Complex Valued Nonlinear Adaptive Filters
Description
More details
Other editions
Additional editions
Persons
Danilo Mandic, Department of Electrical and Electronic Engineering, Imperial College London, London
Dr Mandic is currently a Reader in Signal Processing at Imperial College, London. He is an experienced author, having written the book Recurrent Neural Networks for Prediction: Learning Algorithms, Architectures and Stability (Wiley, 2001), and more than 150 published journal and conference papers on signal and image processing. His research interests include nonlinear adaptive signal processing, multimodal signal processing and nonlinear dynamics, and he is an Associate Editor for the journals IEEE Transactions on Circuits and Systems and the International Journal of Mathematical Modelling and Algorithms. Dr Mandic is also on the IEEE Technical Committee on Machine Learning for Signal Processing, and he has produced award winning papers and products resulting from his collaboration with industry.
Su-Lee Goh, Royal Dutch Shell plc, Holland
Dr Goh is currently working as a Reservoir Imaging Geophysicist at Shell in Holland. Her research interests include nonlinear signal processing, adaptive filters, complex-valued analysis, and imaging and forecasting. She received her PhD in nonlinear adaptive signal processing from Imperial College, London and is a member of the IEEE and the Society of Exploration Geophysicists.
Content
Preface xiii
Acknowledgements xvii
1 The Magic of Complex Numbers 1
1.1 History of Complex Numbers 2
1.1.1 Hypercomplex Numbers 7
1.2 History of Mathematical Notation 8
1.3 Development of Complex Valued Adaptive Signal Processing 9
2 Why Signal Processing in the Complex Domain? 13
2.1 Some Examples of Complex Valued Signal Processing 13
2.1.1 Duality Between Signal Representations in R and c 18
2.2 Modelling in C is Not Only Convenient But Also Natural 19
2.3 Why Complex Modelling of Real Valued Processes? 20
2.3.1 Phase Information in Imaging 20
2.3.2 Modelling of Directional Processes 22
2.4 Exploiting the Phase Information 23
2.4.1 Synchronisation of Real Valued Processes 24
2.4.2 Adaptive Filtering by Incorporating Phase Information 25
2.5 Other Applications of Complex Domain Processing of Real Valued Signals 26
2.6 Additional Benefits of Complex Domain Processing 29
3 Adaptive Filtering Architectures 33
3.1 Linear and Nonlinear Stochastic Models 34
3.2 Linear and Nonlinear Adaptive Filtering Architectures 35
3.2.1 Feedforward Neural Networks 36
3.2.2 Recurrent Neural Networks 37
3.2.3 Neural Networks and Polynomial Filters 38
3.3 State Space Representation and Canonical Forms 39
4 Complex Nonlinear Activation Functions 43
4.1 Properties of Complex Functions 43
4.1.1 Singularities of Complex Functions 45
4.2 Universal Function Approximation 46
4.2.1 Universal Approximation in R 47
4.3 Nonlinear Activation Functions for Complex Neural Networks 48
4.3.1 Split-complex Approach 49
4.3.2 Fully Complex Nonlinear Activation Functions 51
4.4 Generalised Splitting Activation Functions (GSAF) 53
4.4.1 The Clifford Neuron 53
4.5 Summary: Choice of the Complex Activation Function 54
5 Elements of CR Calculus 55
5.1 Continuous Complex Functions 56
5.2 The Cauchy-Riemann Equations 56
5.3 Generalised Derivatives of Functions of Complex Variable 57
5.3.1 CR Calculus 59
5.3.2 Link between R- and C-derivatives 60
5.4 CR-derivatives of Cost Functions 62
5.4.1 The Complex Gradient 62
5.4.2 The Complex Hessian 64
5.4.3 The Complex Jacobian and Complex Differential 64
5.4.4 Gradient of a Cost Function 65
6 Complex Valued Adaptive Filters 69
6.1 Adaptive Filtering Configurations 70
6.2 The Complex Least Mean Square Algorithm 73
6.2.1 Convergence of the CLMS Algorithm 75
6.3 Nonlinear Feedforward Complex Adaptive Filters 80
6.3.1 Fully Complex Nonlinear Adaptive Filters 80
6.3.2 Derivation of CNGD using CR calculus 82
6.3.3 Split-complex Approach 83
6.3.4 Dual Univariate Adaptive Filtering Approach (DUAF) 84
6.4 Normalisation of Learning Algorithms 85
6.5 Performance of Feedforward Nonlinear Adaptive Filters 87
6.6 Summary: Choice of a Nonlinear Adaptive Filter 89
7 Adaptive Filters with Feedback 91
7.1 Training of IIR Adaptive Filters 92
7.1.1 Coefficient Update for Linear Adaptive IIR Filters 93
7.1.2 Training of IIR filters with Reduced Computational Complexity 96
7.2 Nonlinear Adaptive IIR Filters: Recurrent Perceptron 97
7.3 Training of Recurrent Neural Networks 99
7.3.1 Other Learning Algorithms and Computational Complexity 102
7.4 Simulation Examples 102
8 Filters with an Adaptive Stepsize 107
8.1 Benveniste Type Variable Stepsize Algorithms 108
8.2 Complex Valued GNGD Algorithms 110
8.2.1 Complex GNGD for Nonlinear Filters (CFANNGD) 112
8.3 Simulation Examples 113
9 Filters with an Adaptive Amplitude of Nonlinearity 119
9.1 Dynamical Range Reduction 119
9.2 FIR Adaptive Filters with an Adaptive Nonlinearity 121
9.3 Recurrent Neural Networks with Trainable Amplitude of Activation Functions 122
9.4 Simulation Results 124
10 Data-reusing Algorithms for Complex Valued Adaptive Filters 129
10.1 The Data-reusing Complex Valued Least Mean Square (DRCLMS) Algorithm 129
10.2 Data-reusing Complex Nonlinear Adaptive Filters 131
10.2.1 Convergence Analysis 132
10.3 Data-reusing Algorithms for Complex RNNs 134
11 Complex Mappings and Möbius Transformations 137
11.1 Matrix Representation of a Complex Number 137
11.2 The Möbius Transformation 140
11.3 Activation Functions and Möbius Transformations 142
11.4 All-pass Systems as Möbius Transformations 146
11.5 Fractional Delay Filters 147
12 Augmented Complex Statistics 151
12.1 Complex Random Variables (CRV) 152
12.1.1 Complex Circularity 153
12.1.2 The Multivariate Complex Normal Distribution 154
12.1.3 Moments of Complex Random Variables (CRV) 157
12.2 Complex Circular Random Variables 158
12.3 Complex Signals 159
12.3.1 Wide Sense Stationarity, Multicorrelations, and Multispectra 160
12.3.2 Strict Circularity and Higher-order Statistics 161
12.4 Second-order Characterisation of Complex Signals 161
12.4.1 Augmented Statistics of Complex Signals 161
12.4.2 Second-order Complex Circularity 164
13 Widely Linear Estimation and Augmented CLMS (ACLMS) 169
13.1 Minimum Mean Square Error (MMSE) Estimation in c 169
13.1.1 Widely Linear Modelling in c 171
13.2 Complex White Noise 172
13.3 Autoregressive Modelling in c 173
13.3.1 Widely Linear Autoregressive Modelling in c 174
13.3.2 Quantifying Benefits of Widely Linear Estimation 174
13.4 The Augmented Complex LMS (ACLMS) Algorithm 175
13.5 Adaptive Prediction Based on ACLMS 178
13.5.1 Wind Forecasting Using Augmented Statistics 180
14 Duality Between Complex Valued and Real Valued Filters 183
14.1 A Dual Channel Real Valued Adaptive Filter 184
14.2 Duality Between Real and Complex Valued Filters 186
14.2.1 Operation of Standard Complex Adaptive Filters 186
14.2.2 Operation of Widely Linear Complex Filters 187
14.3 Simulations 188
15 Widely Linear Filters with Feedback 191
15.1 The Widely Linear ARMA (WL-ARMA) Model 192
15.2 Widely Linear Adaptive Filters with Feedback 192
15.2.1 Widely Linear Adaptive IIR Filters 195
15.2.2 Augmented Recurrent Perceptron Learning Rule 196
15.3 The Augmented Complex Valued RTRL (ACRTRL) Algorithm 197
15.4 The Augmented Kalman Filter Algorithm for RNNs 198
15.4.1 EKF Based Training of Complex RNNs 200
15.5 Augmented Complex Unscented Kalman Filter (ACUKF) 200
15.5.1 State Space Equations for the Complex Unscented Kalman Filter 201
15.5.2 ACUKF Based Training of Complex RNNs 202
15.6 Simulation Examples 203
16 Collaborative Adaptive Filtering 207
16.1 Parametric Signal Modality Characterisation 207
16.2 Standard Hybrid Filtering in R 209
16.3 Tracking the Linear/Nonlinear Nature of Complex Valued Signals 210
16.3.1 Signal Modality characterisation in c 211
16.4 Split vs Fully Complex Signal Natures 214
16.5 Online Assessment of the Nature of Wind Signal 216
16.5.1 Effects of Averaging on Signal Nonlinearity 216
16.6 Collaborative Filters for General Complex Signals 217
16.6.1 Hybrid Filters for Noncircular Signals 218
16.6.2 Online Test for Complex Circularity 220
17 Adaptive Filtering Based on EMD 221
17.1 The Empirical Mode Decomposition Algorithm 222
17.1.1 Empirical Mode Decomposition as a Fixed Point Iteration 223
17.1.2 Applications of Real Valued EMD 224
17.1.3 Uniqueness of the Decomposition 225
17.2 Complex Extensions of Empirical Mode Decomposition 226
17.2.1 Complex Empirical Mode Decomposition 227
17.2.2 Rotation Invariant Empirical Mode Decomposition (RIEMD) 228
17.2.3 Bivariate Empirical Mode Decomposition (BEMD) 228
17.3 Addressing the Problem of Uniqueness 230
17.4 Applications of Complex Extensions of EMD 230
18 Validation of Complex Representations - Is This Worthwhile? 233
18.1 Signal Modality Characterisation in R 234
18.1.1 Surrogate Data Methods 235
18.1.2 Test Statistics: The DVV Method 237
18.2 Testing for the Validity of Complex Representation 239
18.2.1 Complex Delay Vector Variance Method (CDVV) 240
18.3 Quantifying Benefits of Complex Valued Representation 243
18.3.1 Pros and Cons of the Complex DVV Method 244
Appendix A: Some Distinctive Properties of Calculus in C 245
Appendix B: Liouville's Theorem 251
Appendix C: Hypercomplex and Clifford Algebras 253
C. 1 Definitions of Algebraic Notions of Group, Ring and Field 253
C. 2 Definition of a Vector Space 254
C. 3 Higher Dimension Algebras 254
C. 4 The Algebra of Quaternions 255
C. 5 Clifford Algebras 256
Appendix D: Real Valued Activation Functions 257
D. 1 Logistic Sigmoid Activation Function 257
D. 2 Hyperbolic Tangent Activation Function 258
Appendix E: Elementary Transcendental Functions (ETF) 259
Appendix F: The O Notation and Standard Vector and Matrix Differentiation 263
F. 1 The O Notation 263
F. 2 Standard Vector and Matrix Differentiation 263
Appendix G: Notions From Learning Theory 265
G. 1 Types of Learning 266
G. 2 The Bias-Variance Dilemma 266
G. 3 Recursive and Iterative Gradient Estimation Techniques 267
G. 4 Transformation of Input Data 267
Appendix H: Notions from Approximation Theory 269
Appendix I: Terminology Used in the Field of Neural Networks 273
Appendix J: Complex Valued Pipelined Recurrent Neural Network (CPRNN) 275
J.1 The Complex RTRL Algorithm (CRTRL) for CPRNN 275
J.1.1 Linear Subsection Within the PRNN 277
Appendix K: Gradient Adaptive Step Size (GASS) Algorithms in R 279
K. 1 Gradient Adaptive Stepsize Algorithms Based on ¿J/¿µ 280
K. 2 Variable Stepsize Algorithms Based on ¿J/¿e 281
Appendix L: Derivation of Partial Derivatives from Chapter 8 283
L. 1 Derivation of ¿e(k)/¿w n (k) 283
L. 2 Derivation of ¿e * (k)/¿e(k - 1) 284
L. 3 Derivation of ¿w(k)/¿e(k - 1) 286
Appendix M: A Posteriori Learning 287
M.1 A Posteriori Strategies in Adaptive Learning 288
Appendix N: Notions from Stability Theory 291
Appendix O: Linear Relaxation 293
O. 1 Vector and Matrix Norms 293
O. 2 Relaxation in Linear Systems 294
O.2.1 Convergence in the Norm or State Space? 297
Appendix P: Contraction Mappings, Fixed Point Iteration and Fractals 299
P. 1 Historical Perspective 303
P. 2 More on Convergence: Modified Contraction Mapping 305
P. 3 Fractals and Mandelbrot Set 308
References 309
Index 321
