octoScope_WP_NearFieldMIOChannel_201812
Uploaded by
p.grzelkaoctoScope_WP_NearFieldMIOChannel_201812
Uploaded by
p.grzelkaNEAR-FIELD MIMO CHANNEL MODELING WITH APPLICATIONS TO SMALL
ANECHOIC CHAMBERS
By
Jonathan Tefft
Bachelor of Science in Electrical Engineering, The University of New Hampshire, 2009
Master of Science in Electrical Engineering, The University of New Hampshire, 2011
DISSERTATION
Submitted to The University of New Hampshire
in Partial Ful lment of
the Requirements for the Degree of
Doctor of Philosophy
in
Electrical Engineering
December 2018
This dissertation has been examined and approved in partial ful llment of the requirements
for the degree of Doctor of Philosophy in Electrical Engineering by:
Dissertation Director, Dr. Nicholas J. Kirsch, Associate Professor of
Electrical and Computer Engineering
Dr. Kent A. Chamberlin, Professor of
Electrical and Computer Engineering
Dr. Michael J. Carter, Associate Professor of
Electrical and Computer Engineering
Dr. Qiaoyan Yu, Associate Professor of
Electrical and Computer Engineering
Dr. Joachim Raeder, Professor of
Physics
On November 14, 2018
Approval signatures are on le with the University of New Hampshire Graduate School
Acknowledgements
I would like to acknowledge Dr. Nicholas Kirsch for advising me on this dissertation
work, and for funding me with a Graduate Research Assistantship for multiple years. This
allowed me to continue and complete my work despite numerous changes in direction in my
research.
I would like to acknowledge Octoscope, Inc. and Fanny Mlinarsky for funding much of
this research, and providing a small anechoic chamber for measurements associated with
much of the work in this dissertation. Without this, the work of this dissertation could not
have been produced.
I would also like to acknowledge my committee members Dr. Kent Chamberlin, Dr.
Michael Carter, Dr. Qiaoyan Yu, and Dr. Joachim Raeder, for agreeing to be on my
committee and for taking the time to review my dissertation. I especially would like to
thank Dr. Chamberlin and Dr. Carter for providing advice on several electromagnetic and
mathematical topics included in this dissertation.
I would also like to thank my fellow graduate students in the UNH Wireless Systems Lab
for their moral support over the past several years. Finally, I would like to acknowledge my
parents John and Susan Tet, my partner Avery Parkes, and my friends for their in nite
patience and support. I could not have made it through this arduous process without you.
Jonathan Tet
Contents
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Importance of Small Chamber MIMO Test Solutions . . . . . . . . . 3
1.1.2 Importance of Near-Field MIMO Channel Models . . . . . . . . . . . 5
1.2 Contributions, and Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2. Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1 Fundamental Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.1 SISO channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1.2 MIMO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.1.3 Near-Field Topics and Mutual Coupling . . . . . . . . . . . . . . . . 21
2.2 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.1 MIMO-OTA Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.2 Near-Field MIMO Channels and Capacity . . . . . . . . . . . . . . . 28
iii
2.3 Preliminary Investigations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.3.1 Hertzian Dipole Field-Based Channel Matrix Formulation . . . . . . 34
2.3.2 AWGN Channel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.3.3 Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.3.4 Capacity Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.5 Channel and Capacity Simulation . . . . . . . . . . . . . . . . . . . . 43
2.3.6 Background for Selection of Free-Space Green's Function as Basis . . 45
3. Small Anechoic Chamber Feasibility for MIMO-OTA Testing . . . . . . . 57
3.1 Electromagnetic Field Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.1.1 Number of probe antennas . . . . . . . . . . . . . . . . . . . . . . . . 58
3.1.2 3GPP SCME Channel Model . . . . . . . . . . . . . . . . . . . . . . 59
3.2 Small chamber system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.2.1 Numerical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2.2 Computational Electromagnetic Model . . . . . . . . . . . . . . . . . 65
3.3 Analytical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.4 Electromagnetic Simulation Results . . . . . . . . . . . . . . . . . . . . . . . 69
3.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4. Near-Field Channel Model Framework . . . . . . . . . . . . . . . . . . . . 74
4.1 Formulation of Channel using Dipole Arrays . . . . . . . . . . . . . . . . . . 76
4.1.1 Mathematical Channel Model with Arbitrary Dipole Orientations . . 78
4.1.2 Numerical Computation of Dipole Array Channel . . . . . . . . . . . 86
4.1.3 Numerical Computation Setup and Results . . . . . . . . . . . . . . . 88
4.2 Dipole Array Channel with Spatial Attribute Function Decomposition . . . . 92
4.2.1 Plane Waves as Spatial Attribute Functions . . . . . . . . . . . . . . 94
4.2.2 Point Sources as Spatial Attribute Functions . . . . . . . . . . . . . . 98
4.2.3 Point Source Location . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.2.4 Inner Product of Point Sources . . . . . . . . . . . . . . . . . . . . . 101
4.2.5 Use of Point Source and Plane Waves as Joint Spatial Attribute Functions104
4.3 Fitting of Spatial Attribute Functions to Numerical Channel . . . . . . . . . 104
4.3.1 Point Source Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.3.2 Re nement Using Plane Waves . . . . . . . . . . . . . . . . . . . . . 106
4.4 Validation of Spatial Attribute Function Selection with Numerical Channel . 107
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5. Statistical and Frequency Extensions of Near-Field Channel Model . . . 112
5.1 Accurate Method of Fitting Spatial Attribute Functions to Measured Data . 114
5.1.1 Generalized Least Squares . . . . . . . . . . . . . . . . . . . . . . . . 115
5.1.2 Covariograms and Spatial Dependence . . . . . . . . . . . . . . . . . 116
5.1.3 Complex Geostatistical Covariance Model . . . . . . . . . . . . . . . 118
5.1.4 Geostatistical Regression . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.1.5 Geostatistical Kriging . . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.2 Adapting Spatial Attribute Function Set for a Wideband Model . . . . . . . 128
5.2.1 Algorithm for Fitting Spatial Attribute Function Set . . . . . . . . . 130
5.2.2 Modeling Coe cients as Polynomial Functions of Frequency . . . . . 131
5.2.3 Final Wideband Model Description . . . . . . . . . . . . . . . . . . . 132
5.3 System Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.3.1 Chamber Con guration and Properties . . . . . . . . . . . . . . . . . 133
5.3.2 Antenna Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.3.3 Measurement Equipment . . . . . . . . . . . . . . . . . . . . . . . . . 137
5.4 Measurement Campaign Description . . . . . . . . . . . . . . . . . . . . . . 138
5.4.1 Antenna Orientations and Positions . . . . . . . . . . . . . . . . . . . 140
5.4.2 Chosen Measurement Campaign Positions . . . . . . . . . . . . . . . 142
5.4.3 Resulting Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
5.5 Modeling of Measurement Data . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.5.1 Analysis of Complex Geostatistical Covariance Of Measurements . . . 150
5.5.2 Selected Wideband Frequency Bands for Modeling . . . . . . . . . . . 161
5.5.3 Performance Analysis of Frequency-Scaled Spatial Attribute Functions 162
5.5.4 Spatial Attribute Function Set Extension to Wideband Model . . . . 164
5.5.5 Wideband Model Accuracy . . . . . . . . . . . . . . . . . . . . . . . . 168
5.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
6. Application of Small Chamber Model to MIMO Analysis . . . . . . . . . 173
6.1 Small Anechoic Chamber Channel and Impulse Response . . . . . . . . . . . 173
6.2 Small Anechoic Chamber Spatial Correlation Analysis . . . . . . . . . . . . . 178
6.3 Wideband MIMO Capacity Using Small Chamber Channel Model . . . . . . 185
6.3.1 MIMO Capacity Analysis System Description . . . . . . . . . . . . . 185
6.3.2 MIMO Capacity and Eigenvalue Results . . . . . . . . . . . . . . . . 186
6.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
7. Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . 196
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
List of Tables
3.1 Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
vii
List of Figures
1.1 3GPP Release Timeline for 5G [1] . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Contribution Flowchart for Chapters 4 and 5 . . . . . . . . . . . . . . . . . . 7
2.1 2 Element Dipole and Loop Antenna Array [2] . . . . . . . . . . . . . . . . . 31
2.2 Near-Field 2×2 MIMO Example [2] . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Equivalent Circuit for Near-Field MIMO System [2] . . . . . . . . . . . . . . 32
2.4 Near-Field MIMO Array Re ector [3] . . . . . . . . . . . . . . . . . . . . . . 33
2.5 FDTD Near-Field MIMO Array Simulation [3] . . . . . . . . . . . . . . . . . 33
2.6 2×2 MIMO System Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.7 Capacity with Uniform Allocation vs. SNR and Spacing . . . . . . . . . . . 45
2.8 Dipole Electric Field Formulation . . . . . . . . . . . . . . . . . . . . . . . . 48
2.9 Dipole Mutual Impedance Formulation . . . . . . . . . . . . . . . . . . . . . 53
2.10 Dipole Self Impedance Formulation . . . . . . . . . . . . . . . . . . . . . . . 55
3.1 Small Anechoic Chamber Dimensions . . . . . . . . . . . . . . . . . . . . . . 63
3.2 HFSS Chamber Model of a Small Chamber . . . . . . . . . . . . . . . . . . . 66
3.3 CDF of Re ectivity for Select Con gurations, 700 MHz and 35 degree spread 68
4.1 Chapter 4 Contribution Flowchart . . . . . . . . . . . . . . . . . . . . . . . . 75
viii
4.2 Dipole Mutual Impedance Formulation . . . . . . . . . . . . . . . . . . . . . 78
4.3 Dipole Self Impedance Formulation . . . . . . . . . . . . . . . . . . . . . . . 81
4.4 Antenna Array and Dipole Port De nitions . . . . . . . . . . . . . . . . . . . 82
4.5 Transmission Line Network De nition . . . . . . . . . . . . . . . . . . . . . . 83
4.6 De nition of Parallel Admittance Networks . . . . . . . . . . . . . . . . . . . 84
4.7 Modeled System for Numerical Computation . . . . . . . . . . . . . . . . . . 89
4.8 Magnitude of Channel Vs. Dipole Position . . . . . . . . . . . . . . . . . . . 91
4.9 Phase of Channel Vs. Dipole Position . . . . . . . . . . . . . . . . . . . . . . 91
4.10 Visualization of Point Source Wave Fronts . . . . . . . . . . . . . . . . . . . 100
4.11 Normalized Histogram of Cross-Correlation Vs. Separation . . . . . . . . . . 103
4.12 LogNormal Mean EVM For Combinations of Spatial Attribute Functions . . 108
4.13 Standard Deviation of EVM For Combinations of Spatial Attribute Functions 110
4.14 LogNormal Mean + 1 SD EVM For Combinations of Spatial Attribute Functions110
5.1 Chapter 5 Contribution Flowchart . . . . . . . . . . . . . . . . . . . . . . . . 113
5.2 Octoscope octoBox BOX-38 Interior . . . . . . . . . . . . . . . . . . . . . . . 134
5.3 Octoscope OBS-14 Log Periodic Antenna [4] . . . . . . . . . . . . . . . . . . 135
5.4 Octoscope OBS-14 Array [4] . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
5.5 Octoscope OBS-14 Radiation Patterns [4] . . . . . . . . . . . . . . . . . . . . 136
5.6 Pasternack PE51083 Rubber Duck Antenna [5] . . . . . . . . . . . . . . . . . 137
5.7 Keysight E5063A VNA [6] . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.8 Chamber and Measurement System Diagram . . . . . . . . . . . . . . . . . . 141
5.9 Measured Channel Magnitude (dB), Probe One (Located in Lower Right),
5.15 GHz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.10 Measured Channel Phase (rad.), Probe One (Located in Lower Right), 5.15
GHz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
5.11 LN Mean + σ EVM, Probe One, 5.725 GHz . . . . . . . . . . . . . . . . . . 148
5.12 LN Mean + σ EVM, Probe Two, 5.725 GHz . . . . . . . . . . . . . . . . . . 148
5.13 LN Mean + σ EVM, Probe Three, 5.725 GHz . . . . . . . . . . . . . . . . . 149
5.14 LN Mean + σ EVM, Probe Four, 5.725 GHz . . . . . . . . . . . . . . . . . . 149
5.15 Measured Channel Spatial X-Cov., Mag., Probe One, 5.15 GHz . . . . . . . 152
5.16 Measured Channel Spatial X-Cov., Real, Probe One, 5.15 GHz . . . . . . . . 152
5.17 Measured Channel Spatial X-Cov., Imag., Probe One, 5.15 GHz . . . . . . . 153
5.18 Measured Channel Spatial Power Spectrum, Probe One, 5.15 GHz . . . . . . 153
5.19 Residual Emp. Spatial X-Cov., Mag., Probe One, 5.15 GHz . . . . . . . . . . 154
5.20 Residual Emp. Spatial X-Cov., Real, Probe One, 5.15 GHz . . . . . . . . . . 155
5.21 Residual Emp. Spatial X-Cov., Imag., Probe One, 5.15 GHz . . . . . . . . . 156
5.22 Residual Emp. Spatial Power Spectrum, Probe One, 5.15 GHz . . . . . . . . 157
5.23 Residual Model Spatial X-Cov., Mag., Probe One, 5.15 GHz . . . . . . . . . 158
5.24 Residual Model Spatial X-Cov., Real, Probe One, 5.15 GHz . . . . . . . . . 159
5.25 Residual Model Spatial X-Cov., Imag., Probe One, 5.15 GHz . . . . . . . . . 160
5.26 Empirical LN Mean EVM vs. Freq., Probe One Band Four . . . . . . . . . . 163
5.27 Empirical LN Mean EVM vs. Freq., Probe Four Band Three . . . . . . . . . 163
5.28 Emp. and Poly. Model of β1 Mag. vs. Freq., Probe One Band Two . . . . . 165
5.29 Emp. and Poly. Model of β1 Phase vs. Freq., Probe One Band Two . . . . . 166
5.30 Emp. and Poly. Model of β4 Mag. vs. Freq., Probe Two Band One . . . . . 166
5.31 Emp. and Poly. Model of β4 PHase vs. Freq., Probe Two Band One . . . . . 167
5.32 Emp. and Poly. Model of σ2 vs. Freq., Probe Two Band Four . . . . . . . . 167
5.33 Emp. and Poly. Model of c1 vs. Freq., Probe Three Band Two . . . . . . . . 168
5.34 Modeled LN Mean EVM vs. Freq., Probe One Band Four . . . . . . . . . . . 169
5.35 Modeled LN Mean EVM vs. Freq., Probe Four Band Three . . . . . . . . . . 169
5.36 Increase of LN Mean EVM From Emp. to Model vs. Freq., Probe 1 Band 4 . 170
5.37 Increase of LN Mean EVM From Emp. to Model vs. Freq., Probe 4 Band 3 . 171
6.1 Probe One (Lower Right) Channel Model Mag. Over Test Zone . . . . . . . 174
6.2 Probe One (Lower Right) Channel Model Phase Over Test Zone . . . . . . . 175
6.3 Probe One Channel Impulse Response at Chamber Center . . . . . . . . . . 176
6.4 Probe Two Channel Impulse Response at Chamber Center . . . . . . . . . . 177
6.5 Probe Three Channel Impulse Response at Chamber Center . . . . . . . . . 177
6.6 Probe Four Channel Impulse Response at Chamber Center . . . . . . . . . . 178
6.7 Probe One, Empirical and Polynomial Fit Spatial Correlation vs. Separation 180
6.8 Probe Two, Empirical and Polynomial Fit Spatial Correlation vs. Separation 180
6.9 Probe Three, Empirical and Polynomial Fit Spatial Correlation vs. Separation 181
6.10 Probe Four, Empirical and Polynomial Fit Spatial Correlation vs. Separation 181
6.11 Probe One, Empirical Spatial Correlation CCDFs vs. Separation . . . . . . . 182
6.12 Probe Two, Empirical Spatial Correlation CCDFs vs. Separation . . . . . . . 183
6.13 Probe Three, Empirical Spatial Correlation CCDFs vs. Separation . . . . . . 183
6.14 Probe Four, Empirical Spatial Correlation CCDFs vs. Separation . . . . . . 184
6.15 0.25 λ× 0.25 λ DUT WB Cap. Vs. DUT Center Pos. . . . . . . . . . . . . . 186
6.16 0.5 λ× 0.5 λ DUT WB Cap. Vs. DUT Center Pos. . . . . . . . . . . . . . . 187
6.17 λ×λ DUT WB Cap. Vs. DUT Center Pos. . . . . . . . . . . . . . . . . . . 187
6.18 1.5 λ× 1.5 λ DUT WB Cap. Vs. DUT Center Pos. . . . . . . . . . . . . . . 188
6.19 0.25 λ× 0.25 λ DUT WB Cap. CCDF . . . . . . . . . . . . . . . . . . . . . 189
6.20 0.5 λ× 0.5 λ DUT WB Cap. CCDF . . . . . . . . . . . . . . . . . . . . . . 189
6.21 λ×λ DUT WB Cap. CCDF . . . . . . . . . . . . . . . . . . . . . . . . . . 190
6.22 1.5 λ× 1.5 λ DUT WB Cap. CCDF . . . . . . . . . . . . . . . . . . . . . . 190
λ λ
6.23
4
× 4
DUT Channel Eigenvalue CCDF . . . . . . . . . . . . . . . . . . . . . 192
λ λ
6.24
2
× 2
DUT Channel Eigenvalue CCDF . . . . . . . . . . . . . . . . . . . . . 192
6.25 λ×λ DUT Channel Eigenvalue CCDF . . . . . . . . . . . . . . . . . . . . . 193
3λ 3λ
6.26
2
× 2
DUT Channel Eigenvalue CCDF . . . . . . . . . . . . . . . . . . . . 193
Abstract
NEAR-FIELD MIMO CHANNEL MODELING WITH APPLICATIONS TO SMALL
ANECHOIC CHAMBERS
By
Jonathan Tet
University of New Hampshire
Demand for wireless communications is rapidly expanding, especially in areas of high-
speed communications including 4K video over wireless HDMI replacements and millimeter
wave communications. Multiple-input multiple-output (MIMO) systems leverage spatial
diversity to achieve higher capacity, helping to meet this demand for high-speed communi-
cations. The cost of consumer wireless devices can be lowered by testing in small, aordable
multi-antenna environments. However, there is a lack of knowledge of MIMO in the near
eld. Near- eld (speci cally Fresnel region and near the Fraunhofer distance) MIMO chan-
nels de ne the performance of these systems, therefore a robust method of near- eld MIMO
channel modeling is needed to aid in system design and testing. The contributions of this
dissertation include a method of iteratively decomposing a sampled broadband Fresnel re-
gion and near-Fraunhofer distance SISO channel into a set of spatial functions. The channel
residual of the spatial functions was statistically analyzed to re ne the channel estimate
model and include eects not modeled in selected spatial functions. A robust simulation
and measurement campaign for model veri cation was performed in a small anechoic cham-
ber, and in turn the SISO models were combined into a MIMO channel description used
to evaluate the MIMO performance of a small anechoic chamber environment. In addition
xiii
to developing and evaluating a wideband MIMO channel model, a feasibility analysis was
performed on the ability of small MIMO-OTA anechoic test chambers to reproduce speci c
MIMO test environments in a test zone.
CHAPTER 1
Introduction
1.1 Motivation
Recently, there has been an explosion in the demand for internet access through wireless
communications networks, including smart phones, tablets, and laptops. Applications such
as streaming video have exponentially increased the amount of data consumed by users as well
as the data rate requirements of such use. Wireless communication is a critical broadband
access strategy in this country and around the world, particularly in rural areas. The US
national broadband plan [7] has stated long-term goals including that at least 100 million
US homes should have aordable access to actual download speeds of 100 Mbps and actual
upload speeds of 50 Mbps. . . The United States should have the fastest and most extensive
wireless networks of any nation. . . Every American should have aordable access to robust
broadband service [7]. By 2020, some estimate that the amount of spectrum necessary to
support mobile broadband is 1,720 MHz.
In addition, improvements in spectral e ciency and increase in the density of network
architectures will be necessary to help meet this demand. According to the 3rd Generation
Partnership Project (3GPP), from 2010 to 2025, there is a desired 1,000-fold increase in
capacity desired to meet demand, requiring, for example three times more spectrum, six
times more spectrum e ciency, and 56 times increased network density [8]. The 3GPP is
on track for that level of usage growth, and is addressing this need in its Release 15 and 16
standards, which de ne 5G [1]. Figure 1.1 illustrates this release schedule.
1
CHAPTER 1. INTRODUCTION
Figure 1.1: 3GPP Release Timeline for 5G [1]
Evolving standards and technologies will help to meet these objectives. These standards
will utilize an increase in available spectrum whether as a primary user or opportunistic
secondary user through software-de ned radio. They will utilize state of the art techniques
to optimize spectral e ciency through use of schemes such as multiple input multiple output
(MIMO), massive MIMO, and multi-user (MU)-MIMO. Network densities will increase, with
cells shrinking to reduce inter-cell interference, increase signal to interference plus noise ratio
(SINR) for the communications link, and therefore increase capacity for each user.
Evolving standards that will take advantage of these techniques include 5G. According
to a white paper by The Wireless Association (CTIA) [9], 5G will be able to utilize spectrum
above 24 GHz, or millimeter wave spectrum. This will allow spectrum blocks in excess of
200 MHz, but will require extremely small cells due to the short range of propagation of
millimeter wave signals, and in addition to their inability to penetrate objects including
walls. CTIA states that the United States Federal Communications Commission (FCC) is
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 2
CHAPTER 1. INTRODUCTION
working to provide access to 10,000 MHz of new high-band spectrum for mobile broadband
in its Spectrum Frontiers proceeding. 5G will be 10 times faster and support 100 times more
devices.
5G will need to yield enhanced mobile broadband through capacity enhancement (giga-
bytes per second for applications such as 4K 3D video), support a massive internet of things
through massive connectivity (e.g. sensor networks), and support low latency and ultra-high
reliability (self-driving cars) [8].
In addition to 5G, other high-data rate applications that utilize MIMO, potentially in the
near- eld, include millimeter wave WiFi 802.11ad (more speci cally its extension 802.11ay),
and high de nition video interconnects supporting resolutions up to 4K (which could replace
High-De nition Multimedia Interface (HDMI)).
With the majority of emerging high-data rate standards utilizing MIMO, the importance
of MIMO test solutions to verify that a device under test (DUT) can operate near capacity is
critical to the goal of maximizing spectral e ciency. Ideally, a small, aordable test solution
is desirable. After determining such a system is feasible, the ability to properly model such
a MIMO system, which due to its small nature will be in the near- eld (speci cally the
Fresnel region and near-Fraunhofer distances), is critical to calculating the capacity of the
system and measuring DUT performance. In addition, a general near- eld MIMO channel
model can assist in the development of emerging near- eld MIMO standards for high-data
rate communications.
1.1.1 Importance of Small Chamber MIMO Test Solutions
In an eort to meet the demand for high-speed wireless networking to provide support
for applications such as multimedia streaming and document sharing, multiple-antenna com-
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 3
CHAPTER 1. INTRODUCTION
munications are and will continue to be deployed. Multiple-Input Multiple-Output (MIMO)
systems employ antennas arrays that increase capacity, spectral e ciency, and cell cover-
age [10]. MIMO systems can operate with many signaling techniques, including Bell LAbs
layered Space-Time (BLAST), Alamouti space-time code, vertical BLAST (V-BLAST) and
Turbo-BLAST; all of the techniques leverage the spatial diversity of the antenna arrays at
the transmitter and receiver [11].
Many recent and emerging standards serve to increase the adoption of MIMO technology,
including IEEE 802.11n, 802.11ac and 802.11ax (Wi-Fi), Long Term Evolution (LTE) and
LTE-Advanced (standards developed by the 3rd Generation Partnership Project (3GPP)),
and WiMAX (IEEE 802.16). These standards utilize multiple antennas in both the base
station and the mobile device. Inherently, MIMO systems are dependent on the spatial
orientation of the antennas at both ends of the links and the geometry of the antenna arrays
themselves. This spatial correlation, in addition to the multi-path fading of the wireless
channel, will impact the performance of a wireless communications link [11]. Therefore, a
simple over-the-air (OTA) test which does not take into consideration this correlation or
fading is not su cient for validating the throughput performance of a multiple antenna
system. The development and deployment of next generation devices greatly depends on the
ability to test and analyze the devices under realistic conditions; it is for this purpose that
MIMO-OTA testing platforms are being developed.
Current MIMO-OTA testing platforms such as large anechoic chambers are expensive,
running in excess of one million USD. Recent research has demonstrated that, for a reduced
set of standard MIMO-OTA channels, a small anechoic chamber with a reduced array of
probe antennas can be utilized. Such a small chamber can cost in the range of 200 thousand
USD, a signi cant savings over the large chamber solutions. This can result in smaller
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 4
CHAPTER 1. INTRODUCTION
development costs for electronics manufacturers, which could serve to lower the cost of next-
generation wireless devices for consumers, which could in turn assist in wider adoption of
these technologies. A small chamber is also more portable, allowing it to be easily transported
into a building through doorways, rather than assembled in a single location in a dedicated
room without the option of relocating.
1.1.2 Importance of Near-Field MIMO Channel Models
High-data-rate near- eld MIMO applications include LTE, LTE-A ,millimeter wave sys-
tems such as 5G and 802.11ad, and HDMI replacements. All of these applications either can
be tested in the near- eld (speci cally the Fresnel region and near-Fraunhofer distances) or
operate in the near- eld, and in either case a full characterization of the wireless commu-
nications channel is critical for de ning system performance limits when developing a new
technology and creating test metrics for throughput testing of an existing technology.
An eort to characterize a near- eld MIMO communications channel must include an
understanding of the propagation between the transmitter and receiver, speci cally, the
propagation between each transmitting antenna, and each receiving antenna. This propaga-
tion between antennas is de ned by electromagnetic theory, and can be modeled using elds
in the environment of the transmit and receive antennas (including near- eld aspects of eld
equations). In addition, interactions with the environment that would aect this eld must
be accounted for, including re ections from environmental objects. It is extremely important
to model as many relevant components of the environment as possible (making the model
accurate), while keeping the model general enough to apply to other similar but not iden-
tical scenarios (making a robust model). A useful model framework should also allow for
re nement (thus increasing accuracy) using real-world measurements to capture modeling
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 5
CHAPTER 1. INTRODUCTION
errors and phenomena that may not have been included in the model. An accurate and
robust model will have the largest impact on, and bene t for, the eld of near- eld MIMO
communications, and that impact and bene t are the motivations of this work.
1.2 Contributions, and Outline
The contributions of the research presented herein are:
• demonstrating the ability of a cost-eective small anechoic chamber for MIMO-OTA
testing to reproduce speci ed electric elds corresponding to standardized MIMO-OTA
test cases [12]. This work showed the di culty of recreating a standard MIMO test
environment in a small test chamber; however for certain con gurations of chamber
size, antenna number and spacing, test zone size, and frequency, it was demonstrated
that such an environment can be created in a small test chamber.
• channel models and the corresponding capacity between MIMO arrays located in the
near- eld. This framework utilizes electric eld equations for antennas which include
terms that are normally negligible in the far- eld.
In addition, Fig. 1.2 includes a owchart for the contributions of chapters 4 and 5. This
ow chart will be used as a guide through these chapters. The additional contributions pre-
sented on the owchart (plus an additional contribution related to MIMO channel analysis)
are:
• A generalized numerical model describing the channel between a log periodic antenna
and a dipole, each of arbitrary spacing and orientation, to calculate the channel vs.
receiver dipole position in a test zone.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 6
CHAPTER 1. INTRODUCTION
Motivating How can the measured What is the most accurate
Question What can the small Can this be simplified for a Can the SAF for one
channel be method of fitting SAF to
chamber channel be fixed TX and a test zone frequency be utilized in a
decomposed into measurement data?
modeled as? RX? wide-band model?
chosen SAF?
Free-space
Properties
Properties
Statistical
SA funcitons
approx.
SAF
Theories Used
Use of Generalized Least
Numerical Model of Use of point Eikonal Use Geostatistical Kriging
Use of point sources (free- Squares, Geostatistical
Antenna Interaction (Hallen Surface correlation, to fit SAF coefficients, fit
space Green’s Functions) Regression and
equations, current multidimensional Fourier coefficients to a polynomial
and plane waves as SAF Geostatistical Kriging to
distribution) analysis model vs. frequency
determine SAF weights
Correlation maximization of Modeling coefficients SAF
Contribution
Generalized Formulation Application of Complex
Combination of point point source Eikonal of a central frequency as a
for Log Periodic Array and Kriging with Geostatistical
sources and plane waves surfaces with measured polynomial function of
Arbitrary Antenna Kriging to Spatial Channel
as model SAF data; 2d Fourier transform frequency separation.
Orientation Measurements
of residual
Iteratively determine
Show Numerical Model can Locate point source SAF
Contribution
unbiased best fit of SAF to Model a wide-band
Applied to
Question
Calculate Channel Vs. be represented using point using correlation,
data, as well as spatial channel using a single set of
Receiver Position in Test source and plane wave SAF determine plane wave
statistics of the residual of SAF and polynomial
Zone for Fixed Transmitter with acceptable error vectors with 2D Fourier
the SAF coefficient functions.
Transform of measurement
Chapter 4 Chapter 5
Figure 1.2: Contribution Flowchart for Chapters 4 and 5
• A method of using both point source elds and plane waves as joint spatial attribute
functions (SAFs) to accurately describe the numerical channel model or a measured
channel as a concise sum of functions with acceptable error.
• A method of tting point source functions to a numerical or measured channel using
a phase correlation maximization search, then use two-dimensional Fourier analysis to
re ne the approximation with plane waves.
• Combining methods of Complex Kriging, Geostatistical Regression and Geostatistical
Kriging to statistically model the residual of the SAF function approximation to fur-
ther re ne the channel estimate. Applied this method to an extensive measurement
campaign performed in a small anechoic chamber.
• An extension of the narrowband SAF plus residual model to a wideband channel model,
utilizing frequency-scaled SAFs and polynomial coe cient and parameter models. This
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 7
CHAPTER 1. INTRODUCTION
yields a wideband SISO channel model of the small anechoic chamber which completely
describes the channel as a mathematical function of frequency and receiver position in
a test zone.
• An application of the small anechoic chamber wideband SISO channel models to wide-
band MIMO channel model, and an extensive analysis of the MIMO channel properties
of the modeled small anechoic chamber.
This proposal is organized as follows:
• Chapter 2 presents fundamental background information on near- eld electromagnetics
and MIMO topics, anechoic chambers, and Kriging, as well as an overview of recent
literature in this eld. Also included is a preliminary investigation into simple Hertzian
dipole near- eld MIMO channel models, and
• Chapter 3 presents a computational electromagnetic analysis of the feasibility of small
anechoic chambers for MIMO-OTA eld synthesis.
• Chapter 4 presents a numerical near- eld free-space channel model and a method for
selecting and tting a set of mathematical spatial attribute functions to a measurement
set or sampled numerical model. This results in a narrowband functional model of the
channel.
• Chapter 5 presents two extentions of the narrowband functional channel model, one
re ning the model by using a statistically-informed estimate of an additive residual,
and the other extending the model to a wideband model. These model extensions are
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 8
CHAPTER 1. INTRODUCTION
applied to an extensive small chamber measurement campaign data set and veri ed for
accuracy.
• Chapter 6 presents an extensive application of the small chamber channel model to a
MIMO channel analysis of the small chamber, providing insight as to the performance
of MIMO systems in a small anechoic chamber.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 9
CHAPTER 2
Background
This chapter contains all relevant background to subsequent contributions. Fundamental
concepts from the eld of wireless communications relevant to the research performed for
this dissertation are outlined. A literature review containing recent, relevant research in the
areas of small anechoic chambers and near- eld MIMO channel modeling will be performed.
A section of preliminary investigations is included to show our initial work and analysis that
preceded the contributions in this dissertation.
2.1 Fundamental Concepts
In this section, fundamental concepts in single-input single-output (SISO) and MIMO
wireless communications, anechoic chamber theory, and antennas and electric elds in the
near- eld will be reviewed.
2.1.1 SISO channels
This subsection will serve as a quick review of fundamentals of SISO wireless channels,
consisting of a single transmitting antenna and single receive antenna. Wireless channels
are linear in nature; if we assume the channel to be time-invariant (does not change over
time), causal (outputs can only happen concurrent with or after inputs), and time-dispersive
(input energy is spread out over time at the output), we can describe a relationship between
its input and output as an impulse response h in the time domain, or a transfer function
H in the frequency domain [11]. If the wireless channel is truly time-invariant, then H is
10
CHAPTER 2. BACKGROUND
completely deterministic.
Wireless channels in most real-world situations are time-varying and can be modeled
as such with a time-dependent transfer function T(t,f ). This is due to motion in a real-
world environment, with various re ections of the signal bouncing o and scattering from
moving objects in the environment and adding constantly-varying copies of the signal con-
structively or destructively at the receiver. This phenomenon is known as fading, and the
channel between the transmitter and receiver can change rapidly (known as fast fading) or
slowly (known as slow fading). Unfortunately, this makes both analysis and communications
systems much more complex. However if the channel changes slowly in time in relation to
independent channel accesses (and the channel is su ciently low-bandwidth), i.e. if the chan-
nel exhibits slow fading, it can be modeled as time-invariant (i.e. H can be used in lieu of
T ) for analysis. In other words, in this scenario the channel has a su ciently long coherence
time for the utilized communications scheme [11]. For each access of the channel, however,
H can have dierent random values in this scenario. If H is independent and identically
distributed (I.I.D.) for each channel access, the channel distribution can be characterized
and modelled as a random variable. (Characterization and modeling of channels where the
coherence time is short relative to the communications system (i.e. fast fading) is possible,
but is outside the scope of this dissertation.)
If a channel is considered narrowband (ideally modeled as a single sinusoid at a frequency
f0 ), the impulse response h(t) becomes valid at a single frequency H(f0 ) in the frequency
domain, and it becomes a single complex constant h, or complex gain. This constant can
be multiplied by the transmitted signal in either the frequency domain or time domain to
produce the received signal [11]. This complex constant h will aect the amplitude and phase
of the narrowband signal. Communications systems such as orthogonal frequency division
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 11
CHAPTER 2. BACKGROUND
multiplexing (OFDM), where a wideband channel is divided into independent narrowband
orthogonal sinusoidal channels (subcarriers) with amplitude and phase representing data in
symbols of duration less than the coherence time, are designed in such a way that can utilize
independent channel estimations h for each subcarrier and symbol period(s) [13].
In addition to the eect of the complex channel impulse response, noise is added to the
received signal in a channel model, usually modeled as complex additive white Gaussian
noise (AWGN) (i.e. both the real and imaginary components of the noise follow a Gaussian
distribution, and are spectrally uniform) [11]. Thus, a narrowband SISO channel can be
modeled as:
y(t) = h · x(t) + n(t) (2.1)
2.1.2 MIMO
If both the transmitter and receiver utilize multiple antennas, the system becomes multiple-
input multiple-output (MIMO). Each pair of transmit and receive antennas is assumed to
represent an independent channel that can be modeled as in Eq. 2.1. If we consider the
transmitted signal as a vector ~x of length Nt (the number of transmit antennas), the re-
ceived signal as a vector ~y of length Nr , an impulse response between TX antenna i and RX
antenna j as hji , and an AWGN vector ~n of length Nr , then the vectors ~x and ~y can be
related by the following equation:
~ = Hx(t)
y(t) ~ + n(t)
~ (2.2)
where H, the Nr × Nt channel matrix, consists of elements hji , the complex gain between
each pair of transmit and receive antennas [11]. This can be described as a vector Gaussian
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 12
CHAPTER 2. BACKGROUND
channel [13].
In short, each individual received signal is the sum of a gain-adjusted and phase-shifted
version of each and every transmitted signal plus some additive AWGN. The channel matrix,
determined by environment, determines just how much of each transmitted signal is received
at each antenna. The greater the distance between antennas in the transmit array, and
the greater the distance in antennas in the receive array, the greater the variation in the
combination of signals received at each antenna at the receiver. This spacing (which can be
adjusted and optimized to meet the needs of communications systems), provides a spatial
dimension to the system, or in other words, allows for spatial multiplexing of multiple data
streams between the transmitter and receiver [13]. In certain environments (i.e. certain
channel matrix formulations and distributions), utilizing MIMO can signi cantly increase
the capacity of a communications link over a similar SISO link.
Channel Matrix Analysis and Decomposition
To evaluate the bene ts of MIMO, an analysis of the channel matrix H is performed. H
is a matrix that in the context of Eq. 2.2 can be considered a linear transformation between
the transmit and receive vectors. A linear transformation can be decomposed using singular
value decomposition (SVD) into a product of three matrices: a rotation matrix U (unitary
matrix, i.e. the Frobenius norm of this matrix equals 1), a scaling matrix Λ (diagonal
matrix consisting of the singular values of H), and another rotation matrix VH (another
unitary matrix; H represents the Hermetian operator, i.e. conjugate transpose [13]). If we
pre-multiply the transmit signal vector ~x by V (x(t)
g ), and post-multiply the received signal
vector ~y (and thus also the noise vector) by UH (y(t)
g and n(t)
g ), (which preserves the norm,
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 13
CHAPTER 2. BACKGROUND
and thus power values, of ~x, ~y , and ~ ),
n(t) then the channel can be rewritten as [13]:
y(t)
g = Λx(t)
g + n(t)
g (2.3)
Which, due to the diagonal matrix Λ can be written as a set of parallel Gaussian chan-
nels:
yei = λi xei + n
ei (2.4)
where i is the eigenvalue matrix index ii, of which there are RH values (rank of H). The
power of the transmitted signal xei is equal to Pi , and the received signal power is thus Pi · λ2i .
(The sum of all Pi must equal the total transmitted power Pt .) The power of the noise at
the receiver is N0 . [13] The signal to noise ratio (SNR) of each parallel Gaussian channel,
i.e. the signal power received divided by the noise power received, is:
Pi ∗ λ2i
(2.5)
N0
General MIMO Capacity
From information theory, the Shannon capacity of a SISO channel with normalized band-
width is equal to C = log2 (1 + SN R). Using the derived expression for SNR, the capacity
for each parallel channel becomes:
Pi · λ2i
!
Ci = log2 1+ (2.6)
N0
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 14
CHAPTER 2. BACKGROUND
The total MIMO channel capacity is equal to:
RH
Pi · λ2i
!
X
C= log2 1+ (2.7)
i=1 N0
This formula for instantaneous narrowband MIMO capacity was rst introduced by Fos-
chini and Gans [14]. For a wideband MIMO channel, the wideband capacity becomes an
average of narrowband capacities [11]. The wideband MIMO channel is sampled at regular
frequencies ∆f for some bandwidth W to yield a series of narrowband channel matrices H[f ],
each of which is used in the narrowband capacity formula, and averaged over the range of
frequencies sampled (2.8) [11]:
F
1 X ρ
H
CX = log det I + H̃[f ]H̃ [f ] ∆f (2.8)
W f =1 2 MT x
where F is the number of frequencies for which H is sampled, H is the Hermitian transpose
operator, ρ is the linear transmit SNR, MT x is the number of transmit (probe) antennas,
and I is the identity matrix. H̃ is the normalized channel matrix (2.9):
1
H̃[f ] = v H[f ] (2.9)
M Rx M F
1 Tx X
u
|hi,j,k |2
u X X
t
MRx MT x F i=1 j=1 k=1
where MRx is the number of receiver antennas, and hi,j,k is the (i, j)th element of H[f (k)].
Essentially, this is normalizing H[f ] by the average element magnitude of H[f ] across all
frequencies.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 15
CHAPTER 2. BACKGROUND
Water lling Power Allocation and Capacity
The optimal distribution of power for each independent channel, Pi∗ , can be determined
using a water lling algorithm. Water lling can occur if the transmitter has knowledge of the
channel matrix H. The transmitter determines the eigenvalues, ~λ2 , of the channel matrix.
Then the transmitter initializes a variable µ, which is iteratively updated and will be used
to calculate the optimal power allocation.
N0 Pt
µ = min + (2.10)
~λ2 RH
Another variable p gets initialized, which is the sum of allocated power given the current
value of µ:
RH
!
X N0
p= max µ − ,0 (2.11)
n=1
~λ2
n
the following steps are performed iteratively until the dierence between Pt and p is arbi-
trarily small:
Pt − p
µ=µ+ (2.12)
RH
RH
!
X N0
p= max µ − ,0 (2.13)
n=1
~λ2
n
This iteratively adds power to channels with high eigenvalues rst, then progressively raises
the level of µ thus adding power to subcarriers with su ciently high eigenvalues relative to
µ until all power has been allocated. Once the value of µ is determined that allocates all
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 16
CHAPTER 2. BACKGROUND
power, power is optimally assigned to independent channels as [13]:
N0
P~ ∗ = max 0, µ − (2.14)
~λ2
MIMO Channel Metrics
Two metrics of the channel matrix can indicate how good a channel is for MIMO commu-
nication. First, the rank RH of the channel matrix, or the number of non-zero eigenvalues,
indicates the number of spatial degrees of freedom. If a channel matrix is full-rank, then
it maximizes the dimension of the MIMO channel for the given system with Nt transmit-
ting antennas and Nr receiving antennas [13]. Second, the condition number of the ma-
max (~λ)
trix γ= min (~λ)
, indicates how spread out the singular values of the channel matrix are. If
the singular values are uniformly distributed, then the condition number is equal to 1 and
the matrix is considered well-conditioned. The closer the condition number gets to 1, the
larger the capacity in high SNR conditions [13].
Several properties play into the de nition and characterization of a MIMO channel. De-
terministic (static) properties that contribute to the channel include transmit and receive
array orientation (often used are uniform linear arrays (ULA) for analysis) and antenna spac-
ing, number of resolvable paths in the angular domain (and whether one of these paths is
line of sight) [13]. Analytical methods to determine the rank and condition number of these
deterministic channels is well-studied, particularly in the far- eld. General trends reveal
that small antenna separation in both transmit and receive arrays in the line-of-sight case
yield no spatial degree of freedom gain, large antenna separation in the transmit array or
receive array provide a spatial degree of freedom gain (even in the line-of-sight case), and
multipath channels with large angular separation between multipath components provides
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 17
CHAPTER 2. BACKGROUND
a spatial degree of freedom gain even with low separation in the transmit array or receive
array [13].
Stochastic MIMO Channels
As discussed previously, the channel matrix may be stochastic rather than deterministic
due to the presence of a fading channel. The standard assumptions in these stochastic models
are that the channel exhibits su cient coherence time such that discrete samplings of the
random process (i.e. constant multiplicative complex matrix entries) can be made, and that
each entry of the matrix is I.I.D. The ability of a stochastic channel matrix to represent a
rich multipath environment can be analyzed through transformation of the channel matrix
into an angular domain representation (i.e. pre- and post-multiplying the channel matrix by
unitary matrices constructed from basis vectors for the transmit and receive signal spaces,
or in other words Inverse Discrete Fourier Transform (IDFT) matrices). The elements of
the angular domain representation matrix correspond to transmitter and receiver angular
bins corresponding to direction of departure and direction of arrival respectively, i.e. one
resolvable path. Elements of this matrix above a certain threshold indicate that the channel
has at least one multipath component along this resolvable path. Assume that elements
below the de ned threshold are zero. The rank of the angular channel matrix can be de ned
as the minimum between the number of non-zero rows and non-zero columns [13].
Rayleigh Fading
One common stochastic MIMO channel model is the I.I.D. Rayleigh fading model, where
each element in the channel matrix is I.I.D. circularly-symmetric complex Gaussian. The
distribution of the magnitude of this complex random variable (the square root of sum of
the squares of the real and imaginary components) follows a Rayleigh distribution. If we
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 18
CHAPTER 2. BACKGROUND
investigate the angular domain channel matrix of such a process, we see that the Rayleigh
channel has I.I.D. Gaussian distribution in the angular domain as well, indicating that the
expected value in every resolvable path, i.e. the expected value of a Rayleigh random variable
(which is non-zero by de nition) should be above the threshold. This means that every
resolvable path is utilized between the transmitter and receiver, and the simulated channel
is richly scattered [13]. It should be noted that because each of the two Gaussian processes
that constitute the real and imaginary parts of the channel matrix are zero mean, there is
no dominant (line-of-sight) path between the transmitter and receiver.
Rician Fading
In the case where a line of sight path can be modeled as dominant, but there is also a
richly-scattered channel present, a Rician fading model is appropriate. In essence, Rician
fading is just Rayleigh fading, except the real and imaginary Gaussian random variable
components are not zero-mean, with the real and complex component means equaling the
complex gain of the line-of-sight path between transmitter and receiver. For a Rician channel
model, this complex line-of-sight gain must be determined for each transmit-receive antenna
pair either analytically or experimentally. Then, the ratio of energy in the line-of-sight path
to the energy in the scattered paths κ must be determined [15]. The resulting channel matrix
entry is the following assuming a variance of σ 2 = hijLOS :
s s
κ 1
hij = hij + CN (0, |hijLOS |2 ) (2.15)
κ + 1 LOS κ+1
where CN is the complex normal distribution. It should be noted that when κ is zero, the
channel matrix becomes Rayleigh, and if κ approaches in nity, the channel matrix becomes
deterministic with hij = hijLOS .
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 19
CHAPTER 2. BACKGROUND
MIMO Channel Clusters
Realistic MIMO channel models involve one or multiple multipath clusters impinging on
a receiving array. Each multipath cluster is centered around a central angle of arrival, and
consists of numerous multipath elements, each of which follows a statistical distribution in
time (delay τ ), azimuth angle φ (Power Azimuthal Spread (PAS)), and elevation angle θ.
Each multipath component within a cluster is selected from a fading distribution, which is
correlated to the distributions in time and angle [11]. With a set of randomly generated
multipath components following these distributions, a channel matrix H can be generated
from the sum of their components. Several standardized MIMO channel models have been
developed through the modeling of multipath clusters, including Saleh-Valenzuela, Extended
Saleh-Valenzuela, COST 273, Random Cluster Model (RCM), and others [11].
Mutual Coupling at the Transmitter and Receiver
Mutual coupling between elements of the transmitter can have an eect on the MIMO
system capacity. The relation between the source voltage v~s and terminal (antenna) voltage
~v is [16]:
ZT = ZT (ZT + ZS )−1 (2.16)
~v = ZT v~s (2.17)
where ZS is the diagonal matrix of N source impedances and ZT is the open-circuit, Nt port
transfer impedance matrix at the transmitting end. Therefore ZT in 2.16 becomes part of
the channel used in capacity analysis, i.e. ~y = HZT ~x.
Mutual coupling between elements at the receiver can also have an eect on MIMO system
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 20
CHAPTER 2. BACKGROUND
capacity [17]. The matrix relation (ZR ) between the voltage induced by the antennas without
mutual coupling (HZT ~
x) and with mutual coupling (~y ) is:
Z 1,2 Z 1,N −1 Z 1,N
1 − ZRL ··· − RZL − ZRL
1 1 1
2,1 2,N −1 2,N
− ZR ZR ZR
ZL2
1 ··· − Z L2
− ZL
2
. . .. . .
. . . .
~y =
. . . . . HZT ~
x = ZR HZT ~x (2.18)
Z N −1,1 Z N −1,2 Z N −1,N
− R − ZLR ··· 1 − ZLR
ZL
(N −1) (N −1) (N −1)
N,N −1
Z N,1 ZRN,2
ZR
− ZRL − ZL ··· − ZLN
1
N N
where ZLn is the load impedance of receiver port n, and ZRm,n is the mutual impedance
at the receiver between ports m and n. The resulting channel matrix including mutual
coupling HM , which can be used in subsequent capacity analysis (as described in 2.1.2), is
HM = ZR HZT [2].
2.1.3 Near-Field Topics and Mutual Coupling
In the following subsection, we discuss the distinction between near- and far- eld, and
properties of mutual coupling.
Near- eld Vs. Far Field Ranges
An antenna radiating several modes has three regions of space de ned in terms of
2π
wavenumber k= λ
, the radius of the minimum sphere of the antenna r0 , and N = [kr] + n1
where r is the radius from the center of the antenna, the brackets indicate the largest integer
smaller than or equal to kr, and n1 is an integer which depends on the accuracy desired for
the system, which can be equal to 5 or less for small values of kr [18]. The three ranges of
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 21
CHAPTER 2. BACKGROUND
space are evanescent (r0 ≤ r ≤ N/k ), Fresnel region (near eld) (N/k ≤ r ≤ 4N 2 /πk)),
2
and Fraunhofer region (far eld) (4N /πk ≤ r ≤ ∞). It should be noted that the Rayleigh
distance R (the boundary between near and far elds), for a source with minimum sphere of
2D2
diameter D = 2r0 , is often de ned as R= λ
. This distance is approximate, and making
far- eld assumptions near the Rayleigh distance can still produce signi cant phase error.
Additional Mutual Coupling Topics
Mutual coupling eects between closely-spaced linear antennas can be modeled using
electric and magnetic elds produced at near-distances, which can be obtained by deter-
mining the magnetic vector potential A. In [19], sections 23.1 and 23.2 outline a rigorous
mathematical analysis using sinusoidal current assumptions for determining the near eld for
λ
linear antenna (e.g. general dipole antennas) separations greater than . This antenna sep-
10
aration is more than su cient for use in our research. Mutual impedance and self-impedance
are calculated in section 23.3 using equations relating induced open circuit voltage to inci-
dent electric eld and current. Section 23.5 utilizes the mutual/self impedance matrix along
with driving voltages for each antenna in an array to determine the input currents of each
antenna, which along with the sinusoidal assumption that de nes each element current I(z),
can be used to de ne the electric and magnetic elds produced by the array [19].
2.2 Literature Review
In this section, we reviewed recent literature related to our research eorts. A summary of
this literature was included here to frame the context of our research, and to present the origin
and evolution of some of the techniques and methods used in our research. Our literature
review encompassed the areas of MIMO-OTA testing and near- eld MIMO channels and
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 22
CHAPTER 2. BACKGROUND
corresponding capacity.
2.2.1 MIMO-OTA Testing
This subsection reviews literature on MIMO-OTA testing topics. This literature will
help place the contributions of Chapter 3 of this document in context, and help the reader
understand the state of the art in MIMO-OTA testing. Topics of the literature reviewed
in this subsection include MIMO-OTA chamber types, MIMO-OTA test system calibration,
standard channel emulation in anechoic chambers, full-ring MIMO-OTA plane wave eld
synthesis, and partial-ring MIMO-OTA eld synthesis.
MIMO-OTA Chamber Types
There are several dierent proposed MIMO-OTA systems in the literature: full ring [20,
21] and full-sphere, partial ring [22], and reverberation chamber [23]. All of these systems
are designed to emulate the spatial distribution of power that would occur in a multipath
environment. The spatial distribution of power is dependent on multiple incident waves
of dierent power levels and angles of arrival. The incident power as a function of angle
in a multipath environment can be modeled as multiple clusters, each exhibiting a Lapla-
cian Power Azimuthal Spectrum (PAS). This angular dependence of the incident power is
important as the achievable data rates are inherently tied to the orientation of the antennas.
The full-ring system includes evenly-spaced MIMO-OTA probes in a circle around the
device-under-test (DUT) [21]. This con guration enables the creation of multiple clusters
of energy impinging on the DUT. The generation of multiple clusters makes it possible to
test many dierent wireless environments, but at the cost of a large chamber with a high
number of probes and hardware. Eorts to shrink a full-ring system into a small chamber
with ring radius of 0.5m have been shown to be successful in [24], however these eorts have
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 23
CHAPTER 2. BACKGROUND
dealt with target zone correlation error rather than accuracy of eld synthesis.
The reverberation chamber is an anechoic chamber that has re ectors that redirect the
signals to distribute the power over dierent angles [23]. These facilities are also large and
the control of the power distribution is limited. Finally, the small chamber implementation
considered for our work uses a subset of antennas with respect to the full ring to synthesize
a single cluster, or partial ring. The size and fewer number of probe antennas of the partial
ring small chamber makes it more cost eective than its full-sized or full-ring counterpart,
but the number of wireless environments that can be generated is limited [22]. While a single
cluster does not provide the spatial diversity that results in the highest MIMO data rates,
it does provide a standardized, realistic, and repeatable environment in which to perform
MIMO-OTA testing [25].
Through simulation of a single cluster with Laplacian PAS eld, it has been shown
that a low number of probes can be used to generate a desired electromagnetic eld in a
test zone [25]. Previous work, however, only focuses on minimizing the number of probes,
and does not focus on constraining these probes to a small chamber. This previous work
therefore does not investigate parameters such as chamber dimensions, antenna arrangement,
or frequency.
MIMO-OTA Test System Calibration
A modi ed test zone eld compensation technique for calibrating multi-probe antenna
measurement systems was presented in [26]. This method yielded performance similar to
traditional calibration methods and can be used in environments with high levels of sig-
nal re ections. This technique was expanded in [27], allowing for full 3-D antenna pattern
measurements without an anechoic measuring environment. Small-antenna pattern mea-
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 24
CHAPTER 2. BACKGROUND
surements can be made in non-anechoic environments where multi-re ection amplitudes are
small relative to primary re ections. This technique can also be used to calibrate out the
re ections caused by multiple antennas in an anechoic chamber MIMO-OTA test system.
These methods were extended in [28] to multipath environment synthesis in a spherical test
zone with a limited number of probes. Arbitrary electromagnetic environments can be syn-
thesized with this method (including near- eld eects), including near- eld environments,
due to spherical-wave theory. Spherical wave theory will be outlined in Chapter 3, in section
3.1.1.
Standard Channel Emulation in Anechoic Chambers
The literature has several papers exploring MIMO channel emulation based on various
standards in full-ring anechoic chambers, speci cally the eect of many chamber parameters
(mostly probe antenna con gurations) on numerous channel emulation metrics. The capa-
bility to recreate required radio propagation channel characteristics including power delay
pro le, Doppler spectrum, and spatial auto-correlation in models such as 3GPP SCM, SCME
and IEEE 802.11n in a full-ring MIMO-OTA anechoic chamber test zone was investigated
in [29]. A relation between number of probe antennas and test zone size for accurate spatial
channel emulation was established in this work. Considerations of probe antenna spacing,
radiation patterns, and mutual coupling and their eect on channel emulation accuracy were
further investigated in [30]. The antenna correlation and throughput performance of three
channel models were investigated for three antenna separation con gurations in this work.
An investigation into the ratio of probe ring radius to the test zone radius and its eect on
channel emulation accuracy requirements at various frequencies was investigated in [31].
In addition to probe antenna con guration, the eects of device under test con guration
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 25
CHAPTER 2. BACKGROUND
on MIMO-OTA full-ring test system requirements have been investigated in the literature.
The impact of device under test antenna patterns (speci cally variation in directivity) on the
number of probe antennas necessary for plane wave synthesis (and equivalent induced voltage
(EIV) technique, a technique with almost identical results for large MIMO-OTA setups) and
prefaded signal synthesis was performed in [32]. This impact was measured using metrics of
received voltage and spatial correlation accuracy. Simulation and measurement showed that
higher variation in directivity requires a higher number of probes for comparable accuracy.
Beyond full-ring con gurations, some papers have explored 3-dimensional channel emu-
lation. An investigation into 3-D spatial correlation channel model emulation in a spherical
probe array MIMO-OTA test setup was performed in [33]. A framework to emulate any
spherical incoming power spectrum, with optimal emulation accuracy and low computa-
tional complexity, was presented in this work.
Two methods of MIMO-OTA channel emulation exist in the literature, prefaded signal
synthesis (which relies on correlation within the test zone) and plane wave synthesis, both
of which apply weights to the signal applied to each probe antenna. In [34], mapping of
channel models onto MIMO-OTA probe antennas, speci cally through generation of complex
antenna coe cients through prefaded signal synthesis and plane wave synthesis, is explored.
Evaluation and veri cation through simulation was performed, with the bene ts of each of
the two antenna coe cient generation methods analyzed. The cases where each would be
appropriate were outlined. Plane wave synthesis methods were determined to be superior for
the scenario of a limited antenna ring radius, and therefore would be the antenna coe cient
method of choice for a small chamber environment, the focus of our work.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 26
CHAPTER 2. BACKGROUND
Full-ring MIMO-OTA Plane Wave Field Synthesis
Several papers in the literature investigate plane wave synthesis methods of channel
emulation for MIMO-OTA systems. A review of existing investigations on MIMO-OTA test
methodology is performed in [35], and rules for the required number of probes as a function of
test zone size in wavelength for a desired uncertainty or accuracy in test zone eld synthesis
are presented. The accuracy in test zone eld synthesis is de ned in terms of deviation from a
target electric eld. Probe orientation error and location mismatch and their eect on plane
wave synthesis of an electric eld in a MIMO-OTA full-ring test zone are considered in [36].
A nite-dierence time-domain (FDTD) simulation was used to determine the performance
deterioration caused by these errors. It was shown that radial location mismatch error is
found to have the greatest eect on performance.
The su ciency and advantage of a full-ring system over a full-sphere con guration has
been investigated in the literature. In [37], a 2-D circular array (full-ring) for eld synthesis
in MIMO-OTA testing is investigated in terms of plane wave synthesis analysis, and a case
for a 2-D full-ring test system over a 3-D full-sphere test system is made. Many channels
and channel models have constrained PAS to the horizontal plane. The use of 2-D systems
provides a channel emulation system cost- and complexity-reduction.
Channel emulation through the synthesis of multiple simultaneous plane waves in a test
zone is possible (including cluster models formed through the sum of numerous plane waves
following an angular spectrum), and papers have investigated the requirements of a full-
ring system to synthesize such a eld. A synthesis and measurement of single- and multi-
plane wave elds with multiple angle of arrival (AoA) and weightings is performed in [38],
with phase and power deviations of the synthesized eld investigated. It is shown that the
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 27
CHAPTER 2. BACKGROUND
superposition of multiple plane waves with dierent AoA can cause deep fading, which can
cause phase jumping and inaccuracies in measurements.
Partial-ring MIMO-OTA Field Synthesis
There has been some work in the literature regarding partial-ring anechoic chamber
MIMO-OTA systems. In [25], the number of probe antennas required for generating a single
Laplacian PAS cluster for varying angular spreads and test zone sizes in an anechoic chamber
is established. A Laplacian cluster target eld to model this channel is generated through the
sampling of a Laplacian distribution to determine the amplitude of a series of plane waves
from varying AoA, each with random, uniformly-distributed phase. This paper investigated
partial ring systems, including probe antenna sector angle and number of antennas, and their
eect on synthesized eld accuracy. This paper is the basis of much of our small chamber
work presented in this proposal. It, however, does not discuss how non-ideal eects such as
re ection contributions in a small chamber would aect the eld synthesis accuracy, which
is a focus of our work.
2.2.2 Near-Field MIMO Channels and Capacity
This subsection reviewed literature near- eld MIMO channels and capacity. This litera-
ture helps to place the work in Chapters 4 through 6 of this document in context, and helps
the reader understand the state of the art in near- eld MIMO channel modeling. Topics in-
cluded in this subsection include short-range MIMO channels, near- eld scatterers in MIMO
systems, near- eld MIMO channel modeling, and near- eld MIMO capacity improvement.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 28
CHAPTER 2. BACKGROUND
Short-Range MIMO Channels
Short-range MIMO communications have been described in the literature for over a
decade, with papers going back to at least 2005 found in a literature search [39]. Such
systems are mostly at frequencies in excess of 6 GHz, and are often in the millimeter wave
range. Due to the high frequencies involved in this type of communication, the term `short
range' may or may not actually be in the near eld. Early work in this area in [39] describes
the use of a `spherical-wave model' for analysis of arrays with separation distance in the near-
1
eld. This analysis, however, turns out to be a spherical-geometry approximation using a
r
line-of-sight path loss approximation based on inter-element separation distance rather than
inter-array separation distance, and is thus a far- eld approximation of a near- eld channel
rather than a true spherical-wave analysis in the typical sense. This work shows that using
this approximation produces a more accurate estimation of capacity than assuming a plane
wave (i.e. at wave front) impinging on the entirety of the receive array.
Essentially all papers reviewed in this area utilize the same far- eld line-of-sight path loss
assumptions for capacity analysis [4042]. Also, these papers note that the close spacing of
corresponding antenna elements at the transmitter and receiver result in a strong line-of-sight
path, and relatively low interference from neighboring transmit elements, thus producing a
series of almost parallel data streams without multipath components, with this assumption
being especially valid at higher frequencies. In [40] (and to a less rigorous extent in [41]),
optimal element spacing was investigated and veri ed through measurement. In [42], a short-
range 2×2 ultra-wideband MIMO system is shown to have double the capacity of a SISO
system with the same characteristics, indicating that a parallel-stream assumption for the
ideal case of perfect array alignment is valid.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 29
CHAPTER 2. BACKGROUND
Near-Field Scatterers in MIMO Systems
A couple papers were encountered that modeled the eect of near- eld scatterers near
a MIMO array on capacity. One paper, [43], discussed the eect of near- eld scatterers on
the mutual coupling of a MIMO antenna array and the corresponding eect on channel and
capacity. This paper modi es a previously proposed model for mutual coupling for MIMO
systems to include and characterize the eect of near- eld scatterers. This modi cation is
validated, then used to accurately obtain correlation properties between antenna elements,
and as a result measure the eect on system capacity. Another paper, [44], investigated
through modeling the eects of the equipment case and human user on a MIMO system in
a handset.
Near-Field MIMO Channel Modeling
A small number of papers were found which modeled the near- eld MIMO channel in a 2×
2 system where each array consisted of two elements utilizing some method of orthogonality.
The rst set of these papers utilized a small electric dipole and a small magnetic dipole
in each array. In [2], the 2×2 system with a short electric dipole and small solenoidal
loop (Figure 2.1) is analyzed using the assumption that each element produces a T M10 and
T E10 mode respectively (Figure 2.2), which are orthogonal. An equivalent circuit diagram is
presented in Fig. 2.3. Combining spherical wave propagation of each element and coupling
matrices at the transmitter and receiver, the eective channel matrix is calculated and near-
eld MIMO capacity is determined.
An almost identical approach involving a pair of electric and magnetic dipoles at the
transmitter and receiver is outlined in [45], though this paper neglects the mutual coupling
of the transmit and receive arrays, and lacks the mathematical rigor of [2], but includes an
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 30
CHAPTER 2. BACKGROUND
Figure 2.1: 2 Element Dipole and Loop Antenna Array [2]
Figure 2.2: Near-Field 2×2 MIMO Example [2]
indoor measurement campaign at 1 MHz comparing the ratio of MIMO capacity to SISO
capacity.
The second type of paper found analyzing near- eld MIMO was one utilizing orthogonal
polarizations of two electric dipoles at both the transmitter and receiver. In [46], the au-
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 31
CHAPTER 2. BACKGROUND
Figure 2.3: Equivalent Circuit for Near-Field MIMO System [2]
thors of which also wrote [2], uses an almost point-for-point identical analysis in this paper,
with every pair of antennas using a T M10 mode this time. All other aspects of analysis are
identical, including the mutual coupling formulations, equivalent circuit description, matrix
equations, etc. While this work is similar to the work presented in the Preliminary Inves-
tigations section of this dissertation chapter, there are three distinct dierences. First, the
dipoles modeled in our analysis are parallel and co-planar, and thus co-polarized and not
orthogonally polarized. Second, we compare near- eld to far- eld capacity utilizing a nor-
malized channel, whereas [46] does not normalize and only oers capacity values for a small
range of near- eld distances. Third, our formulation utilizes the Hertzian dipole equation
which encapsulates all propagation modes, while this paper utilizes a single mode spherical
mode approximation of the electric eld, and thus our model is more robust. In short, our
preliminary research could be considered more comprehensive and extensible than that in
[46].
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 32
CHAPTER 2. BACKGROUND
Near-Field MIMO Capacity Improvement
Methods to increase the capacity of near- eld MIMO have been proposed in the literature.
One paper [3] proposes a structure at both the transmit and receive antenna arrays consisting
of a back re ector and two side re ectors (Fig. 2.4), and analyzes capacity through FDTD
simulation, as shown in Fig. 2.5. This method increases the norm and decreases the condition
number of the channel matrix, thus increasing channel capacity.
Figure 2.4: Near-Field MIMO Array Re ector [3]
Figure 2.5: FDTD Near-Field MIMO Array Simulation [3]
Another paper (dissertation [47]) proposes the use of use of two-element dipole arrays
(with separation determining the element's half power beam width(HPBW)) as a single
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 33
CHAPTER 2. BACKGROUND
element in a multi-element transmit array, with this dual-dipole array forming a beam
directed at the corresponding element of the receive array. This con guration almost creates
an independent, parallel connection between each pair of antennas, similar to the concept of
Short Range MIMO previously described. An optimization of element HPBW was performed
to maximize capacity. No multipath was considered in analysis. Secondly, the addition of
metal wires to the inter-array space to cause scattering and increase multipath richness
was introduced. The eect of these wires on capacity was modeled and investigated. This
dissertation suggests that element HPBW optimization in conjunction with objects placed
between the transmit and receive arrays in near- eld MIMO systems can increase capacity
of near- eld MIMO systems.
2.3 Preliminary Investigations
This section contains two preliminary investigations that served as the starting point
of work in this dissertation. The rst investigation serves as a simple initial assessment of
MIMO capacity in the near- eld using Hertzian dipoles. The second serves as an extension of
textbook near- eld dipole interaction equations for simple parallel orientations to arbitrary
orientations. Neither investigation is novel enough to consider a contribution, but both are
important enough to include as background for this document.
2.3.1 Hertzian Dipole Field-Based Channel Matrix Formulation
A MIMO channel at a single frequency can be described as a channel matrix H (a complex
matrix which describes the amplitude gain and phase shift from a vector of applied voltage
signals at the transmit antenna array (~
x) to a vector of induced voltage signals at the received
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 34
CHAPTER 2. BACKGROUND
antenna array (~
y )) and an additive noise vector ~n. This relation is as follows:
~y = H~x + ~n (2.19)
The channel matrix can either be deterministic, a stochastic process, or a sampling of a
stochastic process. Wideband models of MIMO channels involving extensions of the H
matrix to H(ω) can be made, but are beyond the scope of this chapter.
As a rst step into investigating a near- eld MIMO channel model, it is useful to formulate
a channel matrix H which includes near- eld eects. Traditional path loss models used to
calculate values of the H matrix, including the Friis equation and Log-distance path loss
model, neglect these eects. Because the channel matrix is a relation of received voltage vr
and transmit voltage vt , because the electric eld generated by the transmit antenna (using
free-space assumptions) is related to vt , and because vr is proportional to the electric eld
incident on the receive antenna, an electric eld equation which includes near- eld eects
can be used in conjunction with antenna properties to determine the channel matrix.
For an initial investigation, a Hertzian dipole antenna was selected. The following is the
electric eld equation for a Hertzian dipole:
!
~ r) = j η0 it dl k j 1 η0 it dl 1 j
E(~ − 2 − 3 e−jkr sin θejωt θ̂ + 2
− 3 e−jkr cos θejωt r̂ (2.20)
4π r r kr 2π r kr
where ~
r is the position vector in relation to the transmit antenna, η0 is the intrinsic impedance
2π
of free space, k is the wavenumber (k = λ
, where λ is wavelength in free space), r is the
radial component of ~r, θ is the elevation component of ~r, it is the current applied to the
transmitting antenna, and dl is the short length of the Hertzian dipole.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 35
CHAPTER 2. BACKGROUND
First, as a point of comparison, a far- eld channel matrix H will be determined using
the Hertzian dipole equation, followed by a near- eld channel matrix.
Far-Field Channel Matrix Formulation
For the far eld, as r becomes large,
1 1
and 3 terms go to zero. Therefore, ~
E becomes:
r2 r
~ r) = j sin θ η0 it dlk e−jkr ejωt θ̂
E(~ (2.21)
4πr
We can assume voltage applied to the transmitting antenna is vt = Zt it ejωt , where Zt is the
impedance of the antenna, it is the magnitude of the current through the antenna (assumed to
be uniform due to the short length of a Hertzian dipole), and ejωt is the temporal component.
This equation can be arranged to nd current in terms of voltage and impedance:
vt
it = (2.22)
Zt ejωt
Substituting it into the far- eld Hertzian dipole equation:
~ r) = j sin θ η0 vt dlk e−jkr ejωt θ̂
E(~ (2.23)
4πrZt ejωt
Simultaneously, the induced current at the output of the receive antenna ir is proportional
to ~ r)
E(~ with dependence on the frequency of the received signal ω and orientation of the
receive antenna with respect to the incidence of ~ r)
E(~ (both in θr and φr ). This constant
of proportionality can be de ned as a(θr , φr , ω), or in the case of the Hertzian dipole where
orientation in azimuthal angle of incidence φ has no eect on the received current, a(θr , ω).
Therefore we can de ne the following relation between electric eld and current at the
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 36
CHAPTER 2. BACKGROUND
receiver:
~ r)
ir = a(θ, ω)E(~ (2.24)
It follows that the voltage at the receiver vr is the following:
~ r)
vr = Zr ir = Zr a(θ, ω)E(~ (2.25)
Combining this with the ~
E equation in terms of vt :
Zr a(θ, ω)η0 vt dlk −jkr jωt
vr ejωt = j sin θ e e θ̂ (2.26)
4πrZt
Performing a Fourier Transform on this equation:
Zr a(θ, ω)η0 dlk −jkr
2πVr δ(ω − ω0 ) = 2πj e sin θVt δ(ω − ω0 ) (2.27)
4πrZt
Vr
Since H(ω0 ) = Vt
δ(ω − ω0 ), dividing both sides of the previous equation by 2πVt will
yield H(ω0 ):
Zr a(θ, ω0 )η0 dlk −jkr
H(ω0 ) = j e sin θ (2.28)
4πrZt
If we let b(θ, ω0 ) = j Zr a(θ,ω0 )η0 dlk
4πZt
sin θ, then H(ω) can be rewritten as a function of θ, ω
and r:
b(θ, ω0 )e−jkr
H(θ, ω0 , r) = (2.29)
r
This is the channel response between two antennas in the far- eld at frequency ω0 . This is
exactly as expected; power gain is proportional to HH ∗ , which decreases as 1
r2
, which is also
the rate at which the Friis equation (used for far- eld free-space gain) decreases.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 37
CHAPTER 2. BACKGROUND
Next we will apply this result to build the channel matrix H. Figure 2.6 illustrates a
2×2 MIMO system with each Tx/Rx antenna pair labeled with a dierent H(θ, ω, r).
H11(θ1,ω,r11)
H1
2 (θ b(θ1,ω)
Tx1 2 ,ω Rx1
,r1
2)
)
,r 21
1,
ω
(θ
H 21
Tx2 H22(θ2,ω,r22) Rx2 b(θ2,ω)
Figure 2.6: 2×2 MIMO System Model
For each transmit/receive pair of antennas ij , the distance rij is calculated, which is in
turn used (along with ω and receive antenna orientation θj ) to calculate Hij (θj , ω, rij ), which
populate H:
b(θ1 ,ω)e−jkr11 b(θ2 ,ω)e−jkr12
r11 r12
H= (2.30)
b(θ1 ,ω)e−jkr21 b(θ2 ,ω)e−jkr22
r21 r22
For the far eld, r11 ≈ r12 ≈ r21 ≈ r22 , which may result in a high condition number,
especially if θ1 = θ2 . A high condition number may result in minimal gain in capacity
provided by MIMO over a SISO con guration.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 38
CHAPTER 2. BACKGROUND
Near-Field Channel Matrix Formulation
Our next step is to derive H using the same methodology as the previous subsection but
π
without far- eld assumptions. For simplicity of analysis, we assume θ= 2
, resulting in the
r̂ component going to zero due to cos θ.
!
~ r) = j η0 it dl
E(~
k j 1
− 2 − 3 e−jkr ejωt θ̂ (2.31)
4π r r kr
Using Eq. 2.31 and Eq. 2.22 together, we derive the following:
!
~ r) = j η0 vt dl
E(~
k j 1
− 2 − 3 e−jkr ejωt θ̂ (2.32)
4πZt ejωt r r kr
Once again, the received current in Eq. 2.24 is manipulated to obtain an expression for vr
(Eq. 2.25). Substituting Eq. 2.32 into Eq. 2.25, we have:
jZr a( π2 , ω)η0 dl
!
k j 1
vr ejωt
= vt − 2 − 3 e−jkr ejωt θ̂ (2.33)
4πZt r r kr
Performing a Fourier Transform on Eq. 2.33:
jZr a( π2 , ω)η0 dl
!
k j 1
2πVr δ(ω − ω0 ) = 2πVt − 2 − 3 e−jkr δ(ω − ω0 ) (2.34)
4πZt r r kr
Vr
Since H(ω0 ) = Vt
δ(ω − ω0 ), dividing both sides of the previous equation by 2πVt will
yield H(ω0 ):
Zr a( π2 , ω0 )η0 dl
!
k j 1
H(ω0 ) = j − 2 − 3 e−jkr (2.35)
4πZt r r kr
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 39
CHAPTER 2. BACKGROUND
Zr a( π2 ,ω)η0 dl
π
If we let b 2
,ω =j 4πZt
, then H (ω0 ) can be rewritten as a function of θ, ω and r:
!
π π k j 1
H , ω0 , r = b , ω0 − 2 − 3 e−jkr (2.36)
2 2 r r kr
This is the channel response between two antennas in the near- eld at frequency ω0 and
pi 1 1
θ= 2
. In the near- eld, the 2 and 3 terms dominate.
r r
Next we will once again apply this result to build the channel matrix H. Figure 2.6 still
describes the same system in the near- eld. For each transmit/receive pair of antennas ij ,
the distance rij is calculated (being much shorter than in the far- eld case), which is in turn
used (along with ω0 ) to calculate Hij (θj , ω0 , rij ), which populate H:
π k j 1 −jkr11 π k j 1 −jkr12
b 2 , ω0 r11
− 2
r11
− 3
kr11
e b 2
, ω0 r12
− 2
r12
− 3
kr12
e
H=
(2.37)
π k j 1 −jkr21 π k j 1 −jkr22
b 2
, ω0 r21
− 2
r21
− 3
kr21
e b 2
, ω0 r22
− 2
r22
− 3
kr22
e
2.3.2 AWGN Channel
For the AWGN channel, The H in Eq. 2.19 is deterministic, while the additive noise
vector ~n has a Gaussian distribution ñ1 (0, σ 2 )+jn2 (0, σ 2 ). The value σ2 is used to determine
SNR for capacity calculations.
2.3.3 Normalization
Normalization is performed on the channel matrix H by multiplying it by
(− 1 )
1
2
2
kHkF (2.38)
Nt Nr
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 40
CHAPTER 2. BACKGROUND
or equivalently, by multiplying by the square root of the product of the number of receive and
transmit antennas and dividing by the value of the Frobenius norm of H [48]. Normalizing a
matrix by its Frobenius norm forces the sum of its eigenvalues (i.e. the sum of the square of its
singular values) to equal 1, and therefore allows us to investigate distribution of eigenvalues
among the independent parallel channels in the system.
This also allows us to explicitly set the SNR (the ratio of received signal power to noise
power at the receiver) in the capacity equation (de ned in the next section), rather than
having to calculate the received power by multiplying some de ned transmit power by the
squared unnormalized channel matrix. Also, by multiplying by the square root of the product
of number of transmit and receive antennas, we preserve the capacity gains provided by
having multiple independent streams in MIMO, allowing for comparison with other similarly
normalized MIMO or SISO systems.
2.3.4 Capacity Equations
As discussed in the background chapter, the capacity of a single communications channel
is known to be log2 (1 + SN R) (assuming normalized bandwidth). In MIMO communi-
cations, the channel matrix H can be decomposed to form RH (rank of H) independent
channels, each with its own applied power at the transmitter. The applied power is some
fraction of the total transmit power, Pt , and the sum of the powers of each independent
channel is equal to Pt . Each independent channel has a power gain proportional to an eigen-
value of the channel matrix, λ2n , where n is the index of the independent channel. Therefore,
the received power for independent channel n is equal to kn Pt λ2 , where kn is the fraction
of the power Pt applied to independent channel n. The SNR of this single, independent
kn Pt λ2
channel is therefore
N0
where N0 is the noise power of the independent channel. If we
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 41
CHAPTER 2. BACKGROUND
Pt
de ne the overall SNR of the receiver to be , then the independent channel SNR is equal
N0
to SN R · kn λ2 . In short, kn is the normalized power allocation of the transmitter, and the
generalized MIMO capacity equation becomes the sum of the capacity of the independent
channels:
RH
log2 1 + SN R · kn~λ2n
X
C= (2.39)
n=1
The optimal distribution of power ~k between the independent channels depends on the
transmitter channel knowledge. If the MIMO system has no feedback mechanism or time-
division duplexing with which to estimate H, then the optimal power allocation is uniform
1
distribution of power among the independent channels, i.e. kn = Nt
for all n, where Nt is
the number of transmit antennas. Therefore, the capacity of a MIMO channel where the
channel matrix is unknown at the transmitter is:
RH
SN R ~ 2
X
C= log2 1+ λ (2.40)
n=1 Nt n
If the transmitter has a method of accurately determining the channel matrix, it can per-
form eigenvalue decomposition (or singular value decomposition, where eigenvalues are equal
to the square of the singular values) to determine the channel eigenvalues. With this infor-
mation, the transmitter can perform the water lling algorithm (outlined in the background
chapter) to determine the optimal power distribution Pn∗ for each of the RH independent
Pn∗
channels. Normalizing the vector of transmit powers P~ ∗ by kP~ ∗ k yields kn = kP ∗ k
, and
therefore the corresponding MIMO capacity for water lling is:
RH
SN R · Pn∗ ~ 2
!
X
C= log2 1+ λn (2.41)
n=1 kP ∗ k
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 42
CHAPTER 2. BACKGROUND
2.3.5 Channel and Capacity Simulation
In order to investigate the performance of this electric eld-based line-of-sight AWGN
channel model which includes near- eld components, a simulation was constructed in MAT-
LAB. This simulation uses Eq. 2.37 to generate a channel matrix H used in capacity calcu-
lation.
This simulation sweeps over SNR (0 to 15 dB) and array separation (from 0.2 to 20 λ).
First, it calculates the distance between each pair of transmit and receive antennas rij using
π
the antenna separation and array separation value. It is assumed that θ= 2
, and that all
Hertzian dipoles in both receive and transmit arrays are co-planar and parallel. Next, using
the channel matrix equations (Eq. 2.37) from the previous section (note: the near- eld eects
in this equation are negligible for far- eld separation distances, and thus this equation can
be used for both near- and far- eld scenarios. As a result, this equation is used in this
simulation for all array separation distances.), the channel matrix H is calculated.
The purpose of this simulation is to compare the eects of near versus far eld on the
capacity of line-of-sight MIMO, not the eect of increased path loss due to increased array
separation on the capacity, and therefore the channel matrix H must be normalized as
described in the previous section to isolate the eect of near- eld components in the model.
Next, singular value decomposition (SVD) is performed on the normalized channel matrix
Hnorm to determine singular value array ~λ. The rank of H (RH ) corresponds to the number
of non-zero elements in ~λ. (It should be noted that the square of each singular value is equal
to an eigenvalue of Hnorm ). If the channel is unknown at the transmitter, uniform power
distribution is used and MIMO capacity is computed as in equation Eq. 2.40. If the channel is
known at the transmitter through some mechanism, it can calculate channel eigenvalues and
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 43
CHAPTER 2. BACKGROUND
use water lling to allocate power to the independent channels as P ∗, resulting in capacity
de ned by Eq. 2.41.
Our simulation showed that for uniform power allocation, the capacity plot converged
towards a plane as the array separation distance increased, as illustrated in Fig. 2.7. This is
1
to be expected, as in the far- eld, the term dominates in both the electric eld and channel
r
equations, the distances between transmit and receive antenna pairs become proportionally
similar, and the distribution of the eigenvalues becomes such that one value approaches zero
and the rank of the matrix essentially trends to one, indicating a single spatial channel.
With the decomposition of the channel matrix trending towards a single channel, and with
the normalization of the channel matrix removing the eects of path loss associated with
separation distance, the capacity plot trends asymptotically towards a constant when varying
SN R
array separation (a constant which is equal to log2 1 + Nt
). In short, the bene ts of
MIMO disappear for a line of sight AWGN channel in the far- eld, with the capacity trending
towards a SISO value.
In the near- eld, however, (in the range at and below λ = 1), we see that a slight increase
in capacity is observed as the array separation distance decreases, particularly at high SNR
values. This can be attributed to the fact that at lower array separation distances, the
distances between transmit and receive antenna pairs become proportionally less similar,
indicating the signal received at each receive antenna has a dominant component from one
transmit antenna and a less dominant component from another. Therefore in the absence
of re ections and a scattering environment, MIMO systems perform better in the near- eld
than in the far eld.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 44
CHAPTER 2. BACKGROUND
10
8
Capacity (b/s/Hz)
0
20
15 15
10 10
5 5
0 0 SNR (dB)
Array Separation (lambda)
Figure 2.7: Capacity with Uniform Allocation vs. SNR and Spacing
2.3.6 Background for Selection of Free-Space Green's Function as
Basis
The following is adapted from Orfanidis textbook on Electromagnetic Waves and An-
tennas [19], but is adapted for arbitrary-oriented dipoles instead of z-oriented dipoles, non-
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 45
CHAPTER 2. BACKGROUND
parallel dipoles, or coplanar dipoles. It is also adapted to use a 3-term King current distribu-
tion approximation instead of the sinusoidal assumption due to the fact that some antennas
in a log periodic array will be signi cantly longer than a half-wavelength. The goal is to
nd the mutual- and self-impedance at each element in the system in free-space, which can
then be used in a system of equations to nd the voltage at the receiving dipole or dipoles
presented over a load for a unit input. This is the free-space channel from probe input to
receiver output, including near- eld eects and coupling. This channel includes all antenna
eects. Because the far- eld approximation does not hold, the concept of antenna gain pat-
terns and angle of arrival have limited meaning, as electric elds from one element to another
do not follow a typical distribution across the antenna as if the incident eld was a plane
wave with de nite direction of arrival. While re ected wave components may exhibit a plane
wave-like behavior at the receiver due to the increased travel distance of the wave between
transmitter and receiver, the dominant line of sight eld and some re ections may not. In
free space, the only eld of interest will be the line of sight case, meaning the line of sight
case can be analyzed for choice of channel approximation basis in the non-far eld, as well
as the conditions in which this approximation basis holds.
The free-space system that approximates the line of sight conditions of the channel of
interest to be modeled is the rst task in the choice approximation basis function. The
channel of interest to be modeled consists of probe antenna log-periodic arrays each with
fourteen-elements, oriented perpendicular to the X-Y plane and aimed at a point in the center
of the X-Y plane, and receiver dipoles oriented in the Z direction, with centers located along
a plane parallel to the X-Y plane. As a rst approximation, a single log-periodic antenna
and a single dipole will be analyzed.
To model the channel from a log-periodic antenna to a dipole antenna, a relation between
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 46
CHAPTER 2. BACKGROUND
the input of the log periodic antenna and output of the dipole must be established. To
accomplish this, the current distribution on each antenna element and the electric elds of
each antenna must be established.
Each element in a log-periodic antenna is itself a dipole antenna. The length and spacing
of each element in the array is related to the length and spacing of the previous element by a
factor τ. The starting and ending dipole lengths are chosen such that half of the shortest and
longest wavelengths of interest are between the lengths of the shortest and longest elements.
This ensures that for any input frequency in the range of interest will be relatively close in
half-wavelength to the length of one of the array antenna elements. The number of elements
and the initial spacing parameter are determined according to the desired antenna gain,
directivity, and other parameters desired.
In nding the E- eld for a dipole as in gure 2.8, the following variables must be de ned.
The orientation of the dipole is in the r̂m direction. Each half of the dipole has length
hm . The center of the dipole is located at the vector position ~rm0 . The top of the antenna
therefore is at position ~rm0 + hm r̂m , and the bottom is at position ~rm0 − hm r̂m .
The observation point at which the E- eld is observed is vector position w
~. Distance R is
the distance from an arbitrary point on antenna m de ned by ~rm0 + αm r̂m (where αm is a
scalar value between −hm and ~ , ||w
hm ) to the observation point w ~ − ~rm0 − αm r̂m ||. Distance
R0 is the distance from the antenna center to the observation point ||w
~ − ~rm0 ||. Distance R1
is the distance from the antenna top to the observation point, ||w
~ − ~rm0 − hm r̂m ||. Distance
R2 is the distance from the antenna bottom to the observation point, ||w
~ − ~rm0 + hm r̂m ||.
The radial vector from the antenna ρ̂m is in the direction of the observation point and
perpendicular to r̂m , the orientation of the antenna. The observation location vector w
~ can
therefore be expressed as:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 47
CHAPTER 2. BACKGROUND
ẑ
ŷ Em(ρm,zm,Im,hm)
Ezm(ρm,zm,Im,hm)r̂m
x̂ Dipole m w
Eρm(ρm,zm,Im,hm)ρ̂m
r̂ m
zm(w-rm0 ,r̂ m)r̂ m
rm0 +hmr̂ m
rm0 +αmr̂ m
hm
Im rm0
ρ̂m
rm0 -hmr̂ m hm
Figure 2.8: Dipole Electric Field Formulation
w
~ = ~rm0 + zm r̂m + ρm ρ̂m (2.42)
where ~rm0 is the center of the dipole, zm r̂m is an antenna directional component, and ρm ρ̂m
is a radial component. The component zm can be calculated with vector projection as:
~ − ~rm0 ) · r̂m
(w
zm = r̂m (2.43)
r̂m · r̂m
while ρm can be calculated with vector rejection as:
~ − ~rm0 ) · r̂m
(w
~ − ~rm0 −
ρm = w r̂m (2.44)
r̂m · r̂m
With this component de nition of w
~, the distances R, R0 , R1 , and R2 can be rede ned as:
q
R= (zm − αm )2 + ρ2m (2.45)
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 48
CHAPTER 2. BACKGROUND
q
R0 = 2 + ρ2
zm (2.46)
m
q
R1 = (zm − hm ) +2 +ρ2m (2.47)
q
R2 = (zm + hm ) +2 +ρ2m (2.48)
These component de nitions of distances allow for a straightforward evaluation of magnetic
vector potential, and subsequently electric eld expressions.
As adapted from [19] to match the arbitrary vector dipole notation in this analysis, the
r̂m -component of the magnetic vector potential Ar̂m is:
−kR
µ Z hm 0 e 0
Ar̂m (zm , ρm ) = I(αm ) dαm (2.49)
4π −hm R
with zm , ρm , hm , αm , and R as previously de ned, k is the wavenumber at the frequency
of analysis, and µ is the permiability of free space. Using Maxwell's equations, this can be
utilized to nd the r̂m and ρ̂m components of the electic eld. The relevant forms of the
equations used are:
ωµEr̂m = ∂r̂2m Ar̂m + k 2 Ar̂m (2.50)
ωµEρ̂m = ∂ρ̂m ∂r̂m Ar̂m (2.51)
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 49
CHAPTER 2. BACKGROUND
µHφ = −∂ρ̂m Ar̂m (2.52)
∂ρ̂m (ρm Hφ̂m ) = ωρm Er̂m (2.53)
ωEρ̂m = −∂r̂m Hφ̂m (2.54)
Using these relations and the de nition of Ar̂m , in conjunction with the fact that the
r̂m component of the electric eld must vanish along the antenna's surface, along with as-
sumptions the current distribution Im (αm ) is symmetric in αm and is zero at the ends of
the antenna, we can de ne the components of the electric eld at the observation point as a
function of current distribution along the antenna, and the position of the observation point
in relation to the position and orientation of the antenna.
µ Z hm 0 e−kR 0
ωµEr̂m (zm , ρm ) = I(αm )(∂r̂2m + k 2 ) dαm (2.55)
4π −hm R
−kR
Z hm !
0 0 e 0
− 4πωρm Eρ̂m (zm , ρm ) = I(αm )(∂r̂2m 2
+ k ) (zm − αm ) dαm (2.56)
−hm R
Using a dierential identity, de ned end-point conditions, and a rede nition of the integral
into parts, the following equations which express the r̂m and ρ̂m electric eld components at
the observation point w
~ are formed:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 50
CHAPTER 2. BACKGROUND
0− e−kR
"Z
1
Er̂m (zm , ρm ) = I 00 (αm
0 0
) + k 2 I(αm ) dαm0
4πω −hm R
Z hm −kR
e
+ I 00 (αm
0
) + k 2 I(αm0
) dαm 0 (2.57)
0+ R
e−kR0 e−kR1 e−R2
!#
0 0
+ 2I (0+) − I (h) +
R0 R1 R2
e−kR 00 0
"Z
1 0−
0 0 0
Eρ̂m (zm , ρm ) = (zm − αm ) I (αm ) + k 2 I(αm ) dαm
−4πωρm −hm R
Z hm
0 e−kR 00 0 0
0
+ (zm − αm ) I (αm ) + k 2 I(αm ) dαm (2.58)
0+ R
−kR0
e−kR1 e−R2
!#
0 e 0
+ 2I (0+)zm − I (h) (zm − hm ) + (zm + hm )
R0 R1 R2
The nal step for obtaining the accurate electric eld from a dipole is to establish an
accurate approximation of the current along the dipole. A procedure from [19] is utilized
to numerically compute the current distribution of the antenna utilizing Hallen equation
solutions "with point-matching, pulse basis functions, and exact kernel with M=100 upper-
half current samples". The numerical solution is then t to a 3-term approximation de ned
as:
" #
kαm khm
I(αm ) = A1 [sin(k|αm |) − sin(khm )]+A2 [cos(kαm ) − cos(khm )]+A3 cos( ) − cos( )
2 2
(2.59)
with derivatives used in the electric eld equations de ned as:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 51
CHAPTER 2. BACKGROUND
kα0
" #
00 0 0 khm 3
I (αm ) +k 2
I(αm ) 2 2
= −k A1 sin(khm ) − k A2 cos(khm ) − k A3 2
cos( ) − cos( m )
2 4 2
(2.60)
I 0 (0+) = −I 0 (0−) = kA1 (2.61)
1 khm
I 0 (hm ) = −I 0 (−hm ) = kA1 cos(khm ) − kA2 sin(khm ) − kA3 sin( ) (2.62)
2 2
With the component de nition of the electric eld in conjunction with a current ap-
proximation, the complete electric eld from one antenna can be de ned completely along
a second antenna, allowing for calculation of mutual impedance, and thereby allowing one
to nd the voltage or current induced at each antenna from the excitation of one or more of
the antennas in the system.
Let a system of two antennas be de ned as antennas m and n, with arbitrary vector
orientations r̂m and r̂n respectively. Figure 2.9 de nes the the variables in this system.
The two antennas have centers at ~rm0 and ~rn0 respectively, and half-lengths hm and hn
respectively. The observation point is now de ned as a position on antenna n, ~rn0 + αn r̂n
rather than an arbitrary point in space w
~. Using this de nition, the variables zm and ρm
must be rede ned.
~rn0 + αn r̂n = ~rm0 + zm r̂m + ρm ρ̂m (2.63)
once again, where ~rm0 is the center of the dipole, zm r̂m is an antenna directional component,
and ρm ρ̂m is a radial component. zm can be calculated with vector projection as:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 52
CHAPTER 2. BACKGROUND
ẑ
Dipole n (Rx)
ŷ
x̂ Dipole m (Tx)
r̂ m
rm0 +hmr̂ m
zmr̂ m
rm0 +αmr̂ m
hm
Im rm0
ρ̂m
hm
rm0 -hmr̂ m
Figure 2.9: Dipole Mutual Impedance Formulation
(~rn0 + αn r̂n − ~rm0 ) · r̂m
zm (αn ) = r̂m (2.64)
r̂m · r̂m
while ρm can be calculated with vector rejection as:
(~rn0 + αn r̂n − ~rm0 ) · r̂m
ρm (αn ) = ~rn0 + αn r̂n − ~rm0 − r̂m (2.65)
r̂m · r̂m
these de nitions are important as they will allow a projection of the electric eld onto
the orientation vector of the receiving antenna, thus nding the electric eld component
perpendicular to the receive antenna as a function of position on the receive antenna. This
electric eld along the receive antenna, as well as the approximated current along the receive
antenna, can be utilized to nd the mutual impedance of the second antenna on the rst
antenna. First, the induced open-circuit voltage on the antenna n due to antenna m is:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 53
CHAPTER 2. BACKGROUND
1 Z hn ~
Vnm,oc = − (Enm (αn ) · r̂n )In (αn )dαn (2.66)
In −hn
where αn is a scalar from −hn to hn that de nes position on antenna n, In is the current
at the center of antenna n (αn = 0), In (αn ) is the current at position αn on antenna n, and
Enm (αn ) is the electric eld at the point ~rn0 + αn r̂n on antenna n due to antenna m. Using
equations 2.57 and 2.58, Emn (αn ) is de ned as:
~ nm (αn ) = Er̂m (zm (αn ), ρm (αn ))r̂m + Eρ̂m (zm (αn ), ρm (αn ))ρ̂m
E (2.67)
The dot product of this eld with the orientation vector r̂n is therefore the electrical eld
component along antenna n. The current In (αn ), also used in the open-circuit equation, is
found using the 3-term t Hallen current approximation for antenna (n with coe cients A1 ,
A2 and A3 recalculated for antenna n). With the new approximation coe cients, In (αn )
can be de ned using 2.59. Finally, using Im = Im (0) and In = In (0), and the open-circuit
voltage equation 2.66, the mutual impedance is de ned as:
Vnm,oc 1 Z hn ~
Znm = =− (Enm (αn ) · r̂n )In (αn )dαn (2.68)
Im Im In −hn
In addition to mutual impedance between each and every combination of elements in the
system (both within the log-periodic antenna and between the elements of the log periodic
antenna and the receiving dipole antenna), the self impedance of each element must be cal-
culated. (NOTE: this is the part that is resulting in a "bad" half-wave dipole approximation
and should be scrutinized the most.) Figure 2.10 shows a horizontally-magni ed dipole an-
tenna m, with radius am . All other variables remain the same as for the previously presented
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 54
CHAPTER 2. BACKGROUND
electric eld analysis of dipole m. According to [19], the self-impedance can be calculated
with the same formula, treating the surface of the antenna, ~rm0 + αm r̂m + am ρ̂m , as the "sec-
ond antenna" along which the electric eld is calculated for the impedance equation 2.68.
Because the two aspects of the same antenna are oriented identically in the r̂m direction,
only the r̂m component of the E eld will be present in the impedance calculation.
ẑ
ŷ
x̂ Dipole m
rm0 +hmr̂ m r̂ m
am
Ezm(a,z m,Im,hm)r̂m hm
rm0 +amρ̂m+ αmr̂ m
rm0
Im(αn)
ρ̂m
hm
rm0 -hmr̂ m
Figure 2.10: Dipole Self Impedance Formulation
It should be noted that this approach is presented in [19] in the context of a sinu-
soidal current approximation, which only applies to antennas around a half-wavelength in
length. Because elements of the log periodic antenna may be considerably longer than a
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 55
CHAPTER 2. BACKGROUND
half-wavelength at the frequency of analysis, this approximation cannot be used. However,
the current approximation method used in prior sections utilizes assumptions such as the r̂m
component of the electric eld being equal to zero on the antenna surface, which is re ected
in the developed electrical eld equations.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 56
CHAPTER 3
Small Anechoic Chamber Feasibility for
MIMO-OTA Testing
The rst contribution of this work is a cost-eective small anechoic chamber for MIMO-
OTA testing demonstrating the eect of size constraints on the performance of an anechoic
chamber for testing MIMO-OTA signals in a single-cluster environment with Laplacian Power
Azimuthal Spectrum. Typical large anechoic chambers used in MIMO-OTA testing can often
cost in excess of one million USD. A small chamber could be constructed for $200 thousand
USD, a signi cantly more cost-eective solution. Previous work on this problem focuses on
minimizing the number of probes necessary to reproduce such an environment, and does not
focus on constraining these probes to a small chamber. This previous work, therefore, does
not investigate parameters such as chamber dimensions, antenna arrangement, or frequency.
The motivation for this work is to present a system for inexpensive pre-certi cation of next
generation devices that can help meet the demands and calls for greater broadband wireless
access.
The following section includes details on the theory behind the plane wave and Laplacian
PAS Electric eld synthesis. Section 3.2 includes the system design and design of the chamber
used in analysis. Section 3.3 includes the results of the Laplacian PAS channel simulations
performed in MATLAB, and Sec. 3.4 covers the electromagnetic simulation results. The
conclusions and future work are described in Sec. 3.5.
57
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
3.1 Electromagnetic Field Synthesis
In this section, we review previous work to understand the ability of a limited number of
antennas to synthesize a desired electric eld in a test zone. In the rst subsection, spherical
wave expansion theory is reviewed to investigate the number of probe antennas required to
synthesize a eld in a test zone of nite size. In the subsequent subsection, the ability of
an array of antennas to synthesize a Laplacian PAS, and therefore model a single cluster, is
discussed.
3.1.1 Number of probe antennas
A study of spherical wave expansion of a plane wave was performed as an initial investi-
gation into the number of probe antennas required to generate a target E- eld in a test zone
to within a desired degree of accuracy. An arbitrary electromagnetic eld can be expanded
into spherical waves in source-free regions of space limited by spherical surfaces centered at
the origin of a spherical coordinate system [18]. This electric eld expansion is expressed
as:
~ θ, φ) = √k Q(c) ~ (c)
X
E(r, smn Fsmn (r, θ, φ) (3.1)
η csmn
where F~smn
(c)
is the spherical wave function, with c indicating the radial dependency, s indi-
cating the spherical wave function type, m indicating the order of the wave function, and
n indicating the degree of the wave function. Q(c)
smn is the spherical wave coe cient, η is
the admittance of the medium, k is the wave number, and (r, θ, φ) indicates the spherical
coordinate of a point.
In the special case of expansion of a plane wave, many of these spherical wave coe cients
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 58
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
are approximately equal to zero, leaving coe cients for the two fundamental spherical wave
functions with order -1 and 1, and degree 1 to N, or 4N coe cients total. Choice of N is
arbitrary, and should be equal to in nity to contain all degrees (modes) of the expansion.
However, a highly accurate approximation can be achieved if N is chosen to be r0 k , where
k is the wavenumber of the plane wave, and r0 is the radius of the desired spherical surface
boundary. Modes above this value of N are comparatively negligible [18].
In [49], the spherical wave expansion is extended to a cylindrical case, and a 2-D cross
section of the cylindrical solid is analyzed in terms of a spherical wave expansion. Addi-
tionally, [25] relates the number of modes in the spherical wave expansion to the number
of (far- eld) antennas required to accurately synthesize this eld. In the case of the E- eld
perpendicular to the 2-D test zone as described in the paper, to achieve a test zone of radius
r, a minimum number of probes K are required, where K is de ned in Eq. (3.2) as
K = 4πr/λ + 1 (3.2)
where λ is the wavelength of the plane wave. This equation was derived under the assumption
that uniform circular spacing of antennas around the test zone in the far- eld (full-ring test
setup) was utilized, and that plane waves may arrive from any angle. If the direction of plane
wave arrival (or other eld arrival) is limited to a speci c direction or range of directions,
this number may be reduced. It is this reduction that will make a small chamber test setup
possible.
3.1.2 3GPP SCME Channel Model
Our goal is to understand the trade-os of minimizing the chamber size (including min-
imizing the number of probes and separation distance between probes) and the ability to
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 59
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
accurately synthesize a desired eld. To investigate this goal, we need to further discuss the
desired eld to be synthesized.
The 3GPP MIMO-OTA Subgroup (MOSG) [20] is standardizing a channel model based
upon the 3GPP Spatial Channel Model Extended (SCME) [50]. The SCME models a multi-
path fading channel with power-delay parameters with azimuthal-angular dependence. This
angular dependence results from spread of power that is created from a environment based
around a given cluster. The cluster of the power has been shown to have a Laplacian [25]
distribution, which is de ned in unnormalized form by Eq. (3.3)
√
e−| 2φ/σ|
φ ∈ [−π, π)
P (φ) = (3.3)
0
otherwise
where φ is the azimuthal angle, and σ is the standard deviation of the Laplacian distribution.
This equation assumes a cluster arrival from 0 degrees; the equation can be shifted in φ for
a cluster arrival from another angle.
Laitenen et al. [25] generate a target 2-D Laplacian PAS in a test zone by sampling the
unnormalized truncated Laplacian PAS function from 360 azimuthal directions with 1 degree
spacing. This set of sampled amplitudes is applied to 360 plane waves arriving from each
corresponding azimuthal direction. Each plane wave is given a random phase sampled from
a uniform distribution. The sum of the 360 plane waves forms the target E- eld in the test
zone. The target E- eld is uniformly sampled around the perimeter of the test zone, r0 .
The unweighted E- eld, Ek (r0 , φ) contribution of each probe antenna in the probe array is
determined at the same set of sample points, and an E- eld 2-norm matching is performed
to determine the complex coe cients of an each probe antenna, ck . After 2-norm matching,
the synthesized E- eld is approximately equal to the target E- eld as in Eq. (3.4).
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 60
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
~ ~
ESynth (r0 , φ) − ET arget (r0 , φ)
e(r0 ) = 20 log10 ◦ max ◦ (3.5)
~ T arget (r, φ)
0 ≤φ<360 max E
0<r≤r0 ,0◦ ≤φ<360◦
N
~ T arget (r0 , φ) ≈ E
~ Synth (r0 , φ) = ~ k (r0 , φ)
X
E ck E (3.4)
k=1
The re ectivity (error) of the synthesized eld from the target eld is calculated across
the test zone as a function of radius as in Eq. (3.5). This gure of merit will be used to
analyze the performance of small chambers.
In previous work, the size of the test zone and angular spread were varied to determine
the minimum number of OTA probes in various spacings needed to achieve at most -15 dB of
re ectivity in the test zone in each case [25]. Additionally, this previous work assumed a large
chamber without non-ideal eects due to re ections. Our contribution is an investigation
of chamber design with the goal of minimizing the chamber dimensions while considering
practical probe antenna locations and eects of re ections on eld generation.
3.2 Small chamber system
The small anechoic MIMO-OTA test chamber is designed to create a synthetic wireless
environment for one cluster. This cluster models the multi-path behavior as a tapped-delay
line with a Laplacian PAS. The parameters of the multi-path channel are controlled by the
RMS delay spread and the angular spread of the distribution of power, which are incorporated
in the channel emulator. The channel emulator is driven by the 4×4 MIMO signaling and has
N outputs corresponding to the N probe antennas with vertical (or horizontal) polarization.
The method for creating the random channel is described in the section 3.1, which provides
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 61
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
a set of weights that are used to control the probes and generate a desired eld in the test
zone.
The ability to generate any arbitrary eld in the test zone is related to the number of
probes and their con guration. As such, this section describes the parameters of the chamber
for evaluation of dierent test sizes. We assume the 2-D case where the elevation angle is
θ = 90◦ . Figure 3.1 shows the basic arrangement of the chamber, with the width W of the
chamber determined by the vertical span of the N antennas in the diagram, and the height
of the chamber determined by the distance between the test zone and the probe antenna
array A, and the test zone radius r0 . The device under test (DUT) is centered in the test
zone radius and will be rotated in 3-D to evaluate the performance as a function of DUT
orientation to the test eld.
The goal of this work is to determine the feasibility of testing a wide range of devices,
from laptops to mobile phones, in a small anechoic chamber. Our simulations evaluate a wide
range of values to test the corner cases of the system. In Table 3.1, we list the values for the
physical and test parameters that are used in the simulations. Our simulations include two
dierent methods that allow for simulating a wide range of chamber parameters including
dimensions, frequency of operation, and Laplacian PAS spread. Also, one of these methods
(HFSS) provides a method for considering practical chamber eects such as re ections from
the chamber walls. The details of each type of simulation is described in the following
sections.
3.2.1 Numerical Model
A simulation of the system in a vacuum (without a surrounding chamber) was performed
in Mathworks Matrix Laboratory (MATLAB) sweeping over the set of system parameters
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 62
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
Exterior Chamber Boundary
r0
Test Zone
A
Probe Antenna Array
1 2 N
...
W
Figure 3.1: Small Anechoic Chamber Dimensions
listed Table 3.1. In the simulation a target Laplacian PAS was generated as in [25]. 2-
norm matching was performed to determine the complex coe cients of each probe antenna,
and the synthesized eld from the probe antennas with applied weights was generated as
in Eq. (3.4). The re ectivity of the synthesized eld at the test zone perimeter was then
determined (Eq.(3.5)). This simulation utilized a horn antenna radiation pattern with each
antenna pointed at the center of the test zone, which allows for antenna con gurations to
be more like those found in an actual small anechoic chamber. Due to the random nature of
the target E- eld (caused by the random phase of each plane wave), a Monte Carlo analysis
was performed to ensure a certain level of performance with a given system con guration.
The statistical distribution of re ectivity was calculated for each system con guration from
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 63
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
Table 3.1: Simulation Parameters
Parameter Value
Laplacian PAS spread (σ - degrees) 25, 35, 45
Frequency (f - GHz) 0.7, 2.4, 5.9
Test zone radius (r0 - m) 0.10, 0.15, 0.20
Width (W - m) 0.95, 1.00, 2.00
Probe Antennas (N ) 3 ... 6
the analysis.
In addition to the simulation scenarios described above, an additional set of simulations
was performed for comparison to the simulations performed in the following subsection. The
electromagnetic simulation program used in the following subsection did not have preset
horn antenna sources, but did have an ideal Hertzian dipole source that could be utilized.
Also, the complexity and long simulation time of this program limited the number of sys-
tem con gurations that could be simulated. As a result, a subset of simulations was run
in MATLAB for comparison with the electromagnetic simulations using a Hertzian dipole
antenna instead of a horn antenna.
The simulations include Laplacian standard deviation (σ ) of 25, 35 and 45 degrees, test
zone radius r0 of 0.1 m, and two chamber con gurations. Chamber con guration 1 consisted
of 6 antennas, 700 MHz and antenna array width of 2.0 m. Con guration 2 consisted of
3 antennas, 2 GHz, and antenna array width of 0.95 m. Antennas were polarized in the
positive X direction (perpendicular to the test zone, i.e. vertical polarization). Antenna
polarization in the Y direction (parallel to the test zone, i.e. horizontal polarization) was
also considered separately.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 64
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
3.2.2 Computational Electromagnetic Model
The entire anechoic chamber system was also modeled in Ansys's High Frequency Struc-
ture Simulator (HFSS), a nite element method (FEM)-based electromagnetic simulation
program. This allows for validation of the MATLAB simulations in a 3-D system. Two
models were developed in HFSS; one with the antennas and test zone placed in a vacuum
to compare results directly to those generated in MATLAB, and one with a small anechoic
chamber with non-ideal absorbing materials encasing both the antennas and the test zone.
The chamber model shown in Fig. 3.2 consists of an outer layer of stainless steel, and an
inner absorbing material modeled after the Cuming LF-77 absorber. The Cuming LF-77
was modeled as three layers, each of thickness 0.75 inches, and each with complex permit-
tivity and permeability values provided by the manufacturer at frequencies of 2 GHz and
3 GHz. Values of complex permittivity and permeability at 700 MHz were determined by
extrapolation. The space both inside and outside the chamber was modeled as a vacuum.
As mentioned previously, the antennas used for the HFSS simulations performed in this
paper are Hertzian dipoles, polarized rst in the X direction (vertical), and then in the Y
direction (horizontal) in a second simulation set. The interior width of the chamber is the
width of the antenna array plus a space of 0.25 times the antenna spacing added to each
side of the chamber. Antennas are placed 0.1 meters from the bottom of the chamber. The
radius from the center of the test zone to the line of antennas is 1 meter. The center of
the test zone is placed 0.3 meters from the top of the chamber. The depth of the chamber
(front to back) was 1 meter, and the plane containing the probe antenna array and test
zone is located 0.5 meters from both the front and back of the chamber. The test zone over
which the eld is matched is a circle with a radius of either 0.1, 0.15 or 0.2 meters depending
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 65
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
Figure 3.2: HFSS Chamber Model of a Small Chamber
on the simulation. The number of antennas and spacing of the antennas is as de ned in
the MATLAB simulation in the previous section. Figure 3.2 illustrates the chamber as
modeled in HFSS. Note that the Hertzian dipole antennas are illustrated as arrows in the
chamber model, with the tail indicating the location of the antenna, and the cone-shaped
head indicating the polarization of the antenna.
3.3 Analytical Results
In an eort to determine the feasibility of various chamber con gurations in synthesizing
a Laplacian PAS eld of various azimuthal spreads, a MATLAB script performed a Monte
Carlo analysis for each con guration to determine the re ectivity distribution (as detailed
in Sec. 3.2). The mean and standard deviation of the re ectivity was determined to within
95% con dence of being within an error of 0.25 dB.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 66
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
The results of the simulations with horn antenna radiation pattern are analyzed with
-15 dB as the maximum level of re ectivity. General trends of the 324 combinations of
simulation parameters are that lower frequencies, lower test zone radius, smaller Laplacian
spreads, and more antennas yield superior re ectivities. Eighty ve con gurations were
observed with mean re ectivities below -15 dB.
In terms of frequency, 55 con gurations with a frequency of 700 MHz, 29 con gurations
with a frequency of 2.4 GHz, and one con guration with a frequency of 5.9 GHz had a
mean re ectivity of less than -15 dB. Clearly target elds of lower frequencies are easier to
synthesize than those of higher frequencies. This is due to the longer wavelength of lower
frequencies, resulting in lower variation across the test zone, which is easier to synthesize
accurately.
In terms of test zone radius, 58 con gurations with a test zone radius of 0.1 m, 26 con-
gurations with a test zone radius of 0.15 m, and one con guration with a test zone radius
of 0.2 m had a mean re ectivity of less than -15 dB. Target elds of a smaller test zone
are easier to synthesize accurately than those of a larger test zone due to less variation in a
smaller test zone when compared to a larger test zone at a xed frequency of operation.
In terms of number of antennas, 9 con gurations with 3 antennas, 14 con gurations with
4 antennas, 27 con gurations with 5 antennas, and 35 con gurations with 6 antennas had a
mean re ectivity of less than -15 dB. This indicates that target elds can be more accurately
synthesized with a greater number of antennas. This is due to the fact that greater numbers
of antennas can allow for greater variation of the synthesized test eld.
In terms of Laplacian PAS spread, 32 con gurations with a spread of σ = 25 degrees,
29 con gurations with a spread of σ = 35 degrees, and 24 con gurations with a spread of
σ = 45 degrees had a mean re ectivity of less than -15 dB. This does not indicate a strong
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 67
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
1
r0 = 10cm, N = 6, w = 2m
r0 = 15cm, N = 6, w = 2m
r0 = 20cm, N = 6, w = 2m
0.8 r0 = 10cm, N = 3, w = .95m
r0 = 15cm, N = 3, w = .95m
r0 = 20cm, N = 3, w = .95m
0.6
F(x)
0.4
0.2
0
-45 -40 -35 -30 -25 -20 -15 -10 -5 0
reflectivity (dB)
Figure 3.3: CDF of Re ectivity for Select Con gurations, 700 MHz and 35 degree spread
trend between Laplacian PAS spread and accuracy of eld synthesis. This is likely due to the
fact that this eld synthesis depends mostly on the direction of arrival of the power and thus
there is more correlation between antenna spacing and Laplacian spread in terms of eld
synthesis accuracy. For wider Laplacian spreads, wider antenna spacings tend to produce
more accurate eld synthesis, while for narrow Laplacian spreads, smaller antenna spacings
produce a more accurate synthesized eld.
As for the distribution of the re ectivity, the majority of the observed standard deviations
fall between 2 dB and 3 dB in the con gurations with mean re ectivity below -15 dB. Figure
3.3 shows the distribution of two sample con gurations at 700 MHz with a 35 degree spread,
a larger chamber 2m wide with 6 probe antennas, and another chamber 0.95 m wide with 3
probe antennas.
The Laplacian spread of σ = 35◦ is particularly important, as it occurs as the spread
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 68
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
for the Laplacian PAS in numerous standards [20]; chamber con gurations with a spread of
35 degrees are evaluated. For the smaller chamber (antenna array width = 0.95 m) with
3 antennas is considered, the only acceptable con guration is with a test zone radius of
10 cm. With 6 antennas, a test zone radius of 10 cm and 15 cm are acceptable. If the width
is set to 2 m and the number of antennas to 5 or 6, a 15 cm test zone radius is achieved for
5 antennas at 700 MHz and 6 antennas at both 700 MHz and 2.4 GHz. A 10 cm test zone
radius is achieved for 5 antennas and 2.4 GHz.
3.4 Electromagnetic Simulation Results
After investigating the synthesis of the E- eld with horn probe antennas in a test zone in
MATLAB, further investigation was performed on the synthesis of the E- eld using Hertzian
dipole antennas using both MATLAB and Ansys HFSS. First, the simulations were per-
formed in MATLAB as described in section 3.2 for the subset of con gurations that will be
replicated in HFSS. Next, as described previously, the chamber was modeled twice in HFSS,
rst with all parts of the chamber set to a vacuum material (to for direct comparison of
HFSS results to those produced in MATLAB), and then with the materials set to a stainless
steel enclosure and three-layer absorber to model an actual chamber. The simulations were
performed with vertical polarization for both the target eld and the probe antennas.
HFSS simulation results were observed to have a dierent zero-phase point for the eld
generated from each Hertzian dipole antenna with respect to that calculated in MATLAB
using ideal Hertzian dipole equations. As a result, the complex coe cients for each an-
tenna calculated in MATLAB for generating a target eld could not be used directly in
HFSS. Instead, the E- eld contribution of each Hertzian dipole probe antenna was mea-
sured separately at a set of 360 equally-spaced points around the test zone perimeter within
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 69
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
HFSS. Using this eld data along with the target Laplacian PAS eld calculated at the same
points, 2-norm matching was performed to determine the complex weights to be applied to
each Hertzian dipole probe antenna in HFSS. Also, because the resultant HFSS eld data
obtained after applying complex weights to each antenna matched that of the sum of the
complex-weighted eld data measured from each antenna separately (as expected), the re-
sultant synthesized HFSS eld could be calculated outside of HFSS (in MATLAB) directly
from the unweighted eld data in the test zone of each independent antenna. As a result, for
each chamber con guration (solution frequency, antenna array size/con guration, chamber
material), an unweighted eld was measured at a set of test zone points from each antenna
in HFSS, and this eld data was exported.
Table 3.2: Simulation Results
r0 Lap. Freq. Width Matlab, HFSS, HFSS,
(m) N σ (GHz) (m) vacuum vacuum chamber
mean std. mean std. mean std.
ref. dev. ref. dev. ref. dev.
(dB) ref. (dB) ref. (dB) ref.
(dB) (dB) (dB)
0.1 6 25◦ 0.7 2.0 -36.72 2.42 -36.28 2.27 -21.51 4.46
0.1 6 35◦ 0.7 2.0 -34.69 3.52 -34.27 3.41 -22.17 4.47
0.1 6 45◦ 0.7 2.0 -30.79 4.56 -30.32 4.59 -22.42 4.76
0.1 3 25◦ 2.0 0.95 -17.15 3.70 -17.33 3.98 -15.12 2.98
0.1 3 35◦ 2.0 0.95 -14.37 3.65 -13.47 3.58 -12.99 3.17
0.1 3 45◦ 2.0 0.95 -12.27 3.57 -11.39 3.42 -11.08 3.13
From this HFSS measured eld data from each antenna for each chamber con guration,
a MATLAB script was developed which determined complex antenna coe cients using the
method described in the previous paragraph. These coe cients were used to synthesize
the eld using the sum of weighted eld data from each antenna. The re ectivity of the
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 70
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
synthesized eld was determined by comparing it to the target eld at the radius of the test
zone. A Monte Carlo analysis was performed using this procedure to determine the mean
and standard deviation of the re ectivity distribution to within a 95% con dence of being
within 0.25 dB of the mean. The results of this simulation are also presented in Table 3.2.
From the results, one can see that the mean and standard deviation in the MATLAB
simulations and the HFSS vacuum simulations are nearly identical, always within 1 dB of
each other. This validates that the Hertzian dipole equations used in MATLAB are consistent
with the Hertzian dipole simulated in HFSS.
The performance of the chamber was noticeably less than that of the vacuum, which is
expected, as the presence of re ected power (due to imperfect absorption by the absorb-
ing material) makes matching the synthesized eld to the target eld more di cult. The
dierence between the chamber and vacuum cases is most pronounced at 700 MHz, with a
dierence of nearly 10 to 15 dB. The dierence between the chamber and vacuum cases was
less pronounced at 2 GHz, with a dierence of approximately 0.3 to 2.2 dB. This is due to
the reduced contribution of re ected energy to the test zone eld. This indicates that for
2 GHz, MATLAB simulations of Laplacian PAS synthesis in a vacuum will be reasonably
close to that synthesized in an actual chamber, and can be used as a predictor of chamber
performance.
Additional identical simulations were performed with horizontally-polarized Hertzian
dipole antennas located in the same locations as the vertically-polarized antennas. The
target eld was generated using 360 plane waves in the same manner as above, but with
polarizations tangential to the test zone circle in the direction of arrival. In all cases, both
vacuum and simulated chamber, the re ectivity was observed to be approximately 1 dB,
indicating signi cant error. As such, horizontally-polarized target elds are not able to be
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 71
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
reproduced accurately with this antenna con guration. An alternative implementation or
con guration of horizontally-polarized antennas will likely synthesize horizontally-polarized
elds accurately.
3.5 Conclusions
In this chapter, we have demonstrated through simulation that a small anechoic chamber
(built with practical materials) can be successfully utilized as a MIMO-OTA test chamber
for the su ciently-accurate synthesis of a vertically-polarized single-cluster Laplacian PAS
in a test zone under several frequencies and chamber con gurations [12]. This investigation
serves as a rst step to identifying the performance capabilities and limitations of a small
anechoic chamber in terms of test zone size, chamber size, number of probes, and frequency
range.
The con guration of horizontally-polarized antennas such that a horizontally-polarized
target eld (i.e. a target E eld with Ŷ and Ẑ vector components) can be synthesized must
be investigated in future work, as our work showed that horizontally-polarized antennas
co-located with vertically-polarized antennas failed to accurately synthesize a horizontally-
polarized target eld. One potential con guration will be an array of horizontally-polarized
antennas in an array perpendicular to the array of vertically-polarized antennas, thus creat-
ing a cross of antennas at the bottom of the chamber. Another possibility is to create a
more extensive 2-dimensional grid of antennas at the bottom of the chamber for both hor-
izontal and vertical antenna polarizations, which may further increase the accuracy of the
synthesized eld of both polarizations.
In general, the investigation in this paper found that frequencies of 700 MHz and 2.0 GHz,
Laplacian PAS distributions with standard deviations of 25 and 35 degrees, and a test zone
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 72
CHAPTER 3. SMALL ANECHOIC CHAMBER FEASIBILITY FOR MIMO-OTA
TESTING
radius of 10 cm have been shown to produce or nearly produce a su ciently low re ectivity
gure in the test zone. Further study of probe number and chamber dimensions may expand
the capabilities of the chamber.
The use of actual horn antennas rather than ideal Hertzian dipole antennas in HFSS is
an important next step in future work. This will investigate the impact of non-ideal eects
such as coupling and re ections between antennas that will lower the performance of the
system.
One nal area requiring further study is the determination of complex antenna weights
from actual eld measurements around the test zone. Measurements must be performed
using an antenna in the test zone to estimate the contribution of each probe antenna to
the eld in the test zone. Previous work has been performed in this area [22] in order to
overcome near- eld eects, re ections and scattering in a ring chamber in generating a plane
wave. A similar procedure can be employed in a small anechoic chamber to overcome these
same non-ideal properties.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 73
CHAPTER 4
Near-Field Channel Model Framework
The purpose of this chapter is to develop a mathematical framework for describing the
narrowband RF channel within a small anechoic chamber. This framework extends a complex
numerical model into a relatively simple mathematical approximation that can be applied
to both calculated and measured channels. This approximation serves as the basis for the
work in the following chapter, which expands this method to a re ned wideband model. This
model can then be used for the purpose of MIMO system analysis, the ultimate goal of this
work.
The small chamber will most often utilize an array of log-periodic probe antennas with
xed positions and orientations and an array of DUT antennas, which will be for the purpose
of analysis de ned as dipole antennas. As a rst approximation, a SISO channel between
a single log-periodic probe antenna (which consists of an array of co-polarized dipoles) and
a single dipole DUT antenna (of arbitrary orientation relative to the log-periodic array) is
modeled in free space.
Fig. 4.1 shows the breakdown of contributions in chapters 4 and 5. The contributions
for this chapter are emphasized in bold.
The rst contribution of this chapter is the formulation of a numerical method for de-
termining the channel between arrays of dipole antennas with arbitrary orientations and
spacings, which can then be adapted to a general array consisting of the log-periodic an-
tenna (an array of dipoles) and an additional DUT dipole. In the literature investigated in
Chapter 2, every treatment of the near- eld of dipole antennas and mutual-impedance makes
74
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
Motivating How can the measured What is the most accurate
Question What can the small Can this be simplified for a Can the SAF for one
channel be decomposed method of fitting SAF to
chamber channel be fixed TX and a test zone frequency be utilized in a
into chosen SAF? measurement data?
modeled as? RX? wide-band model?
Free-space
SA funcitons
Properties
Properties
Statistical
approx.
SAF
Theories Used
Use of Generalized Least
Numerical Model of Use of point Eikonal Use Geostatistical Kriging
Use of point sources (free- Squares, Geostatistical
Antenna Interaction Surface correlation, to fit SAF coefficients, fit
space Green’s Functions) Regression and
(Hallen equations, current multidimensional Fourier coefficients to a polynomial
and plane waves as SAF Geostatistical Kriging to
distribution) analysis model vs. frequency
determine SAF weights
Correlation maximization Modeling coefficients SAF
Contribution
Generalized Formulation Application of Complex
Combination of point of point source Eikonal of a central frequency as a
for Log Periodic Array and Kriging with Geostatistical
sources and plane surfaces with measured polynomial function of
Arbitrary Antenna Kriging to Spatial Channel
waves as model SAF data; 2d Fourier transform frequency separation.
Orientation Measurements
of residual
Locate point source SAF Iteratively determine
Show Numerical Model
Contribution
using correlation, unbiased best fit of SAF to Model a wide-band
Applied to
Question
Calculate Channel Vs. can be represented using
determine plane wave data, as well as spatial channel using a single set of
Receiver Position in Test point source and plane
vectors with 2D Fourier statistics of the residual of SAF and polynomial
Zone for Fixed Transmitter wave SAF with
Transform of the SAF coefficient functions.
acceptable error
measurement
Chapter 4 Chapter 5
Figure 4.1: Chapter 4 Contribution Flowchart
the assumption of co-polarized dipole antennas, which yields expressions and equations not
directly applicable to arrays of arbitrary orientation. Combined with transmission line the-
ory for feeding the log-periodic antenna, an impedance matrix relating each and every dipole
port is constructed, which can be used to relate currents and voltages, and in turn produce
an expression for the channel from the input of the log-periodic antenna to the output of the
DUT dipole.
The second contribution of this chapter is a justi cation of the usage of plane waves
and point sources as joint basis functions for describing and modeling an arbitrary channel.
Modeling a channel as a function of DUT dipole position in terms of simple basis functions
is far less computationally-complex than computing the channel using the numerical method
described previously. In addition, basis functions that describe the numerical channel within
acceptable accuracy can be utilized to model measured channel data with additional super-
imposed features such as re ections. This novel method of joint basis functions provides
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 75
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
a reduction in the number of terms required to express the channel relative to using plane
waves as basis functions alone.
The nal contribution of this chapter is a method to t the joint basis functions, described
in the previous section, to a numerical channel or measured channel data. A correlation
method for iteratively tting point sources to sampled or calculated data followed by selecting
a series of plane waves by performing Fourier analysis on the residual is utilized as a method
for selecting which basis functions to use for the channel model and determine the complex
weights of each basis function.
The following section (4.1) outlines the adapted numerical method for calculating the
channel between two ports in an arbitrary-orientation dipole array. Section 4.2 outlines the
justi cation and advantage for the usage of point sources and plane waves as joint basis
functions. Section 4.3 proposes a method of tting the joint basis functions from section 4.2
to channel data (calculated or measured), and illustrates the validity of the usage of joint
basis functions using a channel calculated from section 4.1.
4.1 Formulation of Channel using Dipole Arrays
The contributions of this section are as noted the rst column of Fig. 4.1. In this
section, we extend the work of Orfanidis [19], extrapolating the work presented in chapter
25 (Coupled Antennas), Sections 1 through 5 for arbitrary-oriented dipoles instead of Z-
oriented dipoles, non-parallel dipoles, or coplanar dipoles. In addition, a four-term current
distribution approximation for each dipole is utilized instead of the sinusoidal assumption
due to the fact that some elements in a log periodic array will be signi cantly longer than a
half-wavelength.
The motivation of this extension is to nd the mutual- and self-impedance at each element
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 76
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
in the system in free-space, which is used in a system of equations to nd the voltage at
the receiving dipole or dipoles presented over a load for a unit input. This is the free-space
channel from probe input to receiver output, including near- eld eects, coupling, and all
antenna properties. Because the far- eld approximation does not hold (as the short distance
leads to non-planar wave fronts in the test zone, and elds and gains that vary signi cantly
in the proximity of the transmit antennas), the concept of antenna gain patterns and angle
of arrival have limited meaning, as electric elds from one element to another do not follow
a typical distribution across the antenna as if the incident eld was a plane wave with
de nite direction of arrival. While re ected wave components may exhibit a plane wave-like
behavior at the receiver due to the increased travel distance of the wave between transmitter
and receiver, the dominant line of sight eld and some re ections may not. In free space,
the only eld of interest will be the line of sight case, meaning the line of sight case can
be analyzed for choice of channel approximation basis in the non-far eld, as well as the
conditions in which this approximation of basis holds.
The free-space system that approximates the line of sight conditions of the channel to be
modeled is the rst task in the choice approximation basis function. The channel of interest
to be modeled consists of probe antenna log-periodic arrays each with fourteen-elements,
oriented on a plane perpendicular to the X-Y plane and with bore sight aimed at the center
of the projection of the region of possible DUT positions onto the X-Y plane, and DUT
dipoles oriented in the Z direction, with centers located along a plane parallel to the X-Y
plane. As a rst approximation, a single log-periodic antenna and a single dipole (SISO
channel) will be analyzed.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 77
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
4.1.1 Mathematical Channel Model with Arbitrary Dipole Orien-
tations
To model the channel from a log-periodic antenna to a dipole antenna, a mathematical
relation between the input of the log-periodic antenna and output of the dipole must be
established. To accomplish this, the current distribution on each antenna element and the
electric elds of each antenna must be established, and the mutual- and self-impedances of
all elements must be calculated. Refer to Chapter 2, speci cally 2.3, for the procedure of
deriving these values for dipole antennas.
A review of the nal equations and variable de nitions established in the preliminary
sections is reproduced here for clarity. Let a system of any arbitrary two antennas from the
dipole array be de ned as antennas m and n, with arbitrary vector orientations r̂m and r̂n
respectively. Figure 4.2 illustrates the variables used in this system.
ẑ
Dipole n (Rx)
ŷ
x̂ Dipole m (Tx)
r̂ m
rm0 +hmr̂ m
zmr̂ m
rm0 +αmr̂ m
hm
Im rm0
ρ̂m
hm
rm0 -hmr̂ m
Figure 4.2: Dipole Mutual Impedance Formulation
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 78
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
The two dipoles have centers at ~rm0 and ~rn0 respectively, and half-lengths hm and hn
respectively. The eld observation position vector from antenna m is now a position on
antenna n, ~rn0 + αn r̂n (antenna n center position vector plus a scaled version of antenna n
orientation vector r̂n , scaled by αn which can vary from −hn to hn ) rather than an arbitrary
point in space. Using this de nition, the variables zm and ρm , cylindrical coordinates of a
point on antenna n relative to the center of antenna m, can be related to the position on
antenna n with the following equation:
~rn0 + αn r̂n = ~rm0 + zm r̂m + ρm ρ̂m (4.1)
zm and ρm can be calculated in terms of αn with vector projection and rejection as
performed in Chapter 2. These variables can be used to de ne the electric eld from
antenna m along antenna n, Er̂m (zm , ρm ) + Eρ̂m (zm , ρm ), or in terms of ~ nm (αn ) =
αn , E
Er̂m (αn ) + Eρ̂m (αn ). Finally, this electric eld can be projected onto the orientation vector
r̂n to determine the electric eld component along antenna n, Er̂n (αn ).
The current along each antenna (e.g. I(αm ) along antenna m) is determined with a
numerical 3-term approximation as discussed in Chapter 2. Orfanidis de nes a procedure
to numerically compute the current distribution of the antenna utilizing Hallen equation
solutions with point-matching, pulse basis functions, and exact kernel with M = 100 upper-
half current samples [19]. A detailed explanation of this process is provided in Chapter
2.
Using the current distribution and electric eld equations, the mutual impedance of the
antenna n on the antenna m can be calculated. First, the induced open-circuit voltage on
antenna n due to antenna m is:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 79
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
1 Z hn ~
Vnm,oc = − (Enm (αn ) · r̂n )In (αn )dαn (4.2)
In −hn
where In is the short-circuit current at the center of antenna n (αn = 0). Finally, using
Im = Im (0) and In = In (0), and the open-circuit voltage equation 4.2, the mutual impedance
is de ned as:
Vnm,oc 1 Z hn ~
Znm = =− (Enm (αn ) · r̂n )In (αn )dαn (4.3)
Im Im In −hn
In addition to mutual impedance between each and every combination of elements in the
system (both within the log-periodic antenna and between the elements of the log periodic
antenna and the receiving dipole antenna), the self-impedance of each element must be
calculated. Figure 4.3 shows a horizontally-magni ed dipole antenna m, with radius am . All
other variables remain the same as for the previously presented electric eld analysis of dipole
m. The self-impedance can be calculated with the same formula as with mutual-impedance,
treating the surface of the antenna, ~rm0 + αm r̂m + am ρ̂m , as the second antenna along which
the electric eld is calculated for the impedance (4.3) [19]. Because the two aspects of the
same antenna are oriented identically in the r̂m direction, only the r̂m component of the E
eld will be present in the impedance calculation.
The port of each dipole element in the M-element log-periodic antenna plus the port of
the single receiving dipole forms an M +1 port network, as shown in Figure 4.4. Using the
procedures described to determine the mutual- and self- impedance of a pair of arbitrarily-
oriented dipoles, an impedance matrix for the M +1 port network can be constructed, with
each element (i, j) of the matrix describing the trans-impedance of element i on element j.
The trans-impedance matrix can be converted to the trans-admittance matrix, necessary for
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 80
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
ẑ
ŷ
x̂ Dipole m
rm0 +hmr̂ m r̂ m
am
Ezm(a,z m,Im,hm)r̂m
hm
rm0 +amρ̂m+ αmr̂ m
rm0
Im(αn)
ρ̂m
hm
rm0 -hmr̂ m
Figure 4.3: Dipole Self Impedance Formulation
combining two networks in parallel, using the following relation:
YA = ZA −1 (4.4)
In parallel with this network of coupled antennas is a feedline network that feeds each
antenna of the log-periodic array from a single excitation on port 1, as shown in Figure 4.5.
The feedline network does not feed the receiving dipole antenna. Because the coupling of
the feedline network onto the lone dipole and its load is signi cantly less than the coupled
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 81
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
INe+1A
+
-VNe + 1
Figure 4.4: Antenna Array and Dipole Port De nitions
signal from the elements of the log-periodic array, it can be reasonably approximated as
zero admittance from any log-periodic port to the dipole port. The feedline network can be
described using transmission line theory; Stutzman and Thiele [51] present a feedline admit-
tance matrix in Section 14.10 of their text Antenna Theory and Design. This matrix, for the
log-periodic antenna alone, can be de ned as YTLP . Combining this with the zero admit-
tance from the log-periodic array to the dipole antenna, a total network feedline admittance
matrix can be de ned in the following equation:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 82
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
YTLP 0
YT =
(4.5)
0 0
β,Z0
+ I1T - + INeT +
+ V1 V2 V3 VNe
-VNe+1 -
+
- -
d1 d2 dNe-1
Figure 4.5: Transmission Line Network De nition
Finally, the admittance networks YT and YA are de ned as in parallel for the 15 ports,
as in Figure 4.6. Because these two networks are in parallel, the admittance matrices can
simply be added to form the admittance matrix for the entire 15 port network.
Y = Y A + YT (4.6)
The admittance matrix can be inverted to obtain the system impedance matrix Z = Y−1
for the entire network. The column vectors of the voltages and currents of the M +1 port
network, as shown in Figure 4.6, are de ned as V~ and I~ respectively. Applying the basic
relationship between currents and voltages at the ports of the network, we have V~ = ZI~.
If we rearrange this equation such that the right hand side equals zero, using an identity
matrix I, and combining the voltage and current vectors into a single column vector through
vertical concatenation, we have:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 83
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
I1 I1T INe+1T
+ +
V1 -
1 Ne+1 - VNe+1
I2 I2T INeT
+ +
V2 -
2 YT Ne -
VNe
Transmission Line
Admittance
I1A INe+1A
1 Ne+1
I2A INeA
2 YA Ne
Antenna Array
Admittance
Figure 4.6: De nition of Parallel Admittance Networks
~
V
−I Z = ~0 (4.7)
I~
In addition to this set of equations, we must de ne two more to obtain a channel matrix
from this set of equations. The rst is to de ne the input voltage, V1 = 1 applied at port 1.
The added equation row will therefore have a 1 at the column corresponding to V1 in the
column vector, and has a 1 appended to the end of the right-hand side column vector. This
forces V1 to 1 in the system of equations.
Secondly, we know that the current at port M + 1 (IM +1 ), the dipole, times the load
impedance attached to the dipole, ZL , is equal to VM +1 . Therefore, −VM +1 + IM +1 ZL = 0.
The added row to the left hand side matrix should have -1 at the column corresponding
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 84
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
to VM +1 , and ZL in the column corresponding to IN e+1 , and zero appended to the column
vector on the right-hand side.
−I Z ~0
~
V
1 ~0 = 1 (4.8)
I~
~01 ~0ZL
0
If we de ne the left hand side matrix as A, and the right hand side vector as ~b, then the
vector of voltages and currents can be solved for by taking the pseudoinverse of A times ~b.
~
V
= A†~b (4.9)
I~
where A† is the pseudoinverse of A. Finally, the channel can be determined by taking the
ratio of the output voltage, VM +1 , to the input voltage, V1 .
VM +1
H= (4.10)
V1
The channel can be described as a complex function of the positions, orientations, lengths
of all M +1 elements, and frequency, using equations in this section. However, because M
of the elements are included in the log-periodic antenna, these elements are of xed length
relations, related position and spacing, and common orientation. The receiver dipole will
also have a xed length. As a result, the channel is a function of log-periodic position
and orientation, and dipole position and orientation. Due to the con guration of the small
chamber for which the channel modeling is being performed, the log-periodic antennas can
be considered to have xed position and orientation for a small number of discrete locations
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 85
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
in the chamber. The receiver dipole, however, may be located anywhere near the chamber
center. In short, a meaningful channel model for the purposes of our analysis will be a
function of receiver position and orientation only. For simplicity, a single receiver orientation
is considered (Z-oriented) and a xed position height is considered. As such, the channel
model in proceeding sections will be a function of X-Y receiver position and frequency only,
and a separate model will be determined for each of several de ned log-periodic positions
and orientations.
4.1.2 Numerical Computation of Dipole Array Channel
A general procedure to numerically calculate the channel using the equations described
in the previous section is as follows. First, the position and orientation vectors, lengths, and
radii for each dipole in the system are de ned, as well as the frequency of analysis. Next, each
and every pair of antenna elements i, j are selected pair by pair (including cases where i = j ).
The current distribution along antennas i and j are calculated using point matching, pulse
basis functions, and King's 3-term approximation, as discussed in the previous section. This
approximation determines three t coe cients used in standard expressions for accurately
modeling current distribution. The current distribution for antenna i is used in integral
equations de ning the electric eld due to i at points along antenna j. The electric eld
along j due to i is used in conjunction with the current distribution on j in integral equations
to determine the mutual impedance of j on i.
Each of these integrations is performed using the Gauss-Legendre method, a numerical
method of approximating an integral, which rst algorithmically selects a vector of locations
along the antenna being integrated over and corresponding weights. The integrand is then
evaluated at each of the locations in the vector along i. Finally, the vector of integrands
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 86
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
is multiplied by the corresponding weights and summed, with the sum approximating the
integral. For each point αj selected along j , use Gauss-Legendre to calculate the electric eld
due to ~ ji (αj ), (and in turn the entire integrand for the mutual impedance equation),
i on j , E
then use Gauss-Legendre again to calculate the integral in the mutual impedance (4.3). If
i = j, then perform the procedure using the surface of the antenna (at the radius of the
dipole) as antenna j and the center of the antenna as antenna i; both have identical sample
points along the antenna and identical current distribution.
Populate an impedance matrix with the mutual- and self-impedances calculated with
double Gauss-Legendre integration. Take the inverse of this to produce the admittance
matrix of the antenna network, YA . Calculate the transmission line network admittance
matrix between the log periodic elements YTLP using the procedure discussed in the previous
section, and then form the matrix YT as in (4.5). Sum the admittance matrices to determine
the system admittance matrix Y as in (4.6). Take the inverse of Y to obtain the system
impedance network Z. Using (4.8) to relate the voltages and currents of the system, de ne
unit excitation at the input port of the log-periodic antenna, and the dipole voltage equal
to the dipole current times the antenna load, and solve the system of equations using (4.9).
Finally, extract the voltages at the log-periodic input and dipole output, and calculate the
channel as in (4.10).
As explained in the previous section, we wish to model the channel as a function of
DUT dipole position and frequency. Using the procedure just described, the channel can
be calculated for a series of DUT dipole locations comparable to a range of DUT dipole
locations in the chamber being modeled.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 87
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
4.1.3 Numerical Computation Setup and Results
The described numerical computation method is applied to a fourteen-element log-periodic
and dipole antenna of typical position and orientation in a small anechoic chamber. An oper-
ating frequency of 5.45 GHz is chosen, at the center of the 5 GHz ISM bands. An Octoscope
BOX-38 small anechoic chamber has interior dimensions of 81 cm width (x), 55 cm depth
(y ), and 49 cm height (z ). The receiving dipole in this simulation is a z -oriented half-wave
dipole (with λ corresponding to 5.45 GHz), and with a center z coordinate xed at 15.3 cm
(to match the dipole z coordinate in a subsequent measurement campaign). The x and y
positions are varied in an approximately 22 cm by 19 cm rectangular test zone (large enough
to include most small devices under test (DUT)) centered on the center of the chamber
oor. The log periodic antenna is modeled after the probe antenna used in conjunction with
Octoscope's BOX-38 for WiFi testing. It has a value of τ = 0.85, and element lengths from
0.64 cm to 6.2 cm. The coordinate of the tip of the log periodic antenna is approximately
(0.7,-0.44, 0.27) in meters with the bottom/back/left of the chamber de ned as the origin
and the up/away/right directions when facing the front of the chamber are the positive di-
rections of the x, y , and z axes respectively. The orientation of the dipole is boresight in
the direction of the center of the interior bottom oor of the chamber, along a plane per-
pendicular to the oor of the chamber. Fig. 4.7 illustrates the system as described, with
the black outline representing the boundaries of the chamber interior, the green rectangle
representing the boundaries of the test zone along which the blue dipole will move, and a
red fourteen-element log-periodic antenna near the front/right/top corner of the chamber.
As a rst-approximation, the system is modeled as the log periodic and dipole antennas in
free-space, without any interactions with the chamber.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 88
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
0.5
0.4
0.3
z position (m)
0.2
0.1
0
0
-0.2
-0.4
0.7 0.8 0.9
-0.6 0.4 0.5 0.6
y position (m) 0.1 0.2 0.3
0
x position (m)
Figure 4.7: Modeled System for Numerical Computation
The channel between the log periodic input and dipole output is calculated on a grid of
17 evenly-spaced x by 14 evenly-spaced y receiver positions spanning the entire test zone,
resulting in test zone sample spacings of 0.26 λ in the y direction and 0.25 λ in the x
direction. Because the system is only being analyzed with a single frequency excitation,
the free-space electric eld should be limited to a spatial period of λ in any direction, and
therefore sample spacing should be at most 0.5 λ in each direction if discrete Fourier analysis
is to be performed on the sampled eld. This will be important in subsequent sections, so
a sample spacing choice that meets this requirement was selected to meet this requirement.
Using the sample spacing of the channel (which is proportional to the electric eld) of 0.25
λ is a sample density of approximately double the minimum requirement, and is therefore
su cient to prevent aliasing in Fourier analysis. This will also allow the observation of
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 89
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
spatial variation of the electric eld phase structure (or more speci cally, the structure of
the corresponding channel) which will aid in selection of modeling functions.
The following plots show the resultant channel between the log periodic antenna and
dipole antenna for dipole positions over the test area de ned above. Fig. 4.8 shows the
channel magnitude versus dipole position. As expected, if one considers the gain patterns of
dipole antennas and log periodic antennas, one would expect the channel to have lower gain
when the signal propagates away from the log periodic antenna at an angle farther away from
boresight (e.g. when the receiver is below the boresight direction of the antenna) and when
the signal propagates toward the dipole from a direction approaching parallel to the dipole
orientation, even in the case of close proximity. The log periodic antenna is located beyond
the lower right corner of the x and y range of the plot, and this corner of the plot has slightly
decreased gain relative to the rest of the test zone despite closer proximity as expected. The
remainder of the test zone has a path from transmitter to receiver in higher gain regions of
the transmit and receive antenna gain patterns, so we expect that the channel magnitude
would slowly fall o with increased distance between transmit and receive antennas, which
is exactly what happens; the channel gain slowly rolls o as the receiving antenna moves
towards the upper left part of the plot. In short, the magnitude of the numerical channel is
as expected.
Fig. 4.9 shows the channel phase shift versus dipole position. Due to the relatively
coarse sampling of the channel over the test zone, combined with the discontinuity as phase
transitions from −π to π , the interpolated phase plot is quite jagged along the discontinuity.
Despite this, it is clear that the equiphase fronts have a well-de ned curvature, emanating
from a point or region located below the lower right corner of the plot, exactly where the
antenna is located. Therefore the phase of the numerical channel is also as expected.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 90
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
-40
-0.14
-0.16
-40.5
-0.18
-0.2 -41
y position (m)
-0.22
-41.5
-0.24
-0.26 -42
-0.28
-42.5
-0.3
-0.32
-43
0.3 0.35 0.4 0.45 0.5
x position (m)
Figure 4.8: Magnitude of Channel Vs. Dipole Position
-0.14 3
-0.16
2
-0.18
-0.2 1
y position (m)
-0.22
0
-0.24
-0.26 -1
-0.28
-2
-0.3
-0.32 -3
0.3 0.35 0.4 0.45 0.5
x position (m)
Figure 4.9: Phase of Channel Vs. Dipole Position
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 91
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
Due to the presence of curved wave fronts almost as if emanating from a point in space, it
is logical to consider using a complex-weighted eld from a point source as a rst approxima-
tion of the channel. To re ne the channel, a spectral decomposition of the residual from the
point source can be used to select dominant plane wave components and their corresponding
weights to be summed with the weighted point source approximation. The following section
outlines the theory behind this method, a procedure for its implementation, and validation
of the model error relative to the modeled numerically-computed eld.
4.2 Dipole Array Channel with Spatial Attribute Func-
tion Decomposition
The second contribution of this chapter, as noted in the second column of Fig. 4.1,
is presented in this section. We investigate a spatial modeling framework as presented by
Smith in his course on Spatial Data Analysis, speci cally from the section on General Spatial
Prediction Models [52]. Here we model a deterministic global trend µ(~r) (where ~r is spatial
position in the observation region) from the sampled spatial channel H(~r). In the following
chapter we analyze the statistical properties of the residual ε(~r), where ε(~r) = H(~r) − µ(~r).
In [52], the global trend µ(~r) is modeled as the weighted sum of spatial attribute functions.
A spatial attribute function is a smooth, nite and continuous function over a sampled
observation region (in our case, the rectangular test zone). Each spatial attribute function
xk (~r) is weighted by a constant term βk (which is determined through tting of the spatial
attribute functions to the sampled data). A constant term β0 is included in the model to
include any constant mean term over the sampled region. This sum of weighted spatial
PK
attribute functions de nes the global trend as µ(~r) = β0 + k=1 βk xk (~r), where K is the
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 92
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
number of spatial attribute functions. Let X be a matrix where each row corresponds to
each observation position ~rj of J observation positions. The rst column, column 0, of X
is populated with 1+i (corresponding to a complex β0 ). Each subsequent column k is
populated with the value of xk (~rj ). The size of X is therefore J by K + 1. Assign β0 through
βK to a column vector β~ , containing all K +1 complex weights. Using X and β~ , a vector of
global trend values at each observation point can be de ned as in (4.11).
~ = Xβ~
µ (4.11)
This can be used in the expression for the residual as in (4.12).
~ − Xβ~
~ε = H (4.12)
If we assume that β~ should be chosen to minimize the residual ~ε, we can use ordinary
least squares to solve for β~ that best ts the spatial attribute functions to the observed data
at the observation points. This leads to the additional restriction that the spatial attribute
functions are pair-wise linearly-independent over the set of observation points, and thus X0 X
must be non-singular [52]. The least squares solution of β~ is de ned as in (4.13), which uses
the Moore-Penrose pseudoinverse of X.
β~ = (X0 X)−1 X0 H
~ (4.13)
It should be noted that X0 X forms a Gram matrix, each entry of which is in essence an
inner-product over all observation points for each pair of spatial attribute functions (basis
functions; each column of X corresponds to the basis function evaluated at every observation
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 93
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
point). Diagonal-dominant matrices (i.e. ones where for each row the magnitude of the
diagonal element is greater then the sum magnitude of all other elements) have the property
of being non-singular [53]. Therefore choosing spatial attribute functions that make the
Gram matrix diagonal-dominant would make it non-singular. If we can make non-diagonal
elements of the Gram matrix small (either by pair-wise inner product threshold enforcement
or by other properties of chosen functions), then we can make the Gram matrix diagonal-
dominant, and thus non-singular.
4.2.1 Plane Waves as Spatial Attribute Functions
One candidate spatial attribute function is a plane wave. Any function in space can be
described as the sum of an in nite number of three-dimensional plane waves. Each plane
wave is described by a wave vector ~k (de ning spatial frequency and direction of travel),
observation position ~r, and complex weight containing of magnitude and phase information
E0 . The plane wave equation Eqn. 4.14 is:
~
EP W (~r) = E0 e−j k·~r (4.14)
A three-dimensional Fourier Transform of a spatial function will express that function in
terms of its wavenumber spectrum. Mersereau and Speake [54] illustrate the properties and
use of multidimensional Fourier transforms, with discussion in terms of complex exponentials
and general spectra rather than plane waves and wavenumber spectra, however the same
general principles apply. Using spectral theory, if this spatial function is sampled in each
dimension with sample spacings at most equal to the inverse of double the spatial frequency
in any direction, and sampled over the entire volume of the desired signal volume, then the
three-dimensional Discrete Fourier Transform (DFT) of the sampled signal will produce a
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 94
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
discrete wave vector spectrum [54].
The discrete wavevector spectrum has coordinates consisting of the wave vector ~k of a
plane wave, with the function value at that coordinate equal to the complex weight E0 of
−j~k·~
r
the plane wave of that wave vector. Each of the unweighted exponential plane waves (e )
can be considered a spatial attribute function of the sampled signal, and the weights of each
attribute function E0 can be considered a corresponding β as de ned in the previous section;
the sum of weighted plane waves will reconstruct the sampled signal [54].
Certain aspects of discrete, space-limited spatial sampling must be considered in a similar
way to signals sampled in time. The spatial sampled signal should be windowed to reduce
false high-frequency components formed by the DFT algorithm (where the sampled region
itself is assumed periodic itself in space, which would be discontinuous as opposite edges of
the arbitrary sampled region are not continuations of each other. Discontinuities between
assumed periods of a spatially-periodic signal contain frequency content that does not exist
outside the context of assumed periodicity of a truncated sampled signal, and do not exist
in reality). Also, the spatial sampled region may be zero-padded to increase wave vector
resolution, and more accurately determine the wave vector of each spatial attribute function.
Without zero padding, wave vector content between wave vectors in the wave vector domain
will leak to surrounding wave vectors. Zero padding interpolates values in the wave vector
domain, and will reach a peak at the wave vector of the actual dominant spectral component.
The wave vector at the peak can then be extracted. It should be noted that while both
windowing and zero padding will provide a more accurate picture of the wave vector content
of a signal, the original signal has been altered prior to analysis and therefore values of the
DFT output will no longer correspond to the exact complex coe cients for each wave vector
and its corresponding plane wave.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 95
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
Modeling a spatial function as an in nite sum of plane waves, or even a sum of every
plane wave produced by a DFT, produces a cumbersome or impossibly large set of spatial
attribute functions to be a useful channel model. A modest sacri ce of accuracy for a
reduction in number of terms can be achieved by setting a minimum magnitude threshold
for frequency domain components, and extracting only plane wave functions with weights
exceeding this magnitude. Careful selection of this threshold must be performed to meet a
desired or required accuracy, or alternatively, peak spectral components can be iteratively t
and removed until this level of accuracy is achieved. Utilizing the wave vector spectrum of a
zero padded, windowed spatially sampled function allows for increased wave vector accuracy
with each iteration (as well as avoidance of false peaks caused by the discontinuity due to
assumed repetition of the sampled region), and as a result, fewer plane waves can be used
to model a spatial function to a desired degree of accuracy.
In regards to plane waves in two and three dimensions, a plane wave in three dimensions
is de ned by a three-dimensional wave vector, and a three-dimensional plane wave observed
on an arbitrary plane will appear as a two-dimensional plane wave with a wave vector equal
to the projection of the three-dimensional wave vector onto the plane of observation. This
results in a two-dimensional wave vector. If we assume that the observation region is on
the X -Y plane, then the projection of the 3-D wave vector onto this plane is simply the
x- and y- components of the wave vector, with the z -component discarded. The reasoning
for this is that along the X -Y plane, the z component of the position vector is constant.
The constant position z -component occurs in the plane wave exponential multiplied by the
z -component of the wave vector, resulting in a constant additive term in the exponential
function. This constant additive term can be factored out of the plane wave exponential
as a constant phase shift term, leaving the plane wave term in terms of x and y position
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 96
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
only and their corresponding x and y wave vector components, which is the de nition of a
2-D plane wave. Therefore a signal sampled in two-dimensional planar space, which consists
of a sum of 3-D plane waves, can be modeled as a sum of 2-D plane waves. The wave
vectors of these 2-D plane waves on the plane of interest can be determined using a two-
dimensional Discrete Fourier Transform of the spatially-sampled function. Subsequently, the
spatial attribute functions for the channel model can be reduced to 2-D plane waves if the
observation region exists along a 2-D plane, thus reducing the complexity of the problem.
(Also, without extending the sample space to include sampling along a third dimension, no
spectral information along the third dimension can be ascertained.)
One nal attribute of plane waves is how to ensure the pair-wise inner product of two
plane waves is su ciently low over all sample points in the observation region. Wave vectors
extracted directly from the DFT output of the sampled function (without zero padding) will
have integer-related periods over all spatial dimensions, and as a result are guaranteed to
have a pair-wise inner product of zero over all sample points. However, as discussed before,
spectral leakage due to spectral content falling between sampled wave vector points require
multiple plane waves to represent that content, increasing model complexity. If wave vectors
are determined from peaks of the DFT of the zero-padded function, the pair-wise wave vector
spacing between plane waves is no longer guaranteed to produce an inner product of zero
over the sampled region, which may result in a non-diagonal-dominant Gram matrix (which
may not be non-singular). If the peaks of the zero-padded signal are used to determine
wave vectors, then a test to exclude plane waves that have a pair-wise inner product above
a de ned threshold can be implemented. At the very least, a test to ensure a Gram matrix
is non-singular should be implemented upon each additional chosen plane wave.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 97
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
4.2.2 Point Sources as Spatial Attribute Functions
A second candidate spatial attribute function is a point source, the free-space Green's
function solution to the Helmholtz equation [19]. This exponential function has the form:
e−jk||~r−~r0 ||
EP S (~r) = (4.15)
4π||~r − ~r0 ||
where k is the wave number of the point source, ~r is the observation position vector in space,
and ~r0 is the point source position vector in space. The point source function occurs as a
kernel in the integrals found in the dipole channel formulation in Sec. 4.1. Because integrals
can be approximated as weighted Riemann sums of the arguments, it follows that the dipole
electric eld (and thus the dipole channel) could be roughly approximated with one or more
weighted point sources.
Using point sources as an approximation function poses two problems. First is deter-
mining the location of each point source from sampled data. A method must be developed
that can determine if sampled channel points contain an underlying signal that correlates
to a point source at a speci c point. Next is determining if each pair of point sources has
a su ciently low inner-product over the set of sample points, and excluding point sources
that have a high inner product with previously selected point sources. This is important
for creating a non-singular Gram matrix as explained in the previous section. With these
two problems solved, a least-squares method can be utilized to determine the weights of the
point sources that minimize the residual of the estimated channel.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 98
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
4.2.3 Point Source Location
To solve the initial problem of locating one or more point sources, it is helpful to visualize
a point source eld, and its observation on a rectangular plane. Fig. 4.10 illustrates surfaces
of constant phase from a point source. For a selected constant arbitrary phase, a point
2π
source excitation with constant wavenumber k= λ
, will have concentric spherical wavefronts
corresponding with the selected arbitrary phase, each separated by the wavelength λ from
the previous and subsequent wavefront. This is illustrated in Fig. 4.10 by the two red line
segments, each labeled λ. In geometrical optics, these constant-phase surfaces are known as
Eikonal surfaces [51].
If these Eikonal surfaces from a point source are observed on a at rectangular observation
region, as represented by the black rectangle in the gure, arc-shaped wavefronts are observed
(Eikonal curves). The separation of these Eikonal curves along the plane are not λ and are
not uniformly-spaced (as the point source is not co-planar with the observation region, and
therefore the plane is not normal to the Eikonal surfaces). It should be noted that the
observed phase in Fig. 4.9 has arc-shaped Eikonal curves similar to those in 4.10, which
indicates that a point-source spatial attribute function may be a logical choice.
Any Eikonal surface pattern in a planar 2-dimensional sample region can be formed by
a single point in a half-space above the plane of observation, and the re ection of that point
in the half-space below the plane of observation. If we know which half-space in which
the antenna is located, which is where the eld originates and therefore where point source
approximations should be located, we can determine a single, unique point from a set of
Eikonal surfaces.
Keeping the wavenumber k of the analyzed system constant (and thus keeping frequency
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 99
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
-1
-2
2
0 5
-2 3 4
-4 1 2
-6 -1 0
Figure 4.10: Visualization of Point Source Wave Fronts
and wavelength constant), a sampled channel in the rectangular observation region will
consist of complex samples at each point, containing amplitude and phase information. The
expected contribution of a point source at an arbitrary location ~r0 can be calculated at each
and every observation point, and a correlation between the point source contribution and
the sampled channel can be calculated. If the phase fronts (and as a result all phases) in the
point source and an underlying component of the sampled channel align, then the magnitude
of the correlation will be high.
Utilizing this property of high correlation at a likely point source, we can perform a
global search of point sources ~r0 in the upper half-space above the observation plane to
maximize the correlation between the point source and the sampled channel. The value of ~r0
at the maximum is a likely candidate point source. The sampled channel as a result contains
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 100
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
the Eikonal surface structure within the sampled data. This optimization problem can be
expressed as in (4.16):
N
~rˆ0 = arg max | H(~ri )ejk||~ri −~r0 || |
X
(4.16)
r0 |~
~ r0z >rz i=1
where ~rˆ0 is the estimated point source location, ~r0 is an arbitrary point source location, N
is the number of channel sample points, rz is the z-coordinate of the observation region,
and H(~ri ) is the sampled channel at observation point ~ri . A correlation threshold can be
speci ed such that no value of ~rˆ0 is returned if the maximum value of the argument is below
this threshold. This threshold value can be selected through experimentation, and can be
used to prevent the selection of a point source if su cient correlation with the sampled data
is not achieved.
The estimated point source at ~rˆ0 from phase correlation maximization can then be t
to the observed channel data using least squares to obtain a complex coe cient and then
subtracted from the sampled channel data to obtain a residual. Subsequent point sources
can iteratively be extracted from and t to the residual. However, care must be taken when
selecting additional point sources to ensure a low inner-product with other point sources,
which in turn will help prevent a singular Gram matrix.
4.2.4 Inner Product of Point Sources
With the determination of each additional candidate point source, a su ciently-low inner
product between it and every previous point source should be enforced, with new points
with a high inner product being excluded for inclusion in the model. While performing
this calculation for each and every candidate point is possible, an additional constraint on
the optimization problem described in the previous section can be enforced to increase the
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 101
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
chances of nding a candidate point source with a low inner product with other previously-
selected point sources.
Engquist and Zaho [55] investigated the separability of Helmholtz Green's functions at
high frequencies. This paper establishes bounds on the inner-product between two point
sources in an observation region as a function of their separation from each other and the
observation zone. The formulation in this paper was too complex to implement as a con-
straint function for optimization. However, the principle that the inner product of two point
sources decreases over an observation region as the separation increases provided motivation
for performing a numerical analysis and empirical distribution of inner product versus point
source separation distance.
For an observation zone matching the dimensions of the numerical computation sample
λ
zone in Sec. 4.1.3, a grid of sampled observation points with separation of in both X and
20
Y directions was established, and a 3-dimensional grid of potential source points spanning
the width and height of the chamber and the depth of the chamber from a depth of a half-
wavelength behind the observation zone to the rear of the chamber. The source point grid
λ
has a spacing of
3
in the X , Y , and Z directions. Every inner product (i.e. cross-correlation)
of each possible combination of source points in this region was calculated, resulting in over
166 million separations and cross-correlations. Fig. 4.11 is a histogram normalized such that
the integral of each vertical slice equals one (and is thus a PDF for a given separation).
We can see that as point sources increase in separation, the probability of the cross-
correlation being very low increases signi cantly. At a separation of 6λ, the probability of an
cross-correlation being below 0.1 is nearly 1.0. Conversely, when the separation is less than
0.5λ, the probability of the cross correlation being above 0.8 is nearly 1.0. Essentially, this
indicates that a minimum point source separation of somewhere between 0.75λ and 2λ would
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 102
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
Figure 4.11: Normalized Histogram of Cross-Correlation Vs. Separation
be su cient to increase the probability of a low cross-correlation to be signi cant enough to
investigate. Setting a minimum separation too high may exclude valid candidate points that
would increase the accuracy of the model, while setting the minimum separation too low
would increase the number of points excluded due to the inner-product being above a set
threshold, and therefore the computational time of the algorithm searching for new points
that meet this threshold requirement.
Signi cant testing of manual adjustment of the minimum separation and cross-correlation
threshold with measured channel data yielded a minimum separation of 0.75λ and cross-
correlation threshold of 0.08, which allowed the multiple point source correlation algorithm
to run with su ciently-small error and reasonable run time. More detail on the channel
measurements used for this process will be discussed in the following chapter.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 103
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
4.2.5 Use of Point Source and Plane Waves as Joint Spatial At-
tribute Functions
Plane waves and waves from a point source are related. Using the concept of Eikonal
surfaces, plane waves contain planar Eikonal surfaces separated by a wavelength λ, and
have constant amplitude over space. At a large observation distance, over a relatively small
observation region, the local Eikonal surfaces of a point source are approximately planar
due to a small spherical sector solid angle and large spherical sector radius ||~r − ~r0 ||. Also,
the relative change in radius, δ, from the point source over the distant observation region is
small relative to the radius from the point source (i.e. ||~r − ~r0 || >> δ ), and therefore the
1 1
point source coe cients
r−~
4π||~ r0 ||
≈ r−~
4π(||~ r0 ||+δ)
, yielding a nearly-constant amplitude across
the observation zone. As a result, in a small observation region distant from a point source,
the planar Eikonal surfaces and constant amplitude indicate that a distant point source is
approximately a plane wave. Therefore, a joint usage of close point sources and plane waves
is essentially equivalent to using purely point sources (with some close and some very distant)
to model the channel. However, even though these spatial attribute functions are essentially
the same, we can utilize the special shape properties of Eikonal surfaces of near point sources
and the Fourier properties of plane waves to t each to a data set in an observation region
and generate a model.
4.3 Fitting of Spatial Attribute Functions to Numerical
Channel
The third contribution of this chapter is presented in this section, as noted in the third
column of Fig. 4.1. The methods of phase correlation maximization as outlined in Sec. 4.2.3
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 104
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
and spatial spectral decomposition as outlined in Sec. 4.2.1 are combined into a procedure to
model the channel. The following procedure is performed on data sampled in an observation
region that meet the previously-outlined spatial Nyquist sampling requirements.
This method can be utilized to reduce the total number of components (plane wave plus
point sources) required to model a signal (or in this case a channel) when compared to a
pure sum-of-plane waves model, thus reducing model complexity.
4.3.1 Point Source Fitting
The rst part of the procedure involved determining point sources and corresponding
complex weights of the sampled channel. This utilizes methods discussed in Sec. Sec. 4.2.3.
1. Collect the sampled channel data H(r̃) at all measurement points r̃ in the observation
region for a single wavelength λ.
2. De ne a search region for point sources ~r0 for the correlation maximization algorithm.
3. De ne minimum point source separation constraint 0.75λ and maximum inner-product
for point source pairs (0.08).
~rˆ01 = arg max~r01 |~r01z >rz | N ri )ejk||~ri −~r01 || |
P
4. Run a global optimization routine on i=1 H(~ to
determine an initial point source ~rˆ01 .
5. Calculate the contribution of the point source ~rˆ01 at each observation location r̃:
EP S1 (r̃).
6. Use least squares to calculate the complex weight β1 of the point source that best ts
the sampled channel data and minimizes the residual ~ε1 (r̃), H(r̃) = β1 EP S1 (r̃) + ~ε1 (r̃).
7. Calculate the residual ε1 (r̃) after the least squares t.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 105
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
8. Repeat steps 4 through 7 using ~ε1 (r̃) instead of H(r̃), with the optimization in step 4
subject to the constraints de ned in step 3, to calculate ~rˆ02 , EP S2 (r̃), β2 , and ~ε2 (r̃).
9. Repeat step 8 using ~ε2 (r̃) instead of ~ε1 (r̃), then ~ε3 (r̃) instead of ~ε2 (r̃), to calculate as
many point sources ~rˆ0n and complex weights βn as needed, either an arbitrary number
or until a point source is under a de ned correlation threshold. The nal residual ~εn (r̃)
will be used for re nement with plane waves.
4.3.2 Re nement Using Plane Waves
The residual from the point source tting, ~εn (r̃), is then modeled with plane waves
iteratively until a desired accuracy is reached.
1. Using the residual ~εn (r̃) over all measurement points r̃ in the observation region for a
single wavelength λ, apply a 2-dimensional Hanning window and zero pad the data in
both X and Y directions until the desired DFT frequency resolution is obtained.
2. Perform a 2-dimensional DFT on the data from step 1. Calculate the wave vectors for
each point in the frequency domain.
3. Find the wave vector k~1 corresponding to the maximum of the DFT output. This is
the wave vector of the rst plane wave to be t.
~
4. Calculate the contribution of the plane wave EP W1 (r̃) = e−j k1 ·~r at each observation
point r̃.
5. Perform a least squares t of the plane wave to the point source residual ~εn (r̃) to
nd the value βP W1 that minimizes the plane wave residual ~εP W1 (r̃) in the equation:
~εn (r̃) = βP W1 EP W1 (r̃) + ~εP W1 (r̃).
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 106
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
6. Determine the residual ~εP W1 (r̃) after the least squares t.
7. Perform steps 1 through 6 using ~εP W1 (r̃) instead of ~εn (r̃) to nd βP W2 and ~εP W2 (r̃).
8. Repeat step 7 replacing the previous residual with the new residual each time to cal-
culate βP Wn and ~εP Wn (r̃).
9. Stop this procedure when the mean Error Vector Magnitude (EVM), the mean mag-
nitude of the residual, is below a de ned threshold.
4.4 Validation of Spatial Attribute Function Selection
with Numerical Channel
The algorithm de ned in the Sections of 4.3 was performed on the numerical dipole
array channel developed in Sec. 4.1.3. The mean and standard deviation of the Error
Vector Magnitude (EVM) of dierent combinations of point sources and plane waves to
determine the eect of each type of (and number of each) spatial attribute function on
the accuracy of the model. Models consisting of only point sources and only plane waves
were also calculated using just the appropriate half of the algorithm in the previous section.
Finally, the value of mean plus 1 standard deviation of the EVM of each case is observed,
indicating that approximately 84 percent (50 percent plus one σ) of EVM from the model
over the observation area is below the stated value.
Figure 4.12 is a plot of the LogNormal mean EVM for between 0 and 3 point sources
(spherical components) and 0 and 5 plane waves (planar components). For 0 point sources
and 0 plane waves, the error is 0 dB, or 100 percent error, by de nition (nothing modeled).
A model consisting of a single point source and no plane waves provides an EVM of -18.3
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 107
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
dB, meaning that on the average, the error signal contains 1.5 percent of the power of the
channel being modeled. A single plane wave without any point sources, on the other hand,
has an EVM of -8.2 dB. Four plane waves are required to achieve an EVM less than a single
point source. This indicates that a point source is an excellent initial choice of a spatial
attribute function!
0 -5
-5
-10
model mean LN EVM (dB)
-10
-15
-15
-20
-20
-25 -25
-30
-30
0
1
-35 2
0 3
1 4 -35
2 5
3
num. planar components
num. spherical components
Figure 4.12: LogNormal Mean EVM For Combinations of Spatial Attribute Functions
As expected, additional point sources do not decrease the EVM by a signi cant amount.
The point sources are modeling free-space radiation from a relatively small antenna array.
The minimum separation constraint for subsequent point sources after the rst is at least
0.75λ, putting the second or third point source in a region away from the antenna for which
the channel is being modeled. As a result, the second and third point sources had very
low maximum correlations, and had very low magnitude coe cients when t to the residual
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 108
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
data. As a result, they had extremely low contribution to the model, and while they do no
harm to the model, these point sources can be excluded. It should be noted that the system
being modeled in this case is free-space and involves no re ections; additional point sources
will be useful in modeling source re ections, as will be seen in the following chapter.
Finally, we observe combinations of point sources and plane waves, as the plane waves
serve to re ne the point source model. With one point source and two plane waves, a mean
EVM of -24.7 dB is achieved. One point source and 5 plane waves achieves a -26.9 dB EVM.
As the number of plane waves increases, the advantage of the point source component of the
model is still present, though its bene t decreases. The clear bene t of a joint point source
and plane wave model is in models with a relatively low number of terms, and produces
signi cantly lower EVM than a comparable number of plane waves alone.
In terms of standard deviation, models of only a couple plane waves have a standard
deviation of between 1.5 dB and 2 dB, models of only point sources have about 2 dB standard
deviation, and combinations have between 2 dB and 2.5 dB. This is illustrated in Fig. 4.13.
In Fig. 4.14, we observe mean plus σ, meaning that 84 percent of modeled locations
are as accurate as the value in this plot. Because the standard deviations were relatively
constant between 1.5 and 2.5, very little changes in terms of the advantage of point sources
and combinations over plane waves alone, though the gap closes a small amount. Even a
single point source still has 84 percent of EVM below -16.3 dB.
In short, the low EVM for a relatively low number of spatial attribute functions indicates
that point sources, and joint point sources and plane waves are validated as an eective way
of modeling the synthetic numerical dipole array channel with an acceptable accuracy of
between -15dB and -20dB or better. This modeling algorithm will be applied to real channel
measurements in the following chapter.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 109
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
3
2.5
2.5
2
model std LN EVM (dB)
1.5
1.5
1 1
0.5
0.5
0
1
0 2
0 3
1 4 0
2 5
3
num. planar components
num. spherical components
Figure 4.13: Standard Deviation of EVM For Combinations of Spatial Attribute Functions
0 -5
model mean EVM plus one standard dev. (dB)
-5
-10
-10
-15
-15
-20
-20
-25 -25
-30
-30
0
1
-35 2
0 3
1 4 -35
2 5
3
num. planar components
num. spherical components
Figure 4.14: LogNormal Mean + 1 SD EVM For Combinations of Spatial Attribute Functions
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 110
CHAPTER 4. NEAR-FIELD CHANNEL MODEL FRAMEWORK
4.5 Conclusions
The channel modeling algorithm presented in this chapter has been validated to produce
acceptable error for a small number of spatial attribute functions, and therefore is eective
as a modeling tool for relatively simple, accurate description of the channel between a log-
periodic array and a dipole antenna. Given the principles outlined in this chapter, it could in
theory be adapted to other antenna types and environments for investigation and modeling
of other channels, both real and synthetic. In addition, additional variables could be added
to the model, such as transmitter antenna position and orientation, and receiver orientation
and/or vertical position. A combination of point source and plane waves would still be
su cient for description of such channels, however the number of each type of basis function
will vary depending on the environment, including receiver observation zone size, distance
between transmitter and receiver, number of re ections and the power of each re ection,
and physical size of transmitter and receiver arrays. The next chapter will investigate a real,
non-free-space system in which added factors such as re ections will be introduced.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 111
CHAPTER 5
Statistical and Frequency Extensions of
Near-Field Channel Model
The purpose of this chapter is extend the mathematical framework derived in the previous
chapter to incorporate statistical information from the spatial attribute function residual,
as well as extend the model as a function of frequency. This produces a model that has
improved EVM over the spatial attribute function tting method alone, and is broadband.
In addition, this chapter utilizes channel measurements taken inside a small anechoic test
chamber between log-periodic probe antennas in four locations and a receiver dipole antenna
placed along a grid to sample the channel space. The models for each probe antenna resulting
from this measurement campaign can therefore be combined to calculate a MIMO channel
between the probes and an arbitrarily-positioned receiver dipole array. The MIMO analysis
of this channel model will be performed in the next chapter.
Fig. 5.1 shows the breakdown of contributions in chapters 4 and 5. The contributions
for Chapter 5 are emphasized in bold.
The rst contribution of this chapter is to improve the accuracy and EVM of the channel
model presented in the previous chapter utilizing statistical information from the residual of
the tted spatial attribute functions. This is achieved by applying methods of Geostatistical
Regression [56] [52] and Complex Geostatistical Kriging [57], methods taken from the eld
of Geostatistics (so-called due to its original application to Geology and mining). This eld
of analysis focuses on statistical properties of random signals in space (or space and time),
112
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
Motivating How can the measured What is the most accurate
Question What can the small Can this be simplified for a Can the SAF for one
channel be method of fitting SAF to
chamber channel be fixed TX and a test zone frequency be utilized in a
decomposed into measurement data?
modeled as? RX? wide-band model?
chosen SAF?
Free-space
Properties
SA funcitons
Properties
Statistical
approx.
SAF
Theories Used
Use of Generalized Least Use Geostatistical Kriging
Numerical Model of
Use of point sources (free- Use of point Eikonal Squares, Geostatistical to fit SAF coefficients, fit
Antenna Interaction (Hallen
space Green’s Functions) Surface correlation, Regression and coefficients to a
equations, current
and plane waves as SAF multidimensional Fourier Geostatistical Kriging to polynomial model vs.
distribution) analysis determine SAF weights frequency
Correlation maximization of Modeling coefficients SAF
Contribution
Generalized Formulation Application of Complex
Combination of point point source Eikonal of a central frequency as a
for Log Periodic Array and Kriging with Geostatistical
sources and plane waves surfaces with measured polynomial function of
Arbitrary Antenna Kriging to Spatial Channel
as model SAF data; 2d Fourier transform frequency separation.
Orientation Measurements
of residual
Show Numerical Model can Locate point source SAF Iteratively determine Model a wide-band
Contribution
unbiased best fit of SAF
Applied to
Question
Calculate Channel Vs. be represented using point using correlation, channel using a single set
Receiver Position in Test source and plane wave SAF determine plane wave to data, as well as of SAF functions and
Zone for Fixed Transmitter with acceptable error vectors with 2D Fourier spatial statistics of the polynomial coefficient
Transform of measurement residual of the SAF functions.
Chapter 4 Chapter 5
Figure 5.1: Chapter 5 Contribution Flowchart
and the use of these statistics in conjunction with spatial measurements to estimate the
signal at any location. This application is new and novel, as no application of this eld to
spatial channel modeling has been found in the literature. In addition, no use of Complex
Geostatistical Kriging in conjunction with the iterative process of Geostatistical Regression
has been located in the literature.
The second contribution of this chapter is extending our narrow-band model, which de-
termines spatial attribute functions, complex weights, statistics, etc. at a single frequency,
to a broadband model as a function of frequency. Using a common set of frequency-scaled
spatial attribute functions, complex weights for each are determined over a wide range of fre-
quencies, and a polynomial regression is performed to determine a function for each weight as
a function of frequency. This extends our novel Geostatistical channel model to a Wideband
Geostatistical channel model.
Sec. 5.1 outlines the methods of Geostatistical Regression and Complex Geostatistical
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 113
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
Kriging, and combines the two into a complex iterative Geostatistical Regression algorithm
for spatial attribute function determination. Sec. 5.2 presents a method of adapting spa-
tial attribute functions over a wide frequency band, and modeling the complex function
coe cients as a function of frequency, yielding a Wideband Geostatistical model. Sec. 5.3
describes the system being modeled in detail, while Sec. 5.4 explains the measurement cam-
paign performed on that system. Sec. 5.5 applies the Wideband Geostatistical modeling
process on the measurement data to produce a model for the measured system.
5.1 Accurate Method of Fitting Spatial Attribute Func-
tions to Measured Data
The method of ordinary least squares (OLS), utilized in the previous chapter for tting
each spatial attribute component to the sampled channel (or subsequently the residual of the
sampled channel after each tted spatial attribute function (SAF) is subtracted), can be used
to simultaneously estimate the complex coe cients of each SAF. The system can be modeled
with a matrix X, where each column is a spatial attribute function, and each row is a channel
sample location (in other words, each element is the value of an unweighted spatial attribute
function evaluated at a sample location). In addition to the spatial attribute functions, an
additional column is added to the matrix consisting of a unit constant, which will correspond
to a complex additive constant (an extra value to be determined in the vector of complex
weights to be estimated) to force a zero-mean residual. X is multiplied by a column vector of
complex weights for each spatial attribute function, β~ , which can be calculated to minimize
the residual ~ε when subtracted from the vector of sampled data at each sample location, ~y .
This system can be expressed as:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 114
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
~y = Xβ~ + ~ε (5.1)
The residual ~ε is assumed to be normally-distributed with variance σ2 and zero mean.
X
The OLS solution of this equation that minimizes the sum of the residual ~ε is (5.2), the
pseudoinverse of X times the sampled data ~y .
β~ 0 = (XH X)−1 XH ~y (5.2)
H 0
where is the Hermetian transpose, and represents the estimation of a variable, two
notation conventions used through this chapter. This assumes that the Gram matrix, XH X,
is non-singular. In the previous chapter, special care was taken to select SAF that yielded a
non-singular Gram matrix.
It should be noted that OLS is only appropriate for a classical regression model, which as-
sumes that the elements of the residual ~ε are statistically independent [52], thus cov ε = σ 2 In ,
where σ2 is the residual variance, In is the N ×N identity matrix (where N is the length
of ~ε. This means that the residual is unrelated at two separate locations in space, even if
the separation of those locations is extremely small. This assumption is not necessarily true,
and spatial correlation of the residual as a function of separation must be investigated, and
if present, taken into account in the regression process.
5.1.1 Generalized Least Squares
When the correlation matrix C of the residual is known a priori, if non-diagonal elements
are non-zero, then the residual ε is no longer considered spatially-independent. Through
the use of a decomposition of the C matrix into a Cholesky matrix T and its transpose,
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 115
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
C = TTH , (5.1) can be pre-multiplied on both sides by the inverse of T to transform the
equation to one in which the transformed residual term has statistically-independent entries
[52]. This residual independence yields a transformed equation (5.3) in which β0 can be
solved using OLS.
β~ 0 = ((T−1 X)H (T−1 X))−1 (T−1 X)H (T−1 ~y ) (5.3)
Through manipulation and reverse substitution of the Cholesky decomposition [52], the
estimation of β0 (known as Generalized Least Squares (GLS) regression) is expressed in
terms of the correlation matrix C in (5.4).
β~ 0 = (XH C−1 X)−1 XH C−1 ~y (5.4)
This estimate of β~ 0 once again requires C to be known a priori. In most applications,
however, both the complex weights β~ 0 of the SLA and a model of the correlation of the
residual C0 (and associated model parameters) must be simultaneously estimated. This
simultaneous estimate can be performed using an iterative algorithm called Geostatistical
Regression [56] [52], outlined in Section 5.1.4. First, a method of modeling the spatial cross-
correlation of the residual must be established.
5.1.2 Covariograms and Spatial Dependence
A statistical analysis and covariance (and correlation) model of the residual data ~ε must
be performed for use in the iterative Geostatistical Regression process. A useful method
to model spatial dependencies is the variogram (and covariogram). In a covariogram C(h),
the covariance of two processes at two points in space (cov(ε(~
x1 ), ε(~x2 )) is a function of the
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 116
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
separation of the two points in space alone (h = ||~x1 − ~x2 ||), independent of where the points
are [52].
A convenient way to calculate and express the manner in which two processes are expected
to dier in space is to observe the expected squared dierence of the residual E[ε((~x1 ) −
ε(~x2 ))2 ] [52]. Through expansion and manipulation, the de nition of the covariogram can be
utilized, and the expected squared residual dierence becomes 2[C(0) − C(h)]. Normalizing
this by a factor of a half, the variogram can be de ned as γ(h) = C(0) − C(h). Finally, using
C(0) = σ 2 , the relation between the variogram and covariogram becomes C(h) = σ 2 − γ(h).
Residual data can be used to construct an empirical covariogram using the following
procedure [52].
1. For a set of residual data ~ε, calculate the squared dierence of each and every pair of
residual data and the separation h between the points at which the residual is sampled.
2. Divide the span of calculated separations h into evenly-spaced bins from 0 to hmax .
3. Group the squared dierence data into respective bins based on their corresponding
separation h
4. Calculate the mean squared dierence for each bin (thus estimating the expected
squared dierence for each h bin).
The resulting bin mean values versus bin h (normalized by one half ) form the empirical
variogram. It should be noted that bin size and maximum separation hm ax should be
chosen such that each bin contains at least 30 points [52]. This is a general rule of thumb,
and is applied to assure a su cient number of points such that their average approaches the
expected value.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 117
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
The empirical covariogram can be analyzed to choose an appropriate covariogram model.
The empirical covariogram can then be t to a candidate model using OLS, or each bin can
be weighted such that data from bins with more points have a greater weight than data from
bins with fewer points (a process knkown as weighted least squares).
Two issues arise from this procedure as it applies to channel modeling. First, the var-
iogram is de ned only in terms of point separation h, which means that the variogram is
independent of direction of separation. This indicates that the underlying covariogram is
assumed to be isotropic. If the empirical variogram is calculated with bins in two dimensions
(thus in terms of separate x and y separation distances), the variogram can be inspected
for anisotropic behavior. This will likely exhibit as an ellipse with a major and minor axis
and a rotation. Correctional rotation and directional scaling can be applied to the empiri-
cal variogram data to create an equivalent isotropic model. Chapter 2 contains additional
information on isotropy in variograms.
Second, and perhaps most importantly, the variogram models discussed in [56] and [52] all
utilize real-valued data, producing a positive de nite variogram. The residual of the channel
model being developed will be complex-valued. While the general concepts and procedure
remain as described, complex candidate models and empirical variogram methods must be
utilized to apply these methods to a complex channel model.
5.1.3 Complex Geostatistical Covariance Model
Both issues of complex covariance modeling and anisotropy have been addressed by De
Iaco and Posa [57]. The following is a summary of the method extracted from this paper for
use in the iterative Geostatistical Regression procedure in the following section.
For complex random elds, the complex covariance is de ned as:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 118
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
C(~h) = E[(Z(~x1 ) − µ)(Z(~x1 + ~h) − µ)] (5.5)
where Z(~x) = X(~x) + iY (~x) is the complex eld at position ~x, µ is the mean of Z (which
is equal to zero in the context of our residual eld), and ~h is the separation vector (which
it should be noted accounts for separation distance and direction). The covariance function
can be written in terms of two functions, one for the real part and one for the imaginary
part (5.6), the real part is the sum of the autocovariance of X and autocovariance of Y, and
the imaginary part is the dierence of the cross-covariance of Y and X , and cross-covariance
of Y and X [57].
C(~h) = C re (~h) + iC im (~h) = (CX (~h) + CY (~h)) + i(CY X (~h) − CXY (~h)) (5.6)
The imaginary part of this complex function is not a covariance function, which means
an alternate form for expressing a complex covariance must be determined. An admissible
covariance model is [57]:
C(~h; ~c, θ)
~ = cos(~h · ~c)C̃(~h; θ)
~ + i sin(~h · ~c)C̃(~h; θ)
~ (5.7)
where ~h is a separation vector, ~c is a translation vector that relates the real and imaginary
cross-covariances, and θ~ is a parameter vector for a selected real covariance function C̃ . The
real covariance function will be selected after inspection of the empirical covariance and
variogram.
With preliminary variables and equations established, the following procedure from [57]
was followed:
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 119
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
1. Calculate auto-covariance and cross-covariance for the real and imaginary parts of each
and every pair of residual signal ~ε0 , and calculate the separation vector ~h for each pair.
2. Divide the 2-dimensional separation vector space into separation vector bins, and cal-
culate the mean CX (~hi ), CY (~hi ), CY X (~hi ), and CXY (~hi ) for each bin. Calculate C re (~hi )
and C im (~hi ) for each bin ~hi . Observe the behavior of Cre and Cim versus ~hi to select
a C̃(~hi ; θ)
~ which follows the trends of the observed functions.
3. In both [57] and for this procedure, a Gaussian covariance model was selected for
C̃(~h; θ)
~. The form of this model is in (5.8). The parameter vector θ~ consists of the
parameters σ 2 = C(0), the angle corresponding to maximum range θ, maximum range
aθ , and minimum range aφ (range in perpendicular direction to maximum range). In
other words, θ~ = (σ 2 , θ, aθ , aφ ). With the exception of σ2, the elements of θ~ de ne
the anisotropy present in the observed covariance and compensate for it. Note that
~h = (h1 , h2 ), the x and y separation components, in (5.8) and subsequent equations.
2 2
cos(θ)h1 −sin(θ)h2 sin(θ)h1 +cos(θ)h2
− aφ
+ aθ
C̃(~h; (θ, aθ , aφ )) = e (5.8)
This equation, in conjunction with (5.7), produces the complex covariance model
in (5.9).
h i
C(~h; ~c, θ)
~ = σ 2 cos(~h · ~c)C̃(~h; (θ, aθ , aφ )) + i sin(~h · ~c)C̃(~h; (θ, aθ , aφ )) (5.9)
4. Let Nl equal the number of bins ~hi . The translation vector ~c can be estimated through
OLS using C re (~hi ) and C im (~hi ) in conjunction with (5.7). The resulting equation for
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 120
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
estimating ~c0 is (5.10).
−1
PNl 2 PNl PNl C im (~hi )
i=1 h1,i i=1 h1,i h2,i arctan( C re (~h ) )h1,i
~c0 = i=1
i (5.10)
C im (~hi )
PN PNl 2 PNl
i=1 h1,i h2,i i=1 h2,i i=1 arctan( C re (~h ) )h2,i
l
i
5. Use an optimization routine (non-linear least squares or other) to nd the parameter
vector θ~ that minimizes the sum
Nl
θ~0 = arg min [(C re (~hi ) − σ 2 cos(~hi · ~c0 )C̃(~hi ; θ))
~ 2 + (C im (~hi ) − σ 2 sin(~hi · ~c0 )C̃(~hi ; θ))
~ 2]
X
θ~ i=1
(5.11)
6. Using θ~0 and ~c0 in conjunction with (5.9), we now have a complex covariance model
C(~h; ~c0 , θ~0 ).
With this complex covariance model, a full covariance matrix V can be calculated. If
Nl measurement locations (corresponding to the number of elements in the sample vector
y ), then V is an Nl × Nl matrix, with element Vi,j = C(~xi − ~xj ) = C(hi,j ; ~c0 , θ~0 ). The
full covariance matrix V is related to the correlation matrix C by a factor of the variance
(V = σ 2 C) [52]. Because of this relationship, the GLS estimation of β~ 0 (5.4) can be rewritten
in terms of the full covariance matrix instead of the correlation matrix (5.12).
β~ 0 = (XH V−1 X)−1 XH V−1 ~y (5.12)
This GLS equation utilizing a full covariance matrix determined from the complex co-
variance model produces a better estimate of β~ 0 . This better estimate of β~ 0 in turn is applied
to (5.1), which then yields a new residual vector. The new residual vector can be used to
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 121
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
revise the covariance model, and so forth. This directly leads to an iterative re nement of
estimated parameters β~ 0 , ~c0 , and θ~0 , which is presented in the following section.
5.1.4 Geostatistical Regression
Previous work shows a method for iteratively estimating β~ 0 and spherical variogram
parameters θ~0 for real-valued random processes [52]. This procedure is adapted here (with
general progression of the algorithm closely following the source) for a complex Gaussian
covariogram as developed in the previous section, to iteratively estimate β~ 0 , covariogram
parameters θ~0 , and translation vector ~c0 for a complex random process.
1. Use the procedure from the previous chapter to determine a set of plane wave and
point source spatial attribute functions (SAF), and evaluate these functions at all
measurement locations at which the measurement y was taken, thus forming a spatial
attribute matrix X. As before, the rows of X correspond to each sample location, and
each column in X corresponds to a spatial attribute function. An extra row consisting
of a unit-magnitude constant (identical for all rows) is added to force the resultant
residual to be zero-mean.
2. For the system ~y = Xβ~ + ~ε (where ~
beta are the complex coe cients of each SAF, and
~ε is the model residual), perform an initial OLS estimate of β~ 0 (5.13).
β~00 = (XH X)−1 XH ~y (5.13)
Calculate the corresponding residuals (5.14).
~ε00 = ~y − Xβ~00 (5.14)
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 122
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
3. Use the residuals ~ε00 in conjunction with the procedure in Section 5.1.3 to estimate the
empirical complex covariance parameters C re (~hi ) and C im (~hi ) for the separation vector
space divided into bins ~hi .
4. Continue to use the procedure in Section 5.1.3 to t the empirical complex covariance
data to the complex Gaussian covariance model. The estimated parameters to t the
model to the data are the translation vector ~c00 , and the parameter vector vecθ00 =
(σ020 , θ00 , a0θ0 , a0φ0 ). The resulting model is C00 (~h; ~c00 , θ~00 ).
5. Fill in the initial estimate of the full covariance matrix (5.15), where each row i, column
j pair represents a pair of points xi and xj , with separation vector hi,j = xj − xi . Index
i correspond to the row number of spatial attribute matrix X, which corresponds to
the spatial attribute functions evaluated at point xi .
h1,1 ; ~c00 , θ~00 )
C0 (~ C00 (~h1,n ; ~c00 , θ~00 )
0
···
0 0 ~0 . .. .
V00 . .
= V (~c0 , θ0 ) . . .
C00 (~hn,1 ; ~c00 , θ~00 ) · · · C00 (~hn,n ; ~c00 , θ~00 )
σ02 ··· C00 (~h1,n ; ~c00 , θ~00 )
. .. .
. .
=
. . .
(5.15)
C00 (~hn,1 ; ~c00 , θ~00 ) · · ·
σ02
6. Utilize GLS in conjunction with V00 as in (5.12) to estimate a new value of β~10 (5.16).
β~10 = (XH V00−1 X)−1 XH V00−1 ~y (5.16)
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 123
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
Calculate new residuals using the new value of β (5.17).
~ε01 = ~y − Xβ~10 (5.17)
7. Perform Steps 3 and 4 with ~ε01 instead of ~ε00 to determine the translation vector ~c01 , and
the parameter vector θ~10 = (σ120 , θ10 , a0θ1 , a0φ1 ).
8. Let k equal the current iteration number (if immediately following step 7, k = 1), and
k−1 is the previous iteration number (if immediately following step 7, k − 1 = 0).
For each current and previous estimated parameter ~ 0 , ~c0 , σ 20 , θ0 , a0
((β , a0φk ),
k k k k θ k
(β~k−1
0
, ~c0k−1 , σk−1
20 0
, θk−1 , a0θk−1 , a0φk−1 )), calculate the magnitude fractional change of each
parameter p0 (where p0 is any parameter from the list of parameters) from the previous
iteration (k − 1) to the current iteration (k ).
p0k − p0k−1
∆kp0 = (5.18)
p0k−1
If the parameter p0 is a vector with n parameters (as with β~ 0 and ~c0 ), then let ∆kp~0 be
the maximum fractional change between iterations of any element of the vector (5.19).
p0j,k − p0j,k−1
" #
∆kp~0 = max : j = 0, 1, . . . , n (5.19)
p0j,k−1
Finally, the maximum fractional change between iterations is (5.20).
∆k = max ∆kβ~0 , ∆k~c0 , ∆kσ20 , ∆kθ0 , ∆ka0 , ∆ka0 (5.20)
φ θ
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 124
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
9. De ne a maximum fractional change threshold ¯
∆ (for which the maximum fractional
change calculated in the previous step must be under to terminate the algorithm). If the
maximum fractional change is below or equal to this threshold (∆k ¯ ) for the current
≤∆
iteration k, then terminate the algorithm and the nal estimated model parameters
are (β~ 0 , ~c0 , σ 20 , θ0 , a0θ , a0φ ) = (β~k0 , ~c0k , σk20 , θk0 , a0θk , a0φk ). If the maximum fractional change
is above this threshold (∆k ¯ ),
≤∆ increment k by one and repeat Steps 5 through 8
until the maximum fractional change is below or equal to this threshold (∆k ¯)
≤∆ for
the current iteration k, and extract the nal estimated model parameters. In simple
terms, the iterative algorithm will terminate when each and every parameter converges
and changes by a su ciently small amount between iterations.
5.1.5 Geostatistical Kriging
In the previous section, we iteratively estimated the Spatial Attribute Function (SAF)
~ 0 ), translation vector (~c0 ), and residual statistical model parameter vector
complex weights (β
~0 ) from a series of channel measurements. With these channel model parameters in con-
(θ
junction with the sampled channel measurements, we can implement a method of estimating
the channel at any location in the sampled test area using a method known as Geostatistical
Kriging [52]. In Geostatistical Kriging, for a location for which the channel is unknown and
estimation is desired (x0 ), the SAF can be evaluated at x0 forming ~0
X (with capital X rep-
resenting the Spatial Attribute Matrix and lowercase x representing a position in space) and
weighted by the estimated coe cient vector β~ 0 determined in the previous section. Then,
using the residual of nearby measurement locations in conjunction with the residual spatial
0
covariance model, the residual at the unobserved point (ε0 ) can be estimated. The following
procedure, once again closely following that of [52], details this procedure.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 125
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
1. Determine a prediction set for the unobserved point x0 . Only measurements spatially
close to x0 are signi cantly correlated with the residual at x0 . De ne a threshold h0
beyond which measurement residuals will be excluded. This can be determined through
inspection of the complex covariance function determined in the previous section. For
all measurement locations xi in the measurement set {xn }, nd a set {x} which only
includes measurements within a distance of h0 of the unobserved point x0 (5.21).
{x}(x0 ) = {xi ∈ {xn } : ||x0 − xi || < h0 } (5.21)
2. Using the prediction set, create an estimate of the prediction covariance matrix for
the unobserved point x0 united with the prediction set {x} (5.22). Let row/column
one correspond to the unobserved point (and correspond to an index of 0), and all
other rows and columns correspond to points in the prediction set (and correspond to
an index of 1 through i). Populate a matrix V0 where
0
Vi,j = C(h~i,j ; ~c0 , θ~0 ). Let ~v 0 be
de ned as a row vector where element j is de ned by vj = C(h~0,j ; ~c0 , θ~0 ), the covariance
between point xj in the prediction set and the unobserved point x0 .
0
σ
20
~v00H
Vprime =
(5.22)
~v00 V 0
3. Estimate the residual ε00 at the unobserved point x0 through simple Kriging of the
residuals of the prediction set ~ε0 (5.23).
ε00 = ~v00H V0−1 ε0 (5.23)
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 126
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
4. Use the spatial attribute matrix at the unobserved point ~ 0 , complex spatial attribute
X
weights β~ 0 , and estimated residual at the unobserved point ε00 to estimate the channel
at the unobserved point y00 (5.24).
~ 0 β~ 0 + ε00
y00 = X (5.24)
5. Finally, this analysis can produce a prediction standard error (5.25).
q
σ00 = (σ 20 − ~v00H V0−1~v00 ) + (X0 − X H V0−1~v00 )H (X H V0−1 X)−1 (X0 − X H V0−1~v00 )
(5.25)
The process of choosing and tting Spatial Attribute Functions to measured channel
data, then iteratively re ning the complex coe cients of these functions in conjunction with
developing and iteratively re ning a complex covariance model of the residual, results in a
model that can be used in conjunction with channel measurements and the process of Kriging
to mathematically model the channel at any location in the sampled observation zone. One
limitation of this model is that analysis and modeling is performed at a single frequency.
To be useful in a wideband system, this model must be extended to be applicable to a wide
frequency band.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 127
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
5.2 Adapting Spatial Attribute Function Set for a Wide-
band Model
~
In the previous chapter, plane wave (EP W (r̃) = e−j k·~r ) and point source (EP S (~
r) =
e−jk||~r−~r0 ||
) Spatial Attribute Functions (SAF) were selected for use in our channel model.
r−~
4π||~ r0 ||
Both types of function are complex exponentials, containing in the exponential a coe cient
to the independent variable(s). In the case of plane waves, the dependent variable is a
position vector ~r, while in the case of point sources, the dependent variable is radial distance
r = ||~r −~r0 ||. In the exponential, the coe cient to these dependent variable(s) is wavenumber
for point sources, or wavevector ~
2πf 2πf
k= c
k = c
k̂ , where f is frequency in Hz, c is the velocity
of propagation (speed of light in free space), and k̂ is the unit wavevector.
For our narrowband model, which is based on free-space sampling of the channel at a
speci c frequency f0 given a unit excitation of the system at the same frequency, the f term
in the wavenumber or wavevector expression should be equal to f0 . This should be true even
in re ections of the waveforms from surfaces in the system. However, it should be noted that
the 3D plane wave projections onto the X-Y plane that form the 2D plane wave SAF in our
model will only contain the x̂ and ŷ components of k̂ . We would expect in 3D space, the
wavevector would have the form ~k = 2πf
(kx x̂ + ky ŷ + kz ẑ), where kx , ky , and kz are the x, y ,
c
q
and z components of the unit vector k̂ , with kx2 + ky2 + kz2 = 1. This would mean the unit
vector of the projection would be k̂proj = √ k2x x̂ + √ k2y ŷ , and the wavevector coe cient
kx +ky2 kx +ky2
√
2π kx2 +ky2 f
would become . Therefore the apparent frequency of the 2D plane wave would be
c
q
kx2 + ky2 f , and thus each 2D projected plane wave should contain a constant-scaled version
of the excitation frequency f0 . In short, in each SAF, we have a complex exponential with
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 128
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
f0 or a constant-scaled version of f0 in the exponent.
Channel measurements can be performed at a series of narrowly-spaced frequencies to
observe the change in channel versus frequency. In the case of a closed, static, linear system
consisting of materials that have a small change in properties relative to change in frequency,
channel measurements at closely-spaced frequencies will also have a small change in magni-
tude. This is due to the fact that the geometric paths between transmitter and receiver will
remain static independent of frequency, with variation only occuring in the re ection coe -
cients as a function of frequency in re ecting materials in the system. These variations are
small in the system of interest, as the absorbing material properties are relatively constant
in the frequency range of analysis.
Because channel signal paths are essentially constant with small frequency changes, it
is logical that the resultant channel with a small frequency change is a frequency-shifted
version of the underlying components (which are described by SAF), each with coe cients
that change slightly as a function of frequency. Because the wavenumer or wavevector is
proportional to the frequency of excitation and measurement (k ∝ f0 ), then at a shifted
frequency af0 , where a is a shifting constant of proportionality (with a=1 representing no
shift), a wavenumber or wavevector can also be scaled by a.
Using the idea of frequency-shifted SAF as discussed above, a logical extension is to
determine the SAF at a single frequency f0 using the method described in the previous
chapter, then for a shifted frequency f1 = af0 , re-calculate the spatial attribute functions at
each and every measurement point with the exponent scaled by a, then re-estimate all SAF
weights and other estimated parameters. The following section outlines this procedure in
detail.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 129
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
5.2.1 Algorithm for Fitting Spatial Attribute Function Set
The following is a procedure for using the SAF extracted from a single frequency to
estimate model parameters over a wide frequency range.
1. Select an initial frequency f0 corresponding to a frequency used to sample the channel
over an observation region.
2. Utilize the procedure from the previous chapter to extract a set of plane wave and
point source SAF for the frequency f0 .
3. Use the procedure from the previous sections to iteratively estimate (β~ 0 , ~c0 , σ 20 , θ0 , a0θ , a0φ )
at f0 .
4. Select a set of frequencies {fn } near f0 for which the channel was sampled over the
same observation region.
fn
5. For each fn , determine the scaling constant an = f0
. Utilize an to scale the dependent
term of the exponent of each SAF, and evaluate the SAF at each and every sample
location in the observation region. Populate the spatial attribute matrix Xn for fn ,
using the scaled SAF evaluations, where each row of X represents a measurement
location, and each column represents a SAF.
6. Utilize the procedure from the previous sections to iteratively estimate (β~ 0 , ~c0 , σ 20 , θ0 , a0θ , a0φ )
at each and every fn in {fn }.
7. The end result is a set of each estimated parameter for a set of frequencies and for a
single set of SAF.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 130
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
This procedure yields each parameter at a corresponding set of frequencies. These pa-
rameters can be used to estimate the channel at each measurement point at each frequency,
which can then be compared to the measured channel being modeled. The error vector
magnitude (EVM) can be calculated at each sample point in the observation region, and a
mean EVM can be determined for the model at each frequency. An EVM threshold can be
established, and all frequencies with a mean EVM violating this threshold can be excluded
from the model at the selected f0 (and can be included in a model at a separate frequency
and corresponding set of SAF).
Finally, The set of frequencies for which the model produces an acceptable EVM can be
included in the following step, outlined in the next section.
5.2.2 Modeling Coe cients as Polynomial Functions of Frequency
Regression can be used to t a polynomial to the parameter estimates. The previous
section produced an estimate of each model parameter at each selected measurement fre-
quency. Note that each parameter vector consists of multiple model parameters, each of
which should be analyzed separately versus frequency. Additionally, many model parame-
ters are complex-valued, which will not work with a polynomial t. These should be modeled
as two parameters each, one for magnitude, and one for phase angle. Magnitude and angle
are each real values, and therefore can be modeled as a polynomial. For each parameter
to be modeled, we t a polynomial as a function of frequency using the set of estimated
parameters and measurement frequencies.
Each parameter component estimation from the measurement is plotted versus mea-
surement frequency. A polynomial order is selected to ensure the polynomial follows all
signi cant trends in the component estimation versus frequency.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 131
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
The complete list of parameters to be modeled using this polynomial method is β~mag (f )
= (β0mag (f ), β1mag (f ), · · · , βnmag (f )), β~angle (f ) = (β0angle (f ), β1angle (f ), · · · , βnangle (f )), ~c(f )
= (c1 (f ), ~ )
c2 (f )), θ(f = (σ
2
(f ), aφ (f ), aθ (f ), θ(f )).
5.2.3 Final Wideband Model Description
The nal wideband model using the parameters as a function of frequency is an extension
of (5.24) and (5.23), with V and ~v0 expressed as functions of parameters utilized in their
population. The channel estimate H0 between the transmitter and a receiver at location ~x0
at frequency fn is de ned in the following equation (5.26),
~ x0 , fn )β~mag (fn )ej β~angle (fn ) + ~v 0 (~x0 , ~c(fn ), θ(f
H 0 (~x0 , fn ) = X(~ ~ n ))H V0 (~x0 , ~c(fn ), θ(f
~ n ))−1 ε0
0
(5.26)
where ~ x0 , fn )
X(~ is the spatial attribute vector, with each spatial attribute function evalu-
~
ated at position ~x0 and frequency fn , β~mag (fn )ej βangle (fn ) is the vector of complex SAF co-
e cients β~ evaluated at frequency ~ n ))
fn , ~v00 (~x0 , ~c(fn ), θ(f is a vector of modeled covariances
between admissible measurement points ~xi and location ~x0 with elements de ned by C(~xi −
~ n )), V0 (~x0 , ~c(fn ), θ(f
~x0 ; ~c(fn ), θ(f ~ n )) is a matrix of modeled covariances between each pair of
admissible measurement points ~xi and ~xj with elements de ned by ~ n )),
C(~xj − ~xi ; ~c(fn ), θ(f
and ~ε is a vector of residuals at admissible measurement points ~xi at frequency fn .
This modeling procedure is performed for as many subsections of the frequency range
to be modeled as required to ensure that each modeled section is below a de ned EVM
threshold. Let the number of bands be de ned as NB . It also is performed for each individual
transmit antenna of NA transmit antennas. This method therefore requires NA ×NB separate
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 132
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
models to de ne the channel for a wideband MIMO system of NA transmitters by an arbitrary
number of receivers in the observation zone, over NB frequency bands. The remainder
of this chapter is an experimental validation of this statistical wideband channel model
using an extensive measurement campaign. The following chamber is an extension of this
measurement-based model to MIMO system analysis.
5.3 System Description
The octoBox BOX-38 (Fig. 5.2) by Octoscope, Inc. is a semi-anechoic chamber for over
the air (OTA) testing presents a system where the channel is quasi-near eld due to close
proximity of probe antennas (along upper-right side in the gure) and test zone in the center
of the box. A line of sight (LOS) channel in free-space comparable to the channel between a
log periodic probe antenna and a dipole located in the BOX-38 test zone region in the center
of the chamber was simulated and modeled in the previous chapter with accurate results.
It is therefore logical to utilize similar measurements from this chamber for performing and
validating the channel model extension presented in this chapter.
5.3.1 Chamber Con guration and Properties
The octoBox BOX-38 is a small semi-anechoic chamber. The box acts as a Faraday cage,
and its channel between probe antennas and DUT is shielded from signals exterior to the
box. Isolation is rated at >80 dB when using cable connection lters. The box interior
is lined with an absorbing material that attenuates re ections by approximately 20 dB for
frequencies between 700 MHz and 6 GHz. Some elements including the metallic outlet cover
on the back interior wall (not covered by absorbing foam) may produce rst-order re ection
components to the channel, but in general a dominant line of sight (LOS) signal path is
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 133
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
Figure 5.2: Octoscope octoBox BOX-38 Interior
expected.
A plastic frame in the interior allows for the placement of multiple probe antennas,
typically either dipole or log periodic. These antennas are connected to SMA bulkheads un-
derneath the absorbing foam, allowing for antenna excitation or connection to a transceiver.
A power strip in the rear of the chamber allows ltered mains power to be supplied to a
device under test (DUT) and/or a plastic turn table in the test zone of the chamber, allowing
for rotation of the DUT orientation about the bottom center of the chamber. Finally, two l-
tered Ethernet ports are positioned along the rear of the chamber to allow for communication
with the DUT and/or turntable control.
The interior dimensions of the chamber (in the open region excluding the absorbing foam
within the chamber) are 0.49 m high, 0.81 m wide, and 0.55 m deep.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 134
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
5.3.2 Antenna Properties
A set of four custom-designed Octoscope OBS-14 (Fig. 5.3) broadband high-gain log-
periodic antennas are utilized as probe antennas in the small chamber. These are mounted
on the plastic frame along the upper-right edge of the chamber in an array similar to Fig.
5.4. Each element is attached to a gimbal allowing for orientation and boresight adjustment.
Each log-periodic antenna is 14-element, and matches the antenna parameters utilized in
the log-periodic antenna modeled in the previous chapter. The frequency range of operation
for these antennas is 2 GHz to 6 GHz. Antenna is constructed as a microstrip antenna with
dimensions 0.127 m by 0.02 m. Fig. 5.5 shows the gain pattern of the OBS-14 antenna at 3
GHz and 6 GHz. The speci ed impedance of the OBS-14 antenna is 50 Ω [4].
Figure 5.3: Octoscope OBS-14 Log Periodic Antenna [4]
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 135
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
Figure 5.4: Octoscope OBS-14 Array [4]
Figure 5.5: Octoscope OBS-14 Radiation Patterns [4]
The receiver dipole antenna is a Pasternack PE51083 rubber duck dual band antenna
(Fig. 5.6). This antenna has frequency ranges from 2.4 GHz to 2.5 GHz, and 4.9 GHz to
5.825 GHz. It has a length of 0.108 m, a nominal 2 dBi omnidirectional gain and an input
impedance of 50 Ω [5]. Further inspection of the antenna revealed a half-wave sleeve dipole
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 136
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
construction.
Figure 5.6: Pasternack PE51083 Rubber Duck Antenna [5]
5.3.3 Measurement Equipment
Channel measurements were taken with a Keisight E5063A-2H5 ENA Vector Network
Analyzer (VNA) with ECal automatic calibration set (Fig. 5.7). Each port has an input
impedance of 50 Ω. This VNA allows for 2-port S-parameter measurements over frequencies
from 100 kHz to 18 GHz [6]. Each sweep can be used to sample up to 10001 points.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 137
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
Figure 5.7: Keysight E5063A VNA [6]
5.4 Measurement Campaign Description
A measurement campaign was designed to spatially sample the channel from each of
four stationary Log-Periodic probe antennas located along the top right edge of the chamber
interior, with a single sleeve dipole moved along a grid to spatially sample the channel.
Broadband 2 port S-parameters were captured with the VNA over the entire 5 GHz U-NII-
1 through U-NII-2C frequency bands (a frequency range of 5.15 GHz to 5.725 GHz, the most
widely-used frequencies used in 5 GHz WiFi). 1151 frequencies in this range were sampled,
resulting in a frequency domain spacing of 500 kHz. Calibration of the VNA was performed
using a Keysight E-cal set, and was calibrated to the ends of the SMA cables which connect
directly to the SMA bulkheads on the side of the anechoic chamber. All unused anechoic
chamber ports were terminated with 50Ω calibration standards, thus terminating all unused
antennas with an impedance matching the speci ed antenna input impedance. Therefore,
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 138
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
the channel being measured is from the chamber probe SMA input to the chamber DUT
SMA output. The channel measured includes the properties of the transmit and receive
antennas and their interaction with each other through the environment.
As mentioned in previous sections, because the proximity between transmit and receive
antennas is near the far- eld threshold, interactions between the antennas cannot easily be
de ned in terms of far- eld radiation patterns, isolated, and extracted from the measured
channel data. As a result, the channel measurements and channel model will include the
antennas as part of the model, indicating that this model generated from these measurements
will be speci c to this combination of antennas and this setup.
While the value of such a limited, speci c model may at rst appear to be low, a model
speci c to this con guration is still a valuable model in the context of a small chamber, where
the probe antenna locations in the small chamber are xed to a small number of possible
con gurations, receiver antennas are often dipoles, and the receiver antenna locations will
be xed to within a de ned measurement zone. As a result, this channel model will provide
relevant information about the expected channel and associated capacity and spatial corre-
lation in the measurement region. This information is extremely important in validating the
use of a small chamber as a repeatable testing environment.
Channel measurements can be directly derived from S-parameter measurements using
S21
the relation H = 2
, which assumes a match between source, characteristic, and load
impedances in the system. Assuming the VNA has a source of VS with 50Ω impedance at
port 1, and a load impedance of 50Ω at port 2, and the input impedance of the transmit and
receive antennas are also speci ed as 50Ω, then the input voltage into the system is a voltage
divider between the source impedance (50Ω) and the input impedance to the system (also
VS Vout 2Vout
50Ω), and Vin = 2
. Since Vout = Vload and S21 = Vin
, therefore S21 = VS
. Rearranging
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 139
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
Vout S21
this equation, the channel is determined to be H = VS
= 2
. Thus the channel transfer
function H is taken as one-half the measured S21 for the purpose of analysis.
With the de nition of the parameters to be measured, the reasoning for this speci c
measurement campaign established, and the elements making up the system speci ed in the
previous section, the physical setup of the system elements must be established, as well as a
description of measurement campaign positions.
5.4.1 Antenna Orientations and Positions
Let a coordinate system de ning locations within the chamber be established with origin
in the bottom left back interior corner of the chamber, positive X axis along the bottom
back edge of the chamber, negative y axis along the bottom left edge of the chamber, and
positive Z axis along the back left side edge of the chamber. Coordinates are in meters,
with maximum X and Z and minimum Y values de ned by the dimensions of the chamber
interior, and minimum X and Z and maximum Y values all de ned as equal to zero. This
coordinate system will be used to de ne antenna locations within the chamber.
Each of the four Log-Periodic probe antennas (as speci ed in the previous section)
were attached using gimbals to the plastic support structure along the upper right edge
of the chamber interior. Each of the four probes are indexed as 1 through 4, with probe
center coordinates P1 = (0.7355, −0.4590, 0.3026) (near upper front right corner), P2 =
(0.7644, −0.3374, 0.2935), P3 = (0.7373, −0.1837, 0.2913), and P4 = (0.7538, −0.0680, 0.3009)
(near upper back right corner). Each probe antenna had its boresight aimed toward the bot-
tom center of the chamber. Each probe antenna was oriented along a plane perpendicular
to the bottom of the chamber that intersects both the chamber bottom center point and
the probe antenna center point. All four probe antennas were connected to SMA bulkhead
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 140
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
connectors to the exterior of the chamber via short SMA cables tucked behind the foam
absorber along the chamber wall. Fig. 5.8 shows a diagram of the chamber, including probe
antenna positions in the chamber.
back
P4
P3
front
P2
P1
Test Zone
... DUT Antenna
...
... ...
49 X positions
0.22m
Grid
Pitch
0.0158 m
0.0158 m
Figure 5.8: Chamber and Measurement System Diagram
A single sleeve dipole (as speci ed in the previous section) was used as a receiver. The
dipole was kept in a Z orientation (and thus perpendicular to the chamber oor). The dipole
center was located at Z = 0.153, which was kept constant by a plastic support structure
holding the antenna in place. This plastic structure also allowed for precise positioning
of the antenna along an X-Y grid on the chamber oor, allowing for spatial sampling of
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 141
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
the channel. The dipole DUT antenna is shown in Fig. 5.8, along with the directions of
movement represented by arrows in the X and Y directions. The SMA cable connecting the
dipole to the SMA bulkhead to the chamber exterior was tucked behind the foam absorbing
material to minimize the length of cable exposed in the chamber.
5.4.2 Chosen Measurement Campaign Positions
The entire chamber interior bottom surface was measured in a grid with lines perpendic-
ular to the X and Y axes with a pitch of 0.0158 m. This choice of this pitch is explained
after the description of the measurement campaign. Fig. 5.8 illustrates the grid layout on
the bottom of the chamber, along with the pitch of the grid.
For the measurement campaign, the VNA was calibrated using the E-cal for the desired
frequency range, the two SMA cables from the VNA were attached to the port corresponding
to the probe index and the receiver dipole. All other SMA ports were terminated with 50Ω
calibration standards. The measurement campaign consisted of:
1. Placing the dipole antenna at the rst/next intersection point on the grid.
2. Adjusting the SMA cable to minimize the length exposed beyond the absorbing mate-
rial.
3. Closing the chamber door securely.
4. Initiating the VNA to take a measurement over the frequency range speci ed previously.
5. Saving the S parameter data in a le with the probe index and measurement point x
and y index.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 142
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
This procedure was repeated for each measurement position on the chamber oor grid, and
that entire process was repeated for each of the four probe antennas acting as transmitter. A
total of 49 X positions by 28 Y positions by 4 probes, or 5488 measurements were taken (each
containing 2-port S parameters for 1151 frequencies). (Further measurements were taken at
10,001 frequencies between 1 GHz and 6 GHz (with identical spacing), as well as at 6 probe
locations for a total of 8232 broadband measurements, but the scope of this investigation
was reduced to 4 transmitters in the 5 GHz WiFi frequency range. The additional data set
is for an extension of this work to other probe con gurations and the 2.4 GHz WiFi band in
future work.)
Measurement spacing was chosen to be less than that corresponding to the spatial Nyquist
frequency of the highest frequency content in the system. (Sample spacing in space is com-
parable to sample period in time; the spatial sample frequency is inversely proportional to
the spatial sample spacing just as frequency is inversely proportional to the sample period
in time. Just as temporal signals must be sampled with a spacing less than that correspond-
ing to the Nyquist sampling frequency, a spatially-sampled signal must be sampled with a
spacing less than that corresponding to the spatial Nyquist sampling frequency.) With the
highest frequency content in the system originally being considered as 6 GHz, the spatial
sampling spacing must be less than half the corresponding wavelength of 0.050 m, i.e. less
than 0.025 m. A grid spacing of 5/8 in., or approximately 0.0158 m, su ciently met this
requirement with su cient headroom, and therefore was used.
Despite taking measurements over the entire chamber oor, the scope of analysis was
narrowed to a feasible test zone near the center of the chamber. A range about the center of
the chamber, about 22 cm by 19 cm, was chosen for performing the modeling and analysis.
This is illustrated in Fig. 5.8 as a red box around the chamber center oor. For the remainder
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 143
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
of the analysis, this will be known as the test zone. This size will su ciently contain most
small- to mid-sized device under test (DUT) that would be well-suited to a small chamber
test environment. X values therefore fall in the range 0.2856 to 0.5062, while Y values
fall in the range -0.3247 to -0.1357. The method provided can be used in conjunction with
measurements over a larger region to generate a model for a large test zone and DUT.
5.4.3 Resulting Dataset
Using the results of the measurement campaign, we extracted a sampled channel over
the test zone for each of the four probes and 1151 frequencies at 195 measurement points.
Presenting this entire data set of 4604 narrowband channels is not practical or useful in the
scope of this document, but a representative sampled channel magnitude (Fig. 5.9) and
phase plot (Fig. 5.10) for a single frequency (5.15 GHz) and probe (probe one) is presented
here.
The magnitude plot (Fig. 5.9) shows the channel has a channel magnitude varying be-
tween -46 dB and -36 dB over the test zone. The actual sample points are at the intersection
points of the grid, while the surrounding color between sample points is an interpolation of
the magnitude. As Probe 1 is closest to the bottom-right corner of the plot, points in the
lower-right corner of the plot are closest to the probe while points in the upper-left corner
of the plot are farthest.
As expected, points generally farther from the probe have a lower channel magnitude
while points closer to the probe have a generally higher channel magnitude. The bands of
peaks and troughs in the magnitude over the test zone indicate the presence of construc-
tive and destructive interference, likely caused by at least one additive re ection present
within the chamber. These trends are present for other frequencies in the channel measured
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 144
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
-36
-0.14
-37
-0.16
-38
-0.18
-39
-0.2
Channel Mag. (dB)
Y position (m)
-40
-0.22
-41
-0.24
-42
-0.26
-43
-0.28
-44
-0.3
-45
-0.32
-46
0.3 0.35 0.4 0.45 0.5
X position (m)
Figure 5.9: Measured Channel Magnitude (dB), Probe One (Located in Lower Right), 5.15
GHz
from probe one. The channel for probes two, three and four all exhibit similar magnitude
trends proportional to distance from each probe, but have less pronounced constructive and
destructive interference.
The phase plot (Fig.5.10) exhibits primarily concentric Eikonal surfaces with the center
located near probe one. This is expected as the line of sight (LOS) component of the
channel is dominant, and the LOS component is directly from probe one. Due to the spatial
sampling resolution and the linear interpolation of plotted phase between points, the Eikonal
surfaces plotted between sample points do not accurately illustrate the curved nature of the
surfaces. Also, the interpolation does not properly account for phase wrapping by increasing
or decreasing to the point of discontinuity and wrapping around. (For example, if one
−π π
sampled point of phase is and the following is , and the actual phase process between
2 2
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 145
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
-0.14 3
-0.16
2
-0.18
-0.2 1
Channel Phase (rad.)
Y position (m)
-0.22
0
-0.24
-0.26 -1
-0.28
-2
-0.3
-0.32
-3
0.3 0.35 0.4 0.45 0.5
X position (m)
Figure 5.10: Measured Channel Phase (rad.), Probe One (Located in Lower Right), 5.15
GHz
−π
the two points is a line between
2
, and −π , then a discontinuity to π, and a line from π to
pi −pi pi
, the interpolation line would be directly from and with no wrapped phase. Thus
2 2 2
discontinuities between sample points are not captured in a plot consisting of lling between
points with interpolation. Dierent phase dierences between consecutive points bordering a
discontinuity will exhibit dierent interpolation slopes, and the interpolation region between
points covering a series of discontinuities may appear uneven, as is the case in Fig.5.10).
These issues in interpolation cause the Eikonal surfaces to appear jagged in the plot.
To reiterate, this jagged nature of the phase plot is only a side eect of the method of
displaying data. This inaccuracy in presentation should not be confused with introducing
an inaccuracy to the collected data that will be present in our model, which rely on curve
tting and spectral analysis and not linear interpolation. The general trends of this plot, i.e.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 146
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
curved concentric Eikonal surfaces centered around the probe for which the measurement
was taken, are present in all phase plots from all probes at all frequencies.
5.5 Modeling of Measurement Data
With the channel sampled in frequency and space, the rst part of the modeling process
is to determine the minimum number of each type of SAF (point sources and plane waves)
required to achieve a desired level of accuracy (i.e. su ciently-low EVM across the test
zone). In our analysis, a general observed trend was that more SAF were required for a
given level of EVM at higher frequencies than low frequencies, so selecting the number of
SAF through analysis at the highest frequency of the relevant measurement campaign (5.725
GHz) will provide a su cient number of SAF for all other frequencies in the measurement
campaign.
An investigation into the LN mean and LN SD of the EVM over the test zone was
conducted for all possible combinations of 0 to 6 point sources and 0 to 20 plane waves for
each of the four probes at 5.725 GHz using the procedure from the previous chapter. As in
the analysis presented in the previous chapter, a value of LN mean + 1 SD EVM over the
test zone was computed, and presented for each probe and combination of SAF. All possible
combinations of SAF for a probe are presented in each plot (Fig. 5.11 - Fig. 5.14). A
threshold of -15 dB for LN mean + 1 SD EVM was established, and all SAF combinations
with worse EVM performance were truncated from the plots for easy analysis.
Analysis of these plots show that probes two and three (Fig. 5.12 and Fig. 5.12) require
far more SAF components than probes two and four to achieve a LN mean + 1 SD EVM of
-15 dB. The combination of SAF with a LN mean + 1 SD less than -15 dB for all four probes
with the fewest number of components is three point sources (spherical components) and 19
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 147
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
20 -15
18 -15.5
model mean EVM plus one standard dev. (dB)
16 -16
num. planar components
14 -16.5
12 -17
10 -17.5
8 -18
6 -18.5
4 -19
2 -19.5
0 -20
0 1 2 3 4 5 6
num. spherical components
Figure 5.11: LN Mean + σ EVM, Probe One, 5.725 GHz
20 -15
18 -15.5
model mean EVM plus one standard dev. (dB)
16 -16
num. planar components
14 -16.5
12 -17
10 -17.5
8 -18
6 -18.5
4 -19
2 -19.5
0 -20
0 1 2 3 4 5 6
num. spherical components
Figure 5.12: LN Mean + σ EVM, Probe Two, 5.725 GHz
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 148
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
20 -15
18 -15.5
model mean EVM plus one standard dev. (dB)
16 -16
num. planar components
14 -16.5
12 -17
10 -17.5
8 -18
6 -18.5
4 -19
2 -19.5
0 -20
0 1 2 3 4 5 6
num. spherical components
Figure 5.13: LN Mean + σ EVM, Probe Three, 5.725 GHz
20 -15
18 -15.5
model mean EVM plus one standard dev. (dB)
16 -16
num. planar components
14 -16.5
12 -17
10 -17.5
8 -18
6 -18.5
4 -19
2 -19.5
0 -20
0 1 2 3 4 5 6
num. spherical components
Figure 5.14: LN Mean + σ EVM, Probe Four, 5.725 GHz
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 149
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
plane wave components (planar components). It is this combination of 3 point sources and
19 plane waves that will be used for the remainder of the channel modeling at all frequencies
for all probes, as this combination is of su cient accuracy for even the worst-case probe and
frequency.
5.5.1 Analysis of Complex Geostatistical Covariance Of Measure-
ments
After determining the number and type of SAF used, the measurement data for a speci c
frequency and probe can be used in conjunction with the procedure in the previous chapter
to iteratively determine rst the 3 point sources and then the 19 plane waves. These 22 SAF
evaluated at each measurement point in conjunction with a constant term can be t to the
measured data using least squares to obtain an initial estimate for the SAF coe cients, β~ .
This is then used to obtain an estimate of the residual, ~ε. This residual is modeled statistically
as in Sec. 5.1.3. This estimated residual model is then used to re ne the complex coe cients
of the model, which is then used to re ne the residual model, and so forth until convergence
is obtained.
For the purpose of investigating the properties of the empirical and modeled spatial
covariance, the magnitude, real, and imaginary components of the empirically calculated
channel covariance, empirically calculated residual covariance, and model covariance will be
presented for the channel at a single frequency and probe.
First, the magnitude (Fig. 5.15), real (Fig. 5.16) and imaginary (Fig. 5.17) components
of the spatial covariance of the measured channel, and the corresponding spatial power (Fig.
5.18) spectrum are observed for probe one at a frequency of 5.15 GHz. This information is
not used in the channel model, but is useful in comparing characteristics with the empirical
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 150
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
and modeled residual covariance. Only separations for which signi cant signal overlap is
achieved are analyzed, producing an oval-shaped region for which the spatial covariance is
calculated.
The measured channel has a high magnitude of spatial cross-covariance along the general
direction of propagation in the test zone, with the magnitude spreading in direction as
separation distance increases (Fig. 5.15). The real and imaginary plots (Fig. 5.16) and
5.17)) have well-de ned peaks and troughs perpendicular to the expected direction of travel.
Finally, taking the spatial FFT of the windowed complex spatial cross-correlation (Fig. 5.18)
we observe a strong spectral peak with slight spread, at the values of wave vector indicating
wave propagation in the direction away from probe one, as expected. The remainder of the
spectrum is generally 20 dB lower or less than the peak region, and without any signi cant
peaks or noticeable structure. This would indicate that channel components beyond the
LOS component will be signi cantly lower in magnitude.
The estimated empirical residual spatial cross-covariance is observed next, resulting from
subtracting the weighted SAF from the measured channel. The magnitude of the empiri-
cal residual spatial cross-covariance as shown in in Fig.5.19 shows a small region of cross-
correlation at very short separation lengths, and low-level cross-covariance at all other sepa-
rations. Occasional small peaks occur at some large separations (e.g. near a separation vector
of (0.2, 0.03)), but relative to small separations, there are comparatively fewer data points
used for the expected values in these calculations, and therefore the estimated expectation
may deviate from the actual mean.
The real part of the estimated empirical residual spatial cross-covariance as shown in Fig.
5.20 exhibits a peak at zero separation, and rapidly drops o to near-zero cross-covariance
for all other separations. There is a faint low-level pattern in the cross-covariance of low-level
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 151
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
#10 -4
1
y component of separation vector 0.15 0.9
0.1 0.8
Channel spatial cross-cov. mag.
0.7
0.05
0.6
0 0.5
0.4
-0.05
0.3
-0.1
0.2
-0.15 0.1
0
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
x component of separation vector
Figure 5.15: Measured Channel Spatial X-Cov., Mag., Probe One, 5.15 GHz
#10 -4
1
0.15
y component of separation vector
0.1
Channel Real spatial cross-cov.
0.05
0 0
-0.05
-0.1
-0.15
-1
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
x component of separation vector
Figure 5.16: Measured Channel Spatial X-Cov., Real, Probe One, 5.15 GHz
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 152
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
#10 -5
8
0.15
6
y component of separation vector
0.1
Channel Imag. spatial cross-cov.
4
0.05
2
0 0
-2
-0.05
-4
-0.1
-6
-0.15
-8
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
x component of separation vector
Figure 5.17: Measured Channel Spatial X-Cov., Imag., Probe One, 5.15 GHz
-20
30
-30
20
y component of wave vector
-40
Channel spectral mag. (dB)
10
-50
0 -60
-70
-10
-80
-20
-90
-30
-100
-30 -20 -10 0 10 20 30
x component of wave vector
Figure 5.18: Measured Channel Spatial Power Spectrum, Probe One, 5.15 GHz
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 153
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
#10 -6
y component of separation vector 0.15 2
0.1
resid. spatial cross-corr. mag.
1.5
0.05
0
1
-0.05
-0.1 0.5
-0.15
0
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
x component of separation vector
Figure 5.19: Residual Emp. Spatial X-Cov., Mag., Probe One, 5.15 GHz
peaks and troughs similar in direction to those observed in the real and imaginary cross-
covariance of the channel itself, indicating that a very small part of the deterministic part of
the channel remains in the residual that was not modeled by the SAF. For all purposes the
magnitude of this pattern is signi cantly less than the peak, and therefore can be neglected
from consideration in choosing a model cross-covariance function.
The imaginary part of the estimated empirical residual spatial cross-covariance as shown
in Fig. 5.21 exhibits positive and negative peaks at separation vectors of approximately (-
0.01,0) and (0.01,0) respectively. In conjunction with the real peak at zero separation, these
individual imaginary positive and negative peaks are included in the Gaussian complex
covariance model selected previously in Sec. 5.1.3, indicating this model was a good choice
for modeling the residual present in our SAF- tted channel. It should be noted that the
imaginary part of the empirical residual spatial cross-covariance has a similar faint pattern
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 154
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
#10 -6
3
y component of separation vector 0.15
2
0.1
resid. Real spatial cross-corr.
0.05 1
0 0
-0.05 -1
-0.1
-2
-0.15
-3
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
x component of separation vector
Figure 5.20: Residual Emp. Spatial X-Cov., Real, Probe One, 5.15 GHz
with low-level peaks and troughs in directions similar to the real and imaginary channel
spatial cross-covariance plots. Also, the magnitude of the positive and negative imaginary
peaks is lower than the magnitude of the real peak, and is also closer in level to the low-level
pattern, which will have an eect on the tting of the model to the data.
Next, for comparison (and not for use in the model itself ), the spatial power spectrum of
the empirical estimated residual is observed (Fig. 5.22), scaled identically to the spectrum
of the channel (i.e. -100 dB to -20 dB). An approximately white spectrum with relative
uniformity across all wave vectors and no well-de ned peaks shows that the residual is noise-
like and modeling with a cross-covariance model would be practical. Indeed, the spectrum
is relatively uniform, particularly compared to the channel spectrum (Fig. 5.18) which has
well-de ned peak region. The residual spectrum does have more power in the negative X half
space, and does have a couple small peaks near the wave vector corresponding to the peak in
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 155
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
#10 -6
1.5
0.15
1
y component of separation vector
0.1
resid. Imag. spatial cross-corr.
0.5
0.05
0 0
-0.05
-0.5
-0.1
-1
-0.15
-1.5
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2
x component of separation vector
Figure 5.21: Residual Emp. Spatial X-Cov., Imag., Probe One, 5.15 GHz
the channel spectrum, which explains the faint patterns in the real and imaginary residual
cross-correlation plots reminiscent of the channel cross-correlation real and imaginary plots.
It is possible that adding additional plane waves to the SAF set could account for these
peaks and further whiten the residual spectrum, but this would add unnecessary complexity
to the model for the accuracy we wish to achieve.
With the empirical residual spatial cross-covariance calculated for the channel residual,
the procedure from Sec. 5.1.3 can be applied to t the empirical data to a complex Gaussian
covariance model. This tting process yields the covariance at zero distance (directly taken
as the empirical cross-covariance at zero separation), an angle of maximum range (θ ), the
maximum range coe cient aθ , and minimum range coe cient aφ , as well as the translation
vector ~c. These parameters minimize the dierence between the model and the empirical
cross-covariance through the use of a global search algorithm. This procedure resulted in
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 156
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
-20
30
-30
20
y component of wave vector
-40
residual spectral mag. (dB)
10
-50
0 -60
-70
-10
-80
-20
-90
-30
-100
-30 -20 -10 0 10 20 30
x component of wave vector
Figure 5.22: Residual Emp. Spatial Power Spectrum, Probe One, 5.15 GHz
the cross-covariance model yielding Figs. 5.23, 5.24, and 5.25.
Comparing the model cross-covariance magnitude (Fig. 5.23) to empirical cross-covariance
magnitude (Fig. 5.19), we notice that the model has equivalent zero-separation magnitude
to the empirical, and has a similar rate of drop-o in cross-correlation as separation distance
increases. While the empirical separation sampling rate doesn't provide enough resolution
to obtain the maximum and minimum ranges as well as the angle of maximum range by
visual inspection with certainty, it does appear that the angle of maximum range could be
approximately 0 radians as the width of the peak region appears longer along the X axis
than in any other direction. In the model, the angle of maximum range appears to be in the
direction of the LOS signal, perhaps in uenced by the small residual pattern in the direction
of the LOS signal observed in the real and imaginary LOS plots. With the exception of the
angle of maximum range diering slightly (and for reasons that are expected), the model is
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 157
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
reasonably close to the observed empirical magnitude plot.
Figure 5.23: Residual Model Spatial X-Cov., Mag., Probe One, 5.15 GHz
Comparing the model to empirical real cross-covariance plots (Figs. 5.24 and 5.20 respec-
tively), the real component appears to have identical real cross-covariance at zero separation
between the two plots, and both plots rapidly decrease in magnitude to zero in all separa-
tion directions. Once again, due to limited spatial resolution, the shape and orientation of
the empirical cross-covariance peak is not observable by inspection, but the maximum and
minimum range of the model are within a similar short separation range as in the empirical
plot, and thus once again the model is reasonably close to the observed empirical real plot.
Finally, comparing the model to empirical imaginary cross-covariance plots, as shown
in Figs. 5.25 and 5.21 respectively, the separation of the positive and negative peaks near
zero-separation are similar, and the region around these peaks is zero cross-correlation in
both plots. The magnitudes of the two empirical peaks are approximately two orders of
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 158
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
Figure 5.24: Residual Model Spatial X-Cov., Real, Probe One, 5.15 GHz
magnitude higher than the corresponding model peaks, and the model has the axis along
which the two peaks are located in the direction of the small residual pattern (the LOS
direction) instead of along the X axis as in the empirical plot. The direction of the axis
of the peaks once again is likely due to the residual pattern, however the deceased peak
magnitudes are possibly due to either the global search algorithm becoming consistently stuck
in a local minimum when searching for parameters that minimized the dierence between
the model and empirical imaginary components, or the residual signal beyond the two peaks,
which is only slightly lower in magnitude than the two peaks, signi cantly diered from zero
and had a disproportionate eect on the calculation of the imaginary component. Further
investigation of this is necessary, but through experimentation, this global optimization t
model was su ciently accurate for our model and was not pursued further.
With SAF chosen and a su ciently-accurate method for modeling the residual cross-
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 159
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
Figure 5.25: Residual Model Spatial X-Cov., Imag., Probe One, 5.15 GHz
covariance in terms of parameters, Geostatistical Regression (Sec. 5.1.4) is applied to itera-
tively re ne the estimates of the SAF complex coe cients and the residual cross-covariance
model parameters until all coe cients and parameters converge. This collection of SAF,
coe cients, and parameters allow for channel estimation using Geostatistical Kriging (Sec.
5.1.5) at any location in the test zone, for the probe and frequency combination for which
the model was generated.
As an aside, model estimations at measurement points will estimate the residual as the
actual measurement residual, while estimations between measurement points will estimate
the residual based on the spatial cross-covariance model in conjunction with measurement
residuals at nearby points. In a sense, the method of Kriging is statistically-informed inter-
polation. Because model estimates at measurement points are almost exactly equal to the
measurement, the error vector is approximately zero, and EVM at the measurement points
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 160
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
trends towards negative in nity. As such, in subsequent sections, the EVM of the estimated
channel formed by the product of the SAF and complex coe cients β~ after the Geostatistical
Regression process at all measurement points will be used to evaluate model accuracy.
Instead of running this entire procedure at each probe and frequency combination, the
method of determining a set of SAF at a base frequency and re-using a frequency-scaled
version of this set at a wide range of frequencies (as presented in Sec. 5.2) was investigated.
5.5.2 Selected Wideband Frequency Bands for Modeling
A method of applying a set of SAF at an arbitrary base frequency f0 to a band of near
frequencies was speci ed in Sec. 5.2. To eectively apply this method to our dataset, it
was important to determine what bandwidth above or below the base frequency for which
a set of SAF could be re-used while still maintaining acceptable accuracy. We chose various
arbitrary base frequencies for all probes within our dataset, established a set of SAF at
each base frequency, and then tted frequency-scaled versions of the SAF to the channel
measurements at each frequency within 100 MHz of the base frequency. While frequencies
above the base frequency exhibited a rapid increase in EVM, frequencies below each base
frequency generally exhibited a small increase of EVM and in many cases a decrease of EVM.
Further investigation determined that a set of SAF could be used up to 160 MHz below the
base frequency in all cases before a signi cant increase in EVM was observed. As such, we
chose 160 MHz bands to model using SAF calculated at a base frequency at the end of each
band. For each probe, the measurement set was divided into bands from 5.15 GHz to 5.31
GHz (with base frequency 5.31 GHz), 5.31 GHz to 5.47 GHz (with base frequency 5.47 GHz),
5.47 GHz to 5.63 GHz (with base frequency 5.63 GHz), and a nal shorter band from 5.63
Ghz to 5.725 GHz (with base frequency 5.725 GHz).
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 161
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
5.5.3 Performance Analysis of Frequency-Scaled Spatial Attribute
Functions
Using the base frequencies and frequency bands speci ed, the SAF of each base frequency
were scaled to each measurement frequency in the band and used to calculate coe cients and
model parameters using Geostatistical Regression (Sec. 5.1.4), using the procedure speci ed
in Sec. 5.2.1. The measured signal at each frequency in a band, scaled SAF and calculated
coe cients at each frequency were used to calculate the residual and EVM at each frequency.
Because showing a plot of the LN mean EVM of each combination of band and probe (a
total of 16 plots) would be super uous, two such selected plots will be included here.
Figs. 5.26 and 5.27 show two representative probe and band combinations, for probe
one band four, and probe four band three respectively. Probe 1 band 4 has the worst
representative increase in EVM as frequency decreases from the base frequency (still less
than a 2 dB increase). It should be noted that frequencies in this band below 5.63 GHz
are unused as they overlap with band three, and therefore the 5.56 to 5.63 GHz range with
increased EVM does not adversely aect the model accuracy. Probe four band three shows
a general decrease in LN mean EVM as frequency decreases, showing that using the scaled
SAF at these frequencies is even more accurate than their use at the base frequency of this
band!
All other bands have performance between these two relative extremes, and therefore
the use of scaled SAF from a base frequency over a lower band of 160 MHz does not have
a signi cant detriment to model accuracy, while saving signi cant computational cost of
iteratively determining a separate set of SAF at each frequency.
This process has empirically generated all complex SAF coe cients and model param-
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 162
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
-15
-15.5
-16
-16.5
LN Mean EVM (dB)
-17
-17.5
-18
-18.5
-19
-19.5
-20
5.58 5.6 5.62 5.64 5.66 5.68 5.7 5.72
freq (GHz)
Figure 5.26: Empirical LN Mean EVM vs. Freq., Probe One Band Four
-15
-15.5
-16
-16.5
LN Mean EVM (dB)
-17
-17.5
-18
-18.5
-19
-19.5
-20
5.48 5.5 5.52 5.54 5.56 5.58 5.6 5.62
freq (GHz)
Figure 5.27: Empirical LN Mean EVM vs. Freq., Probe Four Band Three
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 163
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
eters at each frequency using scaled SAF. These empirical parameters can be modeled as
polynomial functions of frequency. This process is described and evaluated in the following
section.
5.5.4 Spatial Attribute Function Set Extension to Wideband Model
The procedure in Sec. 5.2.2 (polynomial regression of each parameter versus frequency)
is applied to the magnitude and phase of each of the 23 coe cients in β~ , the values of the
three parameters in θ~, to σ2, and the values of the two parameters in ~c. This is performed
for each probe and each band, for a total of 52 regressions by 16 bands.
The degree of each polynomial parameter regression was selected to closely following
the trends of the modeled parameter over all instances of that parameter for all bands and
probes. For the magnitude and phase models of each value of β~ , a 13th -order polynomial and
a 2nd-order polynomial respectively are t to the magnitude and phase of each value of beta
over the frequencies of each band and probe. For each value of σ 2 , θ~, and ~c, a 13th-order
polynomial model was also t. In the case of the phase of β~ , every plot was observed to
be primarily linear over frequency, so a low-order polynomial was su cient to describe this
parameter. All other parameters occasionally had characteristics that required a higher-
order polynomial to accurately follow the trend of the parameter over all frequencies in a
band.
Once again, due to the large data set, A representative selection of empirical and modeled
parameters versus frequency is presented. Magnitude and phase of a point source coe cient
β1 for probe one band two as in Figs. 5.28 and 5.29 respectively, magnitude and phase of
a plane wave coe cient β4 for probe two band one as in Figs. 5.30 and 5.31 respectively,
values of σ2 for probe two band four as in Fig. 5.32, and values of c1 for probe three band
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 164
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
two as in Fig 5.33.
# 10-3
2.5
Modeled -1 mag.
2.45 Empirical -1 mag.
2.4
- 1 magnitude
2.35
2.3
2.25
2.2
2.15
5.32 5.34 5.36 5.38 5.4 5.42 5.44 5.46
freq (GHz)
Figure 5.28: Emp. and Poly. Model of β1 Mag. vs. Freq., Probe One Band Two
In general, the empirical parameters being modeled range from essentially noise-free
(phase plots in Figs. 5.29 and 5.31) and very low noise (magnitude of β1 in Fig. 5.28) to
relatively constant variance noise following a clear trend (c1 in Fig. 5.33). In all cases, the
polynomial model closely follows the trend of the parameter being modeled. The statistics
of the residual of this polynomial t could be used to determine the error of each model, but
this information is super uous in the context of the goal of this work. Evaluation of this
polynomial model method will be determined empirically through its eect on EVM relative
to the EVM from the use of empirical model parameters.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 165
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
4
Modeled -1 phase
3 Empirical -1 phase
1
- 1 phase
-1
-2
-3
-4
5.32 5.34 5.36 5.38 5.4 5.42 5.44 5.46
freq (GHz)
Figure 5.29: Emp. and Poly. Model of β1 Phase vs. Freq., Probe One Band Two
# 10-4
7.5
Modeled -4 mag.
Empirical -4 mag.
7
- 4 magnitude
6.5
5.5
5
5.15 5.2 5.25 5.3
freq (GHz)
Figure 5.30: Emp. and Poly. Model of β4 Mag. vs. Freq., Probe Two Band One
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 166
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
4
Modeled -4 phase
3 Empirical -4 phase
1
- 4 phase
-1
-2
-3
-4
5.15 5.2 5.25 5.3
freq (GHz)
Figure 5.31: Emp. and Poly. Model of β4 PHase vs. Freq., Probe Two Band One
# 10-6
2.9
Modeled <2
2.8 Empirical <2
2.7
2.6
<2
2.5
2.4
2.3
2.2
2.1
5.58 5.6 5.62 5.64 5.66 5.68 5.7 5.72
freq (GHz)
Figure 5.32: Emp. and Poly. Model of σ2 vs. Freq., Probe Two Band Four
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 167
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
3.5
Modeled c 1
3
Empirical c1
2.5
1.5
c1
0.5
-0.5
-1
-1.5
5.32 5.34 5.36 5.38 5.4 5.42 5.44 5.46
freq (GHz)
Figure 5.33: Emp. and Poly. Model of c1 vs. Freq., Probe Three Band Two
5.5.5 Wideband Model Accuracy
Polynomial model β~ values used in conjunction with frequency-scaled SAF were used to
estimate the channel at all measurement points for all frequencies in each band, and the LN
mean EVM was calculated. Two representative probe and band EVM plots are included
here, Probe one band four as in Fig. 5.34 and Probe four band three as in Fig. 5.35. These
were selected to match the plots of EVM calculated using empirical values of β~ as in Figs.
5.26 and 5.27, for direct comparison so that the eect of parameter polynomial modeling on
EVM can be observed.
The EVM observed in Fig. 5.34 follows the same trend as Fig. 5.26, with small increases
in EVM at around 5.61 GHz, 5.66 GHz, and 5.69 GHz. Likewise, Fig. 5.35 follows the same
trend as Fig. 5.27 with small increases in EVM at about 5.51 GHz, 5.59 GHz, and 5.61 GHz.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 168
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
-15
-15.5
-16
-16.5
LN Mean EVM (dB)
-17
-17.5
-18
-18.5
-19
-19.5
-20
5.58 5.6 5.62 5.64 5.66 5.68 5.7 5.72
freq (GHz)
Figure 5.34: Modeled LN Mean EVM vs. Freq., Probe One Band Four
-15
-15.5
-16
-16.5
LN Mean EVM (dB)
-17
-17.5
-18
-18.5
-19
-19.5
-20
5.48 5.5 5.52 5.54 5.56 5.58 5.6 5.62
freq (GHz)
Figure 5.35: Modeled LN Mean EVM vs. Freq., Probe Four Band Three
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 169
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
The following plots show the calculated dierence at each frequency between the Em-
pirical and Polynomial parameter model LN mean EVM. Fig 5.37 shows the dierences for
probe one band four, and Fig 5.36 shows the dierences for probe four, band three. In both
cases, the increase in LN mean EVM using the polynomial parameter models is between 0
and 0.25 dB, with two to three small frequency ranges spiking up to less than 0.4 or 0.5 dB
increase in EVM. No other LN mean EVM dierence plot for any other probe or band was
observed to have a dierence of greater than 0.5 dB, and in general much lower.
0.5
LN Mean EVM Change, Emp. to Poly. (dB)
0.4
0.3
0.2
0.1
-0.1
5.58 5.6 5.62 5.64 5.66 5.68 5.7 5.72
freq (GHz)
Figure 5.36: Increase of LN Mean EVM From Emp. to Model vs. Freq., Probe 1 Band 4
This result indicates that the use of a polynomial parameter model has an almost negli-
gible eect on model accuracy, and can be used in place of the empirical parameters at each
frequency. This completes the generation and validation of a channel model for the Octobox
small anechoic chamber.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 170
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
0.5
LN Mean EVM Change, Emp. to Poly. (dB) 0.4
0.3
0.2
0.1
-0.1
5.48 5.5 5.52 5.54 5.56 5.58 5.6 5.62
freq (GHz)
Figure 5.37: Increase of LN Mean EVM From Emp. to Model vs. Freq., Probe 4 Band 3
5.6 Conclusions
A complete channel model in terms of receiver position, probe number, and frequency
is the result of the application of the methods contained within this chapter to a set of
broadband spatial channel measurements. The frequency and probe is mapped to a modeled
band, a set of spatial attribute functions for that band is scaled to the selected frequency
and evaluated at the receiver position, and then multiplied by coe cients calculated as a
function of frequency. This description of the channel can be evaluated at any position in
the modeled test zone, at any frequency in the range of analysis. In addition, the residual
of nearby measurement points can be used in conjunction with a statistical model with
parameters calculated as a function of frequency to estimate the residual at the receiver
position, further re ning the channel estimate.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 171
CHAPTER 5. STATISTICAL AND FREQUENCY EXTENSIONS OF NEAR-FIELD
CHANNEL MODEL
While this model is limited to vertically-polarized dipoles along a speci c plane in the
test zone, it could be easily extended to include three-dimensional sampling of a test zone,
orthogonal receiver dipole polarizations, additional probe positions, etc. Each added item
will add an extra dimension to the model space and will require exponentially more measure-
ment points. Without a multi-port VNA, an automated receiver positioning system, and
signi cant computing power for data processing, such additional complexity may make this
added complexity infeasible.
In the next chapter, this model is used in the generation of a MIMO channel with receiver
antennas at arbitrary positions in the test zone, using the channels modeled from each of
the four probes, over the entire 5 GHz WiFi frequency range. An evaluation of capacity
and spatial correlation of arbitrary DUT sizes and con gurations can be performed over the
entire test zone to predict DUT performance within the chamber.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 172
CHAPTER 6
Application of Small Chamber Model to
MIMO Analysis
The purpose of this chapter is to use the channel model developed and tested in the
previous two chapters, speci cally the model developed using the small anechoic chamber
measurement set, in an analysis of expected MIMO performance in the small chamber chan-
nel. The contribution of this chapter is to show how our channel model can be used in
standard MIMO analysis. This chapter revisits the channel magnitude and phase in ad-
dition to illustrating the broadband Channel Impuslse Response (CIR) (Sec. 6.1), then
concisely applies this model to investigate spatial correlation in the test zone for each of the
four separate probe channels (Sec. 6.2). Also, 4×4 MIMO wideband capacity is investigated
over the small chamber test zone as a function of DUT center position using 4 antenna DUTs.
This analysis is performed for several DUT antenna spacings (Sec. 6.3).
6.1 Small Anechoic Chamber Channel and Impulse Re-
sponse
In a small anechoic chamber, such as the one utilized for a measurement campaign as
in the previous chapter, each probe antenna and receiver antenna pair forms a broadband
SISO channel. This channel, as modeled in the previous chapter, is a function of frequency
and receiver position, H(f, x, y). Figs. 6.1 and 6.2 illustrate a realization of the model for
173
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
probe one at a xed f = 5.15 GHz (magnitude and phase respectively), with more densely
sampled X and Y positions in the test zone than in the measurement campaign.
-0.14
-37
-0.16
-38
-0.18
-39
receiver Y position (m)
Channel Model Mag. (dB)
-0.2 -40
-0.22 -41
-0.24 -42
-0.26 -43
-0.28 -44
-0.3 -45
-0.32 -46
0.3 0.35 0.4 0.45 0.5
receiver X position (m)
Figure 6.1: Probe One (Lower Right) Channel Model Mag. Over Test Zone
In addition to observing the model for a xed frequency and observing magnitude and
phase over X and Y positions, the channel can be observed for a xed (X, Y ) position,
varying frequency. This frequency response for a xed (X, Y ) can be considered a transfer
function H(f ). This transfer function at (X, Y ) can then be transformed from the frequency
domain to the time domain to obtain a Channel Impulse Response (CIR), h(t). If the transfer
function H(f ) is evenly sampled with N complex samples over the bandwidth of the channel
model (in our model from 5.15 GHz to 5.725 GHz, a bandwidth of 575 MHz), then the
sampled frequency response is converted to the CIR using the following procedure from [58]:
1. Discard the nal frequency domain sample in the sampled bandwidth B , leaving N − 1
B
samples from fmin to fmax − N −1
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 174
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
-0.14 3
-0.16
2
-0.18
Channel Model Phase (rad.)
receiver Y position (m)
-0.2 1
-0.22
0
-0.24
-0.26 -1
-0.28
-2
-0.3
-0.32 -3
0.3 0.35 0.4 0.45 0.5
receiver X position (m)
Figure 6.2: Probe One (Lower Right) Channel Model Phase Over Test Zone
2. Perform an IFFT shift to rearrange the N −1 frequency domain samples to prepare
the data for an inverse fast Fourier transform (IFFT).
3. Perform an IFFT on the shifted frequency domain samples.
4. Perform a phase shift in the time domain by ej2πfc , where fc is the center frequency
B
of the sampled frequency band (fc = fmin + ). This is the equivalent to shifting the
2
passband spectrum to the baseband spectrum. The result is the CIR.
1
5. Calculate the time domain sample time, tR = B
.
In the end, this entire procedure is encapsulated in a single expression for the CIR
h(t) sampled at multiples of tR , h(ntR ) = it(itshift [H (1 : (N − 1))]) · e−j2πfc [58]. The
command it performs an inverse fast Fourier transform, while itshift shifts the order of the
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 175
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
rst and second halves of the data array to present the frequency response in the expected
order to the it procedure.
The CIR magnitude can be converted to dB, and normalized such that the peak of
the CIR is equal to 0dB. A CIR that decays gradually is indicative of a multipath-rich
environment. Smaller peaks after the initial peak can indicate strong multipath clusters
that have a delayed arrival relative to a line of sight (LOS) component (or relative to other
clusters). Figs. 6.3 - 6.6 illustrate the CIR generated from the channel model for a receiver
at the chamber center for each of the four probes.
-10
-20
CIR power (dB)
-30
-40
-50
-60
0 0.05 0.1 0.15 0.2 0.25 0.3
time ( 7 s)
Figure 6.3: Probe One Channel Impulse Response at Chamber Center
All four CIR plots indicate a sharp peak with rapid decay, with very subtle features
shortly after the peak at a level of about -15 to -28 dB, potentially indicating a couple of
highly-attenuated multipath components. As the absorbing material is rated to attenuate
re ections by about -20 dB, and all multipath distances are extremely short and thus would
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 176
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
-10
-20
CIR power (dB)
-30
-40
-50
-60
0 0.05 0.1 0.15 0.2 0.25 0.3
time ( 7 s)
Figure 6.4: Probe Two Channel Impulse Response at Chamber Center
-10
-20
CIR power (dB)
-30
-40
-50
-60
0 0.05 0.1 0.15 0.2 0.25 0.3
time ( 7 s)
Figure 6.5: Probe Three Channel Impulse Response at Chamber Center
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 177
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
-10
-20
CIR power (dB)
-30
-40
-50
-60
0 0.05 0.1 0.15 0.2 0.25 0.3
time ( 7 s)
Figure 6.6: Probe Four Channel Impulse Response at Chamber Center
have travel time nearly as short as the LOS component, the observed features are as expected.
The CIR can be used to estimate an important characteristic of MIMO channel perfor-
mance, spatial correlation, which is explained in the following section.
6.2 Small Anechoic Chamber Spatial Correlation Analy-
sis
An analysis of spatial correlation is an important tool in understanding the nature and
expected performance of a MIMO channel. As spatial correlation increases between elements
at the transmitter and/or receiver of a MIMO system, the rank of the MIMO channel matrix
is reduced (as elements in channel matrix rows and/or columns become correlated and have
less variation), and the capacity of the channel decreases.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 178
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
In the literature, it is common to utilize a sampled or simulated CIR to obtain spatial
correlation. Ka e et al. specify an equation (6.1) to calculate receiver wideband spatial cor-
RX
relation ri,j given the impulse responses at two receiver locations (i and j ) from transmitter
m, notated him and hjm [59].
RX E[him h∗jm ] − E[him ]E[h∗jm ]
ri,j =q (6.1)
(E[|him |2 ] − |E[him ]|2 )(E[|hjm |2 ] − |E[hjm ]|2 )
In this equation E is the expectation operator over CIR delay time, and ∗ is the complex
cojugation operator.
To observe the eect of spatial correlation over the entire test zone, spatial correlation
was investigated for 31 separation distances ranging from 0 to 4λ. For each separation
distance, 30 random pairs of points in the test zone separated by the separation distance
(with arbitrary uniformly-random orientation) were selected. At each pair of points, the pair
of CIR were evaluated using the procedure described in the previous section. The spatial
correlation between the pair of points was calculated using (6.1). The absolute spatial
correlation was averaged for all 30 pairs of points at each separation distance. This yielded
an empirical estimate of spatial correlation versus spatial separation. This procedure was
repeated for the channels of each of the four probes. Figs. 6.7 - 6.10 show the empirical
spatial correlation versus spatial separation for each of channels from the four probes. A 5th
order polynomial was t to each of the empirical spatial correlation data sets, and are also
included in these plots.
All channels have very similar spatial correlation versus separation, with spatial corre-
lation slowly decreasing from 1 to about 0.7 or 0.8 at a separation of 4λ. This behavior is
typical of a LOS channel and is similar to the receiver spatial correlation plot for an indoor
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 179
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
0.9
0.8
0.7
spatial correlation
0.6
0.5
0.4
0.3
0.2
Empirical spatial corr.
0.1
Poly. fit of spatial corr.
0
0 0.5 1 1.5 2 2.5 3 3.5 4
spatial separation ( 6)
Figure 6.7: Probe One, Empirical and Polynomial Fit Spatial Correlation vs. Separation
0.9
0.8
0.7
spatial correlation
0.6
0.5
0.4
0.3
0.2
Empirical spatial corr.
0.1
Poly. fit of spatial corr.
0
0 0.5 1 1.5 2 2.5 3 3.5 4
spatial separation ( 6)
Figure 6.8: Probe Two, Empirical and Polynomial Fit Spatial Correlation vs. Separation
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 180
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
0.9
0.8
0.7
spatial correlation
0.6
0.5
0.4
0.3
0.2
Empirical spatial corr.
0.1
Poly. fit of spatial corr.
0
0 0.5 1 1.5 2 2.5 3 3.5 4
spatial separation ( 6)
Figure 6.9: Probe Three, Empirical and Polynomial Fit Spatial Correlation vs. Separation
0.9
0.8
0.7
spatial correlation
0.6
0.5
0.4
0.3
0.2
Empirical spatial corr.
0.1
Poly. fit of spatial corr.
0
0 0.5 1 1.5 2 2.5 3 3.5 4
spatial separation ( 6)
Figure 6.10: Probe Four, Empirical and Polynomial Fit Spatial Correlation vs. Separation
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 181
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
LOS environment in [59], which also decreases from 1 to just below 0.8 between 0 and 4λ.
This would indicate that the channel inside the small anechoic chamber behaves like an LOS
dominant channel.
Also of interest, is the distribution of spatial correlation for xed separations. The empir-
ical spatial correlation was calculated as a mean of 30 random location pairs in the chamber.
Rather than averaging the 30 spatial correlations calculated for a set of constant separation,
the distribution of those 30 spatial correlation values can be plotted as a CCDF for each
λ
probe and a selected set of xed separations. Figs. 6.11 - 6.14 each show the CCDF for ,
4
λ 3λ
2
, λ and
2
for the channel corresponding to a probe.
0.9
0.8
P(spatial corr. > abscissa)
0.7
0.6
0.5
0.4
0.3
0.2 0.25 6
0.5 6
6
0.1
1.5 6
0
0.7 0.75 0.8 0.85 0.9 0.95 1
spatial correlation
Figure 6.11: Probe One, Empirical Spatial Correlation CCDFs vs. Separation
λ
In all four plots, the CCDF for is the right-most CCDF curve. For probes two through
4
four, one hundred percent of the observed spatial correlations were above 0.95, and for all
four probes, one hundred percent of the spatial correlations were above 0.925. For probe
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 182
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
0.9
0.8
P(spatial corr. > abscissa)
0.7
0.6
0.5
0.4
0.3
0.2 0.25 6
0.5 6
6
0.1
1.5 6
0
0.7 0.75 0.8 0.85 0.9 0.95 1
spatial correlation
Figure 6.12: Probe Two, Empirical Spatial Correlation CCDFs vs. Separation
0.9
0.8
P(spatial corr. > abscissa)
0.7
0.6
0.5
0.4
0.3
0.2 0.25 6
0.5 6
6
0.1
1.5 6
0
0.7 0.75 0.8 0.85 0.9 0.95 1
spatial correlation
Figure 6.13: Probe Three, Empirical Spatial Correlation CCDFs vs. Separation
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 183
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
0.9
0.8
P(spatial corr. > abscissa)
0.7
0.6
0.5
0.4
0.3
0.2 0.25 6
0.5 6
6
0.1
1.5 6
0
0.7 0.75 0.8 0.85 0.9 0.95 1
spatial correlation
Figure 6.14: Probe Four, Empirical Spatial Correlation CCDFs vs. Separation
λ λ
four in Fig. 6.14, the CCDF curve was nearly as high as the curve. For all probes, in
2 4
λ
almost all cases the CCDF curve was the second right-most CCDF curve throughout. One
2
λ
hundred percent of spatial correlations for were above 0.87 for all four probes.
2
3λ
With the exception of probe four, the CCDF curves of λ and
2
all closely tracked each
3λ
other, with the
2
CCDF and the λ CCDF curves crossing each other in multiple places. Only
3λ λ
probe four had strictly-ordered CCDF curves of to from left to right for all probabilities.
2 4
The following section will investigate the eect of DUT antenna separation and size on
MIMO capacity.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 184
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
6.3 Wideband MIMO Capacity Using Small Chamber
Channel Model
The instantaneous wideband MIMO capacity formula (2.8) can be used in conjunction
with our channel model for each probe. For a 4×4 MIMO system (MRx = MT x = 4),
four receiver antenna locations are selected in the test zone, and the channel hi,j,k from each
of the four probe antennas (j ) is calculated to each of the four receiver locations (i) at a
selection of frequencies (k , with spacing ∆f over a frequency range W, forming F frequency
samples). All values of hi,j,k are used to populate H[f ], which is then normalized as in
(2.9). An arbitrary SNR ρ is chosen for the purposes of analysis (often chosen as the linear
equivalent of 10 dB, 15 dB, or 20 dB in the literature).
6.3.1 MIMO Capacity Analysis System Description
For analysis, 20 frequencies evenly-spaced from 5.15 GHz to 5.725 GHz were selected.
Four types of simulated square DUT antenna arrays were chosen, with antennas located on
the corner of a square. The four square sizes were chosen to be 0.25λ × 0.25λ, 0.5λ × 0.5λ,
λ × λ, and 1.5λ × 1.5λ, allowing for the MIMO capacity to re ect the spatial correlations
investigated in the previous section (as the majority of antenna pairs of the receiver will have
a separation equal to a side length of the square).
Each simulated DUT was moved along an 8×8 grid of DUT center locations within the
test zone, allowing the DUT to fully cover the entire test zone over the course of the 64
measurements. Due to varying DUT sizes, the grid of center locations had to decrease in
size as DUT dimensions increased to allow the simulated DUT to remain in the test zone
at all times. As a result, plots of capacity versus simulated DUT center location will have
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 185
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
smaller X and Y dimensions for simulated DUTs with increased dimensions.
6.3.2 MIMO Capacity and Eigenvalue Results
Figs. 6.15 - 6.18 show the wideband capacity of the four simulated DUT sizes across
the test zone. The X and Y coordinates indicate the center position of the simulated DUT
where that capacity was calculated.
12
-0.14
-0.16
11
-0.18
DUT Center Y Coordinate (m)
Wideband Capacity (bps/Hz)
-0.2 10
-0.22
9
-0.24
-0.26 8
-0.28
7
-0.3
-0.32
6
0.3 0.35 0.4 0.45 0.5
DUT Center X Coordinate (m)
Figure 6.15: 0.25 λ× 0.25 λ DUT WB Cap. Vs. DUT Center Pos.
Common trends across all four DUT sizes include greater capacity on the right side of
the test zone than the left side, generally lower capacity in the bottom left (front left of
the chamber), and capacities across the test zone varying by about 2dB in all cases. The
0.25λ × 0.25λ DUT as in Fig. 6.15 exhibited a capacity range of approximately 6.8 to 8.5
bps/Hz, the 0.5λ × 0.5λ DUT as in Fig. 6.16 exhibited a capacity range of 8.4 to 10 bps/Hz,
the λ×λ DUT as in Fig. 6.17 exhibited a capacity range of 9.2 to 11 bps/Hz, and the
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 186
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
12
-0.14
-0.16
11
-0.18
DUT Center Y Coordinate (m)
Wideband Capacity (bps/Hz)
-0.2 10
-0.22
9
-0.24
-0.26 8
-0.28
7
-0.3
-0.32
6
0.3 0.35 0.4 0.45 0.5
DUT Center X Coordinate (m)
Figure 6.16: 0.5 λ× 0.5 λ DUT WB Cap. Vs. DUT Center Pos.
12
-0.14
-0.16
11
-0.18
DUT Center Y Coordinate (m)
Wideband Capacity (bps/Hz)
-0.2 10
-0.22
9
-0.24
-0.26 8
-0.28
7
-0.3
-0.32
6
0.3 0.35 0.4 0.45 0.5
DUT Center X Coordinate (m)
Figure 6.17: λ×λ DUT WB Cap. Vs. DUT Center Pos.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 187
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
12
-0.14
-0.16
11
-0.18
DUT Center Y Coordinate (m)
Wideband Capacity (bps/Hz)
-0.2 10
-0.22
9
-0.24
-0.26 8
-0.28
7
-0.3
-0.32
6
0.3 0.35 0.4 0.45 0.5
DUT Center X Coordinate (m)
Figure 6.18: 1.5 λ× 1.5 λ DUT WB Cap. Vs. DUT Center Pos.
1.5 × 1.5 DUT as in Fig. 6.18 exhibited a capacity range of 9.8 to 11.8 bps/Hz.
A potentially more useful way of observing each set of 64 capacities is to observe them
as a CCDF rather than plotted versus center coordinate. Figs. 6.19 - 6.22 are plots of the
capacity CCDFs for each of the four simulated DUT con gurations. The capacity ranges
will remain identical to those observed in the previous four plots, but the distribution of the
observed capacities will be apparent.
λ
The
4
× λ4 DUT appears to have a median capacity of 7.7 bps/Hz, the
λ
2
× λ2 DUT has a
median capacity of 9.1 bps/Hz, the λ × λ DUT has a median capacity of 10 bps/Hz, and the
3λ 3λ
2
× 2
DUT has a median capacity of 11 bps/Hz. The increase in capacity by increasing
λ λ λ λ
a DUT size from
4
× 4
to
2
× 2
is more pronounced than further increases in the MIMO
channel present in the small anechoic chamber.
One additional method of characterizing the MIMO channel is to observe the distribution
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 188
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
0.9
0.8
P(Capacity > abscissa)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
7 7.5 8 8.5
Capacity (bps/Hz)
Figure 6.19: 0.25 λ× 0.25 λ DUT WB Cap. CCDF
0.9
0.8
P(Capacity > abscissa)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
8.4 8.6 8.8 9 9.2 9.4 9.6 9.8
Capacity (bps/Hz)
Figure 6.20: 0.5 λ× 0.5 λ DUT WB Cap. CCDF
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 189
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
0.9
0.8
P(Capacity > abscissa)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
9.2 9.4 9.6 9.8 10 10.2 10.4 10.6 10.8 11
Capacity (bps/Hz)
Figure 6.21: λ×λ DUT WB Cap. CCDF
0.9
0.8
P(Capacity > abscissa)
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
9.8 10 10.2 10.4 10.6 10.8 11 11.2 11.4 11.6 11.8
Capacity (bps/Hz)
Figure 6.22: 1.5 λ× 1.5 λ DUT WB Cap. CCDF
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 190
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
of its eigenvalues [60]. An eigenvalue decomposition of an N ×N channel matrix will yield
N eigenvalues. The number of non-zero eigenvalues of the channel matrix H, which is also
the rank of H, indicates the number of spatial degrees of freedom present in the channel [13].
Greater spatial degrees of freedom translate into higher MIMO capacities, especially in chan-
nels where all N eigenvalues are nearly-identical. Practically, eigenvalue decomposition will
likely produce N non-zero eigenvalues in a real channel in almost all circumstances, however
if one or more of these is signi cantly lower than others and/or nearly equal to zero, then
the number of spatial degrees of freedom is eectively lower than N.
Using the channel data generated for the previous set of plots, for each simulated DUT
we have 20 frequencies × 64 center positions = 1280 narrowband MIMO channel matrices.
If we perform an eigenvalue decomposition to obtain the eigenvalues of each, normalize the
eigenvalues such that their linear sum is equal to 1, and then sort the eigenvalues by size, we
can observe a CCDF for each of the four eigenvalues for each simulated DUT con guration.
This allows us to evaluate the distribution of each eigenvalue, and determine if it is too low
to be a spatial degree of freedom for that simulated DUT in the chamber. Figs. 6.23 - 6.26
each plot the CCDF of eigenvalues 1 through 4 (1 to 4 ; the variable will be used for
eigenvectors instead of the traditional λ to avoid confusion with wavelength) for each of the
four simulated DUT con gurations.
λ λ
Fig. 6.23, the
4
× 4
DUT, has a 4 that is likely at least 20 dB smaller than 1 and 3
that is likely 13 dB smaller than 1 . Even 2 is likely at least 5 dB smaller than 1 , indicating
λ
that the channel for the
4
× λ4 DUT practically only has about one or two degrees of spatial
freedom at most. This is a poorly-conditioned channel matrix, and as a results the capacity
will be low, which is exactly what was observed in the corresponding capacity CCDF Fig.
6.19.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 191
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
0.9
0.8
P(Eigenvalue > abscissa)
0.7
0.6
0.5
0.4
0.3
01
0.2 02
03
0.1 04
0
-40 -35 -30 -25 -20 -15 -10 -5 0
Normalized Eigenvalue (dB)
λ λ
Figure 6.23:
4
× 4
DUT Channel Eigenvalue CCDF
0.9
0.8
P(Eigenvalue > abscissa)
0.7
0.6
0.5
0.4
0.3
01
0.2 02
03
0.1 04
0
-40 -35 -30 -25 -20 -15 -10 -5 0
Normalized Eigenvalue (dB)
λ λ
Figure 6.24:
2
× 2
DUT Channel Eigenvalue CCDF
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 192
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
0.9
0.8
P(Eigenvalue > abscissa)
0.7
0.6
0.5
0.4
0.3
01
0.2 02
03
0.1 04
0
-40 -35 -30 -25 -20 -15 -10 -5 0
Normalized Eigenvalue (dB)
Figure 6.25: λ×λ DUT Channel Eigenvalue CCDF
0.9
0.8
P(Eigenvalue > abscissa)
0.7
0.6
0.5
0.4
0.3
01
0.2 02
03
0.1 04
0
-40 -35 -30 -25 -20 -15 -10 -5 0
Normalized Eigenvalue (dB)
3λ 3λ
Figure 6.26:
2
× 2
DUT Channel Eigenvalue CCDF
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 193
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
λ λ
Fig. 6.24, the
2
× 2
DUT, still has a 4 that is likely 15 dB lower than 1 . 3 is
likely about 11 dB smaller than 1 , and 2 is likely about 4dB smaller or less. This is still a
relatively poorly-conditioned channel matrix with about two or possibly three spatial degrees
of freedom.
Fig. 6.25 shows 4 still too low to be considered a degree of freedom, but has 3 and 2
closing in on 1 . This indicates a channel that is closer to three spatial degrees of freedom,
even though the fact that the three eigenvalues are not close in value indicates that this
channel matrix still would not be well-conditioned even for a 3×3 MIMO channel.
Finally, Fig. 6.26 shows the 2 CCDF very close to the 1 CCDF. 3 is also slightly
greater than in the previous plot. 4 is also higher, but is still close to being excluded as a
spatial degree of freedom. This channel could therefore be considered to have three spatial
degrees of freedom or possibly even four, with better conditioning than that of all smaller
DUT con gurations.
This analysis has shown the usefulness of the contributed channel model in extensively
analyzing the MIMO channel characteristics of a system such as the small anechoic chamber.
This information provides feedback to the MIMO performance of such a chamber, helps
validate whether this system meets the criteria of standardized MIMO OTA testing, and can
be used to help improve the chamber design to better meet the needs of a low-cost MIMO
OTA testing solution.
6.4 Conclusions
In this chapter, we have investigated how the small anechoic chamber MIMO channel,
modeled in the previous two chapters, performs by metrics including spatial correlation,
capacity, and channel Eigenvalue distributions. The channel exhibits high spatial correlation
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 194
CHAPTER 6. APPLICATION OF SMALL CHAMBER MODEL TO MIMO ANALYSIS
even at separations up to 4λ, indicating that the MIMO channel in the chamber is LOS-
dominant. Wideband capacity and eigenvalues, calculated over the entire test zone and
frequency range of the model for multiple simulated DUT sizes, further supported the notion
that the chamber MIMO channel is primarily LOS, yielding eigenvalue distributions that even
3λ
with
2
spacing never truly yielded channel matrices that were full-rank in a 4×4 system
or well-conditioned. This analysis of the MIMO performance of a modeled small anechoic
chamber channel simultaneously has showed the usefulness of our channel model and how
it can be applied to performance evaluation of real-world OTA test solutions. Future work
may include investigating the chamber MIMO channel for other DUT con gurations and
orientations, including ULA, rectangular arrays, and larger-dimension arrays.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 195
CHAPTER 7
Conclusion and Future Work
This dissertation has provided an analysis of near- eld MIMO systems of the type that
are present in a small anechoic chamber testing environment. This analysis has included a
feasibility study of standard-compliant electric eld generation in a small anechoic chamber
for MIMO-OTA testing, investigating parameters such as chamber size and probe antenna
array size and spacing. In addition, and more signi cantly, this dissertation contributed
a novel method of modeling a measured or numerical near- eld channel using a sum of
weighted joint spatial attribute functions plus a statistically-modeled residual re nement
term based on neighboring measurements. Model parameters for a series of narrowband
models were t to functions of frequency, thus producing a wideband channel model. This
modeling technique was used for a measurement campaign of a small anechoic chamber, and
the resulting models were utilized in a MIMO system analysis of the small anechoic chamber.
The modeling method in this dissertation achieved an accuracy of below -15 dB error
vector magnitude for at least 84 percent of measurement locations in the de ned test zone
for all frequencies analyzed. This accuracy is prior to application of the modeled residual
term which further increases model accuracy, particularly at locations near measurement
points.
An investigation into MIMO system performance of a small anechoic chamber using a
model obtained from a measurement campaign showed that the channel in the chamber
exhibited high spatial correlation, generally greater than 0.8 for distances of 4λ or less.
This spatial correlation is indicative of a highly line of sight environment, one in which
196
CHAPTER 7. CONCLUSION AND FUTURE WORK
MIMO systems exhibit limited capacity gain from spatial diversity. Further investigation
of capacity and eigenvalue decomposition of the MIMO channel matrix versus a range of
simulated DUT antenna separations over the test zone of the chamber showed that DUTs
λ
with antenna spacing of exhibited approximately a single degree of spatial freedom, while
4
3λ
a separation of exhibited around 3 degrees of spatial freedom, though the third eigenvalue
2
was still well below the rst two, yielding a poorly-conditioned channel matrix.
The results of this dissertation can be used to inform the design or re nement of a small
anechoic chamber environment, with the goal of reducing spatial correlation of the system.
It can also be used to quantify expected DUT performance in a small anechoic chamber,
allowing for test metrics and repeatable testing design.
Future work expanding on these results includes expanding the model to include a larger
test zone in the chamber, extending the model to additional frequency bands for other wire-
less technologies, expanding the test zone to three dimensions to model the channel as a
three-dimensional function of space and frequency, investigating the eect of receiver polar-
ization on the channel and incorporating this information into the channel model. Also, a
measurement campaign varying transmitter position can be performed to investigate trans-
mitter spatial correlation, and potentially extend the model as a function of both receiver
and transmitter location, rather than discrete models for single transmitter locations. Fi-
nally, the eects of coupling between multiple DUT antennas on the channel model can be
investigated, modeled, and incorporated into this channel model. Due to the extremely large
possible con gurations and electromagnetic properties of a DUT, the challenges of deciding
a proper way to incorporate this into a channel model that will be applicable to any DUT
in the chamber proved beyond the scope of this dissertation.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 197
References
[1] 3GPP releases. [Online]. Available: http://www.3gpp.org/speci cations/releases
[2] Y. Tak and S. Nam, Mode-based computation method of channel characteristics for
a near- eld MIMO, IEEE Antennas and Wireless Propagation Letters, vol. 10, pp.
11701173, 2011.
[3] H. Hirayama, G. Matsui, N. Kikuma, and K. Sakakibara, Improvement of channel ca-
pacity of near- eld MIMO, Proceedings of the Fourth European Conference on Antennas
and Propagation, IEEE, pp. 14, April 2010.
[4] Octoscope, Inc. octobox OBS-14 and OBS-15 high gain antenna array datasheet.
[Online]. Available: https://www.octoscope.com/English/Collaterals/Documents/
octoBox_OBS-14_OBS-15_High-Gain_Antenna_Datasheet.pdf
[5] Pasternack. PE51083 rubber duck antenna datasheet. [Online]. Available: https:
//www.pasternack.com/images/ProductPDF/PE51083.pdf
[6] Keysight Technologies. Keysight E5063A ENA vector network analyzer datasheet.
[Online]. Available: https://literature.cdn.keysight.com/litweb/pdf/5991-3614EN.pdf ?
id=2424803
[7] R. Bajaj, S. Bates, B. P., K. Bennett et al. Connecting america: The national broad-
band plan. [Online]. Available: https://transition.fcc.gov/national-broadband-plan/
national-broadband-plan.pdf
198
REFERENCES
[8] The path to 5G: as much evolution as revolution. [Online]. Available: http:
//www.3gpp.org/news-events/3gpp-news/1774-5g_wiseharbour
[9] Unleashing 5G with 'High Band' Spectrum. [Online]. Available: http://www.ctia.org/
docs/default-source/default-document-library/high-band-spectrum-april-2016.pdf
[10] J. Xu, D. Goeckel, and R. Janaswamy, The capacity of MIMO systems with increasing
SNR by electromagnetic analysis, Wireless Communications, IEEE Transactions on,
vol. 8, no. 9, pp. 47524761, 2009.
[11] N. Costa and S. Haykin, Multiple-Input Multiple-Output Channel Models: Theory and
Practice. Wiley, 2010.
[12] J. Tet and N. Kirsch, Small anechoic chambers for MIMO-OTA testing, IEEE Trans-
actions on Instrumentation and Measurement (ready for submission), Aug. 2016.
[13] D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge Uni-
versity Press, 2005.
[14] J. Foschini and M. J. Gans, On the limits of wireless communications in a fading
environment when using multiple antennas, Wireless Personal Communications, vol. 6,
pp. 315335, March 1998.
[15] A. Goldsmith, Wireless Communications. Cambridge University Press, 2005.
[16] R. Janaswamy, Eect of element mutual coupling on the capacity of xed length linear
arrays, IEEE Antennas and Wireless Propagation Letters, vol. 1, no. 1, pp. 157160,
2002.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 199
REFERENCES
[17] H.-S. Lui, H. T. Hui, and M. S. Leong, A note on the mutual-coupling problems in
transmitting and receiving antenna arrays, IEEE Antennas and Propagation Magazine,
vol. 51, no. 5, pp. 171176, Oct 2009.
[18] E. Hansen, Spherical Near-Field Antenna Measurements. Peter Peregrinus Ltd., 1988.
[19] S. Orfanidis. Electromagnetic waves and antennas. [Online]. Available: http:
//www.ece.rutgers.edu/~orfanidi/ewa/
[20] 3rd Generation Partnership Project (3GPP), Inter-Lab/Inter-Technique OTA Per-
formance Comparison Testing for MIMO Devices. Work Group Document:
MOSG120521R9, 2013.
[21] W. Fan, X. Carreno Bautista de Lisbona, F. Sun, J. Nielsen, M. Knudsen, and G. Ped-
ersen, Emulating spatial characteristics of MIMO channels for OTA testing, Antennas
and Propagation, IEEE Transactions on, vol. 61, no. 8, pp. 43064314, 2013.
[22] G. de la Roche, A. Alayn-Glazunov, and B. Allen, LTE-Advanced and Next Generation
Wireless Networks: Channel Modelling and Propagation. Wiley, 2012.
[23] C. Lotback Patane, A. Skarbratt, R. Rehammar, and C. Orlenius, On the use of re-
verberation chambers for assessment of MIMO-OTA performance of wireless devices,
in Antennas and Propagation (EuCAP), 2013 7th European Conference on, 2013, pp.
101105.
[24] I. Carton Llorente, W. Fan, and G. Pedersen, MIMO-OTA testing in small multi-probe
anechoic chamber setups, Antennas and Wireless Propagation Letters, IEEE, vol. PP,
no. 99, pp. 11, 2015.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 200
REFERENCES
[25] T. Laitinen, P. Kyosti, T. Jamsa, and P. Vainikainen, Generation of a eld with a
laplacian-distributed power azimuth spectrum scattered by a single cluster in a MIMO-
OTA test system based on multiple probe antennas, in Proceedings of Asia-Paci c
Microwave Conference (APMC), Dec. 2010.
[26] J. T. Toivanen, T. A. Laitinen, S. Pivnenko, and L. Nyberg, Calibration of multi-probe
antenna measurement system using test zone eld compensation, in 2009 3rd European
Conference on Antennas and Propagation, March 2009, pp. 29162920.
[27] J. T. Toivanen, T. A. Laitinen, and P. Vainikainen, Modi ed test zone eld compensa-
tion for small-antenna measurements, IEEE Transactions on Antennas and Propaga-
tion, vol. 58, no. 11, pp. 34713479, Nov 2010.
[28] J. T. Toivanen, T. A. Laitinen, V. M. Kolmonen, and P. Vainikainen, Reproduction
of arbitrary multipath environments in laboratory conditions, IEEE Transactions on
Instrumentation and Measurement, vol. 60, no. 1, pp. 275281, Jan 2011.
[29] P. Kyösti, J. P. Nuutinen, and T. Jämsä, MIMO OTA test concept with experimen-
tal and simulated veri cation, in Proceedings of the Fourth European Conference on
Antennas and Propagation, April 2010, pp. 15.
[30] Y. Okano, K. Kitao, and T. Imai, Impact of number of probe antennas for MIMO
OTA spatial channel emulator, in Proceedings of the Fourth European Conference on
Antennas and Propagation, April 2010, pp. 15.
[31] P. Kyösti and L. Hentilä, Criteria for physical dimensions of MIMO OTA multi-probe
test setup, in 2012 6th European Conference on Antennas and Propagation (EUCAP),
March 2012, pp. 20552059.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 201
REFERENCES
[32] W. Fan, J. . Nielsen, O. Franek, X. Carreño, J. S. Ashta, M. B. Knudsen, and G. F.
Pedersen, Antenna pattern impact on MIMO OTA testing, IEEE Transactions on
Antennas and Propagation, vol. 61, no. 11, pp. 57145723, Nov 2013.
[33] W. Fan, F. Sun, P. Kyosti, J. Nielsen, X. Carreno, M. Knudsen, and G. Pedersen, 3d
channel emulation in multi-probe setup, Electronics Letters, vol. 49, no. 9, pp. 623625,
April 2013.
[34] P. Kyösti, T. Jämsä, and J. P. Nuutinen, Channel modelling for multiprobe over-the-
air MIMO testing, International Journal of Antennas and Propogation, vol. 2012, pp.
111, Mar 2012.
[35] A. Khatun, T. Laitinen, V. M. Kolmonen, and P. Vainikainen, Dependence of error
level on the number of probes in over-the-air multiprobe test systems, International
Journal of Antennas and Propogation, vol. 2012, p. 6, Mar 2012.
[36] W. Fan, J. . Nielsen, X. Carreño, O. Franek, M. B. Knudsen, and G. F. Pedersen,
Impact of probe placement error on MIMO OTA test zone performance, in Antennas
and Propagation Conference (LAPC), 2012 Loughborough, Nov 2012, pp. 14.
[37] T. Laitinen, J. Toivanen, P. Kyösti, J. P. Nuutinen, and P. Vainikainen, On a MIMO-
OTA testing based on multi-probe technology, in Electromagnetic Theory (EMTS),
2010 URSI International Symposium on, Aug 2010, pp. 227230.
[38] W. Fan, X. Carreño, J. . Nielsen, K. Olesen, M. B. Knudsen, and G. F. Pedersen, Mea-
surement veri cation of plane wave synthesis technique based on multi-probe MIMO-
OTA setup, in Vehicular Technology Conference (VTC Fall), 2012 IEEE, Sept 2012,
pp. 15.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 202
REFERENCES
[39] J.-S. Jiang and M. A. Ingram, Spherical-wave model for short-range MIMO, IEEE
Transactions on Communications, vol. 53, no. 9, pp. 15341541, Sept 2005.
[40] K. Nishimori, N. Honma, T. Seki, and K. Hiraga, On the transmission method for short-
range MIMO communication, IEEE Transactions on Vehicular Technology, vol. 60,
no. 3, pp. 12471251, March 2011.
[41] N. Honma, K. Nishimori, T. Seki, and M. Mizoguchi, Short range MIMO communica-
tion, in 2009 3rd European Conference on Antennas and Propagation, March 2009, pp.
17631767.
[42] N. Gvozdenovic, W. Thompson, M. A. Beach, C. F. Mecklenbräuker, and G. Hilton,
Short range ultra-wideband multiple input multiple output channel measurements,
in 2013 IEEE Wireless Communications and Networking Conference (WCNC), April
2013, pp. 25752578.
[43] M. K. Ozdemir, H. Arslan, and E. Arvas, A mutual coupling model for MIMO systems,
in Wireless Communication Technology, 2003. IEEE Topical Conference on, Oct 2003,
pp. 306307.
[44] H. Kanj, S. Ali, P. Lusina, and F. Kohandani, A modeling approach for simulating
MIMO systems with near- eld eects, in Wireless Technology, 2008. EuWiT 2008.
European Conference on, Oct 2008, pp. 143146.
[45] M. A. Nikravan, D. H. Kwon, H. G. Schantz, and A. H. Unden, Near- eld MIMO
communication utilizing both electric and magnetic eld components, in 2014 IEEE
Antennas and Propagation Society International Symposium (APSURSI), July 2014,
pp. 474475.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 203
REFERENCES
[46] Y. Tak, S. Yun, J. Park, and S. Nam, Analysis of a near- eld mimo based on the polar-
ization diversity by using the mode-based approach, in Antenna Technology and Ap-
plied Electromagnetics (ANTEM), 2012 15th International Symposium on, June 2012,
pp. 14.
[47] D. Zhang, Research on channel capacity characteristics of near- eld MIMO systems,
Ph.D. dissertation, University of Fukui, The address of the publisher, 9 2013.
[48] M. Jensen and J. Wallace, A Review of Antennas and Propagation for MIMO Wireless
Communications, IEEE Transactions on Antennas and Propagation, vol. 52, no. 11,
pp. 2810 2824, Nov 2004.
[49] T. A. Laitinen, P. Kyosti, J.-P. Nuutinen, and P. Vainikainen, On the number of
OTA antenna elements for plane-wave synthesis in a MIMO-OTA test system involving
a circular antenna array, in 4th European Conference on Antennas and Propagation
(EuCAP'lO), Apr. 2010.
[50] D. Baum, J. Hansen, and J. Salo, An interim channel model for beyond-3G systems:
extending the 3GPP spatial channel model (SCM), in 2005 IEEE 61st Vehicular Tech-
nology Conference, vol. 5, May 2005, pp. 31323136.
[51] W. Stutzman and G. Thiele, Antenna Theory and Design. John Wiley and Sons, Inc.,
2012.
[52] T. Smith. Notebook on spatial data analysis (online). [Online]. Available:
http://www.seas.upenn.edu/~ese502/#notebook
[53] R. Horn and C. Johnson, Matrix Analysis. Cambridge University Press, 1985.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 204
REFERENCES
[54] R. Mersereau and T. Speake, The processing of periodically sampled multidimensional
signals, IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. ASSP-31,
pp. 188194, Feb 1983.
[55] B. Engquist and H. Zhao, Approximate separability of the Green's function of the
Helmholtz equation in the high frequency limit, Communications on Pure and Applied
Mathematics, vol. 71, pp. 22202274, Nov 2018.
[56] N. A. C. Cressie, Statistics for Spatial Data. Wiley, 1993.
[57] D. P. S. De Iaco, Wind velocity prediction through complex kriging: Formalism and
computational aspects, Environmental and Ecological Statistics, vol. 23, no. 1, pp.
115139, Mar 2016.
[58] S. Rey and T. Kuerner, IEEE 802.15-16-0207-01-003d how to drive the channel impulse
response from a broadband channel transfer function, IEEE P802.15 Working Group
for Wireless Personal Area Networks, pp. 18, Mar 2016.
[59] P. L. Ka e, A. Intarapanich, A. B. Sesay, J. McRory, and R. J. Davies, Spatial correla-
tion and capacity measurements for wideband MIMO channels in indoor o ce environ-
ment, IEEE Transactions on Wireless Communications, vol. 7, no. 5, pp. 15601571,
May 2008.
[60] J. P. Kermoal, L. Schumacher, P. E. Mogensen, and K. I. Pedersen, Experimental
investigation of correlation properties of mimo radio channels for indoor picocell sce-
narios, in Vehicular Technology Conference Fall 2000. IEEE VTS Fall VTC2000. 52nd
Vehicular Technology Conference (Cat. No.00CH37152), vol. 1, Sept 2000, pp. 1421.
Near-Field MIMO Channel Modeling with Applications to Small Anechoic Chambers 205
You might also like
- TDD Massive MIMO Systems - Channel Estimation, Power Optimization (Elina Nayebi Disse.)No ratings yetTDD Massive MIMO Systems - Channel Estimation, Power Optimization (Elina Nayebi Disse.)172 pages
- Ss1 Third Term Scheme of Work for 2024-2025 Academ_250516_064455No ratings yetSs1 Third Term Scheme of Work for 2024-2025 Academ_250516_06445556 pages
- Introduction To Multiple Antenna Communications and Reconfigurable SurfacesNo ratings yetIntroduction To Multiple Antenna Communications and Reconfigurable Surfaces687 pages
- The Utilization of Software Defined Radios for Adaptive Phased ANo ratings yetThe Utilization of Software Defined Radios for Adaptive Phased A94 pages
- Planar Beam-forming Antenna Array for 60-GHz Broadband CommunicationNo ratings yetPlanar Beam-forming Antenna Array for 60-GHz Broadband Communication184 pages
- Part I of Two Parts: February 2012 Number 2 Ietpak (ISSN 0018-926X)No ratings yetPart I of Two Parts: February 2012 Number 2 Ietpak (ISSN 0018-926X)288 pages
- Channel Estimation in TDD and FDD-Based Massive MIMO Systems (Mirzaei Javad - 202106 - PHD - Thesis)No ratings yetChannel Estimation in TDD and FDD-Based Massive MIMO Systems (Mirzaei Javad - 202106 - PHD - Thesis)159 pages
- Developing A UHF - VHF Phased-Array Satellite Ground StationNo ratings yetDeveloping A UHF - VHF Phased-Array Satellite Ground Station74 pages
- Multi Antenna Synthetic Aperture Radar by Wen-Qin Wang PDFNo ratings yetMulti Antenna Synthetic Aperture Radar by Wen-Qin Wang PDF464 pages
- Create-Accurate-EVM-Measurements-with-the-PNA-X-Series-Network-AnalyzerNo ratings yetCreate-Accurate-EVM-Measurements-with-the-PNA-X-Series-Network-Analyzer47 pages
- Modern Antennas and Microwave Circuits Eindhoven - Universty of TechnologyNo ratings yetModern Antennas and Microwave Circuits Eindhoven - Universty of Technology192 pages
- Addis Ababa University School of Graduate Studies Addis Ababa Institute of Technology100% (1)Addis Ababa University School of Graduate Studies Addis Ababa Institute of Technology85 pages
- Reconfigurable Antenna Array Using The PIN-Diode-Switched PrintedNo ratings yetReconfigurable Antenna Array Using The PIN-Diode-Switched Printed52 pages
- octoScope_WP_UNH_MIMO-OTA_dissertation_summaryNo ratings yetoctoScope_WP_UNH_MIMO-OTA_dissertation_summary5 pages
- radio compliance tests presentation 03 06 2016No ratings yetradio compliance tests presentation 03 06 201624 pages
- Degree Project: Error Rate Performance of Multi-Hop Communication Systems Over Nakagami-M Fading ChannelNo ratings yetDegree Project: Error Rate Performance of Multi-Hop Communication Systems Over Nakagami-M Fading Channel34 pages
- Channel Estimation and Symbol Detection in Massive MIMO Systems UNo ratings yetChannel Estimation and Symbol Detection in Massive MIMO Systems U122 pages
- (D) Antenna Array Signal Processing For Multistatic Radar Systems - Belfiori - 2013No ratings yet(D) Antenna Array Signal Processing For Multistatic Radar Systems - Belfiori - 2013162 pages
- Home Mesh AP Performance Analysis _ Allion LabsNo ratings yetHome Mesh AP Performance Analysis _ Allion Labs12 pages
- 2017-Analysis and Design of Transmitarray AntennasNo ratings yet2017-Analysis and Design of Transmitarray Antennas177 pages
- Altair's Student Guides - Instructor's Manual - CAE For Simulation of Sheet Metal Forming100% (13)Altair's Student Guides - Instructor's Manual - CAE For Simulation of Sheet Metal Forming26 pages
- Noise Variance Estimation in MIMOOFDMtestbedNo ratings yetNoise Variance Estimation in MIMOOFDMtestbed114 pages
- octoScope_WP_Repeatable_Reproducible_Measurement_201804No ratings yetoctoScope_WP_Repeatable_Reproducible_Measurement_2018047 pages
- Antennas and Propagation From A Signal Processing Perspective by Thomas Svantesson 2001No ratings yetAntennas and Propagation From A Signal Processing Perspective by Thomas Svantesson 2001319 pages
- 24 GHZ Patch Antenna Array Design For RADAR PDFNo ratings yet24 GHZ Patch Antenna Array Design For RADAR PDF120 pages
- 11-11-1568-01-00ac-sta-antenna-indication-in-vht-capabilitiesNo ratings yet11-11-1568-01-00ac-sta-antenna-indication-in-vht-capabilities5 pages
- SPSS T-Test Excercise: Data Repurchase IntentionNo ratings yetSPSS T-Test Excercise: Data Repurchase Intention10 pages
- GE6252-Basic Electrical and Electronics EngineeringNo ratings yetGE6252-Basic Electrical and Electronics Engineering13 pages
- Strategies To Make Advertisement More EffectiveNo ratings yetStrategies To Make Advertisement More Effective83 pages
- Fractal Antennas For Wireless Comm System With Source CodeNo ratings yetFractal Antennas For Wireless Comm System With Source Code70 pages
- 2022 Multi-Objective Optimization and Optimal Airfoil Blade Selection For A Small Horizontal-Axis Wind Turbine (HAWT)No ratings yet2022 Multi-Objective Optimization and Optimal Airfoil Blade Selection For A Small Horizontal-Axis Wind Turbine (HAWT)22 pages
- Drake Elkin Method Base Plate Design 181 PDFNo ratings yetDrake Elkin Method Base Plate Design 181 PDF10 pages
- VII Preface - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - .No ratings yetVII Preface - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - . - .5 pages
- Answers To The Learning Activities From Module 8No ratings yetAnswers To The Learning Activities From Module 821 pages
- NCERT Solutions Class 10 Maths Chapter 13 Surface Areas and Volumes - BYJU'SNo ratings yetNCERT Solutions Class 10 Maths Chapter 13 Surface Areas and Volumes - BYJU'S4 pages
- Channel Estimation Techniques Over MIMO-OFDM SystemNo ratings yetChannel Estimation Techniques Over MIMO-OFDM System4 pages
- Determination of Sex: Interest of Frontal Sinus 3D ReconstructionsNo ratings yetDetermination of Sex: Interest of Frontal Sinus 3D Reconstructions5 pages
- Arc Control in Circuit Breakers: Low Contact Velocity 2nd EditionFrom EverandArc Control in Circuit Breakers: Low Contact Velocity 2nd EditionNo ratings yet
