths between

the MU and the i-th RIS is Ni , the AoA of the ni -th path from the MU to the i-th RIS

can be decomposed into the elevation angle 0 ≤ θRa,ni ≤ π in the vertical direction, and

the azimuth angle 0 ≤ ϕRa,ni ≤ π in the horizontal direction, respectively. The array

response vector at the i-th RIS of the ni -th path is written as

(e) (a)
aRa (θRa,ni , ϕRa,ni ) = aRa (ϕRa,ni ) ⊗ aRa (θRa,ni , ϕRa,ni ), (4.3)

where ⊗ denotes the Kronecker product. Moreover, it is formulated as

−j2π(Mx −1)dr sin ϕRa,n

(e) i
aRa (ϕRa,ni ) = [1, · · · , e λc ]T ,
−j2π(Mz −1)dr cos θRa,n cos ϕRa,n
(e) i i
aRa (θRa,ni , ϕRa,ni ) = [1, · · · , e λc ]T , (4.4)

where dr denotes the distance between the elements of the RISs. The channel of MU-RISs

link can be modeled as

Ni
X
hi = αni aRa (θRa,ni , ϕRa,ni )
ni =1

= αM R,i aRa (θM R,i , ϕM R,i )

| {z }
LoS
Ni
X
+ αni aRa (θRa,ni , ϕRa,ni ), (4.5)
ni =2
| {z }
N LoS

where αni denotes the complex channel gain of ni -th path. Moreover, αM R,i , θM R,i , and

ϕM R,i denote the complex channel gain, the elevation AoA, and the azimuth AoA of the

line-of-sight (LoS) path, respectively. As shown in (4.5), channel components of hi can

be categorized into two types, namely LoS and NLoS. LoS path component is the direct

path between the BS and the MU, non-line-of-sight (NLoS) path component consists of

the paths between the RISs and the MU reflected by scatters, e.g., walls, human bodies.
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 61

For the RISs-BS link, it is assumed that the number of propagation paths between

the BS and the i-th RIS is Ji . The AoD of the ji -th path from the i-th RIS to the BS

can be decomposed into the azimuth angle 0 ≤ θRd,ji ≤ π in the horizontal direction

and the elevation angle 0 ≤ ϕRd,ji ≤ π in the vertical direction, respectively. The

array response vector is given by aRd (θRd,ji , ϕRd,ji ), which is similar to the expression of

aRa (θRa,ni , ϕRa,ni ) in (4.3).

Besides, define the azimuth AoA at the BS as 0 ≤ θRB,ji ≤ π, the array response

vector aRB (θRB,ji ) can be written as

−j2π(Nb −1)dr sin θRB,j

i
aRB (θRB,ji ) = [1, · · · , e λc ]T . (4.6)

By using the array response vector aRd (θRd,ji , ϕRd,ji ) and aRB (θRB,ji ), the channel matrix

of the i-th RIS with the BS can be formulated as

Ji
X
Hi = βji aRB (θRB,ji )aH
Ra (θRa,ni , ϕRa,ni ), (4.7)
ji =1

where βji denotes the channel gain of the ji -th path.

4.2.2 Received Signal Model

Accordingly, the received signal model is presented in this subsection. Let et,i ∈ CNy,z ×1

denote the phase shift vector of the i-th RIS at time slot t, which satisfies |[et,i ]ny,z | = 1

for 1 ≤ ny,z ≤ Ny,z . Accordingly, the received signal from the MU via the i-th RIS to

the BS at time slot t could be expressed as

M
X √
y(t) = ( Hi Diag(et,i )hi + hBU ) ps(t) + n(t), (4.8)
i=1

where s(t) denotes the transmitted pilot signal of the MU and n(t) ∈ CNb ×1 is AWGN

following the distribution of CN (0, δ 2 I). Moreover, p is the transmit power of the MU.
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 62

4.3 Two Step 3D Positioning Scheme

In this chapter, the framework of the classical two-step 3D localization scheme is adopted

to estimate the position of the MU. Specifically, the first step is to estimate the channel

parameters (AoAs and TDoAs) from the received signal y(t) in (4.8), and the second step

is to estimate the 3D position of the MU according to the estimated channel parameters.

4.3.1 Step I : Channel Parameters Estimation

Within this subsection, the channel parameters for localization are estimated. Specifi-

cally, the 2D-DFT algorithm [41] is employed to estimate the AoAs associated with the

RISs. Simultaneously, the TDoAs at the RISs are estimated through the transmission

of a pilot signal via each RIS to the BS. The RIS phase is set to all ones during this

process. Such a procedure facilitates the computation of the ToAs for both the MU-

RIS and MU-RIS-BS links. Thereafter, a geometric relationship correlating the channel

parameters with the 3D coordinates of the MU is described. In the concluding part of

this section, a systematic modeling of the estimation, inclusive of its inherent error, is

presented.

[Link] AoA Estimation

According to [57], by employing the 2D-DFT algorithm, the AoAs at the RISs is esti-

mated. Specifically, the estimated azimuth and elevation AoAs at the i-th RIS of the LoS

path. Specifically, in the mmWave band, the contributions of the NLoS path components

to the channel are minimal, given their rapid variations. Consequently, this chapter con-

centrates on estimating the AoA of the LoS path, which sufficiently informs the design of

a subsequent localization algorithm. However, it is important to note that in scenarios

where NLoS paths cannot be ignored, such as in dense urban environments or indoors

where reflections and obstructions are prevalent, NLoS components may significantly

affect the accuracy of localization. In these cases, the impact of NLoS paths must be

carefully considered and accounted for in the localization algorithm to ensure robust
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 63

performance. Consequently, this chapter concentrates on estimating the AoA of the LoS

path, which sufficiently informs the design of a subsequent localization algorithm. can

be given as
 
 λc bni λc ϖ̃2,i 
θ̂M R,i = arcsin  − ,
Nz dr 2πdr

,s
λc bni λc ϖ̃1,i λc qni λc ϖ̃2,i 2
ϕ̂M R,i = arccos(( − ) (1 − ( − ) )), (4.9)
Ny dr 2πdr Nz dr 2πdr

Ny dr cos θM R,i cos ϕM R,i Nz dr sin ϕM R,i

where bni = λc , qni = λc . Besides, ϖ̃1,i ∈ [−π/Ny , π/Ny ]

and ϖ̃2,i ∈ [−π/Nz , π/Nz ] denote the angle rotation parameters.

Consequently, the geometry relationship between the AoA information and the posi-

tion of the MU is introduced. For the MU-RIS-BS link, the available angle information

includes the AoAs (azimuth and elevation angles) at the RISs, the AODs (azimuth and

elevation angles) at the RISs, and the AoAs at the BS. For the MU-BS link, the angle

information for localization is the AoA at the BS. However, since only the azimuths of

the AoAs at the BS can be estimated due to the assumption of ULAs, the 3D coordinates

of the MU cannot be estimated without the elevation of the AoAs. Therefore, the AoAs

at the RISs are used for the localization estimation in this chapter. Moreover, as NLoS

path component usually varies fast and its weight to the channel is marginal, especially

in the mmWave band, this thesis is more interested in LoS path. Hence, this chapter

intends to estimate the path parameter (θM R,i , ϕM R,i ) of the LoS component from the

MU to the RISs, which can be used to derive the position of the MU.

It is assumed that the azimuth AoA θM R,i at the RIS is the angle between the

projection of the wave vector on the x-o-y plane and the y-axis as

xq − xi
θM R,i = arctan , (4.10)
yq − yi

where θM R,i ∈ (0, π).

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 64

The elevation AoA ϕM R,i at the RIS of the MU-RIS link is

zq − zi
ϕM R,i = arctan , (4.11)
sin θM R,i (xq − xi ) + cos θM R,i (yq − yi )

where ϕM R,i ∈ (−π/2, π/2).

[Link] TDoA Estimation

To deduce the TDoA at the RISs, an initial estimation of the ToA at the RIS is imper-

ative. This is achieved by focusing solely on the operation of the i-th RIS while deacti-

vating its counterparts. In this configuration, the MU dispatches a pilot signal directed

towards the BS, enabling the latter to ascertain the ToA corresponding to the MU-RIS-

BS linkage. Taking into consideration the fixed spatial coordinates of both the BS and

the RIS, the ToA affiliated with the RIS-BS link can be computed by dividing the spatial

separation by the universally constant speed of light. Building upon this foundation, the

ToA pertinent to the MU-RIS link is extrapolated by subtracting the previously com-

puted ToA of the RIS-BS link from that of the MU-RIS-BS link. Finally, the desired

TDoA at the RISs is derived by subtracting the MU-BS link’s ToA from the MU-RIS

link’s ToA.

Subsequently, the geometric correlation between the TDoA data and the spatial coor-

dinates of the MU is described. It should be noted that the disparity in link distances can

be efficiently derived by scaling the speed of light with the TDoAs. Thus, the subsequent

discussions will concentrate on these distance differentials, as they are synonymous with

the TDoAs in context.

Let RBU denote the true distance of the MU-BS link, which can be calculated as

q
RBU = (xq − xp )2 + (yq − yp )2 + (zq − zp )2 . (4.12)
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 65

Additionally, the true distance from the MU to the i-th RIS can be expressed as

q
RRU,i = (xq − xi )2 + (yq − yi )2 + (zq − zi )2 . (4.13)

Using RBU as the reference distance, the true distance difference between RRU,i and

RBU can be written as

RB,i = RRU,i − RBU . (4.14)

The true distance difference RB,i will be used in the following localization algorithm

design.

[Link] Estimation of AoAs and TDoAs

It is formulated as the estimated AoAs and TDoAs modeled as

θ̂M R,i = θM R,i + ni ,

ϕ̂M R,i = ϕM R,i + ωi ,

R̂B,i = RB,i + νi , (4.15)

where {θ̂M R,i , ϕ̂M R,i , R̂B,i } denote the estimation of {θM R,i , ϕM R,i , RB,i } and {ni , ωi , νi }

denote the estimation error.

For the sake of illustration, all the estimation localization parameters are collected

in the following vectors

θ̂r = θr + n, ϕ̂r = ϕr + ω, R̂B = RB + ν, (4.16)

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 66

where

θ̂r = Vec(θ̂M R ), θr = Vec(θM R ),

ϕ̂r = Vec(ϕ̂M R ), ϕr = Vec(ϕM R ),

R̂B = Vec(R̂B ), RB = Vec(RB ),

n = Vec(n), ω = Vec(ω), ν = Vec(ν), (4.17)

with Vec(x) represents the vector [x1 , x2 , . . . , xM ]T .

In prevailing literature, the parameters n, ω and ν are commonly assumed to repre-

sent additive zero-mean complex Gaussian noise for the sake of mathematical tractability

[56]. While estimating the TDoA at the RISs, both the time delay and the transmitted

pilot signal inherently exhibit Gaussian randomness, leading to maintain the assumption

that ν signifies the additive zero-mean complex Gaussian noise. However, as indicated

by [57], the PDFs associated with both elevation and azimuth errors display a non-

Gaussian property. Building on this, the variances of ni , ωi can be deduced from the

variance derivation presented in [57]. Moreover, the covariance matrices can be expressed

as

Qn = diag σn2 1 , . . . , σn2 M ,

Qω = diag σω2 1 , . . . , σω2 M ,

(4.18)

Qν = diag σν21 , . . . , σν2M ,

where σn2 i , σω2 i and σν2i denote the variance of ni , ωi and νi , respectively.

4.3.2 Step II: Proposed mWLS Algorithm

Drawing upon the estimated AoAs, TDoAs, and their accompanying non-Gaussian and

Gaussian estimation errors with distinct variances, an mWLS localization algorithm is

introduced in this subsection, yielding a closed-form solution for the MU’s position.
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 67

Initially, pseudolinear equations are formulated using the AoA and TDoA data from

the RISs. Based on these equations, and considering the AoAs’ non-Gaussian noises with

unique variances, a preliminary estimate of the MU’s position is obtained. Subsequent

equations are then constructed based on this preliminary estimate. The refined position

estimation is computed from these equations.

[Link] Pseudolinear equations based on the AoA estimation

First, derive the pseudolinear equations based on the geometry relationship of AoAs at

the i-th RIS. From (4.10), it is formulated as

cos θM R,i (xq − xi ) − sin θM R,i (yq − yi ) = 0. (4.19)

By replacing the true value θM R,i with the estimated θ̂M R,i , it is formulated as

ηθM R,i = cos θ̂M R,i (xq − xi ) − sin θ̂M R,i (yq − yi ), (4.20)

where ηθM R,i denotes the residual error of θM R,i due to the estimation error. For the

sake of analysis, let ğθM R,i = [− cos θ̂M R,i , sin θ̂M R,i , 0]T . Using the definitions of q and

si , (4.20) can be rewritten as

ηθM R,i = ğθTM R,i si − ğθTM R,i q. (4.21)

As the angle estimation algorithms, e.g., 2D-DFT, have good performance [51], [41], thus

it is reasonable to assume that the estimation error is very small. Hence, it is formulated

as sin(ni ) ≈ ni and cos(ni ) ≈ 1, where ni is the estimation error of θM R,i . Consequently,

it is formulated as the following approximations

sin θ̂M R,i = sin(θM R,i + ni ) ≈ sin θM R,i + ni cos θM R,i ,

cos θ̂M R,i = cos(θM R,i + ni ) ≈ cos θM R,i − ni sin θM R,i . (4.22)
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 68

By substituting (4.22) into (4.20), it is formulated as

ηθM R,i ≈ (cos θM R,i − ni sin θM R,i )(xq − xi )

− (sin θM R,i + ni cos θM R,i )(yq − yi )

= −ni RRU,i cos ϕM R,i . (4.23)

In (4.23), (4.19) and (xq −xi ) sin θM R,i +(yq −yi ) cos θM R,i = li1 +li2 = RRU,i cos ϕM R,i .

Next, using (4.11), it is formulated as

0 = cos ϕM R,i (zq − zi ) − sin ϕM R,i (sin θM R,i (xq − xi )

+ cos θM R,i (yq − yi )). (4.24)

By replacing {ϕM R,i , θM R,i } with {ϕ̂M R,i , θ̂M R,i }, it is formulated as

ηϕM R,i = cos ϕ̂M R,i (zq − zi ) − sin ϕ̂M R,i sin θ̂M R,i (xq − xi )

− sin ϕ̂M R,i cos θ̂M R,i (yq − yi ), (4.25)

where ηϕM R,i is the residual error of ϕM R,i . For simplicity, ğϕM R,i = [sin ϕ̂M R,i sin θ̂M R,i ,

sin ϕ̂M R,i cos θ̂M R,i , − cos ϕ̂M R,i ]T is defined. Then, by utilizing the definitions of q and

si , (4.25) can be expressed as

ηϕM R,i = ğϕTM R,i si − ğϕTM R,i q. (4.26)

It is assumed that the estimation error is very small [58], it is formulated as sin(ωi ) ≈ ωi

and cos(ωi ) ≈ 1, where ωi is the estimation error of ϕM R,i . As a result, it is formulated

as the following approximations as

sin ϕ̂M R,i = sin(ϕM R,i + ωi ) ≈ sin ϕM R,i + ωi cos ϕM R,i ,

cos ϕ̂M R,i = cos(ϕM R,i + ωi ) ≈ cos ϕM R,i − ωi sin ϕM R,i . (4.27)
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 69

Then, by substituting (4.27) into (4.25), and performing some mathematical manipula-

tions, it is formulated as

ηϕM R,i ≈ (cos ϕM R,i − ωi sin ϕM R,i )(zq − zi )

− (sin ϕM R,i + ωi cos ϕM R,i )

((xq − xi ) sin θM R,i + (yq − yi ) cos θM R,i )

= −ωi RRU,i . (4.28)

To derive (4.28), it is formulated as used (4.24) and ((zq −zi ) sin ϕM R,i +((xq −xi ) sin θM R,i +

(yq − yi ) cos θM R,i ) cos ϕM R,i ) = di1 + di2 = RRU,i .

Finally, using (4.21), (4.23), (4.26) and (4.28), it is formulated as

ğθTM R,i si − ğθTM R,i q ≈ −ni RRU,i cos ϕM R,i ,

ğϕTM R,i si − ğϕTM R,i q ≈ −ωi RRU,i . (4.29)

[Link] Pseudolinear equations based on the TDoA estimation

Then, the pseudolinear equations are derived based on the TDoA estimation. By taking

the square of both sides of (4.12), it is formulated as

2
RBU = Kq2 + Kp2 − 2xq xp − 2yq yp − 2zq zp , (4.30)

where Kq2 = x2q + yq2 + zq2 and Kp2 = x2p + yp2 + zp2 . Similarly, by taking the square of both

sides of (4.13), it is formulated as

2
RRU,i = Kq2 + Ki2 − 2xq xi − 2yq yi − 2zq zi , (4.31)
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 70

where Ki2 = x2i + yi2 + zi2 . Then, using (4.14), RRU,i = RB,i + RBU can be obtained.

Thus, it is formulated as

2
RRU,i = (RB,i + RBU )2 . (4.32)

By substituting (4.31) into (4.32) and expanding the right hand side of (4.32), (4.32) is

rewritten as

2 2
RB,i + 2RB,i RBU + RBU

= Kq2 + Ki2 − 2xq xi − 2yq yi − 2zq zi . (4.33)

Moreover, using (4.30) and (4.33), it is formulated as the following expression

2
RB,i + 2RB,i RBU

= Ki2 − Kp2 − 2xi,p xq − 2yi,p yq − 2zi,p zq , (4.34)

where xi,p = (xi − xp ), yi,p = (yi − yp ) and zi,p = (zi − zq ). In order to derive the

pseudolinear equations related to TDoA, (4.34) can be transformed into

− xi,p xq − yi,p yq − zi,p zq − RB,i RBU

1 1 1 2
= − Ki2 + Kp2 + RB,i . (4.35)
2 2 2

Additionally, for the sake of analysis, it is defined that

gt,i = −[xi,p , yi,p , zi,p , RB,i ]T ,

u = [xq , yq , zq , RBU ]T ,
1 1 1 2
hi = − Ki2 + Kp2 + RB,i . (4.36)
2 2 2

Accordingly, can derive the pseudolinear equation with respect to the TDoA can be
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 71

formulated as

T
hi − gt,i u = 0. (4.37)

As RB,i is estimated as R̂B,i by applying the TDoA estimation, by replacing RB,i with

R̂B,i , gt,i and hi in (4.37) are re-interpreted as ĝt,i and ĥi , given by

ĝt,i = −[xi,p , yi,p , zi,p , R̂B,i ]T ,

1 1 1 2
ĥi = − Ki2 + Kp2 + R̂B,i . (4.38)
2 2 2

Thus, according to [56], (4.37) can be derived as

T 1
ηt = ĥi − ĝt,i u = RRU,i νi + νi2 , (4.39)
2

where ηt and νi denote the residual error and the TDoA estimation error of RB,i in

(4.15), respectively. As νi ≪ RRU,i is always satisfied in practice, the second order term

on the right hand side of (4.39) can be neglected, and (4.39) can be approximated as

T
ĥi − ĝt,i u ≈ RRU,i νi . (4.40)

[Link] Preliminary Estimation

Then, the preliminary estimation based on the pseudolinear equations (4.29) and (4.40)

is derived, both of which contain the unknown location q.

Hence, by combining (4.29) and (4.40), the following compact form of equations can

be derived as

ĥr − Ĝr u = Br zr , (4.41)

where zr = [nT , ω T , ν T ]T and its covariance matrix is written as

Qr = diag(Qn , Qω , Qν ). (4.42)
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 72

On the left hand side of (4.41), it is formulated as

ĥr = [1T (Ğθr ⊙ S)T , 1T (Ğϕr ⊙ S)T , ĥTt ]T ,

Ĝr = [ĜTθr , ĜTϕr , ĜTt ]T , (4.43)

where

Ğθr = [ğθM R,1 , ğθM R,2 , · · · , ğθM R,M ]T ,

Ğϕr = [ğϕM R,1 , ğϕM R,2 , · · · , ğϕM R,M ]T ,

Ĝθr = [ĝθM R,1 , ĝθM R,2 , · · · , ĝθM R,M ]T ,

Ĝϕr = [ĝϕM R,1 , ĝϕM R,2 , · · · , ĝϕM R,M ]T ,

ĥt = [ĥ1 , ĥ2 , · · · , ĥM ]T , S = [s1 , s2 , · · · , sM ]T ,

Ĝt = [ĝt,1 , ĝt,2 , · · · , ĝt,M ]T ,

ĝθM R,i = [ğθTM R,i , 0]T , ĝϕM R,i = [ğϕTM R,i , 0]T . (4.44)

On the right hand side of (4.41), it is formulated as

Br = [Br,n , Br,ω , Br,ν ]T , (4.45)

where

Br,n = [BθM R,n , O, O]T ,

Br,ω = [O, BϕM R,ω , O]T ,

Br,ν = [O, O, Bt ]T ,

BθM R,n = −diag[RRU,1 cos ϕM R,1 , · · · , RRU,1 cos ϕM R,M ],

BϕM R,ω = −diag[RRU,1 , · · · , RRU,M ],

Bt = diag[RRU,1 , · · · , RRU,M ]. (4.46)

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 73

Based on (4.41), the WLS cost function can be formulated as

f (u) = (ĥr − Ĝr u)T Wr (ĥr − Ĝr u)

= ĥTr Wr ĥr − 2ĥTr Wr Ĝr u + uT ĜTr Wr Ĝr u, (4.47)

where Wr denotes the weight matrix of preliminary estimation.

To derive the estimation of u, the cost function f (u) in (4.47) should be minimized.
∂f (u)
By setting the first-order derivative of f (u) equal to zero, it is formulated as ∂u =

−2(ĜTr Wr ĥr ) + 2(ĜTr Wr Ĝr u) = 0. Hence, u is estimated as

ŭ = [x̆q , y̆q , z̆q , R̆BU ]T = (ĜTr Wr Ĝr )−1 ĜTr Wr ĥr . (4.48)

As the estimation error is correlated, the weight matrix Wr should be equal to the inverse

of the covariance matrix of the estimation error as

Wr = Ω−1 T T T
r , Ωr = E[Br zr zr Br ] = Br Qr Br , (4.49)

where Ωr denotes the covariance matrix of the estimation error. However, as shown

in (4.45) and (4.46), matrix Br contains the true distances {RRU,1 , RRU,2 , · · · , RRU,M },

which remain unknown. Therefore, the weight matrix is initialized by taking the esti-

mated distances as the approximation of true values. Then, the weight matrix can be

iteratively updated by the new estimations.

[Link] Refining Estimation

As shown in (4.12), RBU and q = [xq , yq , zq ]T are related, this relationship is used to

construct a new set of pseudolinear equations to derive the refining estimation. Con-

sidering the estimation error, the relationship between the MU’s preliminary estimated
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 74

position and the BS’s position is

 T
∆
(x̆q − xp )2 , (y̆ q − yp )2 , (z̆ q − zp )2 , R̆BU
2 = ĥ1 . (4.50)

Without the estimation error, the relationship between the MU’s position and the BS’s

position is represented by

 T
∆
(xq − xp )2 , (y q − yp )2 , (z q − zp )2 , RBU
2 = G1 ξ, (4.51)

where
 
I
 3×3 
G1 =  ,
11×3
 T
2 2
ξ = (xq − xp ) , (yq − yp ) , (zq − zp ) 2 = (q − p) ⊙ (q − p) . (4.52)

Using (4.50) and (4.51), the compact form of the pseudolinear equations can be derived

as

ĥ1 − G1 ξ = z1 , (4.53)

where z1 = [z11 , z12 , z13 , z14 ]T = [(x̆q −xp )2 , (y̆q −yp )2 , (z̆q −zp )2 , R̆BU
2 ]T −[(x −x )2 , (y −
q p q

yp )2 , (zq − zp )2 , RBU
2 ]T denotes the residual vector due to the estimation error. Denote

the localization estimation error of the preliminary estimation as ũ = [e1 , e2 , e3 , e4 ]T ,

i.e ., ŭ = u + ũ. Then, the elements of z1 can be represented as the functions of the

elements of error ũ, which can be expressed as

z11 = (x̆q − xp )2 − (xq − xp )2 ≈ 2(xq − xp )e1 ,

z12 = (y̆q − yp )2 − (yq − yp )2 ≈ 2(yq − yp )e2 ,

z13 = (z̆q − zp )2 − (zq − zp )2 ≈ 2(zq − zp )e3 ,

2 2
z14 = R̆BU − RBU ≈ 2RBU e4 . (4.54)
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 75

Based on (4.53), the WLS method can be applyed to estimate ξ, thus the WLS cost

function can be written as

f (ξ) = (ĥ1 − G1 ξ)T W1 (ĥ1 − G1 ξ)

= ĥT1 W1 ĥ1 − 2ĥT1 W1 G1 ξ + ξ T GT1 W1 G1 ξ, (4.55)

where W1 denotes the weight matrix of the refining estimation. To derive the esti-
∂f (ξ)
mation of ξ, f (ξ) in (4.55) should be minimized, leading to ∂ξ = −2(GT1 W1 ĥ1 ) +

2(GT1 W1 G1 ξ) = 0. Therefore, the estimation of ξ is given by

ξ̆ = (GT1 W1 G1 )−1 GT1 W1 ĥ1 . (4.56)

By defining Ω1 = E[z1 zT1 ] as the covariance matrix of z1 , it is formulated as W1 = Ω−1

1 .

Moreover, Ω1 can be further derived as

  
 p 

 

Ω1 = B1 (ĜTr Wr Ĝr )−1 B1 , B1 = 2diag u −   , (4.57)
0 

 

where the proof of (4.57) is given in Appendix A of [59]. Although B1 contains the true

values {xq , yq , zq , RBU }, it can be approximated as

  
 p 

 

B̂1 = 2diag ŭ −   . (4.58)
0 

 

Then, Ω1 can be approximated as Ω̂1 = B̂1 (ĜTr Wr Ĝr )−1 B̂1 . Hence, W1 can be esti-

mated as Ŵ1 = Ω̂−1 −1 T −1

1 = B̂1 (Ĝr Wr Ĝr )B̂1 , and ξ̆ is approximated as

ξ̆ ≈ (GT1 Ŵ1 G1 )−1 GT1 Ŵ1 ĥ1 . (4.59)

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 76

Finally, by using the estimation ŭ in (4.48) and ξ̆ in (4.59), the final closed-form esti-

mation of the position of MU can be expressed as:

q
q̂ = Π ξ̆ + p, Π = diag{sgn(ŭ(1 : 3) − p)}, (4.60)

where sgn(x) denotes the signum function.

4.4 Bias Analysis

In prevailing literature [60], the CRLB is typically applied to Gaussian-distributed data

for localization performance. However, in this chapter, the estimation errors of the

AoAs are non-Gaussian, posing challenges with CRLB analysis. To circumvent this, bias

analysis [55] is employed, a simpler measure of an estimator’s systematic error, offering

a direct insight into primary inaccuracies compared to RMSE. Bias analysis provides a

more intuitive understanding of an estimator’s performance in real-world scenarios where

assumptions of normality do not hold. This approach is particularly useful in complex

environments where signal reflections and non-linearities distort the statistical properties

of received signals. By focusing on the bias, we can tailor the estimation algorithms to

better suit the specific characteristics of the operational environment, thereby enhancing

the overall robustness and reliability of the localization system.

4.4.1 Decomposing Estimation Error Expression

Given that the bias is determined by computing the expectation of the estimation error,

and considering the error terms are interlinked within the expression, it becomes essential

to decompose the expression to accurately derive the expectation of the error.

As the estimation error is inevitable, Ĝr defined in (4.43) consists of two parts: the

true matrix Gr and the error matrix G̃r , i.e., Ĝr = Gr + G̃r . According to (4.43) and

(4.44), to derive the expressions of Gr and G̃r , ĝθM R,i is decomposed, ĝϕM R,i and ĝt,i

into the true values and the estimation error at first.

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 77

First, ĝθM R,i is decomposed into the true value and the estimation error. Using

the approximations in (4.22) and the definition of ĝθM R,i in (4.44), ĝθM R,i is approxi-

mated as ĝθM R,i ≈ [− cos θM R,i , sin θM R,i , 0, 0]T + ni [sin θM R,i , cos θM R,i , 0, 0]T . By defin-

ing ḡθM R,i = [− cos θM R,i , sin θM R,i , 0, 0]T and g1ni = [sin θM R,i , cos θM R,i , 0, 0]T , it is

formulated as

ĝθM R,i ≈ ḡθM R,i + ni g1ni , (4.61)

where ni g1ni denotes the error term.

Similarly, by utilizing the approximations in (4.22), (4.27) and the definition of ĝϕM R,i

in (4.44), ĝϕM R,i is approximated as

ĝϕM R,i ≈ [(sin ϕM R,i + ωi cos ϕM R,i )(sin θM R,i + ni cos θM R,i ),

(sin ϕM R,i + ωi cos ϕM R,i )(cos θM R,i − ni sin θM R,i ),

− (cos ϕM R,i − ωi sin ϕM R,i ), 0]T .

By ignoring the terms ωi ni cos ϕM R,i cos θM R,i and −ωi ni cos ϕM R,i sin θM R,i and defin-

ing ḡϕM R,i = [sin ϕM R,i sin θM R,i , sin ϕM R,i cos θM R,i , − cos ϕM R,i , 0]T , gωi = [cos ϕM R,i

sin θM R,i , cos ϕM R,i cos θM R,i , sin ϕM R,i , 0]T and g2ni = [sin ϕM R,i cos θM R,i , − sin ϕM R,i

sin θM R,i , 0, 0]T , it is formulated as

ĝϕM R,i ≈ ḡϕM R,i + ωi gωi + ni g2ni , (4.62)

where ωi gωi and ni g2ni denote the error terms, respectively.

Similarly, ĝt,i can be decomposed into the true value and the estimation error. Based

on the definition of gt,i in (4.36) and ĝt,i in (4.38), ĝt,i can be rewriten as ĝt,i =

−[xi,p , yi,p , zi,p , RB,i ]T − [0, 0, 0, νi ]T = gt,i + νi [0, 0, 0, −1]T . Letting gν = [0, 0, 0, −1]T ,

it is formulated as

ĝt,i = gt,i + νi gν , (4.63)

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 78

where νi gν denotes the error term.

Therefore, using (4.61), (4.62) and (4.63), Ĝr can be written as

Ĝr = Gr + G̃r . (4.64)

Moreover, using the definition of n, ω and ν in (4.17), G̃r can be written as

G̃r =Ḡn ⊙ [(n11×4 )T , (n11×4 )T , (n11×4 )T ]T

+ Ḡω ⊙ [(ω11×4 )T , (ω11×4 )T , (ω11×4 )T ]T

+ Ḡν ⊙ [(ν11×4 )T , (ν11×4 )T , (ν11×4 )T ]T , (4.65)

where

Ḡn = [GT1n , GT2n , O4×M ]T ,

Ḡω = [O4×M , GTω , O4×M ]T ,

Ḡν = [O4×M , O4×M , GTν ]T ,

G1n = [g1n1 , g1n2 , · · · , g1nM ]T ,

G2n = [g2n1 , g2n2 , · · · , g2nM ]T ,

Gω = [gω1 , gω2 , · · · , gωM ]T ,

Gν = [gν , gν , · · · , gν ]T . (4.66)

In addition, Gr is given by

Gr = [GTθr , GTϕr , GTt ]T , (4.67)

where

Gθr = [ḡθM R,1 , ḡθM R,2 , · · · , ḡθM R,M ]T ,

Gϕr = [ḡϕM R,1 , ḡϕM R,2 , · · · , ḡϕM R,M ]T ,

Gt = [gt,1 , gt,2 , · · · , gt,M ]T . (4.68)

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 79

4.4.2 Bias of Preliminary Estimation

This subsection aims to derive the bias of preliminary estimation ŭ. To derive the bias of

ŭ, the expression of estimation error of ŭ should be derived, which is given by ũ = ŭ − u.

Then, the bias of ŭ is given by taking the expectation of ũ, which is written as E(ũ).

Using (4.48) and the definition of u, ũ can be further derived as

ũ = ŭ − u = (ĜTr Wr Ĝr )−1 ĜTr Wr (ĥr − Ĝr u). (4.69)

Since it is formulated as ĥr − Ĝr u = Br zr in (4.41), ũ = (ĜTr Wr Ĝr )−1 ĜTr Wr Br zr . As it

is formulated as mentioned above, it is necessary to consider the second order term when

analyzing the bias [55]. However, it is formulated as used (4.40) to derive (4.41) rather

than (4.39), which ignores the second order term 12 νi2 . Therefore, to obtain the expression

of ũ and derive the bias of the preliminary estimation, Br zr should be extended to

Br zr + [0T , √12 ν T ]T ⊙ [0T , √12 ν T ]T , where 0 is a 2M × 1 column vector. Then, ũ can be

 

further derived as ũ = (ĜTr Wr Ĝr )−1 ĜTr Wr Br zr + [0T , √12 ν T ]T ⊙ [0T , √12 ν T ]T .
 

Letting η = [0T , √12 ν T ]T and P̂r = ĜTr Wr Ĝr , ũ can be rewritten as

ũ = P̂−1 T
r Ĝr Wr (Br zr + η ⊙ η). (4.70)

According to the definition of Ĝr in (4.64), P̂r can be further derived as

P̂r = (Gr + G̃r )T Wr (Gr + G̃r )

= GTr Wr Gr + G̃Tr Wr Gr + GTr Wr G̃r + G̃Tr Wr G̃r . (4.71)

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 80

Here, G̃Tr Wr G̃r can be neglected2 , thus P̂r can be approximated as

P̂r ≈ GTr Wr Gr + G̃Tr Wr Gr + GTr Wr G̃r = Pr + P̃r , (4.72)

where Pr = GTr Wr Gr and P̃r = G̃Tr Wr Gr + GTr Wr G̃r , respectively. However, if the

definition of P̂r in (4.72) is used to derive (4.70), the inverse matrix P̂−1
r = (Pr + P̃r )
−1

is complex and challenging to derive. Fortunately, the approximation of P̂−1

r can be

derived by utilizing the Newmann expansion [55] when the error level is small, which is

written as
 
P̂−1
r ≈ I − P−1
r P̃ r P−1
r . (4.73)

By substituting (4.73) and (4.64) into (4.70), ũ is derived as

 
ũ ≈ I − P−1
r P̃ r P−1 T
r (Gr + G̃r ) Wr (Br zr + η ⊙ η). (4.74)

For the sake of illustration, using Pr in (4.72), Hr = (GTr Wr Gr )−1 GTr Wr = P−1 T
r Gr Wr

are defined. Then, ũ can be represented as

ũ ≈ Hr Br zr + Hr (η ⊙ η) − P−1 −1
r P̃r Hr Br zr − Pr P̃r Hr (η ⊙ η)

+ P−1 T −1 T
r G̃r Wr Br zr + Pr G̃r Wr (η ⊙ η)

− P−1 −1 T −1 −1 T
r P̃r Pr G̃r Wr Br zr − Pr P̃r Pr G̃r Wr (η ⊙ η). (4.75)

For simplicity, the error terms in (4.75) which are higher than the second order,can be

ignored. As a result, using P̃r in (4.72), ũ is approximated as

ũ ≈ Hr Br zr + Hr (η ⊙ η)

− P−1 T −1 T
r G̃r Wr Gr Hr Br zr − Pr Gr Wr G̃r Hr Br zr

+ P−1 T
r G̃r Wr Br zr . (4.76)
2
If consider G̃Tr Wr G̃r when deriving the expression of ũ, it will be multiplied by Br zr , and the terms
including G̃Tr Wr G̃r in the final expression of ũ is higher than the second order, thus this term can be
neglected here.
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 81

As a result, the bias of preliminary estimation is written as

E[ũ] = E[Hr Br zr + Hr (η ⊙ η) − P−1 T −1 T

r G̃r Wr Gr Hr Br zr − Pr Gr Wr G̃r Hr Br zr

+ P−1 T
r G̃r Wr Br zr ]

= E 1 + E2 + E3 , (4.77)

where E1 = E[Hr Br zr ], E2 = E[Hr (η ⊙ η)] and E3 = E[−P−1 T

r G̃r Wr Gr Hr Br zr −

P−1 T −1 T
r Gr Wr G̃r Hr Br zr + Pr G̃r Wr Br zr ], the detailed derivations of which are given in

Appendix B.1.

4.4.3 Bias of Refining Estimation

In this subsection, it is aimed to derive the bias of q̂ of refining estimation. Similar to the

derivations of the bias of preliminary estimation, it is essential to derive the expression

of estimation error of ξ, and the bias of ξ can be derived by taking the expectation of

the estimation error of ξ. First, denote ξ̃ as the estimation error of ξ, e.g., ξ̃ = ξ̆ − ξ.

Then, using the definition of ξ̆ in (4.59), ξ̃ can be further derived as

ξ̃ = (GT1 Ŵ1 G1 )−1 GT1 Ŵ1 (ĥ1 − G1 ξ). (4.78)

According to (4.53), ξ̃ is derived as

ξ̃ = (GT1 Ŵ1 G1 )−1 GT1 Ŵ1 z1 . (4.79)

Using the derivations of (4.54), the definition of B1 in (4.57) and the definition of ũ

below (4.53), it is formulated as z1 = B1 ũ + ũ ⊙ ũ. Then, ξ̃ can be derived as

ξ̃ = (GT1 Ŵ1 G1 )−1 GT1 Ŵ1 (B1 ũ + ũ ⊙ ũ). (4.80)

By defining P̂1 = GT1 Ŵ1 G1 , it is formulated as ξ̃ = P̂−1 T

1 G1 Ŵ1 (B1 ũ+ ũ⊙ ũ). According

to the definition of Ŵ1 below (4.58), Ŵ1 is composed of the true value and the estimation
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 82

error, which are denoted by W1 and W̃1 , respectively. To further derive the expression

of ξ̃, the expressions of W1 and W̃1 should be derived, which are given by

Ŵ1 = W1 + W̃1 , W1 = B−1 −1

1 Pr B1 ,

W̃1 = B−1 −1 −1 −1
1 P̃r B1 − W1 B̃1 B1 − B̃1 B1 W1 , (4.81)

where B̃1 is given in (B.1). The details of deriving W1 and W̃1 are given in Appendix

B.2.

Similarly, based on the definition of P̂1 , P̂1 can be obtained as a summation of the

true value and the estimation error, which are denoted by P1 and P̃1 , respectively.

Hence, it is formulated as

P̂1 = P1 + P̃1 , P1 = GT1 W1 G1 , P̃1 = GT1 W̃1 G1 . (4.82)

Similar to (4.73), the approximation of P̂−1

1 can be derived as

P̂−1 −1 −1 −1 −1 −1
1 ≈ (I − P1 P̃1 )P1 = P1 − P1 P̃1 P1 . (4.83)

Therefore, using the approximation of P̂−1

1 in (4.83) and the definition of Ŵ1 , ξ̃ can be

derived as

ξ̃ ≈ (P−1 −1 −1 T
1 − P1 P̃1 P1 )G1 Ŵ1 (B1 ũ + ũ ⊙ ũ). (4.84)

For notation simplicity, let H1 = P−1 T

1 G1 W1 , it is formulated as

ξ̃ ≈ H1 (B1 ũ + ũ ⊙ ũ) + P−1 T

1 G1 W̃1 (B1 ũ + ũ ⊙ ũ)

− P−1
1 P̃1 H1 (B1 ũ + ũ ⊙ ũ)

− P−1 −1
1 P̃1 P1 G1 W̃1 (B1 ũ + ũ ⊙ ũ). (4.85)

By ignoring the error terms higher than the second order and substituting P̃1 in (4.82)
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 83

into (4.85), it is formulated as

ξ̃ ≈ H1 (B1 ũ + ũ ⊙ ũ) + P−1 T −1

1 G1 W̃1 B1 ũ − P1 P̃1 H1 B1 ũ. (4.86)

By assuming that P2 = I4×4 − G1 H1 , it is formulated as

ξ̃ ≈ H1 (B1 ũ + ũ ⊙ ũ) + P−1 T

1 G1 W̃1 P2 B1 ũ. (4.87)

Finally, by taking expectation of ξ̃, the bias is given by

E[ξ̃] = H1 B1 E[ũ] + H1 E[ũ ⊙ ũ] + P−1 T

1 G1 E[W̃1 P2 B1 ũ]. (4.88)

To further derive the bias, it is introduced Proposition 4. 1 as follows.

Proposition 4. 1 For the vector ã ∈ CM ×1 , the expectation of ã⊙ã can be expressed

as a column vector ca containing the diagonal elements of Ωa , which is the expectation

of the second order moment of ã.

Proof: Please see Appendix B.3. ■

Using Proposition 4. 1, E[ũ ⊙ ũ] can be derived as cu containing the diagonal

elements of Ωu . Furthermore, the details of deriving Ωu are given in Appendix A of [59].

Therefore, the bias of ξ̃ can be further derived as

E[ξ̃] = H1 (B1 E[ũ] + cu ) + P−1 T

1 G1 E[W̃1 P2 B1 ũ], (4.89)

where the derivations of E[W̃1 P2 B1 ũ] are given in Appendix B.4.

Then, the estimation error of MU’s position is defined as q̃ = q̂ − q. Then using the

definition of ξ in (4.52), ξ̆ can be derived as

ξ̆ = (q̂ − p) ⊙ (q̂ − p) = q̃ ⊙ q̃ + 2q̃ ⊙ (q − p) + ξ. (4.90)

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 84

By utilizing ξ̃ = ξ̆ − ξ, it is formulated as ξ̃ − q̃ ⊙ q̃ = 2q̃ ⊙ (q − p). Then, by assuming

that Bq q̃ = 2q̃ ⊙ (q − p), where Bq = 2diag{(q − p)}, it is formulated as the expression

of q̃ given by

q̃ = B−1
q (ξ̃ − q̃ ⊙ q̃). (4.91)

Then, the bias of refining estimation can be derived as

E[q̃] = E[B−1
q (ξ̃ − q̃ ⊙ q̃)]. (4.92)

Using Proposition 4. 1, it is formulated as E[q̃ ⊙ q̃] = cq , where cq is a column vector

formed by the diagonal elements of Ωq , and the details of deriving Ωq can be found in

Appendix B.5. Therefore, (4.92) can be further derived as

E[q̃] = B−1
q (E[ξ̃] − cq ), (4.93)

where E[ξ̃] is given in (4.89).

4.5 Simulation Results

This section presents simulation results to evaluate the performance of the proposed

localization algorithm aided by multiple RISs. Moreover, the MU, the BS, and the RISs

are assumed to be placed in a 3D area. The location of the BS is p = [10, 12, 12]T ,

while the locations of three RISs are s1 = [2, 20, 2]T , s2 = [−12, −16, 58]T , and s3 =

[−10, −10, 50]T , respectively. The phase shift matrix of the RIS is set to a unit matrix

(all ones). The MU’s location is assumed to be within a reasonable 3D range of [0, 30]

meters in the x-axis, [0, 50] meters in the y-axis, and [0, 60] meters in the z-axis, ensuring

effective simulation. The experiments are conducted in the millimeter-wave frequency

band of 92 GHz. The following results are obtained by averaging over 10,000 random

estimation error realizations. The localization accuracy is assessed in terms of the MSE

and the bias. The TDoA and AoA estimation error are assumed to follow the Gaussian

distribution. In the figures showing the simulation results, the log scale for the error
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 85

level is utilized to indicate the wide range of the levels tested.

-20

-22

n
-24

-26

10*log10( )
-28

-30

-32

-34

-36

-38

-40
5 10 15 20 25 30
SNR (dB)

Figure 4.2: Standard deviation of ni and ωi versus SNR (dB).

Fig.4.2 shows the standard deviations of ni and ωi as the functions of SNR (dB). It

is observed that the standard deviations decrease with SNR increasing, which validates

that improved communication setting is capable of enhancing angle estimation accuracy.

Besides, it is observed that the standard deviation σω is smaller than σn . This implies

that the estimation accuracy for ϕM R,i is superior to that of θM R,i . This phenomenon

is hypothesized to be due to the sequential estimation process in 2D-DFT. Initially,

ϕM R,i is estimated, followed by θM R,i . The sequential nature of this process leads to an

accumulation of errors, resulting in a greater estimation error for θM R,i .

-20

-22
n
-24

-26
10*log10( )

-28

-30

-32

-34

-36

-38

0 10 20 30 40 50 60 70
Ny (Nz)

Figure 4.3: Standard deviation of ni and ωi versus Ny (Nz ).

Fig.4.3 illustrates the standard deviations of ni and ωi as the functions of the RIS
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 86

size, denoted as Ny and Nz . Furthermore, it is shown that once Ny and Nz exceeds 32,

increasing the panel size of the RIS does not significantly improve the accuracy of angle

estimation. This indicates that there is a diminishing return on the benefits of increasing

the number of RIS elements. In practical communication scenarios, it may be sufficient

to maintain a certain number of RIS elements without continually expanding them.

-45

-50 2 RISs,Gaussian
2 RISs,non-Gaussian
3 RISs, non-Gaussian
-55 3 RISs,Gaussian
10*log10(MSE)

-60

-65

-70

-75

-80
5 10 15 20 25 30
SNR (dB)

Figure 4.4: The MSE comparison of the proposed algorithm in the cases of
Gaussian and non-Gaussian versus SNR (dB).

Fig. 4.4 clearly demonstrates the MSE comparisons in cases of Gaussian and non-

Gaussian AoA estimation errors. As shown in the figure, the MSE decreases with SNR

as expected. Furthermore, it is evaluated that the localization accuracy of the systems

employing 3 RISs is much higher than that of the systems employing 2 RISs, which

effectively demonstrates the importance of the number of the RISs.

Remark 1. Notably, with 3 RISs, the difference between the Gaussian and non-Gaussian

AoA estimation errors becomes negligible, suggesting that increasing the number of RISs

not only enhances overall accuracy but also mitigates the impact of non-Gaussian error

distributions.

More importantly, it can be observed from Fig. 4.4 that the MSE obtained from

the localization algorithm using practical non-Gaussian errors does not align with that

using Gaussian errors. This suggests that in designing practical localization algorithms,

it is preferable to base them on actual non-Gaussian errors. It is important to note

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 87

that algorithms designed using actual non-Gaussian errors are more adept at reflecting

practical error characteristics. This contrasts with algorithms based on Gaussian errors,

which typically rely on idealized assumptions of normal distribution, potentially lead-

ing to inaccuracies in practical applications. The inconsistency observed in the MSE

results between these two approaches underscores the gap between theoretical models

and actual scenarios. This further substantiates the advantage of designing an algorithm

based on non-Gaussian errors, as it can more accurately reflect and adapt to the error

characteristics encountered in practical applications.

-40
Geometry
-45 2 RISs, Prop.
3 RISs, Prop.
-50
3 RISs, WLS
10*log10(MSE)

-55

-60

-65

-70

-75

-80
5 10 15 20 25 30
SNR (dB)

Figure 4.5: The MSE comparison of the proposed algorithm with other algo-
rithms versus SNR (dB).

As shown in Fig.4.5, when comparing the performance in terms of MSE, the geom-

etry algorithm [60], which relies on geometric relationships for localization estimates,

shows approximately 10 times higher MSE compared to the WLS and the proposed algo-

rithms. On the other hand, the proposed algorithm consistently outperforms the WLS

algorithm, achieving approximately 5 times lower MSE. These results strongly demon-

strate the superior performance of the proposed algorithm within the framework. These

findings reinforce the effectiveness and superiority of the proposed algorithm within the

framework.

Based on the angle estimation expressions presented in (4.9), it is evident that the

estimation accuracy is influenced by the number of RIS elements. Accordingly, simula-

tion results are provided in Fig.4.6, comparing the MSE across various algorithms. The
Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 88

-50
Geometry
2 RISs, Prop.
-55 3 RISs, Prop.
3 RISs, WLS

-60

10*log10(MSE)
-65

-70

-75

-80
0 10 20 30 40 50 60 70
Ny (Nz)

Figure 4.6: The MSE comparison of the proposed algorithm with other algo-
rithms versus RIS size Ny (Nz ).

results, as depicted in Fig.4.6, reveal that the proposed algorithm demonstrates supe-

rior performance over other algorithms. Additionally, it observed that beyond a certain

threshold, specifically when Ny and Nz exceeds 16, the increase in the number of RIS

elements does not significantly enhance the precision of the localization system. Further-

more, the use of 3 RISs does not notably improve the localization accuracy compared

to a configuration with 2 RISs. This suggests a point of diminishing returns in terms of

localization accuracy relative to the number of the RISs and the number of RIS elements

employed. This phenomenon may be due to the fact that the optimization of the system

is already approaching its limits at lower RIS numbers, or that other factors (e.g., signal

interference, hardware limitations, etc.) are starting to dominate at higher RIS numbers.

Fig.4.7 and Fig.4.8 compare the bias of the proposed localization algorithm aided by

2 and 3 RIS, respectively, thereby validating the accuracy of the theoretical analysis on

the algorithm’s bias. Specifically, Fig.4.7 focuses on contrasting the bias in cases of Gaus-

sian and non-Gaussian AoA estimation errors. The simulation results, as illustrated in

Fig.4.7, align more closely with the theoretical outcomes for non-Gaussian AoA estima-

tion errors, rather than Gaussian ones. This alignment suggests that for a more accurate

depiction of localization performance in real-world wireless systems, it is preferable to

consider practical non-Gaussian AoA estimation errors in performance analyses. Addi-

Chapter 4. RIS-Aided Localization Algorithm Design: Leveraging AoA and TDoA 89

-26
2 RISs, Gaussian, Theo.
2 RISs, Prop., Theo.
-28
2 RISs, Prop., Simu.
3 RISs, Prop., Theo.
-30 3 RISs, Prop., Simu.
3 RISs, Gaussian, Theo.

10*log10(bias)
-32

-34

-36

-38

5 10 15 20 25 30
SNR (dB)

Figure 4.7: The bias comparison of the proposed algorithm in the cases of
Gaussian and non-Gaussian versus SNR (dB).

-35.5
2 RISs, Prop., Theo.
-36
2 RISs, Prop., Simu.
3 RISs, Prop., Theo.
-36.5
3 RISs, Prop., Simu.
-37
10*log10(bias)

-37.5

-38

-38.5

-39

-39.5

-40

-40.5
0 10 20 30 40 50 60 70
Ny (Nz)

Figure 4.8: The bias comparison of the proposed algorithm versus RIS size Ny
(Nz )

tionally, these findings underscore the advantage of the bias analysis approach, which

proves to be more tractable in practical scenarios.

4.6 Summary

This chapter investigated the algorithm design and analysis for RIS-aided localization

IoT system aided by multiple RISs, tackling the non-Gaussian AEE. The mWLS algo-

rithm is designed to derive the closed-form expression of the position of the MU by

leveraging the different error property of AoA and TDoA. Finally, the unique bias anal-

ysis is proposed to evaluate the performance of the proposed algorithm.

Chapter 5

RIS-Aided Localization Algorithm

Design: Employing Fingerprint

5.1 Introduction

Currently, wireless localization algorithms aided by the RIS have been studied by some

researchers, including two-step localization algorithms and fingerprint based algorithms

[61, 62]. However, two-step localization algorithms are sophisticated and require channel

estimations.

Without estimating the channel parameters, fingerprint based localization algorithms

directly predict the position by using the fingerprint (e.g., RSSI). For example, [62]

regarded RSSI as a type of fingerprint to predict the MUs aided by the RIS. Hence, this

chapter proposes to employ fingerprint-based method to estimate multi-user. However,

RSSI-based fingerprint localization algorithms may be unstable due to the fast fading

fluctuation. Recently, some researchers proposed to use the CSI as the fingerprint, due to

its potential to enhance the localization accuracy compared with RSSI [26, 63, 64]. The

authors of [63] proposed to use CSI as a type of fingerprint in a SISO Wi-Fi localization

system. The angle delay channel power matrix CSI fingerprint was adopted in [64] in a

90
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 91

massive MIMO orthogonal frequency-division multiplexing (OFDM) system to estimate

the 2D position of the MU. The authors of [26] proposed a CSI-based 3D localization

method for the MIMO-OFDM system using a CNN. However, due to the obstacles in

indoor environment, the direct links between the MUs and the AP may be blocked.

Hence, this chapter investigate the scenarios where the RIS is used to reconstruct an

alternative communication link between the MUs and the AP. Moreover, as the CSI

fingerprint of the cascaded AP-RIS-MU channel contains more features than the other

types of CSI fingerprint, this chapter propose a novel RCNR learning algorithm to extract

more features and protect the integrity of features.

Against the above background, the main contributions are summarized as follows:

1) For the fingerprint based mmWave localization system, this chapter proposes a new

type of fingerprint, STCRV, which consists of multipath channel characteristics.

The STCRV has more features than the other types of CSI fingerprint, e.g., the

AoA and the AoD at the RIS, and is closely related to the position of the MU.

Moreover, the STCRV can be used as an alternative fingerprint for localization

when the direct link between the MU and the AP is blocked by obstacles.

2) Utilizing the STCRV as the wireless localization fingerprint, this chapter propose

a novel RCNR learning algorithm to estimate the 3D positions of the MUs. The

RCNR algorithm can be used to study the mapping between these channel charac-

teristics and the position of the MU, thereby enabling precise localization. Specif-

ically, STCRV is processed by a data processing block at first. Then, a normal

convolution block is then used to extract the features of the output from the data

processing block. Consequently, four residual convolution blocks are utilized to

further extract more features caused by the cascaded AP-RIS-MU channel and

protect the integrity. Finally, the 3D position is estimated through a regression

block.

3) Simulation results are provided to evaluate the performance of the proposed STCRV
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 92

fingerprint and the RCNR learning algorithm. The proposed algorithm outper-

forms the CNN and the WKNN in terms of CDF.

5.2 System Model and Problem Formulation

This chapter consider an RIS-aided 3D massive TDD mmWave localization system,

where the MUs send pilot signals to the BS to locate the positions of the MUs aided by

an RIS. In addition, it is assumed that the direct channels between the BS and the MUs

are blocked by some obstacles, such as thick walls.

BS

Figure 5.1: RIS-aided multiuser localization system model.

As shown in Fig. 5.1, the BS is placed on the left side of the wall with the center

located at p = [xp , yp , zp ]T . Moreover, the BS is assumed to be equipped with a UPA

with Mx,z = Mx × Mz antennas, where Mx and Mz denote the numbers of antennas

along the x-axis and the z-axis, respectively. Additionally, the RIS is assumed to be

placed at the front side of the wall with the center located at s = [xs , ys , zs ]T . The

UPA-based RIS has Ny,z = Ny × Nz reflecting elements along the y-axis and the z-axis,

respectively. Furthermore, there are Nu MUs , each of which is equipped with a single

antenna. The MUs are located at ui = [xi , yi , zi ]T , i = 1, 2, · · · , Nu .

It is assumed that the number of propagation paths between the ith MU and the RIS

is Pi , the AoA of the pth path from the ith MU to the RIS can be decomposed into the
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 93

Figure 5.2: The structure of the RCNR learning algorithm.

elevation angle 0 ≤ θp,i ≤ π in the vertical direction, and the azimuth angle 0 ≤ ϕp,i ≤ π

in the horizontal direction. As a result, the array response vector at the RIS can be

expressed as [65]

(e) (a)
aRa (θp,i , ϕp,i ) = aRa (θp,i ) ⊗ aRa (θp,i , ϕp,i ), (5.1)

where ⊗ denotes the Kronecker product. Moreover, it is formulated as

−j2πdr cos θp,i −j2π(Ny −1)dr cos θp,i

(e)
aRa (θp,i ) = [1, e λc , ..., e λc ]T , (5.2)

and

−j2πdr sin θp,i cos ϕp,i −j2π(Nz −1)dr sin θp,i cos ϕp,i
(a)
aRa (θp,i , ϕp,i ) =[1, e λc ,··· ,e λc ]T , (5.3)

where dr and λc denote the distance between the adjacent elements of the RIS and the

carrier wavelength, respectively. Then, the channel from the ith MU to the RIS, denoted

as gi , can be modeled as

Pi
X
gi = αp,i aRa (θp,i , ϕp,i ), (5.4)
p=1

where αp,i denotes the complex channel gain of the pth path.

Similarly, it is assumed that the number of propagation paths between the BS and

the RIS is Nj . The AoD of the jth path from the RIS to the AP can be decomposed

into the elevation angle 0 ≤ θj ≤ π in the vertical direction and the azimuth angle
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 94

0 ≤ ϕj ≤ π in the horizontal direction. Hence, the array response vector can be

expressed as aRd (θj , ϕj ), which is similar to the expression of aRa (θp,i , ϕp,i ) in (5.1).

For the RIS-BS link, the AoA of the jth path can be decomposed into the elevation

angle 0 ≤ ψj ≤ π in the vertical direction and the azimuth angle 0 ≤ ωj ≤ π in the

horizontal direction. Therefore, the array response vector aB (ψj , ωj ) can be written as

(e) (a)
aB (ψj , ωj ) = aB (ψj ) ⊗ aB (ψj , ωj ), (5.5)

with

−j2πdb cos ψj −j2π(Mx −1)db cos ψj

(e)
aB (ψj ) = [1, e λc , ..., e λc ]T , (5.6)

and

−j2πdb sin ψj cos ωj

(a)
aB (ψj , ωj ) =[1, e λc , ...,
−j2π(Mz −1)db sin ψj cos ωj
e λc ]T , (5.7)

where db is the antenna spacing at the BS.

By using the array response vector aRd (θj , ϕj ) and aB (ψj , ωj ), the channel matrix of

the BS with the RIS can be formulated as

Nj
X
H= βj aB (θj , ϕj )aH
Rd (ψj , ωj ), (5.8)
j=1

where βj denotes the channel gain of the jth path.

Denote Ψt ∈ CNy,z ×Ny,z as the phase shift matrix of the RIS in time slot t. It is

assumed that the MUs transmit pilot sequences of length τ via the RIS to the BS.

During the uplink transmission of the ith MU, in time slot t, 1 ≤ t ≤ τ , the received
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 95

signal from the ith MU at the BS can be written as

√
yi (t) = HΨt gi psi (t) + ni (t), (5.9)

where si (t) denotes the pilot signal from the ith MU, ni (t) ∈ CMx,z ×1 ∼ CN (0, δ 2 I)

represents AWGN with power δ 2 at the BS. Here, p denotes the transmit power of the

ith MU.

Remark 2. In this chapter, we assume that the BS is capable of effectively handling

interference cancellation during the uplink transmission of pilot signals. After success-

fully canceling the interference, the proposed algorithm can process with multi-users. This

assumption allows us to focus on the fingerprint-based algorithm design. However, how

pilot signals from multiple MUs interact, represents an interesting avenue for further

research.

According to the expression of the received signal from the ith MU, the STCRV of

the ith MU is given as

hi = HΨt gi , (5.10)

which consists of the channel parameters (e.g., channel gain and AoA/AoD). The STCRVs

are unique for different positions, hence the STCRVs can be regarded as a new type of

CSI fingerprint of the MU. According to the definition of directly localization algorithms

[26], the STCRVs can be used to estimate the positions of the MUs. Therefore, by

denoting the estimated position of the ith MU as ûi = [x̂i , ŷi , ẑi ]T , it is formulated as

ûi = f (hi ), (5.11)

where f (·) denotes the complex non-linear function between the STCRV and the esti-

mated position of the ith MU. Hence, the regression problem of estimating the positions
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 96

of the MUs can be formulated as

Nu
1 X
min (ûi − ui )2 . (5.12)
ûi Nu
i=1,··· ,Nu i=1

5.3 Residual Convolution Network Regression Learning Algo-

rithm

According to the regression problem in (5.12), the closed-form expression of the 3D

position of the MU is not available by using traditional optimization methods. Hence, a

regression learning algorithm to predict the positions of MUs is developed in this section.

Deep learning based method, popular in image recognition of computer science, can be

applied to the STCRV fingerprint localization since the STCRV can be seen as an image.

Therefore, it is proposed a novel RCNR learning algorithm to represent the function f (·)

in (5.11) and solve the regression problem (5.12).

5.3.1 Regression-oriented Positioning

In computer science, CNN is often used for image classification with the last layer being

activated by a softmax function [66]. As an alternative to the classification function,

CNN can also be viewed as a regression function if the softmax layer is replaced by

a fully connected layer containing an activation function. As an advanced network of

CNN, residual convolution network (RCN) can also be used as regression function by

substituting a fully connected layer into the last softmax layer.

5.3.2 The structure of the RCNR learning algorithm

Fig. 5.2 shows the details of the proposed algorithm. As shown in Fig. 5.2, the proposed

algorithm consists of a data processing block, a normal convolution block, four residual

convolution blocks, and a regression block. The descriptions of these blocks will be

introduced as follows.
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 97

(a) Data processing (DP) block. (b) Regression block.

Figure 5.3: Two blocks used for the proposed algorithm.

[Link] Data Processing Block

The data processing (DP) block is designed to process the input STCRV. As shown Fig.

5.3(a), the DP block includes two layers: a data decomposing (DD) layer, and a data

reshape (DR) layer.

First, since the STCRV is a complex vector, the DD layer is designed to decompose
(r) (i)
hi into two parts, the real value vector hi and the imaginary value vector hi , which

can be expressed as

(r) (i)
hi = hi + jhi . (5.13)

Then, according to the expression of the STCRV in (5.10), the STCRV consists of

the features of the horizontal angle domain and the vertical angle domain of the BS.

To further extract these features, the DR layer is designed to reshape the two vectors
(r) (i)
hi ∈ CMx,z ×1 and hi ∈ CMx,z ×1 as two space channel response matrices (SCRM),
(r) (i)
which are denoted as Hi ∈ CMx ×Mz and Hi ∈ CMx ×Mz , respectively. To be specific,
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 98

(a) Normal convolution (NC) block. (b) Residual convolution (RC) block.

Figure 5.4: Two kinds of convolution block used for the proposed algorithm.

these two vectors are reshaped by arranging the elements of vector into Mx rows and

Mz columns.

[Link] Normal Convolution Block

The normal convolution (NC) block is designed to extract the features of the SCRM. As

shown in Fig. 5.4(a), the NC block consists of three layers: a convolution (Con) layer, a

batch normalization (BN) layer, and a max pooling (MP) layer.

Due to the large number of antennas at the BS, the SCRM has a large dimension.

Therefore, if the SCRM is input directly into a DNN consisting of fully connected layers,

a large number of weight parameters of the DNN should be trained. To improve the

efficiency of the training neural network, the Con layer is proposed [66]. The Con layer

has multiple filters sliding over it for a given input SCRM so that the features of SCRM

can be extracted [66]. As a result, the number of weight parameters to be trained can

be reduced. Moreover, the BN layer is designed to normalize the input data so that the

convergence rate can be improved. Furthermore, the MP layer is designed to reduce the

complexity of further layers, which is similar to reducing the resolution in the field of
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 99

computer science.

[Link] Residual Convolution Block

In the field of computer science, increasing the number of the Con layers will allow the

neural network to extract more features [66]. However, according to the experiment in

[25], when the neural network reaches a certain depth, the problem of gradient explosion

and gradient disappearance will appear, which leads to a worse optimization effect and

lower accuracy of the proposed neural network. Hence, to improve the estimation accu-

racy, the residual convolution network is proposed to enable the deeper neural network

to train well and obtain a better optimization effect [25].

As the considered STCRV is the CSI that consists of the cascaded BS-RIS-MU chan-

nel. Hence, the STCRV contains more features than the other types of CSI fingerprint.

Inspired by the residual convolution network of computer science, the residual convolu-

tion (RC) block is designed to further extract the features of STCRV and protect the

integrity of features in the proposed RCNR learning algorithm. As shown in Fig. 5.4(b),

the RC block includes three Con layers and two BN layers.

As illustrated in Fig. 5.4(b), the RC block starts with two Con layers. Each Con

layer is followed by a BN layer and a ReLU activation function. Then it skip these 2

convolutional operations through the cross-layer datapath and add the input directly

before the final ReLU activation function. As a result, the integrity of the features is

protected and the degradation of the neural network can be solved.

[Link] Regression Block

The regression block is designed to output the estimated positions of the MUs. As shown

in Fig. 5.3(b), the regression block includes an average pooling (AP) layer and a fully

connected (FC) layer.

The AP layer is used to reshape the output of the RC blocks for the final FC layer

by taking the average of each feature from the RC block [66]. Adding the AP layer
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 100

between the RC block and the FC layer avoids the large number of weight parameters

introduced by the FC layer. As a result, the AP layer reduces overfitting while improving

the convergence rate.

The FC layer is used to combine the features from the NC blocks and output the

estimated positions of the MUs. The FC layer multiplies the input by a parametric

weight matrix and then adds a bias vector. By using Ω to denote the output before the

FC layer, the estimated position after the FC layer can be written as

ûi = Wvec{Ω} + b, (5.14)

where W and b are the parametric weight matrix and bias vector respectively that can

be learned together with the training of the RCNR learning algorithm.

5.4 Simulation Results

In this section, simulation results are provided to evaluate the performance of the

proposed RCNR algorithm. The software Wireless Insite [67] is used to simulate the

mmWave localization system aided by the RIS. The carrier frequency is f = 90 GHz.

For the BS, the number of antennas is set to Mx,z = Mx × Mz = 255 × 255. Besides,

the center of the BS is located at p = (−10, −5, 2.5 m), and the distance between the

antennas db is set to db = λ/2 = 1.67 × 10−3 m. For the RIS, the number of the elements

is set to Ny,z = Ny × Nz = 255 × 255. Moreover, the center of the RIS is located at

s = (−5.10, −1.43, 2 m) and the distance of the elements is set to dr = λ/2 = 1.67×10−3

m. For the MUs, it is assumed that the MUs are uniformly distributed in the grid of 9.6

m in length and 5.8 m in width. Furthermore, it is formulated as set up three grids in

total, and the heights of the grids are 1.4 m, 1.5 m, and 1.6 m, respectively. In addition,

the distance of the MUs is 0.2 m and the transmit power of the MUs is 10 dBm. During

the training phase, the total number of MUs is 25800, while for testing, the number of

MUs is 20. The phase shifts of the elements at the RIS are set to a unity matrix. The
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 101

1
STCRV
0.9 CSI amplitude
ADCPM

Cumulative distribution function

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Estimation error (m)

Figure 5.5: The CDF of estimation error with different fingerprints.

1

0.9
RCNR
Cumulative distribution function

0.8 WKNN
CNN
0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Estimation error (m)

Figure 5.6: The CDF of estimation error with different algorithms.

training of the proposed algorithm achieves convergence after 100 epochs.

To demonstrate the superiority of the proposed STCRV fingerprint, it is compared

with two benchmarks, namely the CSI amplitude in [63] and the angle delay channel

power matrix (ADCPM) in [26]. As illustrated in Fig. 5.5, the results show that when

the STCRV fingerprint is used as the dataset, the proposed RCNR algorithm achieves the

highest localization accuracy, with 90% of the estimation error within 0.25 m. This level

of precision not only signifies the algorithm’s practical effectiveness but also aligns with

and satisfies the stringent requirements of current 5G standards for indoor positioning,

making it highly suitable for advanced telecommunications applications. In comparison,

Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 102

0.9

0.8
Imperfect, 2 =10 -2

Cumulative distribution function

0.7 Imperfect, 2 =10 0
Perfect
0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5
Estimation error (m)

Figure 5.7: The CDF of estimation error with different kinds of CSI error.

the CSI amplitude fingerprint and the ADCPM fingerprint have lower estimation accu-

racy, with 50% and 82% accuracy, respectively. Therefore, Fig. 5.5 clearly demonstrates

the superiority of the proposed STCRV fingerprint.

To evaluate the performance of the proposed RCNR algorithm, the estimation error

CDF comparison of the RCNR algorithm, CNN algorithm in [63], and WKNN algorithm

in [64] is presented. As shown in Fig. 5.6, when the proposed RCNR algorithm is

employed, the estimation error at the 90% point is 0.25 m. However the estimation

errors at the 90% point are 0.7 m and 0.9 m for the WKNN algorithm and the CNN

algorithm, respectively. It means that the proposed RCNR algorithm outperforms the

WKNN and the CNN algorithm.

To investigate the impact of the CSI estimation error on the considered systems and

the algorithm, the performance versus different CSI estimation conditions is evaluated

in Fig. 5.7. It is assumed that the estimation error of the STCRV is i.i.d. complex

Gaussian random variable with zero-mean and variance σ 2 . The estimation errors at

90% are 0.25 m, 0.3 m, and 0.45 m for σ 2 = 0, σ 2 = 10−2 , and σ 2 = 100 . It implies that

the CSI estimation error degrades the estimation accuracy.

Moreover, the impact of the phase shift matrix on the performance of the proposed

system and algorithm is studied in Fig. 5.8. As illustrated in Fig. 5.8, the estimation
Chapter 5. RIS-Aided Localization Algorithm Design: Employing Fingerprint 103

1
unity matrix
0.9 random matrix

0.8

Cumulative distribution function

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5
Estimation error (m)

Figure 5.8: The CDF of estimation error with different phase shift matrices.

error at the 90% point is 0.25 m for the RIS with a unity matrix, whereas it increases to

0.35 m for the random matrix. The simulation results suggest that the proposed system

and algorithm are sensitive to the choice of the phase shift matrix of the RIS.

Remark: In this letter, a unity matrix is assumed for the scenario that the RIS is

located at the center of an indoor environment with symmetrical areas on both sides

where the MUs need to be located. However, exploring the optimal phase shift for other

scenarios in the RIS-assisted localization system would be an interesting future work.

5.5 Summary

This chapter studies the MU localization problem in MIMO TDD mmWave systems

aided by the RIS. This chapter derives the expression for STCRV at the BS as a new

type of fingerprint. In addition, by using the STCRV fingerprint as input, this chapter

also proposes a novel RCNR algorithm to predict the 3D position of the MU.
Chapter 6

RIS-Aided Localization
Algorithm Design: Overcoming
Faulty Elements

6.1 Introduction

Consider a more general scenario where some of the RIS elements may have been dam-

aged (refer to as faulty elements in this chapter) due to the environments conditions.

In such case, localization algorithms designed under the assumption of a perfect RIS

without faulty elements will no longer be reliable and may result in low localization

accuracy. Furthermore, the RIS faulty elements can reduce the dimension of the RIS

received signal, leading to the localization information lost. This, in turn, significantly

degrades the performance of the localization algorithm based on RIS information.

To mitigate the harmful impact of the faulty elements and regain the lost localization

accuracy, it is crucial to first identify the faulty elements. Detecting faulty elements is

equivalent to determining the status of each element of the RIS, classifying them as

either faulty or non-faulty. This task, when applied to all elements of the RIS, becomes

104
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 105

a multi-label binary classification problem. However, due to the passive nature of the

RISs, conventional faulty detection algorithms designed for antenna are not capable of

determining the number and the places of the faulty elements.

Given the presence of the faulty elements, the localization algorithm design based on

the RIS information remains unexplored. To the best of the authors’ knowledge, only

[68] has addressed the scenario in which the RIS contains faulty elements. However,

the authors of [68] designed the corresponding localization algorithm based on the BS

information. Besides, this work maximized the estimation accuracy by assuming the

fault of each element follows a specific distribution model. However, in practice, the

faults of the RIS elements may not follow a specific fault distribution. Furthermore,

knowledge of a specific distribution is difficult to acquire by the network manager, and

its estimation may require a significant number of pilot signals.

Besides, it is worth noting that this challenge is difficult to address using the tradi-

tional optimization methods. Deep learning (DL)-aided algorithms have emerged as a

promising approach [21–25]. DL algorithms are widely applied in wireless communica-

tion localization due to their ability to model complex signal propagation in challenging

environments [26], and their capacity for strong nonlinear mapping, which translates

observed signal characteristics into precise location coordinates [27]. However, the con-

sidered challenge remains non-trivial even when resorting to the powerful DL tools. First,

since the RIS is commonly comprised of a large number of elements, the detection of

faulty elements across the RIS is very challenging with existing DL algorithms. Sec-

ond, a specific neural network (NN) has to be designed for reconstructing a complete

high-dimensional signal. Third, the existing DL-aided localization algorithms require

extensive training epochs and large datasets to ensure accurate estimation, making their

deployment in dynamic environments unfeasible.

To address this challenge, this chapter proposes an effective localization algorithm

based on the RIS information in the presence of faulty elements. To mitigate the harmful

impact of the faulty elements, this chapter first employ transfer learning to design an
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 106

algorithm for detecting faulty RIS elements. Then, to reconstruct the complete high-

dimensional RIS information, this chapter initially recover the signal received by the

non-faulty elements and subsequently restore the lost information from these identified

faulty elements. Finally, to effectively localize the MU with smaller dataset and less train-

ing steps compared to the traditional training-from-scratch DL algorithm, this chapter

employ transfer learning to design a localization algorithm. The main contributions are

summarized as follows:

1) Given an RIS possessing some unknown faulty elements, transfer learning is har-

nessed to design a TPTL algorithm, addressing the faulty element detection prob-

lem. To cope with the limitations of multi-label classification, in Phase I, the well-

trained DenseNet 121 is leveraged to pinpoint faulty sub-arrays (SAs). Building

on this, in Phase II, the faulty reflecting elements contained within the previously

identified faulty SAs is detected by transferring the NN of Phase I.

2) This chapter reconstruct the high-dimensional signal received at the RIS which

contains rich information and can be utilized for high-accuracy localization. Specif-

ically, based on the low-dimensional BS signal, this chapter employ both a CNN

and a VAE to recover the signal received at the non-faulty RIS elements and to

restore the lost information at the faulty RIS elements.

3) This chapter propose a TEDS localization algorithm, where the signal reconstruc-

tion is performed in Stage I and the reconstructed signals are stored as fingerprints.

In Stage II, DenseNet 121 is transferred to estimate the coordinates of the MU.

By fully exploiting the RIS’s high-dimensional information, the proposed TFDS

algorithm effectively improves the localization performance.

4) As a benchmark algorithm, a TEDF localization algorithm is designed, which

only exploit the BS information (e.g., received signal at the BS). The comparison

between this algorithm and the TEDS algorithm is provided to better understand

the benefits of using RIS information for localization.

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 107

5) Through accurate faulty element detection and the effective signal reconstruc-

tion, simulation results demonstrate that the localization performance of the pro-

posed algorithm can benefit from the high-dimensional RIS information even with

faulty elements. Additionally, the proposed TEDS localization algorithm showcases

remarkable robustness against unoptimized phase shifts and s SNR variations.

6.2 System Model and Problem Formulation

Consider an RIS-aided UL localization system, where an MU transmits pilot signals to

the BS. The objective is to determine the MU’s location with the aid of an RIS with

some faulty elements but it is unknown which elements are damaged. Furthermore, the

BS is equipped with a UPA having M1 × M2 = M antennas, and the MU is equipped

with a single antenna. Moreover, the RIS is equipped with a UPA having N1 × N2 = N

reflecting elements.

6.2.1 Geometry and Channel Model

The BS is located at coordinates pb = [xb , yb , zb ]T , while the center of the RIS is situated

at pr = [xr , yr , zr ]T . The true position of the MU is denoted as pu = [xu , yu , zu ]T .

Typically, once RIS and BS are deployed, their positions pr and pb , are known and

remain fixed. To motivate the deployment of an RIS, a classical scenario is considered

where the LoS path between the BS and the MU is blocked [14]. Such blockages can

arise from numerous factors, including buildings, vehicles, or natural obstructions [69].

Next, the MU-RIS and RIS-BS links are modeled. For any antenna array, given an

elevation angle θ ∈ (0, π] and an azimuth angle ϕ ∈ (0, π], the array response vector can

be generally described as

ax (θ, ϕ) = a(e) (a)

x (θ) ⊗ ax (θ, ϕ), (6.1)
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 108

where x represents the specific link and ⊗ denotes the Kronecker product, and

 T
−j2π(Ne −1)dx cos θ
a(e)
x (θ) = 1, . . . , e λc , (6.2)
 T
−j2π(Na −1)dx sin θ cos ϕ
a(a)
x (θ, ϕ) = 1, . . . , e λ c , (6.3)

where Ne and Na denote the number of elements in the elevation and azimuth planes,

respectively, dx denotes the distance between adjacent array elements, and λc is the

carrier wavelength.

[Link] MU-RIS link

Considering P propagation paths between the MU and the RIS, and utilizing the general

array response vector defined in (6.1), the channel response of the MU-RIS link, gur , can

be expressed as

P
X
gur = αp aRa (θp , ϕp ), (6.4)
p=1

where aRa denotes the array response vector of the RIS and αp represents the channel

gain of the p-th path.

[Link] RIS-BS link

Assuming J propagation paths between the RIS and the BS, the channel matrix of the

RIS-BS link, Hrb , can be modeled as

J
X
Hrb = βj aB (θj , ϕj )aH
Rd (ψj , ωj ), (6.5)
j=1

where βj represents the channel gain of the j-th path and aB is the array response

vectors of the BS.

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 109

6.2.2 Faulty Element Model

The RIS panel is passive, lacking the self-diagnostic capabilities of active antennas. Given

this absence of direct feedback, establishing an external detection mechanism becomes

crucial. For this purpose, drawing from the faulty antenna model in [70], a corresponding

model for the RIS elements is introduced as follows.

[Link] RIS Faulty Element Model

Without faulty elements, the RIS phase shift vector is defined as ω = [ω1 , · · · , ωN ]T ∈

CN ×1 . Actually, some of the reflecting elements may be damaged (refer to as faulty

elements in this chapter) due to some unknown environment conditions, leading to the

situation that they cannot effectively receive/reflect any signals 1 .

It is assumed that the received signal strength at the n-th element of the RIS is
(r)
denoted as yn . Besides, a threshold parameter ζr ≪ pr to determinate the fault status

of the n-th element is employed, where pr represents the minimum transmit power to

the reflecting elements, and ζr is assumed to be exceedingly small. Accordingly, the fault

status is mathematically defined as


 (r)
0, yn ⩽ ζr


Bn = (6.6)
 (r)
1, yn > ζr ,

 n ∈ {1, · · · , N }.

(r)
As shown in (6.6), there are two distinct cases: First, when yn ⩽ ζn , this indicates that

the strength of the RIS received signal is below or equal to the threshold ζn , suggesting

that this element is damaged. Consequently, it is assigned Bn = 0 in such cases. Con-

(r)
versely, when the received signal at the RIS exceeds the threshold ζn , i.e., yn > ζn , this

signifies that the n-th element is operating correctly. Consequently, it is assigned Bn = 1

in these situations.
1
There may be other metrics to define whether an RIS element is damaged or not, which will be left
to the future work.
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 110

By integrating the statuses of all elements into a vector as follows

B = [B1 , · · · , BN ]T . (6.7)

Accordingly, the RIS phase shifts profile can be mathematically formulated as follows

ϖ = ω ⊙ B = [ω1 B1 , · · · , ωN BN ]T = [ϖ1 , · · · , ϖN ]T , (6.8)

where ⊙ denotes the Hadamard product.

6.2.3 Signal Model

The UL signal received by the RIS can be expressed as

yr = gur s. (6.9)

Accordingly, The UL signal received by the BS can be expressed as

y = Hrb diag(ϖ)gur s + n, (6.10)

where diag(ϖ), s, and n denote the phase shift matrix of the RIS, the pilot signal

transmitted by the MU, and the zero-mean additive white Gaussian noise, respectively.

6.3 Problem Formulation

Building upon the received signal y, defined in (6.10), the objective is to design a local-

ization algorithm for the scenario where the RIS contains deterministic faulty elements.

To this end, it is essential to detect the faulty elements, which enables to mitigate their

harmful impact on the localization accuracy. Unlike active components that can self-

diagnose or report inconsistencies, the potential faults or damages within the RIS are

invisible without external intervention. Detecting and localizing these faulty elements

requires a deeper understanding and innovative methodologies, presenting a non-trivial

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 111

Figure 6.1: Network structure of TPTL algorithm.

challenge in the field. In the following subsection, a faulty elements detection problem

will be formulated.

6.3.1 Faulty Element Detection Problem

First, let B̂ denote an estimate of B, given by

B̂ = F(y) = [B̂1 , . . . , B̂N ]T , (6.11)

where F(·) denotes the non-linear function between y and B̂, and B̂n represents the

estimated status of the n-th RIS element, which is either 1 (functioning) or 0 (faulty).

Consequently, the detection problem is given by

(P1) : min LB (F (y) , B) , (6.12)

B̂

where LB is the loss function for predicting the element status. To further illustrate the

proposed problem and the related challenge, the following remarks are made.

Remark 3. Given the nature of the received signal y at the BS, y can be regarded as an

image, drawing parallels between problem (P1) and tasks in computer science (CS) [27].

Since the status of the individual elements, Bn , is either 1 or 0, the estimation of the

fault status corresponds to a binary classification problem. Furthermore, considering that

the RIS comprises multiple elements, (P1) can be regarded as a multi-label classification
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 112

problem. To generate label data for detecting faulty elements in an RIS, a methodical

approach is used. Firstly, a maximum threshold for the number of faulty elements within

the RIS is established. Following this, a substantial number of binary arrays are randomly

generated. Each array represents the operational status of the RIS elements, with ‘0’

denoting a faulty element and ‘1’ indicating a normal, functioning element. The count

of ‘0’s in each array does not surpass the previously set maximum number of faulty

elements.

Remark 4. Contemporary research on multi-label classification has only advanced to

handle datasets with up to 100 labels [71]. For instance, datasets like “MS-COCO”

encompass 122, 218 images spanning 80 labels [72]. However, the number of the reflecting

elements of an RIS can easily surpass 100 (e.g., N1 × N2 > 10 × 10). In other words, the

number of the labels that need to be classified will exceed 100 and reach thousands, posing

a significant challenge for detecting faulty elements using the existing DL algorithms.

To address the challenge in Remark 2, it is proposed to divide the RIS panel into

multiple SAs, each containing a same number of reflecting elements. Then, Problem

(P1) is decomposed into two sub-problems. The first sub-problem is to detect the SA

containing the faulty elements (referred to as the “faulty SA”), represented by Problem

(P1-a). The second sub-problem is to detect the faulty elements within the faulty SA,

represented by Problem (P1-b). These two sub-problems will be sequentially introduced

in the following.

[Link] Detection of Faulty SA

To begin with, it is introduced the procedures for dividing and labeling the SAs. The

RIS with N elements is divided into K SAs, with each SA containing Ns = ⌊N/K⌋

reflecting elements 2 . If the k-th SA contains one or more faulty reflecting elements,

it is classified as a faulty SA. Conversely, if the k-th SA does not contain any faulty

elements, it is determined that the k-th SA to be operating correctly. Then, by denoting

2
⌊x⌋ denotes the function that output the greatest integer less than of equal to x.
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 113

(faulty)
the number of faulty elements of the k-th SA as Nk , the status of the k-th SA can

be mathematically defined as

 (faulty)
0, Nk ≥ 1,


Ck = (6.13)
 (faulty)
1, Nk

 = 0, k ∈ {1, · · · , K}.

By collecting the status of all SAs in vector form, it is formulated as C = [C1 , · · · , CK ]T .

The estimate of C is thus defined as Ĉ = X (y) = [Cˆ1 , · · · , CˆK ]T , where X (·) denotes a

non-linear function and Cˆk denotes an estimate of Ck .

Building upon this, the problem of detecting the faulty SAs can be formulated as

follows

(P1-a) min LC (X (y) , C) , (6.14)

Ĉ

where LC denotes the specific loss function for estimating the status of all SAs.

[Link] Detection of Faulty Elements of SA

After identifying the faulty SAs, the next step is to detect the faulty elements within

a given faulty SA. Similar to (6.7), the statuses of the faulty elements within the k-th
(k)
faulty SA can be organized in a vector, denoted as Bsub = [B1 , · · · , BNs ]T . Thus, the
(k) (k) (k) (k) (k)
estimate of Bsub can be written as B̂sub = [B̂1 , · · · , B̂Ns ]T , where B̂ns , ns ∈ {1, · · · , Ns },

represents the estimated status of the ns -th element within the k-th faulty SA. Besides,

it is formulated as H(·) to denote the complex non-linear function between the received

signal y and the estimated statuses of faulty elements within the k-th faulty SA.

Consequently, detecting the faulty elements within the k-th faulty SA can be formu-

lated as follows

 
(k)
(P1-b) min LBsub H (y) , Bsub , (6.15)
(k)
B̂sub
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 114

where LBsub denotes the specific loss function for predicting the statuses of the faulty

elements within the k-th faulty SA.

Similar to (P1), these two sub-problems can also be regarded as a kind of multi-label

classification problem. In the field of CS, the multi-label classification problem can be

solved by using a CNN [27]. Notably, the parameters of the CNN have to be trained

with a large number of data sets before being used in the online phase [27]. However,

training two CNN networks to solve the above problems is inefficient. To address this

challenge and to save time and computation resources, it is proposed to leverage transfer

learning algorithms [21] to fine-tune pre-trained CNN networks.

6.3.2 Localization Problem Under Faulty Elements

The localization accuracy of existing localization algorithms designed under assumption

of a perfect RIS is expected to deteriorate in the presence of faulty RIS elements. There-

fore, building upon the obtained knowledge on the faulty elements, i.e., B̂, it is crucial

to develop a localization algorithm rather that can correct the harmful impact of faulty

elements. By defining the estimate of the position of the MU as p̂u , the relationship

with the knowledge of faulty elements can be expressed as

p̂u = Q(y, B̂) = [x̂u , ŷu , ẑu ]T , (6.16)

where Q(·) represents the intricate non-linear mapping between received signal y and

the estimated position of the MU, p̂u . In the next step, the localization problem is

addressed, labeled as (P2). This problem aims to accurately predict the 3D position of

the MU by minimizing the loss function, Lp . Mathematically, it is formulated as

 
(P2) min Lp Q(y, B̂), pu . (6.17)
p̂u

Traditional convex optimization and channel estimation methods, as highlighted in

[40], are not capable of accurately estimating pu . Therefore, this chapter introduces a
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 115

Figure 6.2: Network structure of DenseNet 121.

fingerprint-based algorithm for this purpose. After the faulty elements have been identi-

fied the complete high-dimensional RIS signal is reconstructed from the low-dimensional

BS signal. This reconstructed RIS received signal then acts as a unique fingerprint

for estimating pu . For comparison, the BS received signal is also utilized as another

fingerprint to estimate pu .

6.4 Faulty Element Detection Algorithm Design

In this section, it is aimed to solve Problem (P1) to estimate the fault statuses of the

RIS elements which serves as the foundation for the design of the proposed localization

algorithms. To this end, a TPTL algorithm is proposed, whose the network structure

is depicted in Fig. 6.1. TPTL employs transfer learning in two distinct phases to

sequentially tackle problems (P1-a) and (P1-b).

6.4.1 Adoption of Transfer Learning

Transfer learning [21] is particularly advantageous for tasks like detecting faulty elements

within the RIS. One of the primary reasons behind its suitability is the analogous nature

of received signals to images. Received signal y can be viewed as an image, also with

features to be extracted and recognized. This similarity harness the power of models

pre-trained for image recognition tasks and adapt them for faulty elements detection.

Transfer learning can efficiently use data from image domains to facilitate faulty elements

detection.

Hence, a TPTL algorithm is proposed to sequentially address problems (P1-a) and

(P1-b). As shown in Fig. 6.1, during Phase I, the classical DenseNet 121 [24] is utilized
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 116

as the base model for transfer learning, adapting the input and output modules for

Problem (P1-a). Upon completing Phase I training, the trained NN is saved to serve

as the base model for transfer learning in Phase II. Furthermore, in Phase II, necessary

adjustments are made to both the input and output modules of the well-trained model

to better fit Problem (P1-b). To facilitate an intuitive understanding of the employed

approach, the details of DenseNet 121 will be introduced in the following subsection.

6.4.2 The Architecture of DenseNet 121

DenseNet [24], inspired by the foundational principles of ResNet [25], reuses the features

from earlier layers. However, different from the skip connections which might bypass a

couple of layers, DenseNet establishes connections between every layer and every other

layer, ensuring a more robust feature propagation. This approach, while reducing the

parameter count, allows DenseNet to potentially surpass the performance of ResNet.

The design of DenseNet 121 is illustrated in Fig. 6.2, and its architecture comprises

three main components: Dense Modules, Transition Modules, and an Output Module,

which will be introduced as follows.

[Link] Dense Module

As depicted in Fig. 6.2, the Dense Module comprises multiple Dense Blocks. The details

of a Dense Block are shown in Fig. 6.3. Each Dense Block starts with a batch normal-

ization (BN) layer, followed by a layer rectified linear unit (ReLU) activation function

layer, and concludes with a 3 × 3 convolutional (Con) layer. A notable characteristic of

the Dense Module is its connectivity: Every block incorporates the feature maps of all

preceding blocks within the module. This design ensures efficient feature reuse and a

rich flow of information throughout the network [24].

[Link] Transition Module

From Fig. 6.2, it is observed that the Transition Module, which acts as a connector

between Dense Modules, starts with a 1×1 Con layer, followed by a 2×2 average pooling
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 117

Figure 6.3: Network structure of a Dense Block.

layer, effectively reducing the spatial dimensions of the feature maps and streamlining

them for either the next Dense Module or the concluding output layer.

[Link] Output Module

This module employs average pooling to aggregate the feature maps. Subsequently, these

integrated features are channeled through a linear layer. A sigmoid activation function

then finalizes the process, ensuring that the output values lie between 0 and 1.

Having detailed the architecture of DenseNet 121, the proposed TPTL algorithm will

be explained, which is structured in two distinct phases. These phases are elaborated in

the following subsections.

6.4.3 Phase I: Transfer Learning of DenseNet 121

To extract the features from the received signal y and tackle sub-problem (P1-a), transfer

learning [21] with a pre-trained DenseNet 121 [24] is employed. Though the internal

layers of DenseNet 121 are effective at feature extraction due to their extensive pre-

training on datasets such as ImageNet [73], adjustments are necessary to cater to the

specific nuances of sub-problem (P1-a) and the nature of signal y. As shown in Fig. 6.1,

an Input Module is introduced to preprocess the received signal, while the Output Module

is configured to classify the different statuses of the SAs on the RIS panel. Additionally,

choosing an appropriate loss function for this binary classification problem is crucial.

The details will be discussed in the subsequent paragraphs.

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 118

[Link] Input Module

The received signal at the BS, i.e., y, is first processed by the Input Module so that it

can be fed into the transferred DenseNet 121 for classifying the statuses of the SAs [21].

To this end, signal y is decomposed as follows

y = R(y) + jI(y), (6.18)

where R(y) and I(y) are the real and imaginary parts, respectively. In this way, it is

ensured that both the real and imaginary information contributes to enhancing signal

differentiation and facilitating feature extraction.

Thereafter, both R(y) and I(y) can be reshaped into matrices, which are denoted

as MRy ∈ RM1 ×M2 and MIy ∈ RM1 ×M2 , respectively. The primary motivation for this

particular reshaping strategy is to effectively extract meaningful features from the UPA

of the BS, which spans both the horizontal and vertical dimensions. Then, the reshaped

matrices of the real and imaginary parts are concatenated, resulting in a 3D tensor

represented as  
MRy  2×M1 ×M2
Ty =  ∈R . (6.19)
MIy

However, the values of M1 and M2 may not be directly compatible with DenseNet 121.

Hence, an upsampling technique is applied:

Tup = Upsample(Ty ) ∈ R2×256×256 . (6.20)

The upsampling technique is employed to adjust the dimensions of the tensor Ty to

the desired dimensions of 2 × 256 × 256. It typically involves pixel value interpolation

to expand the image or tensor size, redistributing pixel values accordingly. To make

Tup suitable for processing by DenseNet 121 [24], which expects an input tensor of the

form (3, 256, 256), a Con Layer is introduced to transform the bi-channel tensor to a
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 119

tri-channel one:

TDenseNet = ConLayer(Tup ). (6.21)

[Link] Transferred DenseNet 121

TDenseNet is input to the well-trained DenseNet 121 without the first Con layer and Output

Module. While the internal layers of DenseNet 121 are frozen during the early stages

of training to maintain their feature extraction capabilities, they are gradually unfrozen

later, enabling minor adjustments to adapt better to the specific task of detecting faulty

SAs. This ensures that the network leverages the generic feature extraction capabilities

of the pre-trained model while also fine-tuning specific layers to the dataset and problem

definition for SA detection.

[Link] Output Module

In the typical use case, DenseNet 121 is employed for single-label image classification

using a softmax layer to determine category probabilities [74]. However, the task in

(P1-a) entails a multi-label binary classification problem concerning the statuses of the

SAs. Specifically, it is formulated as K SAs, and for each one, the goal is to determine if

it is faulty or not. The status of a given SA is independent of the statuses of the other

SAs. To accommodate this, the linear layer of the Output Module is modified to produce

K outputs, which is aligned with the number of SAs under consideration. Furthermore,

the softmax layer with the sigmoid layer [74] is replaced, which allows for individual

probability estimation for each label. Defining the output from the linear layer by xp1 ,

the probability that the k-th SA is faulty (represented as Cˆk = 0) can be expressed as

P (Cˆk = 0|xp1 ) = Fsig (wkT · xp1 + bk ), (6.22)

where Fsig denotes the employed sigmoid function, wk is the weight vector associated

with the k-th SA, and bk represents the corresponding bias.

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 120

[Link] Loss Function

For addressing the complexities of the multi-label classification task, it is imperative to

employ a suitable loss function. The multi-label soft margin loss in [75] is employed

as the specific loss function for addressing the multi-label classification problem in this

chapter. The application of the multi-label soft margin loss is motivated by its design,

specifically tailored for multi-label scenarios [74]. Unlike conventional loss functions, the

multi-label soft margin loss was constructed for problems where input samples might

simultaneously belong to multiple labels, providing the means to individually estimate

label probabilities. The mathematical formulation of the loss function is given by

K
1 X
Lmsml =− [Ck log(Cˆk ) + (1 − Ck ) log(1 − Cˆk )], (6.23)
K
k=1

where K represents the total number of SAs (or labels) in the RIS. For the k-th SA, Ck

denotes the true label, while Cˆk represents the predicted label.

[Link] Training for Faulty SA Detection

Once the Input Module, Output Module, and loss function adjustments are established

for the SA classification, the next step is to fine-tune the pre-trained DenseNet 121 on

the specific dataset for detecting faulty SAs. This means that training the model on

a new dataset specifically focuses on detecting faulty SAs. The utilization of weights

initialized from prior training on comprehensive datasets, such as ImageNet [73], offers

a foundation that aids faster convergence and potentially better generalization [21].

The primary objective during the training step is to optimize the network weights

by minimizing the loss LC , which is specifically designed for the faulty SA detection

problem (P1-a). By training and converging the network, Problem (P1-a) of estimating

the faulty SAs of the RIS can be solved. The training process involves iteratively adjust-

ing the network weights based on the gradient of the loss with respect to the weights.
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 121

Mathematically, the optimization problem for training in Phase I can be expressed as

min LC (θP1-a ), (6.24)

θP1-a

where θP1-a represents the parameters (or weights) of the designed network for the detec-

tion of faulty SAs in Phase I.

[Link] Output after Training

Upon training convergence, the network outputs undergo the sigmoid activation function

as detailed previously. This produces probabilities for each status of the faulty SA. The

classification based on a threshold of 0.5 is given by


0 if P (Cˆk = 0|xp1 ) ≥ 0.5;


Cˆk = (6.25)

1

otherwise.

Besides, the threshold can be tailored to meet precision-recall needs or practical require-

ments 3 .

6.4.4 Phase II: Transfer Learning of ‘Phase I’ Network

Problem (P1-b) is similar to Problem (P1-a), since they have the same prior knowledge

y and are both muti-label classification tasks. Hence, transfer learning is employed in

this phase to reduce the training time and cost. Specifically, the well-trained NN from

Phase I is adopted as the base model for transfer learning in Phase II.

If F1 denotes the feature space shaped in Phase I and F2 represents the desired

feature space for Phase II, the intent is to discern a transformation function ϕ : F1 → F2
3
To adjust the classification threshold, precision-recall curves can be used to find a balance between
precision and recall suitable for the application. A higher threshold is used to prioritize precision, while
a lower threshold is chosen to favor recall.
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 122

that minimizes the divergence between these spaces, which is expressed as

ϕ∗ = arg min∥F1 − ϕ(F2 )∥2 . (6.26)

ϕ

Leveraging the NN pre-trained in Phase I as the base model, modifications tailored

for Phase II are introduced, especially in the Output Module. By doing so, it can be

ensured that the foundational features and representations learned in Phase I remain

intact while the network is trained to better align with the task of Phase II. The network

structure of Phase II is shown in Fig. 6.1. Details of the modifications are presented in

the following.

[Link] Input Module of Phase II

Although Phase II directly transfers the input module from Phase I, it is essential

to unfreeze this module before training. This is due to the distinct tasks associated

with each phase. This operation aids the network to adapt more effectively to the new

problem.

[Link] Output Module

To tackle Problem (P1-b), it is imperative to make adjustments to the linear layer within

the Output Module. Specifically, the output dimension of this linear layer needs to be

adjusted to Ns . Hence, the Output Module can yield the probabilities associated with

the fault statuses of the Ns reflecting elements in the k-th SA. Denoting the output of

this modified linear layer by xp2 , the probability of the ns -th reflecting element in the
(k)
k-th SA being faulty (denoted as B̂ns = 0) can be expressed as

P (B̂n(k)
s
= 0|xp2 ) = Fsig (wnTs · xp2 + bns ), (6.27)

where wns denotes the weight vector associated with the ns -th reflecting element and

bns represents the corresponding bias.

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 123

[Link] Training for Phase II

Once the adjustments have been made, the next step is fine-tuning the pre-trained NN of

Phase I on another specific dataset for detecting faulty reflecting elements within each

faulty SA. This means that it is formulated as to train the model on a new dataset for

detecting faulty reflecting elements.

The primary objective during the training step is to optimize the network weights

by minimizing the loss LBsub , which is specifically designed for the faulty reflecting ele-

ments detection Problem (P1-b). By training and converging the network using this

optimization, Problem (P1-b) of estimating the faulty reflecting elements can be solved.

Mathematically, the optimization problem for training in Phase II can be formulated as

min LBsub (θP1-b ), (6.28)

θP1-b

where θP1-b denotes the parameters (or weights) of the network for this specific sub-

problem. Using the gradient of the loss with respect to these weights, the training

process iteratively adjusts the network weights until convergence.

Crucially, for Phase II, the internal layers of the NN transferred from Phase I are ini-

tially frozen during the early stages of training to retain their feature extraction prowess.

However, as training progresses, these layers are gradually unfrozen, allowing for better

adaptation to the specific task of detecting faulty reflecting elements. This allows the

network to use the pre-trained model’s general features and fine-tune specific layers for

detecting faulty reflecting elements.

[Link] Output after Training

Upon training convergence, the network outputs are fed into a the sigmoid activation

function. This produces probabilities for each status of the reflecting elements within
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 124

the SA. A threshold, typically set at 0.5, determines the binary classification:


(k)
0 if P (B̂ns = 0|xp2 ) ≥ 0.5,


B̂n(k)
s
= (6.29)

1

otherwise.

The threshold can be adjusted based on specific precision-recall requirements or based

on the practical requirements.

6.4.5 Trained TPTL Algorithm Usage Guide

After training of the TPTL algorithm, the well-trained models are integrated into a

‘sealed black box’. Below are the instructions for using the TPTL algorithm to detect

the position of faulty elements on the RIS.

[Link] Phase I (Diagnosis of Faulty SAs)

The MU acts as a diagnostic device, transmitting a pilot signal towards the BS via the

RIS. The received signal at the BS is then input to the well-trained model of Phase I to

output the statuses of the SAs within the RIS panel.

[Link] Phase II (Diagnosis of Faulty Elements within Detected SA)

After identifying a faulty SA, all other SAs are turned off, and the MU sends another pilot

signal to the BS. The received signal at the BS is then input to the well-trained model

of Phase II to estimate the faulty statuses of the elements within this faulty SA. This

method effectively utilizes the comprehensive information provided by the BS received

signal to distinguish between varying faulty scenarios across SAs, thereby demonstrating

the versatility and robustness of the proposed diagnostic approach in Phase II.

By sequentially addressing Problem (P1-a) and Problem (P1-b) with the TPTL algo-

rithm, the overall Problem (P1) is resolved.

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 125

Figure 6.4: Network structure of TEDS algorithm.

6.5 RIS Signal Reconstruction and Localization

After having identified the faulty elements, the primary aim is to mitigate their detri-

mental effects and fully realize the potential of high-dimensional RIS information for

localization. Addressing this, the proposed TEDS algorithm is introduced, whose net-

work structure is depicted in Fig. 6.4. Deriving a closed-form solution for the 3D coor-

dinates of the MU using traditional optimization methods is generally impossible [40].

Fingerprint-based methods emerge as an enticing solution, utilizing the received signal as

a unique fingerprint to determine the MU’s location [64]. The proposed TEDS algorithm

employs the fingerprint-based method, exploiting the high-dimensional RIS information

which includes rich information beneficial for localization. As shown in Fig. 6.4, in Stage

I, the focus is on reconstructing the complete high-dimensional RIS received signal based

on the lower-dimensional BS received signal, while Stage II employs the reconstructed

RIS information to accurately localize the MU.

6.5.1 Stage I: RIS Signal Reconstruction

To enhance localization performance, it is imperative to restore the information lost

due to the identified faulty elements. To this end, the primary objective of Stage I is

the reconstruction of the complete high-dimensional signal received at the RIS. To be

specific, signal reconstruction involves two key processes: recovering the signal from the
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 126

non-faulty elements and restoring the signal that was lost due to faulty elements. As

shown in the top half of Fig. 6.4, Stage I is based on an integrated approach using a

CNN followed by a VAE [76]. Specifically, as shown in Fig. 6.4, given y, the CNN is

employed to recover the signal from the non-faulty elements. Then, given the identified

faulty elements and the signal from the non-faulty elements, the signal lost by the faulty

elements is restored with the VAE. This chapter will first explain the rationale behind

choosing the integrated CNN-VAE network, and subsequently outline the corresponding

operational details. Finally, the integrated CNN-VAE network is described.

[Link] Rationale for adapting CNN-VAE Network

Rationale for using CNN CNNs are good at capturing complex function mappings

and predicting continuous outputs, especially with a regression component [27]. They

effectively extract spatial features from signals using their Con layers [25].

Rationale for using VAE VAE renowned for their adept data reconstruction from

deficient inputs, emerge as a viable solution for image restoration [76]. Applying to

the reconstruction of the high-dimensional RIS received signal, VAE can capitalize on

spatial correlations and the information of non-faulty elements. This enable VAE to learn

the latent distribution of the input incomplete RIS signals, subsequently generating the

complete RIS received signal in alignment with this distribution [76].

Rationale for Integrating CNN and VAE The integration of the CNN and VAE

models provides significant benefits. Training them jointly allows shared feature extrac-

tion, improving model performance. This approach saves computational resources by

avoiding multiple passes in separate networks and uses a single loss function for consis-

tent training, leading to better signal reconstruction.

[Link] Recovering Signal Recieved from Non-Faulty Elements

To reproduce the incomplete received signal at the RIS with faulty elements, as illustrated

in Fig. 6.4, a combination of a CNN and a regression module is employed, where the
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 127

CNN is specifically utilized to estimate the signal, denoted as yr . However, due to the

architecture of the CNN, its output cannot contain zero values. This leads to a potential

misalignment between the CNN’s output and the expected incomplete received signal.

To overcome this challenge, status vector B is employed, where each element of the

vector is either 0 or 1. By multiplying the CNN’s output with this status vector, the

incomplete received signal at the RIS is successfully reproduced.

CNN Architecture As depicted in Fig. 6.4, the proposed CNN has two Convolution

Modules followed by an Output Module. Each Convolution Module encompasses a Con

layer, a ReLU activation function, and a max-pooling operation. The final Output

Module consists of a linear layer with dimension N1 × N2 .

CNN Training During training, the primary goal is to condition the CNN, using

received signal y from the BS, to reproduce the received signal at the RIS, denoted as

yr . Besides, the output of the employed CNN is defined as ŷr , which is given by

ŷr = FCNN (y; θcnn ) ⊙ B, (6.30)

where θcnn denotes the set of the parameters (or weights) of the CNN, and FCNN repre-

sents a function mapping y to ŷr . The objective of this training is to minimize the error

caused by the reproduced signal as follows

min ∥yr − ŷr ∥22 . (6.31)

θcnn

Upon optimizing this objective, the CNN is conditioned to best recover the incomplete

received signals at the RIS panel.

[Link] Restoring Lost Information from Faulty Elements

After recovering the incomplete RIS received signals through the CNN, ŷr undergoes

further reconstruction using a VAE as shown in Fig. 6.4.

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 128

VAE Network Architecture According to Fig. 6.4, the VAE begins with an encoder

that learns the distribution of the input ŷr through a series of linear layers. This process

results in the establishment of a latent space, which represents the essential features of

the data by determining the latent mean µ and latent variance σ 2 . From this latent

space, the encoder samples a set of latent variables. The decoder, which consists of

another sequence of linear layers, uses these sampled latent variables to reconstruct the

output, denoted as ȳr , which is an approximation of the input ŷr intended to be as

close to the original input as possible. To be specific, the decoder randomly generates

new instances within the latent space that are used to reconstruct the input data.

Training of VAE With the VAE structure in place, the reconstructed signal ȳr can be

obtained by feeding ŷr through the encoder to get the latent variables, and subsequently

passing these variables through the decoder. The specific operations are shown in Fig.

6.4. Mathematically, the reconstruction process can be represented as follows

ȳr = FVAE (ŷr ; θvae ), (6.32)

where θvae denotes the set of parameters (or weights) of the VAE, and FVAE represents

a function mapping ŷr to ȳr . Hence, the optimization problem for this training is

formulated as

min ∥ŷr − ȳr ∥22 + KL(Lv ), (6.33)

θvae

where KL and Lv denote the Kullback-Leibler divergence and the latent variables [76].

The inclusion of the KL-term acts as a regularization mechanism, ensuring that the

latent variables adhere to a certain distribution. Through this VAE-based reconstruction

process, this chapter aims to achieve a more precise and noise-resistant reconstruction

of the RIS received signal.

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 129

[Link] Integrated CNN-VAE Network

The proposed integrated network combines the capabilities of CNN and VAE. The CNN

is responsible for recovering the incomplete RIS received signal. Following this, the VAE

reconstructs the complete RIS received signal.

Training During training, the model is conditioned using the low-dimensional BS sig-

nal, y, to reconstruct complete high-dimensional RIS signal, ȳr . Simultaneously, the

network incorporates the statuses of the faulty elements, B̂, to guide the learning pro-

cess. The objective during training is defined as

 

∥yr − ŷr ∥22 + ∥ŷr − ȳr ∥22 + KL(Lv ).

 
min (6.34)
θcnn ,θvae  

Output Upon convergence of training, the network’s parameters θcnn , θvae are opti-

mized to reconstruct the RIS received signal. The model can then process any low-

dimensional input y to generate a higher-dimensional output ȳr .

6.5.2 Stage II: Enhanced Localization using Reconstructed Signal ȳr

This stage builds upon the reconstructed RIS information, ȳr , obtained in Stage I. As

shown in the lower half of Fig. 6.4, the primary goal of this phase is to achieve enhanced

localization using the reconstructed RIS information as well as the statuses of the ele-

ments. Similar to how to tackle the faulty element detection problem in the previous

section, it is proposed to utilize the pre-trained DenseNet 121 for transfer learning to

handle the localization problem described in (6.17). By doing so, the knowledge of the

model acquired from previous tasks [27] can be harnessed. This approach, grounded in

the success of transfer learning for image classification [21], is satisfied due to its smaller

training times and often achieves better results comparing to the traditional training-

from-scratch methods [27]. Hence, a transfer learning approach is employed to localize

the MU at Stage II. As depicted in Fig. 6.4, the reconstructed signal ȳr is input as a
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 130

Figure 6.5: Network structure of TEDF algorithm.

type of fingerprint to the transferred DenseNet 121 to output the 3D location of the MU.

Before the training process, some modifications of the transferred DenseNet 121 are

added to specifically address Problem (P2). The details are given as follows.

[Link] Input Module

To extract features from ȳr , it is divided into its real and imaginary parts, represented

as R(ȳr ) and I(ȳr ). These parts are then merged and reshaped based on the UPA

antenna setup at the RIS. Since N1 and N2 may not fit the NN architecture’s input

size, interpolation techniques are used to make the signal dimensions compatible with

DenseNet. Lastly, a Con layer adjusts the tensor size to meet the DenseNet 121 input

requirements.

[Link] Transferred DenseNet 121

After preprocessing the received signal ȳr , the Input Module outputs the preprocessed

tensor to the transferred DenseNet 121. Notably, the internal layers of the transferred

DenseNet 121 are initially frozen during the early training stage and gradually unfrozen

as training progresses.

[Link] Output Module

Conventionally, DenseNet 121 employs a softmax layer after the linear layer at the Output

Module to yield category probabilities. To cater to the considered regression problem,

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 131

the softmax layer is replaced with a linear layer. This layer comprises three neurons,

corresponding to the MU’s coordinates.

[Link] Selection of the Loss Function

Predicting 3D coordinates is more effectively tackled using regression than classifica-

tion, given that these coordinates exist in a continuous space rather than discrete cat-

egories [26]. Trying to classify each coordinate as a unique category would drastically

increase the number of classes, increasing complexity and computational demands [64].

For regression tasks, the MSE loss is often favored [21]. It measures the discrepancy

between predicted and actual values and has continuous derivatives that facilitate opti-

mization. Furthermore, larger errors are penalized more by MSE, which makes the

model highly sensitive to significant deviations. Given the importance of precision in 3D

coordinate prediction, regression with MSE emerges as a sound and efficient option.

[Link] Transfer Learning for Localization

Leveraging the foundational weights and configurations from pre-trained models, the

focus shifts to training the DenseNet 121 for the task of localization. At this stage, the

primary goal is to solve Problem (P2) which is mathematically represented as

min Lp (θloc ) , (6.35)

θloc

where θloc stands for the parameters (or weights) specific to the localization problem in

the transferred DenseNet 121.

[Link] Localization with well-trained DenseNet

Upon convergence of the training phase, the DenseNet 121 model stands ready to provide

estimates for the 3D position of the MU. As shown in Fig. 6.4, for every reconstructed

RIS singal ȳr , the network outputs a 3D vector, [x̂u , ŷu , ẑu ]T , which is an estimate of the

MU’s coordinates. Therefore, Problem (P2) has been effectively solved by the proposed
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 132

TEDS algorithm.

6.6 Transfer-Enhanced Direct Fingerprint Algorithm

To be able to show more light on the gain achieved by using RIS information and the

impact of faulty RIS elements on localization, this section propose a benchmark algo-

rithm, named TEDF algorithm, which only exploits the BS signal. As depicted in Fig.

6.5, the BS received signal y as a type of fingerprint is directly input to the transferred

DenseNet 121 to predict the 3D location of the MU. In contrast to TEDS algorithm,

this localization algorithm, without relying on RIS information, may suffer from some

performance loss. Therefore, the comparison between these two kinds of algorithm can

further highlight the necessity of faulty element detection and the benefits of exploit-

ing the high-dimensional RIS signal for localization. The details of this algorithm are

provided in the following.

6.6.1 Transfer Learning for Localization

The TEDF algorithm is also based on the DenseNet 121 model, leveraging its power-

ful transfer learning capabilities. However, in contrast to the method discussed in the

previous section, the input data to this model is the received signal y at the BS. The

fundamental rationale remains unchanged; the DenseNet is capable of deriving intricate

spatial relationships from the signal, providing insights into the MU’s position.

6.6.2 Training Using BS Received Signal

During the training phase, the model learns to map the received signal y to the actual

MU position without knowledge of the faulty elements. In a training process, the model

is optimized to reduce discrepancies between the predicted position and the actual MU

position, using the same loss functions and methodologies as discussed earlier.
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 133

Figure 6.6: Train and test loss of the Phase I of TPTL algorithm.

Figure 6.7: Train and test loss of the Phase II of TPTL algorithm.

6.6.3 Localization with Well-trained TEDF Algorithm

After convergence of the training phase, the proposed TEDF algorithm is capable of

estimating the 3D position of the MU. For every received BS singal y, the network

outputs a 3D vector, [x̂u , ŷu , ẑu ]T , which corresponds to the MU coordinates.

6.7 Simulation Results

In this simulation, this chapter considers a setup in a 3D space, involving a MU, an RIS,

and a BS. The carrier frequency is f = 90 GHz. The RIS is composed of N1 × N2 = 81

elements, with adjacent elements spaced at dr = λ/2 = 1.67 × 10−3 m. The center of

the RIS is located at pr = (15, 0, 2) m. The BS comprises M1 × M2 = 16 antennas, with

an antenna spacing of db = λ/2 = 1.67 × 10−3 m. The center of the BS is positioned

at pb = (0, 10, 1.5) m. The MU-RIS link includes P = 10 propagation paths, and the

RIS-BS link comprises J = 10 propagation paths. The transmit power of the MU is 10

dBm. The phase shifts of the elements at the RIS are set to a unity matrix.
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 134

TPTF, DenseNet
OPTF, ResNet
0.9 OPTF, DenseNet
TPTF, ResNet
0.8

Cumulative distribution function

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0.7 0.75 0.8 0.85 0.9 0.95 1
Accuracy

Figure 6.8: The CDF of detection accuracy with different kinds of algorithm.

0.955

0.95

0.945

0.94
Accuracy

0.935

0.93

0.925
N =10
fau
0.92 N
fau
=15
Nfau =20
0.915
0 5 10 15 20 25 30
SNR (dB)

Figure 6.9: Detection accuracy versus SNR (dB) across different Nfau .

During the training stage of the proposed TPTL algorithm for detecting the faulty

elements of the RIS, it is assumed that the MU is located at pu = (30, 6, 0.5) m, and

the number of faulty elements within the RIS is assumed to be no more than 15, i.e.,

Nfaul ≤ 15. Furthermore, it is assumed that there are K = 9 SAs, each containing

Ns = 9 reflecting elements. The dataset consists of 20, 000 randomly generated fault

scenarios, split into training and testing sets at a ratio of 0.8 : 0.2. During the training

stage of the proposed TEDF and TEDS algorithms for estimating the location of the

MU, it is assumed that the MU is uniformly distributed within three grids measuring

10 m in length, 10 m in width, and heights of 0.5 m, 1.5 m, and 2 m, respectively. The

dataset consists of 60, 000 samples, each sample has the MU’s coordinates, the condition

of faulty elements on the RIS panel, and the corresponding signals received by both the
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 135

BS and RIS. The dataset is split into training and testing sets at a ratio of 0.8 : 0.2.

Fig. 6.6 and Fig. 6.7 showcase the train and test loss curves of the two phases of

the TPTL algorithms. Fig. 6.8 presents the CDF of faulty element detection accuracy 4

for different algorithms. The blue curve, labeled ‘TPTF DenseNet’, signifies the TPTF

algorithm transferring the DenseNet 121 architecture. Conversely, the purple curve,

labeled ‘TPTF ResNet’, represents the TPTF algorithm using ResNet 18. Furthermore,

the yellow and red curves, labeled ‘OPTF DenseNet’ and ‘OPTF ResNet’ respectively,

represent the benchmarks of the one phase transfer learning (OPTF) algorithm trans-

ferring DenseNet 121 and ResNet 18. As depicted in Fig. 6.8, the proposed TPTF,

by sequentially detecting the faulty SAs and internal faulty elements, optimally har-

nesses DenseNet 121 or ResNet 18 for targeted faulty element detection, thus amplifying

accuracy. Contrarily, the OPTF algorithm does not first detect the SAs and then the

elements within the SAs, resulting in the number of elements (or labels) to be classified

surpassing the performance capacity of DenseNet (or ResNet). Consequently, the perfor-

mance of the OPTF algorithm is inferior. Significantly, the proposed ‘TPTF DenseNet’

achieves commendable detection outcomes, with a 80% probability of attaining a 91%

detection accuracy, surpassing ‘TPTF ResNet’, ‘OPTF DenseNet’, and ‘OPTF ResNet’,

which have the accuracy of 76%, 74%, and 71% at the identical confidence interval. In

summary, Fig. 6.8 emphasizes the superior detection performance of the propose ‘TPTF

DenseNet’ algorithm.

Fig. 6.9 presents the detection accuracy of the proposed TPTF algorithm under

different SNRs, showcasing its performance in test scenarios with a specified number of

faulty elements, Nfau = 10, 15, 20. These numbers are exclusive to the testing phase. As

delineated in Fig. 6.9, the TPTF algorithm maintains remarkable efficacy, achieving a

92% detection accuracy at an SNR of 10 dB even faced with 20 faulty elements. Moreover,

in challenging conditions with an SNR as low as 0 dB, the detection accuracy remain
4
The accuracy metric in the study requires identifying all faulty elements for a correct detection.
The dataset consists of 20, 000 randomly generated fault scenarios, split into training and testing sets at
a ratio of 0.8 : 0.2.
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 136

0.955

0.95

0.945

Accuracy
0.94
Nfau=10, p u= (15,6,2) m
Nfau=10, p u = (40,6,0) m
0.935 Nfau=10, p u = (30,6,0.5) m
Nfau=15, p u = (30,6,0.5) m
Nfau=15, p u = (40,6,0) m

0 5 10 15 20 25 30
dB

Figure 6.10: Detection accuracy versus SNR (dB) across different positions
and Nfau .

robust at approximately 91.7%, 93.6%, and 94.5% for 20, 15, and 10 faulty elements,

respectively. This underscores the robustness of the proposed TPTF detection algorithm

in low SNR scenarios 5 .

Fig. 6.10 delves into the TPTL algorithm’s effectiveness in identifying faulty ele-

ments under variable MU locations, particularly focusing on pu = (15, 6, 2) m and

pu = (40, 6, 0) m with up to 10 faulty elements, extending to pu = (40, 6, 0) m for

scenarios encompassing up to 15 faulty elements. It is observed that there exists a

marginal improvement as the distance between the MU and the RIS-BS system narrows,

especially in environments characterized by lower SNR values. However, the detection

accuracy remains the same when SNR is larger than 20 dB. This reveals that changes

in the MU’s position do have some effect, but it’s not significant. Moreover, as the SNR

increases, this impact diminishes.

The essence of these observations lies in their testament to the TPTL algorithm’s

versatile performance across diverse spatial arrangements and signal conditions. This

demonstrates not just the algorithm’s capability to adjust its performance based on the

proximity and configuration of the MU relative to the BS and RIS, but also its robustness

and reliability in a variety of operational environments. This versatility is paramount

5
Exploring additional evaluation metrics such as precision, recall, and F-score represents an interest-
ing avenue for future work, promising to provide deeper insights into model performance.
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 137

Figure 6.11: Test loss and train loss of TEDF algorithm.

Figure 6.12: Test loss and train loss of Stage I of TEDS algorithm.

Figure 6.13: Test loss and train loss of Stage II of TEDS algorithm.

for ensuring the algorithm’s applicability and efficacy in real-world deployments, where

environmental and situational variability is the norm.

Fig. 6.11, Fig. 6.12, and Fig. 6.13 showcase the train and test loss curves of two algo-

rithms, i.e., TEDF and TEDS, employed for MU coordinate estimation. Specifically, Fig.

6.11 displays the iterative convergence behaviors of train and test losses for the TEDF

algorithm. It is evident that by approximately 20 epochs, both train and test losses

for the TEDF algorithm have converged. Although there are minor fluctuations in the

test loss afterwards, their magnitude is negligible. Additionally, in Stage I of the TEDS
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 138

algorithm, the CNN-VAE network is employed, which uses low-dimensional signals, y,

to reconstruct complete high-dimensional RIS received signal at the RIS, ȳr . Fig. 6.12

presents the train and test losses of the proposed CNN-VAE network. During training,

there is a rapid decline in both train and test losses initially, followed by a more gradual

reduction. This demonstrates the complexity of reconstructing higher-dimensional infor-

mation from low-dimensional information as well as the challenge of restoring signals

expected to be received by faulty elements. The losses appear to stabilize around the

500-th iteration. Lastly, Fig. 6.13 represents the convergence behavior for train and test

loss in Stage II of the TEDS algorithm, where losses converge around the 15-th epoch

with negligible fluctuations. Comparing Fig. 6.13 and Fig. 6.11 which employ transfer

learning with Fig. 6.12, the advantages of transfer learning are evident. By leveraging

transfer learning, the number of required train epoches is significantly reduced while

maintaining accuracy, making it highly suitable for the complex and dynamic environ-

ments anticipated in 6G. This offers valuable insights for future RIS-aided localization

systems.

Fig. 6.14 employs normalized mean square error (NMSE) to compare the localization

accuracy of the TEDF and TEDS algorithms for Nfau = 10 and Nfau = 15. The curves

labeled ‘TEDF, Nfau = 10’ and ‘TEDF, Nfau = 15’ correspond to scenarios using BS

information as a fingerprint, while ‘TEDS, Nfau = 10’ and ‘TEDS, Nfau = 15’ represent

the utilization of RIS information as a fingerprint for the same Nfau values. The figure

clearly shows that employing high-dimensional RIS information significantly improves

localization performance compared to using lower-dimension BS information, even in

the presence of faulty elements. This underscores the advantage of leveraging high-

dimensional RIS information. Furthermore, the figure demonstrates that the localization

performance remains relatively stable with minimal degradation when the number of

faulty elements increases from 10 to 15, emphasizing the robustness of the proposed

algorithms against faulty element variations.

Fig. 6.14 also presents the curves representing scenarios where RIS information is
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 139

used as a fingerprint but without processing through the CNN-VAE network. These

scenarios are for RIS with no faulty elements and RIS containing 15 faulty elements,

labeled as ‘RIS Inf., Nfau = 0’, and ‘RIS Inf., Nfau = 15’, respectively. As a comparison

in the picture, the curves ‘TEDS, Nfau = 10’ and ‘TEDS, Nfau = 15’ show a minor gap

with the ‘RIS Inf., Nfau = 0’ curve but a significant gap with the ‘RIS Inf., Nfau =

15’curve. This observation suggests the effectiveness of the proposed CNN-VAE network

in reconstructing the complete high-dimensional RIS information and demonstrates that

the localization accuracy can be regained with the proposed algorithms.

Furthermore, Fig. 6.14 also presents the scenarios that train the TEDS algorithm

with the same loss but without the detection result, which is labeled as ‘w/o Detection,

Nfau = 15’. It is shown that even though without the detection result of the faulty

elements, the proposed integrated CNN-VAE network can still reconstruct the essential

localization-related information. Additionally, Fig. 6.14 includes a curve labeled ‘Opt.

RIS, Nfau = 0’, which represents scenarios where the information of a RIS without faulty

elements is used as a fingerprint, with its phase shifts optimized to maximize the received

SNR [69]. As a comparison, there exists marginal gap between ‘Opt. RIS, Nfau = 0’ and

‘RIS Inf., Nfau = 0’, highlighting the robustness of the RIS information based localization

algorithm against the unoptimized phase shifts. It is conjectured that the reason for this

phenomenon lies in that the localization accuracy is dominated by the high-dimensional

RIS information, and therefore the phase shift design does not play an important role.

Fig. 6.15 utilizes NMSE to compare the performance of the TEDF and TEDS algo-

rithms under two distinct SNR conditions, i.e., SNR = 10 dB and SNR = 30 dB. The

red, purple, green, and orange curves respectively represent ‘TEDF, SNR = 10 dB’,

‘TEDF, SNR = 30 dB’, ‘TEDS, SNR = 10 dB’, and ‘TEDS, SNR = 30 dB’. As depicted

in Fig. 6.15, the localization performance of the TEDF algorithm exhibits sensitivity to

SNR variations, leading to a measurable reduction. In contrast, the TEDS algorithm

showcases remarkable robustness against SNR variations, as evidenced by the near over-

lap of the red and green curves. It is conjectured that the reason for this phenomenon
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 140

0.9

Cumulative distribution function

0.8

0.7

0.6
TEDF, Nfau= 15
0.5 TEDF, Nfau= 10

0.4 RIS Inf., Nfau= 15

TEDS, Nfau= 15
0.3
TEDS, Nfau= 10
0.2 RIS Inf., Nfau= 0
Opt. RIS, N fau=0
0.1
w/o Detection, Nfau= 15
0
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
NMSE

Figure 6.14: Comparison of TEDF and TEDS algorithms when Nfau = 10, 15.

0.9

0.8
Cumulative distribution function

0.7

0.6

0.5

0.4

0.3

0.2
TEDF, SNR=10 dB
TEDF, SNR=30 dB
0.1 TEDS, SNR=10 dB
TEDS, SNR=30 dB
0
0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055
NMSE

Figure 6.15: Comparison of TEDF and TEDS algorithms when SNR = 10, 30
dB.

lies in that the employed CNN-VAE network can effectively against the variations of

SNR. Such observations underscore the superiority of the proposed TEDS algorithm,

suggesting that the localization algorithm remains highly effective even in diverse and

intricate communication scenarios. This further supports the notion that localization

performance can indeed benefit from the high-dimensional RIS information even with

the faulty elements.

Fig. 6.16 compares the NMSE of the proposed TEDS algorithm and TEDF algorithm

with the near field and far field effects. As shown in Fig. 6.16, it can be observed that

when employing the TEDF algorithm, the localization performance under near-field

conditions significantly surpasses that in the far-field. However, with the application
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 141

0.9

Cumulative distribution function

0.8

0.7

0.6

0.5

0.4

0.3 NF, TEDS, N fau=15

0.2 FF, TEDS, N =15
fau
FF, TEDF, N fau=15
0.1
NF, TEDF, N fau=15
0
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05
NMSE

Figure 6.16: NMSE Comparison between the far field and near field effect
when employing the TEDS algorithm and TEDF algorithm.

of the TEDS algorithm, while localization accuracy in the near-field still outperforms

the far-field, the advantage is not as pronounced. It is inferred that this phenomenon is

attributable to the richer localization information available in the near-field. Nonetheless,

the dominant role of the high-dimensional RIS information reconstructed by the TEDS

algorithm mitigates the disparity in gains between the near and far fields. Moreover,

Fig. 6.16 also corroborates the effectiveness of the proposed algorithm in both near-field

and far-field conditions, underscoring its robustness and adaptability across different

propagation environments.

Fig. 6.18 compares the NMSE of the proposed TEDS algorithm with the integrated

CNN-VAE network and with the CNN-Interpolation method. As shown in Fig. 6.18, the

CNN-Interpolation method shows some improvement over the TEDF algorithm. How-

ever, there is still a significant gap when compared to the TEDS algorithm that utilizes

the CNN-VAE network. This comparison validates the superiority of the proposed TEDS

algorithm, especially in addressing the capability of restoring the lost RIS information.

Fig. 6.17 compares the NMSE of the proposed TEDS algorithm with the RCNR

algorithm in the last chapter, the proposed TEDS algorithm outperforms the RCNR

algorithm and the classical ‘AlexNet’ algorithm, which validate the effectiveness of the

proposed TEDS algorithm.

Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 142

Cumulative distribution function (CDF)

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2 ResNet opt.

AlexNet
0.1 TEDS

0
0.005 0.01 0.015 0.02 0.025
NMSE

Figure 6.17: NMSE Comparison between TEDS algorithm and RCNR algo-
rithm.

0.9
Cumulative distribution function

0.8

0.7

0.6

0.5

0.4

0.3

0.2
TEDS, CNN-Interpolation
0.1 TEDF
TEDS, CNN-VAE
0
0 0.01 0.02 0.03 0.04 0.05
NMSE

Figure 6.18: NMSE Comparison between the Integrated CNN-VAE method

and the CNN-Interpolation method.

6.8 Summary

This chapter investigates the localization algorithm design based on the high-dimensional

RIS information in the presence of faulty RIS elements. This chapter designs a TPTL

algorithm for faulty element detection. The TEDS algorithm is proposed to clear their

harmful impact, unleashing the gain of RIS on localization. At Stage I, the com-

plete high-dimensional RIS received signal is reconstructed from the low-dimensional

BS received signal. At Stage II, this reconstructed RIS information is input to the

transferred DenseNet to estimate the location of the MU. Further, a low-complexity

TEDF algorithm is designed for localization which only utilizes the low-dimensional BS
Chapter 6. RIS-Aided Localization Algorithm Design: Overcoming Faulty Elements 143

information.
Chapter 7

RIS-aided Localization Algorithm

Design: Tracking in Near-field

7.1 Introduction

To overcome the high path-loss attenuation and fully unleash the gain of RIS, the RIS is

expected to be constructed on an extremely large scale and equipped with large numbers

of passive and low-cost reflecting elements. However, with a large physical dimension,

the communication between an RIS and users will be located in the near field, rather

than the conventional far field [6, 55, 77]. Recognizing this, recent research [18, 31–35]

has significantly advanced the design of RIS-aided localization algorithms in the presence

of the near-field effect. These studies have introduced various methodologies to enhance

localization accuracy, from creating virtual line-of-sight (VLoS) paths using RIS [31] to

employing advanced optimization techniques for precision improvement [32]. Efforts have

also been made to explore the theoretical limits of holographic RIS-aided localization

system [18] and address specific challenges such as phase-dependent amplitude variations

[34] and bi-static sensing for fixed transmitter and receiver systems [35].

Besides the research limitation concerning the near field, another important challenge

144
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 145

is about mobile localization aiming to accommodate the movement of MU. Applications

such as robotic vacuum cleaners and smart machinery in factories are particularly reliant

on precise mobile localization to function effectively. While the static approaches devel-

oped in earlier works [18, 31–35] laid a strong foundation, their applicability is some-

what constrained in scenarios featuring mobile users. Addressing this need, more recent

research [78–80] have switched the focus towards designing the mobile tracking algo-

rithms. Specifically, the authors in [78] delved into jointly optimizing the RIS reflection

coefficients and BS precoding to adeptly track MUs in scenarios with multi-RIS setups.

Additionally, recognizing challenge that integrated the mobility and LoS obstructions,

Mei et al. in [79] developed a multi-user tracking system using the extended Kalman

filter (EKF). Meanwhile, Rahal et al. in [80] presented a three-step algorithm for simul-

taneously estimating MU position and velocity, accounting for Doppler effects in the

near-field.

However, the above-mentioned mobile tracking algorithms mainly rely on the signals

received at the BS, i.e., BS information. Nevertheless, the high cost and power consump-

tion associated with radio-frequency chains often result in a limited number of antennas

at the BS. This limitation reduces the dimension of the BS information, thereby dimin-

ishes the localization accuracy. However, an important yet unexplored potential is the

use of rich location-related information embedded in the signals received by the RIS,

i.e., RIS information. Due to its composition of a vast number of reflecting elements,

the RIS panel can catch a much higher dimension of useful information. Actually, the

idea of using RIS information for localization is perfectly matched with the newly pro-

posed concept, the XL-RIS. With an extremely large size, near-field effect exhibits and

therefore each individual reflecting element on XL-RIS could collect unique AoA and

ToA information. Exploiting the large number of diverse AoA and ToA patterns across

the entire XL-RIS, a detailed and nuanced dataset is established and realize localization

with high granularity and accuracy.

Despite the attractive benefits, the localization design based on XL-RIS information
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 146

is indeed a highly challenging task [13]. This challenge arises from the complexity of

reconstructing high-dimensional XL-RIS information. On the one hand, the passive

nature of XL-RIS poses a challenge in actively acquiring the signals. On the other hand,

when employing the low-dimensional BS information to obtain the high-dimensional

XL-RIS information, the result is not unique. This is because the same BS information

corresponds to multiple distinct XL-RIS information. The uncertainty in the signal

reconstruction process requires innovative algorithms to accurately capture the XL-RIS

information.

Furthermore, for mobile tracking tasks, the works in [78–80] utilized the velocity to

model the trajectory of the MUs, employing traditional tracking algorithms like the EKF

for predicting trajectory of the MU. However, traditional tracking algorithms face chal-

lenges in complex environments, reducing localization accuracy. Specifically, the EKF

linearizes MUs’ nonlinear trajectories and thereby introduces the linearization error,

especially exacerbated in indoor settings with multi-path and near-field effects, signif-

icantly lowering precision. Additionally, traditional tracking algorithms inadequately

leverage time-series data, particularly in recognizing long-term dependencies, restricting

prediction accuracy and model adaptability.

Against the above-mentioned background, this chapter investigates the mobile track-

ing problem with complex environments and near-field effects, and proposes an effective

algorithm to enable the use of XL-RIS information. This chapter first employs the

data-driven methods. i.e., CNN, to design an XL-RIS-IR algorithm. Building on this,

This chapter proposes a comprehensive data-driven framework for MU tracking, which

includes a Feature Extraction Module and a Mobile Tracking Module. The contributions

are outlined as follows:

1) This chapter proposes an XL-RIS-IR algorithm for reconstructing XL-RIS infor-

mation from BS information. The XL-RIS-IR algorithm includes a data process-

ing module for preprocessing low-dimensional BS information, a DenseNet module

for feature extraction, and an output module to produce the reconstructed high-
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 147

dimensional RIS information.

2) As the first component of the proposed tacking framework, this chapter designs

a Feature Extraction Module for extracting the essential features from the recon-

structed XL-RIS information. This module encompasses three distinct extractors:

a CNN extractor for extracting the spatial features, a time and frequency (T&F)

extractor aimed at extracting time and frequency domains features, and a near-

field AoAs extractor for capturing the AoAs features within XL-RIS. Finally, these

features are then combined into a final feature vector for mobile tracking.

3) The Mobile Tracking Module, serving as the second component of the proposed

tacking framework, employs an Auto-encoder (AE) configured with a stacked Bi-

LSTM encoder and a conventional LSTM decoder. This sophisticated AE processes

the time-varying sequence derived from the final feature vector to accurately predict

the MU’s 3D coordinates for the upcoming time slot.

4) Simulation results reveal that the tracking accuracy of the proposed tacking frame-

work is enhanced by utilizing the high-dimensional XL-RIS information. More-

over, the proposed framework exhibits considerable robustness against variations

in SNR. Additionally, numerical analyses confirm that the framework outperforms

established benchmark models in MSE metrics, highlighting its superior tracking

precision.

7.2 System Model

Consider an XL-RIS aided UL localization system where an MU transmits the pilot

signals to the BS via the XL-RIS. The BS is equipped with a UPA comprising M1 ×M2 =

M antennas, while the MU has a single antenna. Additionally, the XL-RIS is equipped

with a UPA comprising N1 × N2 = N reflecting elements.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 148

7.2.1 Geometry and Channel Model

This subsection introduces the geometry configuration of the system, for the BS, the

XL-RIS, and the MU. The BS is positioned at a known location pBS ∈ R3 , while the

center of the XL-RIS is denoted by pRIS ∈ R3 , and the location of the n-th RIS element

is represented by pn ∈ R3 for 1 ≤ n ≤ N . The MU’s location, denoted by p ∈ R3 , is

unknown and needs to be estimated.

In this system, the MU is assumed to lie in the Fresnel (radiative) near-field region of

the XL-RIS. This region is defined by the distance d between the MU and the XL-RIS,

which satisfies the following condition:

r
D3 2D2
0.62 ≤d≤ , (7.1)
λ λ

where λ is the carrier wavelength, and D represents the XL-RIS aperture size, defined

as the largest distance between any two elements of the XL-RIS. This condition ensures

that the MU is located within a specific range that allows for effective interaction with

the radiative fields emitted by the XL-RIS, facilitating precise estimation of the MU’s

location.

Given the near-field condition, the communication channel between the MU and XL-

RIS is modeled. This channel is characterized by a multipath scenario, where the signal

propagation can be divided into two primary components: LoS and NLoS. The LoS

component represents the direct propagation path from the RIS to the MU without

any obstruction. Conversely, the NLoS component encompasses multiple indirect paths

where the signal is reflected by various scatterers in the environment, such as walls,

furniture, and human bodies.

For the LoS path from the MU to the XL-RIS, the steering vector a(p) ∈ CN ×1 is

introduced, which models the signal phase variations across the RIS elements, given the

MU’s position p. This steering vector is derived with respect to the XL-RIS’s center,
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 149

denoted by pRIS , which is given as

 
2π
[a(p)]n = exp −j (∥p − pn ∥ − ∥p − pRIS ∥) , (7.2)
λ

for n ∈ {1, · · · , N }, where pn represents the position of the n-th element on the XL-RIS.

In addition, the NLoS paths, which result from reflections of Ms scatterers, are mod-

eled through steering vectors that account for the additional path lengths. Specifically,

for the ms -th scatterer, the steering vector is given by


2π
[ams (p)]n = exp − j (∥p − pn ∥ + ∥qms − pn ∥

λ


+ ∥p − qms ∥ − ∥p − pRIS ∥), (7.3)



for ms ∈ {1, · · · , Ms } and n ∈ {1, · · · , N }, where qms denotes the position of the ms -th

scatterer.

Therefore, the overall complex channel from the MU to the XL-RIS, incorporating

both LoS and NLoS components, is thus represented as

Ms
X
h(p) = αa(p) + αms ams (p), (7.4)
ms =1

where α and αms denote the complex channel attenuation coefficients for the LoS path

and each NLoS path through the ms -th scatterer, respectively.

For the communication channel between the BS and the XL-RIS, it is assumed that no

NLoS paths exist due to the BS being situated close to the XL-RIS [81]. The normalized

channel gain for the LoS path between BS and the XL-RIS, is represented by G, whose
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 150

(m, n)-th element is given by [45]

 
 2π
G(m, n) = expj rm,n , (7.5)

λ

where rm,n denotes the path length between the m-th antenna element of the BS and the

n-th reflecting element. Letting β denote the common large-scale path loss attenuation,

the complex channel gain of the channel between the BS and the XL-RIS is then given

by

H = βG. (7.6)

7.2.2 Signal Model

To motivate the deployment of the XL-RIS, it is assumed that the direct path between

the BS and the MU is blocked, e.g., due to the buildings, cars or trees. Letting yr denote

the XL-RIS received signal, it is formulated as

yr = h(p)s, (7.7)

where s denotes the pilot signal transmitted by the MU. Then, defining the phase shift

vector of the XL-RIS as ω = [ω1 , · · · , ωN ]T ∈ CN ×1 , the UL signal received by the BS

can be expressed as

y = Hdiag(ω)h(p)s + n, (7.8)

where diag(ω) and n denote the phase shift matrix of the XL-RIS and the zero-mean

additive white Gaussian noise, respectively.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 151

Figure 7.1: The structure of the XL-RIS-IR algorithm.

7.3 XL-RIS Information Reconstruction Algorithm

Traditionally, localization algorithm design are based on the BS information, e.g., y. the

estimated position of pu is defined as p̂u , and then the localization problem is formulated

as

min ||H(y) − pu ||2 , (7.9)

p̂u

where H(·) denotes the non-linear function between y and p̂u . The above problem

can be solved with several localization algorithms which have been extensively investi-

gated [31, 32, 40]. However, constraints such as cost and deployment conditions often

limit the number of antennas, resulting in low-dimensional BS information that provides

insufficient localization features, thereby potentially reducing localization accuracy. In

contrast, XL-RIS can achieve larger sizes at a lower cost, enabling them to contain

more sufficient location information, which can be used to enhance localization preci-

sion. However, due to the passive nature of RISs, it is not feasible to directly obtain

XL-RIS information. This necessitates the reconstruction of XL-RIS information to fully

leverage its potential for improving localization accuracy.

By defining the reconstructed high-dimensional XL-RIS information as ŷr , the RIS

information reconstruction problem can be formulated as

min ∥yr − ŷr ∥22 , (7.10)

ŷr
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 152

which cannot be solved with traditional convex optimization methods, since yr is unknown.

Hence, this chapter design a deep learning-based XL-RIS-IR algorithm to solve Problem

(7.10). As shown in Fig. 7.1, the XL-RIS-IR algorithm includes three modules: data

processing module, DenseNet module, and output module, whose details are given in the

following.

7.3.1 Data Processing Module

The data processing module is designed to process the input BS information y. Since

y is a complex vector, it is divided into its real and imaginary parts, represented as

R(y) and I(y), respectively. Then, since y consists of the features of the horizontal

angle domain and the vertical angle domain of the BS, the two vectors R(y) ∈ CM ×1

and I(y) ∈ CM ×1 are reshaped into matrices, denoted as MRy ∈ RM1 ×M2 and MIy ∈

RM1 ×M2 , respectively. Then, the reshaped matrices of the real and imaginary parts are

concatenated in a 3D tensor as
 
MRy  2×M1 ×M2
Ty =  ∈R . (7.11)
MIy

However, the values of M1 and M2 may not be directly compatible with CNNs’ archi-

tectures. Hence, an upsampling technique is applied:

Tup = Upsample(Ty ) ∈ R2×256×256 . (7.12)

The upsampling technique is employed to adjust the dimensions of the tensor Ty to the

desired dimensions of 2 × 256 × 256. It typically involves pixel value interpolation to

expand the image or tensor size, redistributing pixel values accordingly. To make Tup

suitable for processing by CNNs, which expects an input tensor of the form (3, 256, 256),

a Con layer is introduced to transform the bi-channel tensor to a tri-channel one:

TCon = Con(Tup ). (7.13)

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 153

Figure 7.2: Network structure of DenseNet.

7.3.2 DenseNet Module

DenseNet [24], inspired by the foundational principles of ResNet [25], reuses the features

from earlier layers. However, different from the skip connections which might bypass a

couple of layers, DenseNet establishes connections between every layer and every other

layer, ensuring a more robust feature propagation. This approach, while reducing the

parameter count, allows DenseNet to potentially surpass the performance of ResNet.

DenseNet is distinguished by their excellent capability in feature extraction and is

used for extracting features from the low-solution images. Besides, the intrinsic hierar-

chical structure of DenseNet allows for the map from the low-level features to high-level

features. Therefore, DenseNet is adopted as a main module for reconstructing RIS infor-

mation.

The design of DenseNet is illustrated in Fig. 7.2, and its architecture comprises two

main components: Dense Modules and Transition Modules, which will be introduced as

follows.

[Link] Dense Module

As depicted in Fig. 7.2, the Dense Module comprises multiple Dense Blocks. The

details of a Dense Block are shown in Fig. 7.3. Each Dense Block starts with a batch

normalization (BN) layer, followed by a layer rectified linear unit (ReLU) activation
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 154

Figure 7.3: Network structure of a Dense Block.

function layer, and concludes with a 3×3 Con layer. A notable characteristic of the Dense

Module is its connectivity: Every block incorporates the feature maps of all preceding

blocks within the module. This design ensures efficient feature reuse and a rich flow of

information throughout the network [24].

[Link] Transition Module

From Fig. 7.2, it is observed that the Transition Module, which acts as a connector

between Dense Modules, starts with a 1×1 Con layer, followed by a 2×2 average pooling

layer, effectively reducing the spatial dimensions of the feature maps and streamlining

them for either the next Dense Module or the concluding output layer.

7.3.3 Output Module

After the Dense module, an output module is designed to output the reconstructed XL-

RIS information. Within the output module, a ReLU layer is used to introduce the

non-linearity, enabling the network to capture and express the complex patterns of the

high-dimensional RIS information. Besides, the ReLU layer can address the vanishing

gradient problem, facilitating better learning in deep neural networks.

Following the ReLU layer, a linear layer is employed for directly outputting the desired

high-dimensional XL-RIS information. This layer serves to consolidate the complex fea-

tures extracted by prior layers and transforms them into a specific high-dimensional

format, such as the required XL-RIS information, effectively aligning the network’s out-

put with the target data structure.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 155

7.3.4 Training of XL-RIS-IR Algorithm

Building upon the detailed structure of the XL-RIS-IR algorithm, the direct training

process is initiated. This involves inputting the received signal y into the proposed algo-

rithm. The training aims to adjust the algorithm’s parameters to achieve an optimized

output, namely best reconstructing the XL-RIS information. As training progresses and

converges, the algorithm efficiently transforms the low-dimensional BS information into

the high-dimensional XL-RIS information.

7.4 Employing XL-RIS Information For Tracking

Once obtaining the XL-RIS information, it can be used to estimate the 3D coordinates of

the MU. Hence, in this section, the time-varying sequence consisted by the reconstructed

RIS Information, ŷr , is employed to determine the MU’s location when the MU is moving

within a 3D indoor space.

7.4.1 Tracking Problem Formulation

The movement of the MU causes variations in the XL-RIS information, ŷr , acting as a

unique fingerprint that aids in determining the MU’s 3D coordinates. Therefore, this

chapter aims to exploit the temporal dynamics of the RIS information to accurately

track the MU’s movements.

To accurately document the dynamic nature of the MU’s movement, a time-varying

sequence is assembled that reflects the changes in ŷr across a specified interval. This

sequence, indexed by time slot t, represents the moments when the MU transitions to a

subsequent position. Formally, the sequence is expressed as [ŷr (t), ŷr (t + 1), . . . , ŷr (t + T − 1)],

covering T discrete time slots. It serves to chronicle the evolution of RIS information as

the MU traverses various locations.

For predicting the MU’s future location at time step T , it is considered that the

XL-RIS information, ŷr , is a dynamic fingerprint that evolves over time with the MU’s
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 156

movement. Therefore, it is formulated as

p̂u (T ) = F([ŷr (t), ŷr (t + 1), . . . , ŷr (t + T − 1)]), (7.14)

where F(·) is the nonlinear mapping function that transforms the time-varying sequence

of ŷr into the estimated position, p̂(T ), of the MU at the subsequent time slot. Accord-

ingly, the mobile localization problem can be formulated as

min ||Q(p̂u (T ) − pu (T )||2 . (7.15)

p̂u (T )

aiming to minimize the discrepancy between the estimated position of the MU and its

actual position at time T , thereby enhancing the accuracy of MU tracking.

7.4.2 Tracking Framework Assisted by XL-RIS Information

In addressing the challenges of tracking the MU, this chapter introduces a comprehensive

framework designed to accurately track the MU location in the near-field scenario. The

proposed framework comprises two main modules: the Feature Extraction Module and

the Localization Module.

[Link] Feature Extraction Module

The Feature Extraction Module is pivotal in transforming raw XL-RIS information into

a meaningful representation that encapsulates spatial and temporal cues essential for

localization. This module integrates three key components: a CNN extractor, a T&F

feature extractor and a near-field AoAs extractor.

CNN extractor The CNN extractor is adept at capturing spatial features from the

XL-RIS information by leveraging its hierarchical structure. Through Con layers, the

CNN learns to identify patterns and textures in ŷr that are indicative of the MU’s 3D

coordinates, making it a powerful tool for spatial feature extraction.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 157

Time & frequency (T&F) feature extractor The T&F feature extractor is designed

to extract time domain and frequency domain characteristics to capture a comprehensive

representation of the XL-RIS information.

Near-field AoAs extractor To better leverage the rich AoA information available

in near-field scenarios, it is formulated as designed a specialized AoAs extractor. This

extractor is aimed at capturing AoAs features within the XL-RIS panel.

[Link] Localization Module

Within the Localization Module, an innovative architecture featuring an Auto-encoder

(AE) with a stacked Bi-LSTM in the encoder and a standard LSTM in the decoder is

utilized to forecast the MU’s future locations.

The AE architecture employs a stacked Bi-LSTM layer within the encoder to delve

into the temporal patterns and dependencies in the XL-RIS information, capturing both

past and future contexts for a comprehensive understanding. The decoder, equipped with

an LSTM layer, then reconstructs the future position from this encoded representation,

facilitating precise prediction of the MU’s location at any given time T .

By integrating these two modules, the framework is specifically designed to enhance

the utilization of XL-RIS information for MU tracking. Further details and insights into

the XL-RIS-assisted tracking framework will be elaborated in the following two sections.

7.5 XL-RIS Information Based Feature Extraction

As the first module of the proposed tracking framework, the Feature Extraction Mod-

ule includes three kinds of extractors, which are CNN extractor, T&F extractor, AoA

extractor, respectively, whose details are given as follows.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 158

7.5.1 CNN Extractor

To efficiently handle the high-dimensional reconstructed XL-RIS information ŷr , and

extract potential location-related features, a CNN-based extractor is designed in this

subsection, which includes several blocks. The details of the designed module are given

as follows.

[Link] Data Processing Block

To accurately handle the preprocessing step for the reconstruction of high-dimensional

XL-RIS information ŷr , before feature extraction via a CNN block, it is crucial to detail

the preparation process of the input data. This preprocessing involves separating the

XL-RIS information into its real and imaginary components, denoted as R(ŷr ) and

I(ŷr ), respectively. Given that yr encapsulates features from both the horizontal and

vertical angle domains of the XL-RIS, the vectors R(yr ) ∈ CN ×1 and I(yr ) ∈ CN ×1 are

reshaped into matrices, denoted as MRyr ∈ RN1 ×N2 and MIyr ∈ RN1 ×N2 , respectively.

These matrices are then concatenated to form a 3D tensor represented as

 
M
 Ryr  2×N1 ×N2
T yr = ∈R . (7.16)
MIyr

This 3D tensor, Tyr , serves as the input to the CNN, which is designed to extract

location-related features from the complex XL-RIS information.

[Link] CNN Block

The CNN architecture processes the input tensor Tyr through several layers, each designed

to perform specific operations that contribute to feature extraction. The operation in a

Con layer can be described as

Fl = fReLU (Wl ∗ Tyr + bl ), (7.17)

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 159

where Fl is the feature map produced by the l-th layer, Wl represents the weights of the

filters, bl is the bias term, ∗ denotes the convolution operation, and fReLU = max(0, x)

denotes the ReLU function, which is widely used as a nonlinear activation function in

CNNs.

Consequently, the pooling layer is employed to reduce the dimension of each feature

map while retaining the most important information, typically using max pooling. This

procedure can be formulated as

Pl = fMaxPool (Fl ), (7.18)

where Pl represents the pooled feature map.

Then, the flattening layer is employed to convert the 2D feature maps into a 1D

feature vector, preparing the data for processing by fully connected layers, which can be

expressed as

gflattened = fFlatten (PL ), (7.19)

where gflattened is the flattened feature vector, and L is the last Con layer.

Finally, the fully connected layer is applied to allocate weights to the flattened fea-

ture vector to produce the final feature vector T , encapsulating the critical information

extracted from the input. This can be represented as

fCNN = WFC · gflattened + bFC , (7.20)

where WFC and bFC are the weights and bias of the fully connected layer, respectively.

To provide a clear understanding of the CNN block used for feature extraction from

the preprocessed XL-RIS information, the following Table 7-A outlines each layer’s def-

inition and corresponding dimensions.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 160

Table 7-A: CNN Architecture for Feature Extraction

Layer Definition Dimension
Input Preprocessed Tensor 2 × N1 × N2
(1)
Conv Layer 1 Convolution + ReLU Activation Nf × H1 × W1
(1)
Pooling Layer 1 Max Pooling Nf × H1′ × W1′
(2)
Conv Layer 2 Convolution + ReLU Activation Nf × H2 × W2
(2)
Pooling Layer 2 Max Pooling Nf × H2′ × W2′
(2)
Flattening Layer Flatten Feature Maps Nf · H2′ · W2′
Fully Connected Layer Dense Layer Nf

(1) (2)
In this table, Nf , Nf denote the number of filters in the first and second convolu-

tional layers, respectively. H1 , W1 , H2 , W2 represent the dimensions of the feature maps

after each convolutional layer before pooling. H1′ , W1′ , H2′ , W2′ are the dimensions of the

feature maps after pooling. Nf is the dimension of the final feature vector fCNN .

7.5.2 T& F Feature Extractor

In addition to the spatial features extracted by the CNN, feature set are enhanced with

frequency domain characteristics and time statistical measures to capture a comprehen-

sive representation of the XL-RIS information. Herein, the feature extraction steps are

detailed, including normalization, frequency domain transformation, and time statistical

analysis.

[Link] Normalization

Before feature extraction, normalization is a critical preprocessing step to ensure con-

sistent scale across different signals, facilitating more effective analysis and comparison.

Normalization adjusts the amplitude of ŷr to a common scale, enhancing the robustness

of subsequent feature extraction steps. The process can be mathematically represented

as
ŷr − min(ŷr ) · 1
ŷr′ = , (7.21)
max(ŷr ) − min(ŷr )

where ŷr′ is the normalized signal, 1 is a all-ones vector with the same dimensions as ŷr ,

and min(ŷr ) and max(ŷr ) are the minimum and maximum values of ŷr , respectively.
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 161

[Link] Frequency Domain Transformation

Performing the fast fourier transform (FFT) and extracting information in the frequency

domain are essential for wireless localization. Firstly, the frequency domain analysis

helps in identifying the presence of distinct path components in a multi-path environ-

ment, where signals from a single source reach the receiver via multiple paths due to

reflections, diffractions, and scattering. Each path can have different lengths, and conse-

quently, different phase shifts when observed in the frequency domain, providing valuable

information about the relative distances and angles of arrival, which are directly related

to the position of the source. Besides, the spectral content of a signal, as revealed by the

FFT, can be used to extract features that are sensitive to the localization of the MU. For

example, the Doppler shift, a change in frequency due to the relative motion between the

transmitter and receiver, can be precisely identified in the frequency domain, offering

insights into the velocity and direction of movement of the source relative to the MU.

Hence, FFT is applied to ŷr′ to analyze its frequency components, crucial for under-

standing the signal’s spectral properties. This procedure is formulated as

ŷr(FFT) = fFFT (ŷr′ ) ∈ CNyf ×1 , (7.22)

(FFT)
where ŷr denotes the frequency domain representation of the normalized signal ŷr′ ,
(FFT)
Nyf denotes frequency points of the ŷr . To quantify the signal’s energy distribu-

tion across the frequency spectrum, the spectral energy is computed using the squared

magnitude of the FFT output, which is expressed as

e = |ŷr(FFT) |◦2 ∈ CNyf ×1 , (7.23)

where (·)◦2 denotes the element-wise square operation, transforming the complex signal

into a vector of real-valued energy measurements across the frequency spectrum.

This energy vector e is then normalized to a probability distribution, which is given

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 162

by
e
p(ŷr(FFT) ) = PN , (7.24)
yf
nyf =1 e

allowing for the calculation of spectral entropy, which can be used as an essential feature

of the reconstructed RIS information.

[Link] Feature Extraction

Several features are extracted from both the time-domain and frequency-domain repre-

sentations of the signal to capture its essential characteristics.

Frequency Domain Features The spectral energy and entropy are computed as key

features to capture the signal’s frequency distribution. First, the spectral energy is given

by
Nyf
X
EFFT = e. (7.25)
nyf =1

Then, the spectral entropy is formulated as

Nyf
X
HFFT = − p(ŷr(FFT) ) log p(ŷr(FFT) ). (7.26)
nyf =1

Time Domain Features Except the frequency domain features, time domain fea-

tures are also essential to the localization, especially for the MU is moving with time.

To extract the time domain features, the statistical time domain features are employd,

including mean, variance. They are calculated directly from ŷr′ , which is given as

µ = fmean (ŷr′ ),

σ 2 = fvar (ŷr′ ), (7.27)

where µ is the mean, σ 2 is the variance.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 163

[Link] Feature Vector Construction

Finally, the extracted features are concatenated to form a comprehensive feature vector

fT&F , which is written as

fT&F = [µ, σ 2 , EFFT , HFFT ]T , (7.28)

providing a multi-dimensional representation of ŷr for further localization.

7.5.3 AoAs Feature Extraction

In near-field localization applications, extracting AoA information within the XL-RIS

panel is crucial as the near-field effects offer richer and more detailed AoA data compared

to far-field scenarios. Hence, this subsection provide an approach to estimate the AoAs

from the reconstructed XL-RIS information ŷr .

[Link] Angle Estimation

Given the large number of different AoAs at different reflecting elements within the XL-

RIS to be estimated in near-field conditions, which increases the estimation complexity

and causes signal energy to be dispersed, it is challenging to accurately estimate the AoA

across the entire XL-RIS panel. Consequently, the XL-RIS is partitioned into sub-arrays

for angle estimation to manage the computational complexity and concentrate the signal

energy within each sub-array, thus enhancing the accuracy and efficiency of AoA esti-

mations. By doing so, the wavefront impinging on each sub-array can be approximated

as a planar wavefront due to the limited physical dimension of each sub-array, minimiz-

ing the impact of wavefront curvature. This approximation allows for the effective use

of high-resolution AoA estimation techniques like MUSIC, which are predicated on the

assumption of planar wavefronts, thus enabling accurate AoA estimation within each

sub-array’s constrained spatial domain.

It is assumed that there are K sub-arrays within the XL-RIS. To estimate K AoAs

using the MUSIC algorithm across K sub-arrays. For the k-th sub-array, where k ∈
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 164

{1, · · · , K}, it is formulated as the steering vector of the k-th sub-array expressed as

(e) (a)
ak (θk , ϕk ) = ak (θk ) ⊗ ak (θk , ϕk ), (7.29)

where

−j2π(Nk,e −1)dx cos θk T

 
(e)
ak (θk ) = 1, . . . , e λc , (7.30)
 −j2π(Nk,a −1)dx sin θk cos ϕk
T
(a)
ak (θk , ϕk ) = 1, . . . , e λc , (7.31)

where Nk,e and Nk,a denote the numbers of elements within the k-th sub-array in the

elevation and azimuth directions, respectively, dx denotes the distance between adjacent

array elements, and λc is the carrier wavelength.

The received signal at the k-th sub-array of the XL-RIS can be expressed as

yr,k = αak (θk , ϕk )s, (7.32)

where ak (θk , ϕk ) denotes the array response vector. The signal s corresponds to the trans-

mitted signal. Due to the passive nature of XL-RIS, yr,k cannot be directly observed.

To address this challenge, the reconstructed XL-RIS information of the k-th sub-array

is employed, which is denoted as ŷr,k . Utilizing this information, it is able to estimate

the AoAs (e.g., θk , ϕk ) across the K sub-arrays. The details are given as follows.

Signal Covariance Matrix First, by employing the FFT technique to ŷr,k , it is

formulated as

(FFT) ′
ŷr,k = fFFT (ŷr,k ),

′ ŷr,k − min(ŷr,k ) · 1
ŷr,k = , (7.33)
max(ŷr,k ) − min(ŷr,k )
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 165

Then, the frequency covariance matrix is given as follows:

Rf,k
Nyf
1 X (FFT) (FFT) (FFT) (FFT)
= (ŷr,k [n] − ȳr,k )(ŷr,k [n] − ȳr,k )H , (7.34)
Nyf
nyf =1

(FFT) (FFT)
where ȳr,k is the mean vector of ŷr,k .

Eigen Decomposition For each sub-array k, an eigen decomposition on Rf,k is per-

formed, which is formulated as

Rf,k = EΛE H , (7.35)

where E is a matrix whose columns are the eigenvectors of Rf,k , and Λ is a diago-

nal matrix containing the eigenvalues in descending order. H denotes the conjugate

transpose.

Identifying Signal and Noise Subspaces Following the results of decomposition,

it is proceeded to determine the signal and noise subspaces, represented by Esignal and

Enoise , respectively.

MUSIC Spectrum The MUSIC spectrum for estimating θk and ϕk is given by the

inverse of the projection of the steering vector onto Enoise . For each potential pair of

(Θk , Φk ), MUSIC spectrum can be derived as

1
PMUSIC,k (Θk , Φk ) = H
. (7.36)
ak (Θk , Φk )H Enoise,k Enoise,k ak (Θk , Φk )

Estimation of θk and ϕk The AoAs, θk and ϕk , are estimated by identifying the

peaks in the MUSIC spectrum, which is given as

(θ̂k , ϕ̂k ) = arg max PMUSIC,k (Θk , Φk ), (7.37)

(Θk ,Φk )
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 166

where (θ̂k , ϕ̂k ) are the estimated elevation and azimuth angles for the k-th sub-array.

[Link] Feature Vector Construction

Upon obtaining the AoAs estimates, (θ̂k , ϕ̂k ), from the MUSIC algorithm for each of

the K sub-arrays, these estimations are aggregated into a structured feature set. The

collective feature set F, representing the AoA estimates across all sub-arrays, can be

constructed as
n o
FAoA = (θ̂1 , ϕ̂1 ), (θ̂2 , ϕ̂2 ), . . . , (θ̂K , ϕ̂K ) . (7.38)

Alternatively, for numerical and analytical convenience, these AoA estimates can be

organized into a vector form fAoA , which is given as

h i
fAoA = θ̂1 , ϕ̂1 , θ̂2 , ϕ̂2 , · · · , θ̂K , ϕ̂K . (7.39)

7.5.4 Final Feature Vector Construction

After combing the feature vector fCNN , the T&F feature vector fT&F and the AoAs

feature vector fAoA , it is formulated as the final feature vector, denoted as fFinal , which

is given by

fFinal = [fCNN , fT&F , fAoA ] , (7.40)

which will be fed into the following Mobile Tracking Module.

7.6 XL-RIS Information Based Tracking Algorithm

In this section, the final feature vector fFinal is employed for tracking the MU using

the data-driven DL based framework. Data-driven DL based framework is now widely

designed and employed for MU localization and channel estimation. However, all these

DL-based localization work ignore the time sequence of the recieved signal when the MU

is moving on a specific trajectory. In another word, the adjacent observations of the

time-varying recieved signal can be further utilized for more precise estimation.
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 167

There are two widely employed techniques called recurrent neural network (RNN)

and LSTM on solving the time-varying tasks, such as natural language processing. Both

techniques consider the information from the previous entered data and the currently

entered data to estimate. Specifically, RNN has feedback loops to maintain information

over time. However, it is difficult for RNN on learning long-term temporal dependencies

due to the vanishing gradient problem. Differently, LSTM introduces input and forget

gates for better preservation of long-term dependencies on dealing gradient flow.

Therefore, this section aims to develop an AE equipped with a stacked Bi-LSTM

encoder and a standard LSTM decoder, referred to as the Stacked Bi-LSTM AE algo-

rithm. This design is intended to accurately predict the 3D coordinates of the MU in

the subsequent time slot.

7.6.1 LSTM

Before presenting the details of the Stacked Bi-LSTM AE algorithm, the details of LSTM

network are given as follows.

The LSTM network, an enhancement over traditional RNN, introduces a gradient-

based learning algorithm capable of connecting past information to the current task.

LSTMs process time-sequence data in a chain-like structure, moving in a forward direc-

tion. The unique aspect of LSTM architecture compared to standard RNNs is its hidden

layers within each LSTM cell. The details of LSTM network are given as follows.

[Link] Forget Gate

The first layer of the LSTM network is called forget gate which is also named forget

layer, which is formulated as

 
ft = σ Wf x(t) + W̃f x̃(t − 1) + bf , (7.41)
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 168

where σ denotes the sigmoid activation function, x̃(t − 1) denotes the information passed

from previous layer, x(t) represents the input feature vector at time step t, while W̃f ,

Wf and bf denote the weight and bias of forget gate, respectively.

[Link] Input Gate

Consequently, the second layer is called input layer which is also named input gate, which

can be expressed as
 
it = σ Wi x(t) + W̃i x̃(t − 1) + bi , (7.42)

where W̃i , Wi and bi denote the weight and bias of input gate, respectively.

[Link] Cell Input State

Then, to update the cell input state, e.g., ct , tanh function is employed and calculated

as
 
ct = tanh Wc x(t) + W̃c x̃(t − 1) + bc . (7.43)

where W̃c , Wc and bc denote the weight and bias of cell input state, respectively.

[Link] Output Gate

To decide the next hidden state, the output gate ot is utilized, which is written as

 
ot = σ Wo x(t) + W̃o x̃(t − 1) + bo , (7.44)

where W̃o , Wo and bo denote the weight and bias of output layer, respectively.

[Link] Output State

Apart from the cell input state cg,t in the hidden layers, there are two other cell states

in the structure: the previous cell output state c(t − 1) fed in the current LSTM cell,

and the current cell output state c(t) passed to the next LSTM cell. The output state
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 169

at the current time t can be formulated as

c(t) = ft ⊙ c(t − 1) + it ⊙ ct . (7.45)

where ⊙ denotes the Kronecker product.

[Link] Output Layer

The conventional input layer is x(t) at the time slot t. Finally, the output layer can be

formulated as

x̃(t) = ot ⊙ tanh(c(t)). (7.46)

7.6.2 Stacked Bi-LSTM

To accurately capture the dynamic nature of the input time-varying feature sequence, the

Bi-LSTM is employed when constructing encoder. This design choice is pivotal because

it allows the model to maintain the temporal continuity and dependencies between each

time step in the sequence. By integrating both forward and backward LSTM cells, the

architecture ensures that information from both past and future contexts is utilized,

facilitating a more comprehensive understanding of the temporal patterns. This bidirec-

tional approach is especially beneficial for tasks like mobile localization tracking, which

recognizing the sequence’s temporal evolution.

Furthermore, it has been proved that by stacking multiple hierarchical models, the

performance can be improved progressively. Hence, a stacked structure is adopted where

the output from the lower layer is then fed as the input to the upper layer with Ls ≥ 2

Bi-LSTM layers. The workflow of the stacked Bi-LSTM considers both forward and

backward directions and deeper structure with T time slots. During T time slots, it is

formulated as the time sequence given by Ffea = [fFinal (t), fFinal (t+1), · · · , fFinal (t+T )].

By employing this time varying sequence from the past T time slots, the location of the

MU at T + 1 time slot can be estimated via the design AE Bi-LSTM. The details of the

AE Bi-LSTM are given as follows subsections.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 170

7.6.3 Encoder Construction for Stacked Bi-LSTM AE

In this subsection, the details of the encoder constructed by the stacked Bi-LSTM are

presented, which is designed to extract the time-varying sequence Ffea .

[Link] The first Bi-LSTM Layer

In a Bi-LSTM layer, the input time-varying feature sequence is processed to capture

both past (forward) and future (backward) dependencies within the sequence. The first

Bi-LSTM is employed for estimating the location of the MU at the T + 1 time slot by

using the time-varying sequence Ffea , whose details are given as follows.

Forward Pass in Bi-LSTM The forward pass processes the input sequence from

time t to time t + T , capturing information as it moves forward in time.

Backward Pass in Bi-LSTM The backward pass processes the sequence in reverse,

from time t + T to time t, capturing future context for each time step. The operations

are analogous to those of the forward pass but applied in reverse order.

[Link] Stacked Ls Bi-LSTM Layers

After introducing the first Bi-LSTM layer, the construction of the Encoder in an AE Bi-

LSTM setup involves stacking additional Bi-LSTM layers to further refine and process

the features. This stacked structure enables the model to capture more complex temporal

dependencies in the sequence data, enhancing its predictive capabilities for tasks like

estimating the MU location based on historical data.

Following the first Bi-LSTM layer, Ls Bi-LSTM layers are incorporated. Each sub-

sequent layer receives the output from all time slots of its predecessor as input. Con-

sequently, each Bi-LSTM layer builds upon the patterns and dependencies identified in

the time-varying sequence by the layer before it, extracting increasingly abstract features

from the sequence.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 171

Figure 7.4: 3D random walk example.

Within each Bi-LSTM layer, temporal steps are processed akin to the first layer,

employing forward and backward LSTM units to assimilate information from past and

future contexts, respectively. Thus, the model’s output at each time step synthesizes cur-

rent input features with contextual insights drawn from both preceding and succeeding

steps.

To mitigate overfitting, Dropout layers can be interspersed between Bi-LSTM layers.

These Dropout layers randomly nullify a fraction of the neurons’ activations during

training, thereby enhancing the model’s generalization ability.

The output generated by the last Bi-LSTM layer in the Encoder is typically used to

represent a compressed and high-level feature representation of the entire input sequence.

This representation captures the core information and patterns of the input sequence,

providing necessary context for the Decoder.

7.6.4 Decoder Construction for Stacked Bi-LSTM AE

After constructing the encoder of the Stacked Bi-LSTM AE, the next step is to build the

decoder to predict the location of the MU at the T + 1 time slot. The decoder utilizes
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 172

Figure 7.5: 3D wave example.

Figure 7.6: 3D spiral example.

the high-level features generated by the encoder to perform this prediction.

[Link] Initialization

The encoder processes the entire input sequence and abstracts the input through its

hierarchy into a high-level representation that captures the essence of the input data,

including timing dependencies and patterns. Hence, the final hidden state generated by
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 173

Figure 7.7: Test and train losses of random walk trajectory.

Figure 7.8: Test and train losses loss of spiral trajectory.

Figure 7.9: Test and train losses loss of wave trajectory.

the encoder is employed for the initial hidden state of the decoder. By initializing the

decoder with this state, a strong starting point for the generation process is provided,

ensuring that the decoder has all the context needed to generate accurate and relevant

output.
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 174

[Link] Layers Construction for Decoder

The decoder consists of one LSTM layers, one dropout layer, one dense layer, where the

details of each layers are given as follows.

LSTM Layer In the context of predicting the location at time T + 1, a single LSTM

layer is chosen for the decoder instead of a Bi-LSTM or a Stacked Bi-LSTM configuration

due to the nature of the prediction task. Bi-LSTM layers, which process information in

both forward and backward directions, are more suited to situations where the entire

sequence is known and the context from future steps is as important as the past, which is

not the case when predicting future locations based solely on past and current informa-

tion. Stacked LSTMs, while powerful for capturing complex patterns in more intricate

sequences, increase model complexity and computational cost, which might not be nec-

essary for tasks where the encoder has already provided a condensed and meaningful

representation of the input sequence. Thus, a single LSTM layer is efficient and suffi-

cient for leveraging the encoded context to predict the next location, minimizing the risk

of overfitting while maintaining model simplicity and interpretability.

Dropout Layer Dropout is included as a regularization technique. By randomly

omitting portions of the layer’s inputs, dropout forces the network to learn more robust

features that are useful in conjunction with many different random subsets of the other

neurons. This reduces the likelihood of overfitting, especially when training data is

limited or when the architecture is complex relative to the amount of training data.

Dense Layer Dense layer is used to produce a specific output size from the LSTM

layer’s output, matching the dimensionality of the prediction task (e.g., 3D coordinates

for location). The linear activation in the output Dense layer allows for the direct

prediction of continuous values. The linear activation in the output Dense layer allows

for the direct prediction of continuous values.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 175

7.7 Simulation Results

In this simulation, a setup in a 3D space is considered, involving an MU, an RIS, and a

BS. The RIS is composed of N1 × N2 = 2 × 50 = 100 elements. The center of the RIS

is located at pr = (6, 0, 2) m. The BS comprises M1 × M2 = 4 × 4 = 16 antennas. The

center of the BS is positioned at pb = (0, 5, 1.5) m. The MU-RIS link includes P = 10

propagation paths. During the training phase, the transmit power of the MU is set to

30 dBm, while in the testing phase, the transmit power of the MU varies from 0 dBm

to 30 dBm. The experiments are conducted within the mmWave frequency of 200 GHz,

with the MU confined to a space of 8.8 × 8.8 × 3 m3 , ensuring the setup adheres to the

near-field conditions as derived from the formula for Fresnel regions. Specifically, the

MU is positioned such that it remains within the critical near-field range of the RIS,
q
3 2
calculated as between 0.62 Dλ and 2Dλ , enhancing the precision and reliability of the

localization simulations. The phase shift matrix of the RIS is configured to a unit matrix

(all ones), simplifying the analysis while focusing on fundamental interactions between

the MU and the RIS.

To simulate the movement trajectories of objects, it is formulated as defined three

types of paths: random walk, wave, and spiral, to model the possible trajectories that

objects may follow in reality. These trajectories are generated with a random starting

point within a space limited to 10 × 10 × 3 m3 . Over 10000 such paths are generated,

each comprising 11 steps. For each step along a trajectory, the signal received at both

the BS and RIS are generated, denoted by yt and yr , respectively. Besides, the principles

behind the generation of the three types of trajectories are as follows:

1. Random Walk Trajectory: Start by selecting a random starting point within

the indoor space. For each step in the trajectory, generate a unit vector pointing

in a random direction within 3D space, and move along this unit vector’s direction

by a step size randomly chosen from the range (0, 1]. An example of random walk

is given in Fig. 7.4.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 176

10

10*log10(MSE)
Proposed, BS inf.
w/o NF feature, RIS inf.
-5 w/o CNN feature, RIS inf.
Proposed, RIS inf.
Real RIS inf.
-10

-15

-20
-5 0 5 10 15 20 25 30
SNR (dB)

Figure 7.10: The MSE comparison of different inputs for proposed algorithm
versus SNR (dB).

2. Wave Trajectory: To generate a wave trajectory, start by picking a random

point. Then draw this path along the x-axis, setting its length and dividing it

into steps. Then, based on the wave’s amplitude of 2 and wavelength of 5, a sine

function is employed to modify the y-coordinates. This creates the familiar peaks

and troughs of a wave. Throughout, the z-coordinate is kept constant, ensuring

the wave moves in a two-dimensional plane. Fig. 7.5 presents an example of the

wave trajectory.

3. Spiral Trajectory: Create a spiral path by using the formula r = a + bθ. Here, r

is how far each point is from the random start point, a = 0.1 is the starting distance

from the center, and b = 3 decides how much this distance increases. θ increases

step by step, making the spiral trajectory. By changing from polar coordinates

(using r and θ) to regular 3D coordinates, this spiral in a 3D space can be drawed.

Fig. 7.6 presents an example of the generated spiral trajectory.

Fig. 7.7 displays the iterative convergence behaviors of training and test losses for the

random walk trajectory. It is evident that by approximately 100 epochs, both training

and test losses for the random walk trajectory have converged. Although there are minor

fluctuations in the test loss afterwards, their magnitude is negligible. Furthermore, Fig.

7.8 and Fig. 7.9 depict the convergence trends of training and testing losses for the
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 177

10

-10

10*log10(MSE)
-20 Random walk, BS information
Wave, BS information
Spiral BS information
Random walk, RIS information
-30 Wave, RIS information
Spiral, RIS information

-40

-50
-5 0 5 10 15 20 25 30
SNR (dB)

Figure 7.11: The MSE comparison of different types of trajectory versus SNR
(dB).

spiral and wave trajectories, respectively. For both trajectory types, the training and

testing losses converge by the 15-th epoch, exhibiting minimal fluctuations thereafter.

Due to the regular patterns of the spiral and wave trajectories, the proposed algorithm

is capable of quickly learning their movement patterns. In contrast, the random walk,

characterized by its unpredictable nature, requires more epochs for the neural network

to learn. However, the eventual convergence of the random walk as well demonstrates

the effectiveness of the proposed algorithm.

Fig. 7.10 shows the mean squared error (MSE) comparison among three scenarios:

without near-field feature extraction, without CNN feature extraction, and with feature

extraction. Fig. 7.10 reveals that utilizing near-field feature extraction alone yields better

results than relying solely on CNN feature extraction. This underscores the necessity of

extracting near-field features, suggesting that CNN may not fully capture all relevant

localization information on its own. Besides, as shown in Fig. 7.10, compared to the

curve without near feature extraction, incorporating the near-field feature extraction

significantly enhances performance, highlighting the value of tailored near-field feature

extraction in improving localization accuracy.

Fig. 7.10 also displays the MSE performance curve for localization using actual RIS

information, labeled as ‘Real RIS inf.’. The closeness of this curve to the ‘Proposed, RIS
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 178

10

10*log10(MSE)
-5
100 elements, BS information
400 elements, BS information
-10 100 elements, RIS information
400 elements, RIS information

-15

-20

-25
-5 0 5 10 15 20 25 30
SNR (dB)

Figure 7.12: The MSE comparison of different number of reflecting elements

within the RIS versus SNR (dB).

inf.’ curve, which represents localization using reconstructed RIS information, validates

the effectiveness of the proposed algorithm for XL-RIS information reconstruction. It

shows that the localization accuracy achieved with the reconstructed information closely

approximates that achieved with the actual RIS information, supporting the reliability of

the reconstruction approach in mimicking real RIS information for localization purposes.

Fig. 7.11 presents the comparison of MSE for different trajectories versus the SNR

in dB. As depicted in Fig. 7.11, the MSE values associated with both the wave and

spiral trajectories are significantly lower than those of the random walk trajectory. This

distinction arises because the wave and spiral trajectories exhibit structured movement

patterns, so that the deployed Stacked Bi-LSTM AE network can analyze more effec-

tively. Besides, it is also noteworthy that utilizing the received signal at the BS, e.g.,

BS information, results in substantially poorer performance compared to the algorithms

when XL-RIS information is used. This further validates the efficacy of employing high-

dimensional RIS information for near-field tracking, as proposed in the study. Moreover,

as shown in Fig. 7.11, when employing the high-dimensional RIS information, the accu-

racy maintains a certain level of accuracy even under low SNR conditions with a random

walk trajectory, which demonstrates robustness of proposed algorithms against low SNR.
Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 179

10

10*log10(MSE)
AE-Stacked-Bi-LSTM, BS information
AE-LSTM, RIS information
AE-Stacked-LSTM, RIS information
-5 AE-Stacked-Bi-LSTM, RIS information

-10

-15

-20
-5 0 5 10 15 20 25 30
SNR (dB)

Figure 7.13: MSE performance across SNR (dB) for localization algorithms
with and without feature extraction.

As shown Fig. 7.12, it is evident that increasing the number of reflective elements

in the RIS panel enhances localization accuracy, for both cases employing BS informa-

tion and RIS information. However, it can be observed that increasing the number of

reflecting elements from 100 to 400, the improvement of localization accuracy when using

BS information is not significantly pronounced. Besides, it necessitates a higher SNR

to achieve better localization results. In contrast, when employing RIS information,

as shown in Fig. 7.12, the accuracy with 400 reflective elements is significantly higher

than that with 100. This emphasizes the importance of using XL-RIS information for

improved localization accuracy in near-field scenarios.

Fig. 7.13 showcases a comparison of MSE across SNR for different algorithms. As

shown in Fig. 7.13, ‘AE-Stacked-LSTM’ refers to an architecture that incorporates mul-

tiple layers (stacked) but does not employ bidirectional processing, while ‘AE-LSTM’

utilizes a simpler, single-layer structure without the stacking or bidirectional enhance-

ments. The observed results clearly demonstrate that the proposed algorithm, referred

to as ‘AE-Stacked-Bi-LSTM’, significantly surpass the other algorithms in localization

performance. This superiority suggests that the bidirectional processing, when combined

with a stacked approach, effectively captures the temporal dependencies and spatial char-

acteristics of the RIS information, thus enhancing the predictive accuracy and proving

the proposed methodology’s efficacy.

Chapter 7. RIS-aided Localization Algorithm Design: Tracking in Near-field 180

7.8 Summary

This chapter introduces a mobile tracking framework leveraging the high-dimensional

signal received from XL-RIS information. An XL-RIS-IR algorithm is presented to

reconstruct the high-dimensional XL-RIS information from the low-dimensional BS infor-

mation. Building on this, a comprehensive framework is proposed for mobile tracking,

consisting of a Feature Extraction Module and a Mobile Tracking Module. Besides, it has

been validated that the proposed framework exhibited the robustness to SNR variation.
Chapter 8

Conclusion

In this chapter, the contributions of this thesis are summarised and some possible future

work is presented.

8.1 Summary of Contributions and Insights

This thesis develops effective localization algorithms for 6G RIS-aided wireless systems,

tailored to address practical challenges. Chapter 3 proposes a comprehensive framework

that analyzes the AEE using its PDF and designs a corresponding localization algorithm

to mitigate the impact of AEE. Chapter 4 integrates AoAs and TDoAs at RISs, utilizing

their distinct error properties to improve localization accuracy, with specific performance

analyses in terms of bias. In Chapters 6 and 7, the utilization of high-dimensional RIS

information is explored. Chapter 6 presents an effective detection algorithm designed

to identify and mitigate the impact of faulty elements. Chapter 7 shifts focus to the

mobility of the MU, where an effective tracking algorithm is developed.

In Chapter 3, a comprehensive framework is developed to analyze angle estimation

errors and to design a 3D localization algorithm for mmWave systems. Initially, the AoAs

at the RISs are estimated using the 2D-DFT algorithm. The angle estimation error is

subsequently analyzed based on the PDF properties inherent to the 2D-DFT algorithm.

181
Chapter 8. Conclusion 182

To simplify the complex geometric representation of this error PDF, a first-order linear

approximation of trigonometric functions is applied. Additionally, the variance of the

error is derived using this PDF. Finally, the TSWLS algorithm is employed to estimate

the 3D position of the MU using the estimated AoAs and the calculated non-Gaussian

variance.

Chapter 4 investigates the algorithm design and analysis for an RIS-aided localization

IoT system, utilizing multiple RISs to address the non-Gaussian AEE and Gaussian

TDoA error. The mWLS algorithm is designed to derive a closed-form expression for

the position of the MU by leveraging the distinct error properties of the AoA and TDoA.

Finally, a unique bias analysis is proposed to evaluate the performance of the algorithm.

In Chapter 5, the multiple MUs localization problem in MIMO TDD mmWave sys-

tems aided by the RIS is investigated. To address the property of RIS-aided localization

system, Chapter 5 derives the expression for STCRV as a new type of fingerprint. A

novel RCNR learning algorithm is proposed to simultaneously predict the 3D positions

of the multiple MUs by using the STCRV fingerprint as input.

Chapter 6 explores the design of a localization algorithm that utilizes high-dimensional

RIS information in scenarios with faulty RIS elements. A TPTL algorithm is developed

for detecting faulty elements. Subsequently, the TEDS algorithm is introduced to miti-

gate their adverse effects, thereby enhancing the localization accuracy. At Stage I, the

complete high-dimensional RIS received signal is reconstructed from the low-dimensional

BS received signal. At Stage II, this reconstructed RIS information is processed using

a transferred DenseNet model to estimate the location of the MU. Additionally, a low-

complexity TEDF algorithm, which relies solely on the low-dimensional BS information,

is developed for localization.

Chapter 7 introduces a mobile tracking framework that leverages high-dimensional

signals received from XL-RIS. The XL-RIS XL-RIS-IR algorithm is presented to recon-

struct the high-dimensional XL-RIS information from the low-dimensional BS data.

Chapter 8. Conclusion 183

Building on this, a comprehensive framework for mobile tracking is proposed, which

includes a Feature Extraction Module and a Mobile Tracking Module. Furthermore, it

has been validated that this framework demonstrates robustness to variations in NR.

8.2 Future Research

In this section, some possible extensions of this thesis are proposed in the following.

8.2.1 Passive Sensing

Passive sensing is emerging as a pivotal direction for future RIS-aided localization tech-

nologies due to its potential advantages in cost-efficiency, privacy protection, and system

performance. This thesis investigates the scenarios where the BS proactively sends pilot

signals for localization, integrating seamlessly with the passive sensing approach. Utiliz-

ing existing ambient wireless signals, such as Wi-Fi or cellular signals, for environmental

sensing, passive sensing avoids the need for additional transmitters or frequent signal

emission, thereby reducing both hardware and operational costs [82–84]. Moreover, it

involves no additional signal transmission from the sensing devices themselves, which

minimizes interference with external systems and enhances privacy, a crucial benefit for

applications requiring stringent confidentiality, such as indoor monitoring and user loca-

tion privacy. The integration of RIS enhances this approach by dynamically controlling

signal paths to improve coverage and signal reachability, particularly in NLoS conditions

where traditional sensors fail. This capability not only extends the effective monitoring

area but also ensures high-quality sensing even in traditionally challenging environments.

Through these strengths, passive sensing with RIS support stands out as a robust solution

for advanced localization scenarios, capitalizing on the rich environmental information

available through ambient signals.

Chapter 8. Conclusion 184

8.2.2 Active RIS

This thesis investigates the RIS-aided localization systems aided by a passive RIS, focus-

ing on the limitations imposed by the inherent “double path loss” that signals suffer

when reflected by a passive RIS, traversing both the BS-RIS and RIS-user paths with-

out amplification, thus significantly weakening them. To address these challenges and

enhance system performance, the introduction of an active RIS is proposed and inves-

tigated in [85, 86]. Unlike passive RIS, which only adjusts phase shifts using low-power

components like PIN diodes or varactors, an active RIS also amplifies the received signals,

compensating for the losses incurred over the dual paths and restoring signal strength to

practical levels. This capability not only mitigates the double path loss attenuation but

also supports advanced transmission schemes such as the two-timescale transmission,

crucial for massive MIMO systems. Consequently, active RIS emerges as a promising

direction for future development in RIS-aided localization systems [87], promising to

expand their applicability and effectiveness in complex and adverse environments where

maintaining signal integrity is critical. The transition towards active RIS technologies

reflects a pivotal shift towards more robust, reliable, and precise localization capabilities,

potentially revolutionizing the landscape of wireless communications.

8.2.3 Reinforcement Learning

A promising future direction for RIS-aided localization systems involves the application

of reinforcement learning (RL) to optimize RIS phase shifts in real-time, enhancing com-

munication efficiency and localization accuracy simultaneously [88–90]. Reinforcement

learning, a type of machine learning, operates by learning to make decisions from direct

interaction with the environment through a process of trial and error, using a reward sys-

tem to identify favorable outcomes. In the context of this thesis, which does not explore

phase optimization, RL can be employed to dynamically adjust the RIS’s phase settings

in response to changes in the MU’s position, modeled as a Markov decision process. This

method uses feedback from the localization algorithms to iteratively improve the phase

shift strategy, aiming to maximize communication quality while minimizing localization

Chapter 8. Conclusion 185

error. Integrating RL into RIS-aided systems holds the potential to significantly enhance

both the adaptability and effectiveness of localization and communication processes in

complex and dynamic wireless environments.

8.2.4 Fundamental Limitation of RIS Information

This thesis explores the utilization of RIS information for localization, leveraging the

low-dimensional BS information to reconstruct the high-dimensional RIS information.

However, this reconstruction process inherently faces a fundamental limit due to the

information loss during signal transmission from the BS information to the RIS infor-

mation, which significantly impacts the achievable precision in localization. A critical

future research direction is the application of information theory to rigorously study

and define these fundamental limits. Information theory can provide a mathematical

framework to quantify the maximum amount of information. This framework will help

in understanding the extent of information degradation and its effect on localization

accuracy.

8.2.5 Extending Tracking Beyond Predefined Areas

This thesis investigates the application of RIS technology for enhanced tracking within

a fixed simulation area. However, the mobility of MUs introduces significant challenges

when they move beyond the predefined boundaries. This movement can lead to unfore-

seen variations in signal reception, complicating the effectiveness of the established local-

ization framework. Future research can therefore focus on predicting and modeling user

movements that extend outside the predetermined area. This involves developing adap-

tive signal models that can dynamically adjust to changes in user location and optimize

signal strength accordingly.

Further investigation could explore techniques for enhancing signal reach and integrity

in extended areas, possibly through the strategic deployment of additional RIS panels

or the use of advanced algorithms that adjust the phase and amplitude of the RIS
Chapter 8. Conclusion 186

dynamically. Such adaptations would aim to ensure consistent localization accuracy

even as users move into previously unconsidered zones, thus expanding the practical

applicability of RIS-based systems in real-world scenarios.

8.2.6 Optimization of RIS Phase Shift Matrix

In this thesis, the investigation of localization algorithms across different scenarios assumes

that the RIS phase shift matrix is set to all ones, without delving into the specific effects

that different RIS configurations might have on localization accuracy. The simplification

of assuming an all-one matrix allows for focusing on the primary mechanisms and per-

formance of the localization algorithms themselves. However, the phase shifts in an RIS

are crucial as they directly influence beamforming and, therefore, the precision of signal

direction and localization.

Future research could fruitfully explore the impact of varying RIS phase shift config-

urations on localization precision. Such studies would be critical to understanding how

subtle changes in the RIS matrix can alter the propagation patterns of the transmitted

signals, potentially leading to either enhancements or degradation in localization accu-

racy. Analyzing these effects would involve detailed simulations and theoretical models

to determine optimal phase settings under various environmental and operational condi-

tions. This line of inquiry would significantly enrich the field by linking RIS technology

more closely with practical localization tasks, potentially leading to advancements in the

accuracy and reliability of RIS-enhanced localization systems.

Chapter 8. Conclusion 187

References
[1] C.-X. Wang et al., “On the Road to 6G: Visions, Requirements, Key Technologies,
and Testbeds,” IEEE Comm. Surv. Tutor., vol. 25, no. 2, pp. 905–974, 2023.
[2] H. Chen et al., “RISs and Sidelink Communications in Smart Cities: The Key
to Seamless Localization and Sensing,” IEEE Commun. Mag., vol. 61, no. 8, pp.
140–146, 2023.
[3] C.-X. Wang et al., “Pervasive Wireless Channel Modeling Theory and Applications
to 6G GBSMs for All Frequency Bands and All Scenarios,” IEEE Trans. Veh.
Technol., vol. 71, no. 9, pp. 9159–9173, 2022.
[4] H. Chen et al., “A Tutorial on Terahertz-Band Localization for 6G Communication
Systems,” IEEE Commun. Surv. Tutor., vol. 24, no. 3, pp. 1780–1815, 2022.
[5] K. Meng et al., “Throughput Maximization for UAV-Enabled Integrated Periodic
Sensing and Communication,” IEEE Trans. Wireless Commun., vol. 22, no. 1, pp.
671–687, 2023.
[6] H. Wymeersch et al., “Radio Localization and Sensing Part II: State-of-the-Art and
Challenges,” IEEE Commun. Lett., vol. 26, no. 12, pp. 2821–2825, 2022.
[7] J. He et al., “Beyond 5G RIS mmWave Systems: Where Communication and Local-
ization Meet,” IEEE Access, vol. 10, pp. 68 075–68 084, 2022.
[8] C.-X. Wang et al., “6G Wireless Channel Measurements and Models: Trends and
Challenges,” IEEE Veh. Technol. Mag., vol. 15, no. 4, pp. 22–32, 2020.
[9] T. Wu, C. Pan, Y. Pan, S. Hong, H. Ren, M. Elkashlan, F. Shu, and J. Wang,
“Two-step mmwave positioning scheme with RIS-Part II: Position estimation and
error analysis,” 2022.
[10] J. He et al., “3D Localization With a Single Partially-Connected Receiving RIS:
Positioning Error Analysis and Algorithmic Design,” IEEE Trans. Veh. Technol.,
vol. 72, no. 10, pp. 13 190–13 202, 2023.
[11] Y. Sun et al., “A 3D Non-Stationary Channel Model for 6G Wireless Systems
Employing Intelligent Reflecting Surfaces With Practical Phase Shifts,” IEEE Trans.
Cogn. Commun. Netw., vol. 7, no. 2, pp. 496–510, 2021.
[12] J. Huang et al., “Reconfigurable Intelligent Surfaces: Channel Characterization and
Modeling,” Proc. IEEE, vol. 110, no. 9, pp. 1290–1311, 2022.
[13] C. Pan et al., “An Overview of Signal Processing Techniques for RIS/IRS-Aided
Wireless Systems,” IEEE J. Sel. Topics Signal Process., vol. 16, no. 5, pp. 883–
917, 2022.
[14] K. Zhi, C. Pan, H. Ren, and K. Wang, “Uplink achievable rate of intelligent reflecting
surface-aided millimeter-wave communications with low-resolution adc and phase
noise,” IEEE Wireless Communications Letters, vol. 10, no. 3, pp. 654–658, 2021.
[15] G. Zhou et al., “A Framework of Robust Transmission Design for IRS-Aided MISO
Communications With Imperfect Cascaded Channels,” IEEE Trans. Signal Pro-
cess., vol. 68, pp. 5092–5106, 2020.
Chapter 8. Conclusion 188

[16] K. Zhi et al., “Two-Timescale Design for Reconfigurable Intelligent Surface-Aided

Massive MIMO Systems With Imperfect CSI,” IEEE Trans. Inf. Theory, vol. 69,
no. 5, pp. 3001–3033, 2023.
[17] T. Wu et al., “3D Positioning Algorithm Design for RIS-aided mmWave Systems,”
2022.
[18] X. Gan et al., “Near-Field Localization for Holographic RIS Assisted mmWave
Systems,” IEEE Commun. Lett.s, vol. 27, no. 1, pp. 140–144, 2023.
[19] H. Chen, P. Zheng, M. F. Keskin, T. Al-Naffouri, and H. Wymeersch, “Multi-RIS-
Enabled 3D Sidelink Positioning,” 2023.
[20] Q. Hu, F. Gao, H. Zhang, S. Jin, and G. Y. Li, “Deep Learning for Channel Esti-
mation: Interpretation, Performance, and Comparison,” IEEE Trans. Wirel. Com-
mun., vol. 20, no. 4, pp. 2398–2412, 2020.
[21] S. J. Pan and Q. Yang, “A Survey on Transfer Learning,” IEEE Trans. Knowl.
Data Eng., vol. 22, no. 10, pp. 1345–1359, 2010.
[22] Z. Chen et al., “Dynamic Task Software Caching-Assisted Computation Offloading
for Multi-Access Edge Computing,” IEEE Trans. Commun., vol. 70, no. 10, pp.
6950–6965, 2022.
[23] ——, “Knowledge-aided federated learning for energy-limited wireless networks,”
IEEE Trans. Commun., vol. 71, no. 6, pp. 3368–3386, 2023.
[24] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely Connected
Convolutional Networks,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit.,
2017, pp. 2261–2269.
[25] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recog-
nition,” in Proc. of the Conf. Comput. Vis. Pattern Recognit. (CVPR), June
2016.
[26] C. Wu et al., “Learning to Localize: A 3D CNN Approach to User Positioning in
Massive MIMO-OFDM Systems,” IEEE Trans. Wireless Commun., vol. 20, no. 7,
pp. 4556–4570, 2021.
[27] Z. Li et al., “A Survey of Convolutional Neural Networks: Analysis, Applications,
and Prospects,” IEEE Trans. Neural Netw. Learn. Syst., vol. 33, no. 12, pp.
6999–7019, 2022.
[28] K. Keykhosravi et al., “SISO RIS-Enabled Joint 3D Downlink Localization and
Synchronization,” in IEEE ICC, 2021, pp. 1–6.
[29] E. Čišija et al., “RIS-Aided mmWave MIMO Radar System for Adaptive Multi-
Target Localization,” in 2021 IEEE Statistical Signal Processing Workshop (SSP),
2021, pp. 196–200.
[30] K. Keykhosravi et al., “RIS-Enabled SISO Localization Under User Mobility and
Spatial-Wideband Effects,” IEEE J. Sel. Topics Signal Process., vol. 16, no. 5, pp.
1125–1140, 2022.
[31] S. Huang et al., “Near-Field RSS-Based Localization Algorithms Using Reconfig-
urable Intelligent Surface,” IEEE Sens. J., vol. 22, no. 4, pp. 3493–3505, 2022.
Chapter 8. Conclusion 189

[32] X. Zhang et al., “Hybrid Reconfigurable Intelligent Surfaces-Assisted Near-Field

Localization,” IEEE Commun. Lett., vol. 27, no. 1, pp. 135–139, 2023.
[33] M. Luan et al., “Phase Design and Near-Field Target Localization for RIS-Assisted
Regional Localization System,” IEEE Trans. Veh. Technol., vol. 71, no. 2, pp.
1766–1777, 2022.
[34] C. Ozturk et al., “RIS-Aided Near-Field Localization Under Phase-Dependent Ampli-
tude Variations,” IEEE Trans. Wireless Commun., vol. 22, no. 8, pp. 5550–5566,
2023.
[35] R. Ghazalian et al., “Bi-Static Sensing for Near-Field RIS Localization,” in GLOBE-
COM 2022, 2022, pp. 6457–6462.
[36] H. Ye et al., “Power of DeepLearning for Channel Estimation and Signal Detection
in OFDM Systems,” vol. 7, no. 1, pp. 114–117, 2018.
[37] H. Huang et al., “Deep Learning for Super-Resolution Channel Estimation and DOA
Estimation Based Massive MIMO System,” IEEE Trans. Veh. Technol., vol. 67,
no. 9, pp. 8549–8560, 2018.
[38] Y. Yang, F. Gao, X. Ma, and S. Zhang, “Deep Learning-Based Channel Estimation
for Doubly Selective Fading Channels,” IEEE Access, vol. 7, pp. 36 579–36 589,
2019.
[39] U. Bhattacherjee, C. K. Anjinappa, L. Smith, E. Ozturk, and I. Guvenc, “Localiza-
tion with Deep Neural Networks using mmWave Ray Tracing Simulations,” in 2020
SoutheastCon, 2020, pp. 1–8.
[40] T. Wu et al., “Fingerprint-Based mmWave Positioning System Aided by Recon-
figurable Intelligent Surface,” IEEE Wirel. Commun. Lett., vol. 12, no. 8, pp.
1379–1383, 2023.
[41] G. Zhou, C. Pan, H. Ren, P. Popovski, and A. L. Swindlehurst, “Channel Estimation
for RIS-Aided Multiuser Millimeter-Wave Systems,” IEEE Trans. Signal Process.,
vol. 70, pp. 1478–1492, 2022.
[42] K. Zhi, C. Pan, H. Ren, K. K. Chai, and M. Elkashlan, “Active RIS Versus Passive
RIS: Which is Superior With the Same Power Budget?” IEEE Commun. Lett.,
vol. 26, no. 5, pp. 1150–1154, 2022.
[43] E. Basar, M. Di Renzo, J. De Rosny, M. Debbah, M.-S. Alouini, and R. Zhang,
“Wireless communications through reconfigurable intelligent surfaces,” IEEE Access,
vol. 7, pp. 116 753–116 773, 2019.
[44] A. Papazafeiropoulos, C. Pan, P. Kourtessis, S. Chatzinotas, and J. M. Senior,
“Intelligent reflecting surface-assisted MU-MISO systems with imperfect hardware:
Channel estimation, beamforming design,” 2021.
[45] F. Bohagen, P. Orten, and G. E. Oien, “Design of optimal high-rank line-of-sight
MIMO channels,” IEEE Transactions on Wireless Communications, vol. 6, no. 4,
pp. 1420–1425, 2007.
[46] F. Bohagen, P. Orten, and G. Oien, “Construction and capacity analysis of high-
rank line-of-sight MIMO channels,” in IEEE Wireless Communications and Net-
Chapter 8. Conclusion 190

working Conference, 2005, vol. 1, 2005, pp. 432–437 Vol. 1.

[47] A. Shahmansoori, G. E. Garcia, G. Destino, G. Seco-Granados, and H. Wymeersch,
“5G Position and Orientation Estimation through Millimeter Wave MIMO,” in 2015
IEEE Globecom Workshops (GC Wkshps), 2015, pp. 1–6.
[48] G. Ghatak, R. Koirala, A. De Domenico, B. Denis, D. Dardari, B. Uguen, and
M. Coupechoux, “Beamwidth Optimization and Resource Partitioning Scheme for
Localization Assisted mm-Wave Communication,” IEEE Trans. Commun., vol. 69,
no. 2, pp. 1358–1374, 2021.
[49] W. Wang and W. Zhang, “Joint Beam Training and Positioning for Intelligent
Reflecting Surfaces Assisted Millimeter Wave Communications,” IEEE Trans. Wire-
less Commun., vol. 20, no. 10, pp. 6282–6297, 2021.
[50] Y. Wang and K. C. Ho, “An asymptotically efficient estimator in closed-form for 3D
AOA localization using a sensor network,” IEEE Transactions on Wireless Com-
munications, vol. 14, no. 12, pp. 6524–6535, 2015.
[51] D. Fan, F. Gao, Y. Liu, Y. Deng, G. Wang, Z. Zhong, and A. Nallanathan, “Angle
domain channel estimation in hybrid millimeter wave massive MIMO systems,”
IEEE Transactions on Wireless Communications, vol. 17, no. 12, pp. 8165–8179,
2018.
[52] T. Wu, C. Pan, Y. Pan, S. Hong, H. Ren, M. Elkashlan, F. Shu, and J. Wang, “Two-
step mmwave positioning scheme with RIS-Part I: Angle estimation and analysis,”
2022.
[53] A. Fascista, M. F. Keskin, A. Coluccia, H. Wymeersch, and G. Seco-Granados,
“RIS-aided joint localization and synchronization with a single-antenna receiver:
Beamforming design and low-complexity estimation,” 2022.
[54] C. Pan, G. Zhou, K. Zhi, S. Hong, T. Wu, Y. Pan, H. Ren, M. D. Renzo, A. Lee Swindle-
hurst, R. Zhang, and A. Y. Zhang, “An Overview of Signal Processing Techniques
for RIS/IRS-Aided Wireless Systems,” IEEE J. Sel. Topics Signal Process., vol. 16,
no. 5, pp. 883–917, 2022.
[55] K. C. Ho, “Bias reduction for an explicit solution of source localization using
TDOA,” Signal Processing, IEEE Transactions on, vol. 60, no. 5, pp. 2101–2114,
2012.
[56] Y. T. Chan and K. C. Ho, “A simple and efficient estimator for hyperbolic location,”
IEEE Transactions on Signal Processing, vol. 42, no. 8, pp. 1905–1915, 2002.
[57] T. Wu, C. Pan, Y. Pan, S. Hong, H. Ren, and M. Elkashlan, “Two-step mmWave
positioning scheme with RIS-Part I: Angle estimation and analysis,” To appear.
[58] C. Hu, L. Dai, S. Han, and X. Wang, “Two-timescale channel estimation for recon-
figurable intelligent surface aided wireless communications,” IEEE Transactions on
Communications, vol. 69, no. 11, pp. 7736–7747, 2021.
[59] T. Wu, C. Pan, Y. Pan, S. Hong, H. Ren, M. Elkashlan, F. Shu, and J. Wang, “3D
Positioning Algorithm Design for RIS-aided mmWave Systems,” 2022. [Online].
Chapter 8. Conclusion 191

Available: [Link]
[60] H. Wymeersch and B. Denis, “Beyond 5G wireless localization with reconfigurable
intelligent surfaces,” in ICC 2020 - 2020 IEEE International Conference on Com-
munications (ICC), 2020, pp. 1–6.
[61] Y. Pan, C. Pan, S. Jin, and J. Wang, “RIS-Aided Near-Field Localization and
Channel Estimation for the Sub-Terahertz System,” 2022. [Online]. Available:
[Link]
[62] C. L. Nguyen, O. Georgiou, G. Gradoni, and M. Di Renzo, “Wireless Fingerprint-
ing Localization in Smart Environments Using Reconfigurable Intelligent Surfaces,”
IEEE Access, vol. 9, pp. 135 526–135 541, 2021.
[63] H. Chen, Y. Zhang, W. Li, X. Tao, and P. Zhang, “ConFi: Convolutional Neural
Networks Based Indoor Wi-Fi Localization Using Channel State Information,” IEEE
Access, vol. 5, pp. 18 066–18 074, 2017.
[64] X. Sun et al., “Single-Site Localization Based on a New Type of Fingerprint for
Massive MIMO-OFDM Systems,” IEEE Trans. on Veh. Tech., vol. 67, no. 7, pp.
6134–6145, 2018.
[65] C. Wang, Z. Lv, X. Gao, X. You, Y. Hao, and H. Haas, “Pervasive channel modeling
theory and applications to 6G GBSMs for allfrequency bands and all scenarios,”
IEEE Trans. Veh. Technol., accepted for publication, doi, vol. 10.
[66] Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, “A Survey of Convolutional Neural
Networks: Analysis, Applications, and Prospects,” IEEE Trans. Neural Networks
Learn. Syst., pp. 1–21, 2021.
[67] [Link]
[68] C. Ozturk, M. F. Keskin, V. Sciancalepore, H. Wymeersch, and S. Gezici, “RIS-
aided Localization under Pixel Failures,” 2023.
[69] Q. Wu and R. Zhang, “Intelligent Reflecting Surface Enhanced Wireless Network via
Joint Active and Passive Beamforming,” IEEE Trans. Wireless Commun., vol. 18,
no. 11, pp. 5394–5409, 2019.
[70] B. Jalal, O. Elnahas, and Z. Quan, “Efficient DOA Estimation Under Partially
Impaired Antenna Array Elements,” IEEE Trans. Veh. Technol., vol. 71, no. 7, pp.
7991–7996, 2022.
[71] S. Liu et al., “Query2label: A simple transformer way to multi-label classification,”
arXiv preprint arXiv:2107.10834, 2021.
[72] C. Consortium, “Common Objects in Context (COCO) - Dataset Download,” https:
//[Link], 2023.
[73] J. Deng et al., “ImageNet: A Large-Scale Hierarchical Image Database,” in CVPR,
2009, pp. 248–255.
[74] M.-L. Zhang et al., “Multilabel Neural Networks with Applications to Functional
Genomics and Text Categorization,” IEEE Trans. Knowl. Data Eng., vol. 18,
no. 10, pp. 1338–1351, 2006.
[75] P. Yang et al., “SGM: Sequence Generation Model for Multi-Label Classification,”
Chapter 8. Conclusion 192

arXiv preprint arXiv:1806.04822, 2018.

[76] J. Sun et al., “Learning Sparse Representation With Variational Auto-Encoder for
Anomaly Detection,” IEEE Access, vol. 6, pp. 33 353–33 361, 2018.
[77] R. Wang, Z. Xing, and E. Liu, “Joint location and communication study for intel-
ligent reflecting surface aided wireless communication system,” 2021.
[78] S. Palmucci, A. Guerra, A. Abrardo, and D. Dardari, “Two-Timescale Joint Precod-
ing Design and RIS Optimization for User Tracking in Near-Field MIMO Systems,”
IEEE Trans. Signal Process., vol. 71, pp. 3067–3082, 2023.
[79] Y. Mei, R. Wang, E. Liu, and I. Soto, “Multi-User Tracking in Reconfigurable
Intelligent Surface Aided Near-Field Wireless Communications System,” Applied
Sciences, vol. 14, no. 1, p. 205, 2023.
[80] M. Rahal, B. Denis, M. F. Keskin, B. Uguen, and H. Wymeersch, “RIS-Enabled
NLoS Near-Field Joint Position and Velocity Estimation under User Mobility,”
arXiv preprint arXiv:2312.09720, 2023.
[81] Y. Pan, C. Pan, S. Jin, and J. Wang, “RIS-Aided Near-Field Localization and
Channel Estimation for the Terahertz System,” IEEE J. Sel. Topics Signal Process.,
vol. 17, no. 4, pp. 878–892, 2023.
[82] H. Fu, S. Abeywickrama, L. Zhang, and C. Yuen, “Low-Complexity Portable Pas-
sive Drone Surveillance via SDR-Based Signal Processing,” IEEE Communications
Magazine, vol. 56, no. 4, pp. 112–118, 2018.
[83] Y. Sun, S. Abeywickrama, L. Jayasinghe, C. Yuen, J. Chen, and M. Zhang, “Micro-
Doppler Signature-Based Detection, Classification, and Localization of Small UAV
With Long Short-Term Memory Neural Network,” IEEE Transactions on Geo-
science and Remote Sensing, vol. 59, no. 8, pp. 6285–6300, 2021.
[84] P. Samczynski et al., “5G Network-Based Passive Radar,” IEEE Transactions on
Aerospace and Electronic Systems, vol. 58, no. 2, pp. 1432–1444, 2022.
[85] Z. Zhang, L. Dai, X. Chen, C. Liu, F. Yang, R. Schober, and H. V. Poor, “Active
RIS vs. Passive RIS: Which Will Prevail in 6G?” IEEE Trans. Commun., vol. 71,
no. 3, pp. 1707–1725, 2023.
[86] K. Zhi, C. Pan, H. Ren, K. K. Chai, and M. Elkashlan, “Active RIS Versus Passive
RIS: Which is Superior With the Same Power Budget?” IEEE Commun. Lett.,
vol. 26, no. 5, pp. 1150–1154, 2022.
[87] G. Mylonopoulos, C. D Andrea, and S. Buzzi, “Active Reconfigurable Intelligent
Surfaces for User Localization in mmWave MIMO Systems,” in 2022 IEEE 23rd
International Workshop on Signal Processing Advances in Wireless Communication
(SPAWC), 2022, pp. 1–5.
[88] R. Luo, W. Ni, H. Tian, and J. Cheng, “Federated deep reinforcement learning
for RIS-assisted indoor multi-robot communication systems,” IEEE Trans. Veh.
Technol., vol. 71, no. 11, pp. 12 321–12 326, 2022.
[89] Y. Wang, I. W.-H. Ho, S. Zhang, and Y. Wang, “Intelligent Reflecting Surface
enabled Fingerprinting-based Localization with Deep Reinforcement Learning,” IEEE
Chapter 8. Conclusion 193

Trans. Veh. Technol., 2023.

[90] C. Huang, R. Mo, and C. Yuen, “Reconfigurable intelligent surface assisted mul-
tiuser MISO systems exploiting deep reinforcement learning,” IEEE J. Sel. Areas
Commun., vol. 38, no. 8, pp. 1839–1850, 2020.
[91] I. S. Gradshteyn and I. M. Ryzhik, “Table of integrals, series, and products,” Math-
ematics of Computation, vol. 20, no. 96, p. 1157, 2007.
[92] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge University Press,
1985.
194
Chapter A. Derivations in Chapter 3 195

Appendix A

Derivations in Chapter 3

A.1 Derivation of D1,i

Firstly, we derive the expression of D11 as

" 2 #
B2 
X̂(V̂ − θ̃RA Û )
Z
β1
D11 = α22 − 2
θ̃RA dθ̃RA
B1 8abŶ Û + θ̃RA V̂
B2 B2 2 − Û θ̃ 3
X̂α22 X̂β12
Z Z
2 3 V̂ θ̃RA RA
= (V̂ θ̃RA − Û θ̃RA )dθ̃RA − dθ̃RA
8abŶ B1 8abŶ B1 (Û + V̂ θ̃RA )2
B2
X̂α22
Z
2 3
= (V̂ θ̃RA − Û θ̃RA )dθ̃RA
8abŶ B1
  
V̂ 2
 X̂β12 
Z B2 θ̃ Z B2 3
θ̃RA
Û RA
− dθ̃RA − dθ̃RA 

 V̂ θ̃ V̂ θ̃
8abŶ Û 0 (1 + RA
)2 0 (1 + RA
) 2
Û Û
 
V̂ 2
2
X̂β1 
Z B1 θ̃ Z B1 3
θ̃RA
Û RA
− dθ̃RA − dθ̃RA 


8abŶ Û 0 (1 + V̂ θ̃RA )2 0 (1 + V̂ θ̃RA )2
Û Û
  B2
4 3
X̂α22  Û θ̃RA V̂ θ̃RA
= − −

4 3
 
8abŶ
B1
    
X̂β12 B24 V̂ X̂β12 V̂ V̂B23
 − · · 2 F1 2, 4; 5; − B2  + · · 2 F1 2, 3; 4; − B2 
    
8abŶ Û 4 Û 8abŶ Û 2 3 Û
    
X̂β12 B4 V̂ X̂β12 V̂ B3 V̂
−− · 1 · 2 F1 2, 4; 5; − B1  + · 1 · 2 F1 2, 3; 4; − B1 ,
    
8abŶ Û 4 Û 8abŶ Û 2 3 Û

(A.1)
Chapter A. Derivations in Chapter 3 196

 
xµ−1 dx
Ru
where =2 F1 ν, µ; 1 + µ; −βu is a generalized hypergeometric series [91].
 
0 (1+βx)v

Then, D12 in (3.63) is derived as

B3
X̂(V̂ − θ̃RA Û ) 2
Z
D12 = θ̃RA dθ̃RA
B2 2b
Z B3 2 − Û θ̃ 3 )
X̂(V̂ θ̃RA RA
= dθ̃RA
B2 2b
  B3
3 4
X̂  V̂ θ̃RA Û θ̃RA
= − . (A.2)

2b 3 4
 
B2
Chapter A. Derivations in Chapter 3 197

Finally, D13 in (3.63) is given by

 2 
B4
X̂(V̂ − θ̃RA Û ) 
Z
β2 2 2
D13 =  − α1 θ̃RA dθ̃RA


B3 8abŶ Û + θ̃RA V̂
B4 3 + V̂ θ̃ 2 B4
X̂β22 −Û θ̃RA X̂α12
Z Z
RA 3 2
= dθ̃RA − (−Û θ̃RA + V̂ θ̃RA )dθ̃RA
8abŶ B3 (Û + V̂ θ̃RA )2 8abŶ B3
  
V̂ 2
 X̂β22 
Z B4 θ̃ Z B4 3
θ̃RA
Û RA
= dθ̃RA − dθ̃RA 

 V̂ θ̃ V̂ θ̃
8abŶ Û 0 (1 + RA )2 0 (1 + RA )2
Û Û
 
V̂ 2
2
X̂β2 
Z B3 θ̃ Z B3 3
θ̃RA
Û RA
− dθ̃RA − dθ̃RA 

 V̂ θ̃ V̂ θ̃
8abŶ Û 0 (1 + RA
) 2 0 (1 + RA
) 2
Û Û

X̂α12 B4
Z
3 2
− (−Û θ̃RA + V̂ θ̃RA )dθ̃RA
8abŶ B3
 
2 Z B4 2 2 Z B4 3
 X̂ V̂ β2 θ̃RA X̂β2 θ̃RA
=  dθ̃RA − dθ̃RA 

2
8abŶ Û 0 (1 + RA )2 V̂ θ̃ 8abŶ Û 0 (1 + RA )2 V̂ θ̃
Û Û
 
B 2 B 3
 X̂ V̂ β22 X̂β22
Z Z
3
θ̃RA 3
θ̃RA
− dθ̃RA − dθ̃RA 

2
8abŶ Û 0 (1 + V̂ θ̃ 2 8abŶ Û 0 (1 + V̂ θ̃ 2
RA
) RA
)
Û Û

X̂α12 B4
Z
3 2
− (−Û θ̃RA + V̂ θ̃RA )dθ̃RA
8abŶ B3
    
X̂β22 B44 V̂ X̂β22 V̂ B43V̂
=  − · · 2 F1 2, 4; 5; − B4  + · · 2 F1 2, 3; 4; − B4 
    
8abŶ Û 4 Û 8abŶ Û 2 3 Û
    
X̂β22 B4 V̂ X̂β22 V̂ B3 V̂
−− · 3 · 2 F1 2, 4; 5; − B3  + · 3 · 2 F1 2, 3; 4; − B3 
    
8abŶ Û 4 Û 8abŶ Û 2 3 Û
  B4
4 3
X̂α12 Û θ̃RA V̂ θ̃RA
− − + . (A.3)
 
4 3

8abŶ
B3
Chapter A. Derivations in Chapter 3 198

A.2 Derivation of D2,i

Firstly, D21 in (3.64) is derived as

  2 
B2
X̂(V̂ − θ̃RA Û )  2 
Z
β1
D21 = α2 −   θ̃RA dθ̃RA
 
B1 8abŶ Û + θ̃RA V̂
B2 B2 2 + V̂ θ̃
X̂α22 X̂β12 −Û θ̃RA
Z Z
2 RA
= (−Û θ̃RA + V̂ θ̃RA )dθ̃RA − dθ̃RA
8abŶ B1 8abŶ B1 (Û + V̂ θ̃RA )2
B2
X̂α22
Z
2
= (−Û θ̃RA + V̂ θ̃RA )dθ̃RA
8abŶ B1
  
V̂
2
 X̂β1 
Z B2 θ̃ Z B2 2
θ̃RA
Û RA
− dθ̃RA − dθ̃RA 

 V̂ θ̃ V̂ θ̃
8abŶ Û 0 (1 + RA
)2 0 (1 + RA
) 2
Û Û
 
V̂
X̂β12  B1
Z θ̃ Z B1 2
θ̃RA
Û RA
− dθ̃RA − dθ̃RA 

 V̂ θ̃ V̂ θ̃
8abŶ Û 0 (1 + RA )2 0 (1 + RA )2
Û Û
  B2
3 2
X̂α22  Û θ̃RA V̂ θ̃RA
= − +

3 2

8abŶ
B1
    
X̂β12 B3 V̂ X̂β12 V̂ B2 V̂
−  − · 2 · 2 F1 2, 3; 4; − B2  + · 2 · 2 F1 2, 2; 3; − B2 
    
8abŶ Û 3 Û 8abŶ Û 2 2 Û
    
X̂β12 B3 V̂ X̂β12 V̂ B2 V̂
−− · 1 · 2 F1 2, 3; 4; − B1  + · 1 · 2 F1 2, 2; 3; − B1 .
    
8abŶ Û 3 Û 8abŶ Û 2 2 Û

(A.4)

Then, D22 in (3.64) is written as

B3
X̂(V̂ − θ̃RA Û )
Z
D22 = θ̃RA dθ̃RA
B2 2b
Z B3
X̂ 2
= (V̂ θ̃RA − Û θ̃RA )dθ̃RA
2b B2
  B3
3
Û θ̃RA V̂ 2
θ̃RA
X̂ 
= − + . (A.5)

2b 3 2

B2
Chapter A. Derivations in Chapter 3 199

Finally, D23 in (3.64) could be formulated as

 2 
B4
X̂(V̂ − θ̃RA Û ) 
Z
β2 2
D23 =  − α1 θ̃RA dθ̃RA


B3 8abŶ Û + θ̃RA V̂
B4 2 + V̂ θ̃ B4
X̂β22 −Û θ̃RA X̂α12
Z Z
RA 2
= dθ̃RA − (−Û θ̃RA + V̂ θ̃RA )dθ̃RA
8abŶ B3 (Û + V̂ θ̃RA )2 8abŶ B3
    
X̂β22 B43 V̂ X̂β22 V̂
V̂ B42
=  − · · 2 F1 2, 3; 4; − B4  + · · 2 F1 2, 2; 3; − B4 
    
8abŶ Û 3 Û 8abŶ Û 2 2 Û
    
X̂β22 B3 V̂ X̂β22 V̂ B2 V̂
−− · 3 · 2 F1 2, 3; 4; − B3  + · 3 · 2 F1 2, 2; 3; − B3 
    
8abŶ Û 3 Û 8abŶ Û 2 2 Û
  B4
3 2
X̂α12 Û θ̃RA V̂ θ̃RA
− − + . (A.6)
 
3 2

8abŶ
B3

′
A.3 Derivation of D1,i

′ in (3.66), it could be expressed as

As for D11

  2 
B3
X̂(V̂ − θ̃RA Û )  2 
Z
′ β1   2
D11 = α2 −   θ̃RA dθ̃RA
B1 8abŶ Û + θ̃RA V̂
B3 B3 3 + V̂ θ̃ 2
X̂α22 X̂β12 −Û θ̃RA
Z Z
3 2 RA
= (−Û θ̃RA + V̂ θ̃RA )dθ̃RA − dθ̃RA
8abŶ B1 8abŶ B1 (Û + V̂ θ̃RA )2
  B3
4 3
X̂α22 Û θ̃RA V̂ θ̃RA
= − + −
 
4 3

8abŶ
B1
    
X̂β12 B4 V̂ X̂β12 V̂ B3 V̂
 − · 3 · 2 F1 2, 4; 5; − B3  + · 3 · 2 F1 2, 3; 4; − B3 
    
8abŶ Û 4 Û 8abŶ Û 2 3 Û
    
X̂β12 B4 V̂ X̂β12 V̂ B3 V̂
−− · 1 · 2 F1 2, 4; 5; − B1  + · 1 · 2 F1 2, 3; 4; − B1 .
    
8abŶ Û 4 Û 8abŶ Û 2 3 Û

(A.7)
Chapter A. Derivations in Chapter 3 200

′ in (3.66) is given by
Then, D12

B2
X̂ Ẑ(V̂ − θ̃RA Û )
Z
′ 2
D12 = θ̃RA dθ̃RA
B 2aŶ (Û + θ̃RA V̂ )2
3    
X̂ Ẑ B24 V̂ X̂ Ẑ V̂ V̂ B23
=− · · 2 F1 2, 4; 5; − B2  + · · 2 F1 2, 3; 4; − B2 
    
2aŶ Û 4 Û 2aŶ Û 2 3 Û
    
X̂ Ẑ B34 V̂ X̂ Ẑ V̂ B33 V̂
−− · · 2 F1 2, 4; 5; − B3  + · · 2 F1 2, 3; 4; − B3 .
    
2aŶ Û 4 Û 2aŶ Û 2 3 Û

(A.8)

′ in (3.66) is derived as
Finally, D13

 2 
B4
X̂(V̂ − θ̃RA Û ) 
Z
′ β2 2 2
D13 =  − α1 θ̃RA dθ̃RA


B2 8abŶ Û + θ̃RA V̂
B4 3 + V̂ θ̃ 2 B4
X̂β22 −Û θ̃RA X̂α12
Z Z
RA 3 2
= dθ̃RA − (−Û θ̃RA + V̂ θ̃RA )dθ̃RA
8abŶ B2 (Û + V̂ θ̃RA )2 8abŶ B2
    
X̂β22 B44 V̂ X̂β22 V̂ B43
V̂
=  − · · 2 F1 2, 4; 5; − B4  + · · 2 F1 2, 3; 4; − B4 
    
8abŶ Û 4 Û 8abŶ Û 2 3 Û
    
X̂β22 B4 V̂ X̂β22 V̂ B3 V̂
−− · 2 · 2 F1 2, 4; 5; − B2  + · 2 · 2 F1 2, 3; 4; − B2 
    
8abŶ Û 4 Û 8abŶ Û 2 3 Û
  B4
4 3
X̂α12 Û θ̃RA V̂ θ̃RA
− − + . (A.9)
 
4 3

8abŶ
B2
Chapter A. Derivations in Chapter 3 201

′
A.4 Derivation of D2,i

′ in (3.67) is formulated as
Firstly, D21

  2 
B3
X̂(V̂ − θ̃RA Û )  2 
Z
′ β1
D21 = α2 −   θ̃RA dθ̃RA
 
B1 8abŶ Û + θ̃RA V̂
  B3
3 2
X̂α22 Û θ̃RA V̂ θ̃RA
= − + −
 
3 2

8abŶ
B1
    
X̂β12 B33 V̂ X̂β12 V̂ B32 V̂
 − · · 2 F1 2, 3; 4; − B3  + · · 2 F1 2, 2; 3; − B3 
    
8abŶ Û 3 Û 8abŶ Û 2 2 Û
   
X̂β12 B3 V̂ X̂β12 V̂ B2 V̂
−− · 1 · 2 F1 2, 3; 4; − B1  + · 1 · 2 F1 (2, 2; 3; − B1 ).
   
8abŶ Û 3 Û 8abŶ Û 2 2 Û

(A.10)

′ in (3.67) is written as
Then, D22

B2
X̂ Ẑ(V̂ − θ̃RA Û )
Z
′
D22 = θ̃RA dθ̃RA
B3 2aŶ (Û + θ̃RA V̂ )2
    
X̂ Ẑ B23 V̂ X̂ Ẑ V̂ B22
V̂
=− · · 2 F1 2, 3; 4; − B2  + · · 2 F1 2, 2; 3; − B2 
    
2aŶ Û 3 Û 2aŶ Û 2 2 Û
    
X̂ Ẑ B33 V̂ X̂ Ẑ V̂ B32 V̂
−− · · 2 F1 2, 3; 4; − B3  + · · 2 F1 2, 2; 3; − B3 .
    
2aŶ Û 3 Û 2aŶ Û 2 2 Û

(A.11)
Chapter A. Derivations in Chapter 3 202

′ in (3.67) is obtained as
Finally, D23

 2 
B4
X̂(V̂ − θ̃RA Û ) 
Z
′ β2 2
D23 =  − α1 θ̃RA dθ̃RA


B2 8abŶ Û + θ̃RA V̂
B4 2 + V̂ θ̃ B4
X̂β22 −Û θ̃RA X̂α12
Z Z
RA 2
= dθ̃RA − (−Û θ̃RA + V̂ θ̃RA )dθ̃RA
8abŶ B2 (Û + V̂ θ̃RA )2 8abŶ B2
    
X̂β22 B43 V̂ X̂β22 V̂ B42V̂
=  − · · 2 F1 2, 3; 4; − B4  + · · 2 F1 2, 2; 3; − B4 
    
8abŶ Û 3 Û 8abŶ Û 2 2 Û
    
X̂β22 B3 V̂ X̂β22 V̂ B2 V̂
−− · 2 · 2 F1 2, 3; 4; − B2  + · 2 · 2 F1 2, 2; 3; − B2 
    
8abŶ Û 3 Û 8abŶ Û 2 2 Û
  B4
3 2
X̂α12 Û θ̃RA V̂ θ̃RA
− − + . (A.12)
 
3 2

8abŶ
B2
Appendix B

Proofs and Derivations in

Chapter 4

B.1 Derivation of E[ũ]

By utilizing (6.16), recall that zr = [nT , ω T , ν T ]T , Br zr can be expanded as (Br,n n +

Br,ω ω + Br,ν ν). Then, it is formulated as

E1 = E[Hr Br zr ] = Hr E[Br,n n + Br,ω ω + Br,ν ν] = Hr Br,n E[n] + Hr Br,ω E[ω],

where E[n] and E[ω] denote the expectation of n and ω, respectively.

For E2 , it is formulated as

1 1
E2 = E[Hr (η ⊙ η)] = Hr E[[0, · · · , 0, ν12 , · · · , νM
2 T
] ] = Hr qu ,
2 | {z } 2
2M

where
1
qu = [0, · · · , 0, σν21 , · · · , σν2M ]T .
2 | {z }
2M

203
Chapter B. Proofs and Derivations in Chapter 4 204

Then, E3 can be derived as

E3 = E[P−1 T −1 T −1 T
r G̃r Wr Br zr ] − E[Pr G̃r Wr Gr Hr Br zr ] − E[Pr Gr Wr G̃r Hr Br zr ]

= E31 − E32 − E33 ,

where the detailed derivations of E31 , E32 and E33 are given as follows.

For E31 , as it is formulated as G̃r in (4.65) and Br zr = (Br,n n + Br,ω ω + Br,ν ν), E31

can be derived as

E31 = E[P−1 T T T T
r (Ḡn ⊙ [(n11×4 ) , (n11×4 ) , (n11×4 ) ])Wr Br,n diag(n)1M ×1 ]

+ E[P−1 T T T T
r (Ḡω ⊙ [(ω11×4 ) , (ω11×4 ) , (ω11×4 ) ])Wr Br,ω diag(ω)1M ×1 ]

+ E[P−1 T T T T
r (Ḡν ⊙ [(ν11×4 ) , (ν11×4 ) , (ν11×4 ) ])Wr Br,ν diag(ν)1M ×1 ].

To further derive the expression of E31 , it is introduced Proposition 4. 2 and Propo-

sition 4. 3 as follows.

Proposition 4. 2 For vectors α ∈ CM ×1 , β ∈ CN ×1 and matrix G ∈ CM ×N , and

corresponding diagonal matrices diag(α) and diag(β) with these vectors as their main

diagonals, it has (αβ T ) ⊙ G = diag(α)Gdiag(β).

Proof: Please see reference [92]. ■

Proposition 4. 3 For a vector x ∈ CM ×1 with its covariance matrix given by Qx ,

it is formulated as

E[(A ⊙ [(x11×N )T , (x11×N )T , (x11×N )T ])Bdiag(x)]

= A(B ⊙ [Qx , Qx , Qx ]T ),

where A ∈ CN ×3M and B ∈ C3M ×M .

Proof: Please see reference [59].

Chapter B. Proofs and Derivations in Chapter 4 205

B.2 Derivation of W1 and W̃1

First, based on the definition of Ŵ1 below (4.58) and the definition of P̂r above (4.70),

it is formulated as Ŵ1 = B̂−1 −1

1 P̂r B̂1 . Then, according to the definition of B1 in (4.57)

and B̂1 in (4.58), it is formulated as

B̂1 = B1 + B̃1 , B̃1 = 2diag{ũ}. (B.1)

If B̂1 = B1 + B̃1 is used to derive Ŵ1 , the inverse matrix B̂−1

1 = (B1 + B̃1 )
−1 would be

complex and challenging. Fortunately, using the Newmann expansion [55], the approxi-

mation of B̂−1 −1 −1 −1
1 ≈ B1 − B1 B̃1 B1 can be derived. As a result, Ŵ1 is given by

Ŵ1 = (B−1 −1 −1 −1 −1 −1
1 − B1 B̃1 B1 )(Pr + P̃r )(B1 − B1 B̃1 B1 ). (B.2)

By neglecting the error terms that are higher than the first order, Ŵ1 can be approxi-

mated as

Ŵ1 ≈ B−1 −1 −1 −1 −1 −1 −1 −1 −1 −1
1 Pr B1 + B1 P̃r B1 − B1 Pr B1 B̃1 B1 − B1 B̃1 B1 Pr B1

= W1 + W̃1 , (B.3)

where W1 and W̃1 are defined as

W1 = B−1 −1
1 Pr B1 ,

W̃1 = B−1 −1 −1 −1
1 P̃r B1 − W1 B̃1 B1 − B̃1 B1 W1 . (B.4)

B.3 Proof of Proposition 4. 1

Before proving Proposition 4. 1, Proposition 4. 4 is introduced as follows.

Proposition 4 . 4 The Hadamard product of two vectors a ∈ CN ×1 and b ∈ CN ×1 ,

is the same as matrix multiplication of one vector by the corresponding diagonal matrix
Chapter B. Proofs and Derivations in Chapter 4 206

of the other vector:

a ⊙ b = diag(a)b.

Proof: please see reference [92]. ■

Let E[ã⊙ ã] denote the expectation of ã⊙ ã, using Proposition 4. 4, it is formulated

as E[ã ⊙ ã] = E[diag(ã)ã] = E[diag(ã)diag(ã)1M ×1 ]. Using Proposition 4. 2, E[ã ⊙ ã]

can be rewritten as

E[ã ⊙ ã]

= E[((ããT ) ⊙ IM ×M )1M ×1 ]

= (Ωa ⊙ IM ×M )1M ×1 = ca ,

where ca denotes the vector containing the diagonal elements of Ωa . Here, the proof of

Proposition 4. 1 is completed.

B.4 Derivations of E[W̃1 P2 B1 ũ]

Using the definition of W̃1 in (B.4), E[W̃1 P2 B1 ũ] can be derived as

E[W̃1 P2 B1 ũ] = E[(B−1 −1 −1 −1

1 P̃r B1 − W1 B̃1 B1 − B̃1 B1 W1 )P2 B1 ũ]

= E[B−1 −1 −1 −1
1 P̃r B1 P2 B1 ũ] − E[W1 B̃1 B1 P2 B1 ũ] − E[B̃1 B1 W1 P2 B1 ũ].

Let a1 = E[B−1 −1 −1 −1
1 P̃r B1 P2 B1 ũ], a2 = −E[W1 B̃1 B1 P2 B1 ũ] and a3 = −E[B̃1 B1 W1 P2 B1 ũ],

it is formulated as E[W̃1 P2 B1 ũ] = a1 + a2 + a3 , while a1 , a2 , a3 are given as follows.

Using the definition of P̃r below (4.72), a1 can be derived as

a1 = E[B−1 −1 −1 T T −1
1 P̃r B1 P2 B1 ũ] = E[B1 (G̃r Wr Gr + Gr Wr G̃r )B1 P2 B1 ũ].

Using the definition of ũ in (4.76) and ignoring the error terms that are higher than the

second order, a1 can be approximated as E[B−1 T T −1

1 (G̃r Wr Gr +Gr Wr G̃r )B1 P2 B1 Hr Br zr ].
Chapter B. Proofs and Derivations in Chapter 4 207

For notation simplicity, it is assumed that B−1

1 P2 B1 Hr = P3 . Then, by applying

Proposition 4. 2 and Proposition 4. 3, a1 can be derived as

a1 = B−1 T
1 (Ḡn ((Wr Gr P3 Br,n ) ⊙ Q̄n )1M ×1

+ ḠTω ((Wr Gr P3 Br,ω ) ⊙ Q̄ω )1M ×1

+ ḠTν ((Wr Gr P3 Br,ν ) ⊙ Q̄ν )1M ×1

+ GTr ((Wr Ḡn P3 Br,n ) ⊙ Q̄n )1M ×1

+ GTr ((Wr Ḡω P3 Br,ω ) ⊙ Q̄ω )1M ×1

+ GTr ((Wr Ḡν P3 Br,ν ) ⊙ Q̄ν )1M ×1 ).

Moreover, for a2 , using the definition of B̃1 in (B.1), it is formulated as

a2 = −E[W1 B−1
1 B̃1 P2 B1 ũ]

= −2W1 B−1
1 E[diag(ũ)P2 B1 diag(ũ)14×1 ].

Using Proposition 2, it is formulated as

a2 = −2W1 B−1 T
1 E[(ũũ ) ⊙ (P2 B1 )]14×1

= −2W1 B−1
1 (Ωu ⊙ (P2 B1 ))14×1 .

As Ωu is a diagonal matrix, a2 can be further derived as

a2 = −2W1 B−1 −1
1 diag(Ωu (P2 B1 ))14×1 = −2W1 B1 p4 ,

where p4 represents the column vector formed by the diagonal elements of P4 = P2 B1 Ωu .

Similar to the derivations of a2 , a3 can be derived as

a3 = −E[B−1 −1
1 B̃1 W1 P2 B1 ũ] = −2B1 p4,W1 ,

where p4,W1 represents the column vector formed by the diagonal elements of W1 P4 .
Chapter B. Proofs and Derivations in Chapter 4 208

B.5 Derivation of the Covariance Matrix of q̂

First, the covariance matrix of ξ is derived, which is denoted by Ωξ . Using the definition

of ξ̃, its covariance matrix Ωξ is given by

Ωξ = E[ξ̃ ξ̃ T ] ≈ GT1 Ŵ1 G1 )−1 GT1 Ŵ1 Ŵ1−1 Ŵ1T G1 ((GT1 Ŵ1 G1 )−1 )T = (GT1 Ŵ1 G1 )−1 ,

where E(z1 zT1 ) ≈ Ω̂1 = Ŵ1−1 is given below (4.58).

Furthermore, by neglecting the second order term in (4.91), the q̃ can be approxi-

mated as B−1
q ξ̃, therefore, the covariance matrix of q̂ can be derived as

Ωq = E[q̃q̃T ] = B−1 T −1 −1
q (G1 Ŵ1 G1 ) Bq .