-
Notifications
You must be signed in to change notification settings - Fork 0
/
AASP.tex
1314 lines (1132 loc) · 77.7 KB
/
AASP.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
% % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % %
% LaTeX4EI Template for Cheat Sheets Version 1.0
%
% Authors: Emanuel Regnath, Martin Zellner
% Contact: [email protected]
% Encode: UTF-8, tabwidth = 4, newline = LF
% % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % % %
% ======================================================================
% Document Settings
% ======================================================================
% possible options: color/nocolor, english/german, threecolumn
% defaults: color, english
\documentclass[english]{latex4ei/latex4ei_sheet}
% set document information
\title{Adaptive and Array\\ Signal Processing}
\author{Johannes Kunz} % optional, delete if unchanged
\myemail{[email protected]} % optional, delete if unchanged
\mywebsite{www.latex4ei.de} % optional, delete if unchanged
\let\T\relax
\DeclareMathOperator{\T}{\textsf{\textit{T}}} % Zufallsvariable X
\DeclareMathOperator{\Bias}{Bias} % Zufallsvariable X
\DeclareMathOperator{\argmax}{argmax}
\DeclareMathOperator{\rang}{rang}
\renewcommand{\vec}[1]{\underline{\boldsymbol{#1}}}
%\renewcommand{\vec}[1]{\boldsymbol{#1}}
\newcommand{\x}{\textit{x}}
\newcommand{\vx}{\underline{\textbf{\textit{x}}}}
\newcommand{\VX}{\underline{\textbf{\textit{X}}}}
\newcommand{\y}{\textit{y}}
\newcommand{\vy}{\underline{\textbf{\textit{y}}}}
\newcommand{\VY}{\underline{\textbf{\textit{Y}}}}
\newcommand{\VA}{\underline{\textit{A}}}
\newcommand{\A}{\textit{A}}
\newcommand{\va}{\underline{\textbf{\textit{a}}}}
\newcommand{\VB}{\underline{\textit{B}}}
\newcommand{\danger}[1]{\textcolor{red}{#1}}
\newcommand{\comment}[1]{\textcolor{red}{#1}}
\newcommand{\vfsymbol}[1][]{\mbox{\setlength{\unitlength}{0.1em}%
\begin{picture}(15,18)%
\put(6,2){\circle*{4}}%
\put(6,4){\line(0,1){12}}%
\put(6,18){\circle{4}}%
\put(13,7){#1}
\end{picture}%
}%
}%
\newcommand{\fsymbol}[1][]{\mbox{\setlength{\unitlength}{0.1em}%
\begin{picture}(30,10)%
\put(6,4){\circle{4}}%
\put(6,4){\line(1,0){12}}%
\put(21,4){\circle*{4}}%
\put(14,7){#1}
\end{picture}%
}%
}%
\newcommand{\finvsymbol}[1][]{\mbox{\setlength{\unitlength}{0.1em}%
\begin{picture}(30,10)%
\put(6,4){\circle*{4}}%
\put(6,4){\line(1,0){12}}%
\put(21,4){\circle{4}}%
\put(14,7){#1}
\end{picture}%
}%
}%
\usepackage{multirow}
% ======================================================================
% Begin
% ======================================================================
\begin{document}
\IfFileExists{git.id}{\input{git.id}}{}
\ifdefined\GitRevision\mydate{\GitNiceDate\ (git \GitRevision)}\fi
% Title
% ----------------------------------------------------------------------
\maketitle % requires ./img/Logo.pdf
\section{Complex Analysis}
\begin{sectionbox}
\subsection{Derivatives of Non-analytic functions}
\begin{emphbox}
$h: \mathbb{C} \ni z \mapsto h(z) \in \mathbb{C}$\\
$f: \mathbb{R}^{2} \ni(x, y) \mapsto f(x, y) \in \mathbb{C}$, $z=x+jy$\\
$g: \mathbb{C}^{2} \ni\left(z_{1}, z_{2}\right) \mapsto g\left(z_{1}, z_{2}\right) \in \mathbb{C}$, $x=\frac{z+z^{*}}{2}, y=\frac{z-z^{*}}{2 \mathrm{j}}$
\end{emphbox}
Derivative in direction $\varphi$:\\
$\frac{\mathrm{d} h}{\mathrm{~d} z}=\left(\frac{\partial f}{\partial x} \cos (\varphi)+\frac{\partial f}{\partial y} \sin \varphi\right) \mathrm{e}^{-\mathrm{j} \varphi}$, $\mathrm{d} z=\mathrm{e}^{\mathrm{j} \varphi} \mathrm{d} t, \quad \varphi, \mathrm{d} t \in \mathbb{R}$\\
\begin{emphbox}
$\frac{\partial g}{\partial z}=\frac{1}{2}\left(\frac{\partial f}{\partial x}-\mathrm{j} \frac{\partial f}{\partial y}\right), \quad \frac{\partial g}{\partial z^{*}}=\frac{1}{2}\left(\frac{\partial f}{\partial x}+\mathrm{j} \frac{\partial f}{\partial y}\right)$
\end{emphbox}
\end{sectionbox}
\begin{sectionbox}
\subsection{Analytic functions}
Definition:
\begin{emphbox}
$\frac{\mathrm{d} h}{\mathrm{~d} z}$ independent of $\varphi$\\
$\forall(x, y) \in \mathbb{R}^{2}: \quad \frac{\partial f}{\partial x}+j \frac{\partial f}{\partial y}=0$\\
$\frac{\mathrm{d} h}{\mathrm{~d} z}=\frac{\partial f}{\partial x}=-\mathrm{j} \frac{\partial f}{\partial y}$\\
$\frac{\partial g}{\partial z^{*}} \equiv 0$
\end{emphbox}
\begin{itemize}
\item Ansatz for obtaining the Lemma: compute derivative of $\frac{\mathrm{d} h}{\mathrm{~d} z}$ wrt. to $\varphi$ and set it to 0.
\item If $g()$ depends on $z^*$ it is not analytic
\end{itemize}
\end{sectionbox}
\begin{sectionbox}
\subsection{Minimization of $h(z)=g\left(z, z^{\star}\right) \in \mathbb{R}$}
\begin{emphbox}
Necessary condition for an extremum:\\
$\frac{\partial g}{\partial z^{\star}}=0$
\end{emphbox}
Direction of steepest descent: $z \leftarrow z-\mu \frac{\partial g}{\partial z^{\star}}, \quad \mu>0$\\
\begin{emphbox}
Useful derivatives:\\
$\frac{\partial\left(\boldsymbol{z}^{\mathrm{H}} \boldsymbol{p}+\boldsymbol{p}^{\mathrm{H}} \boldsymbol{z}\right)}{\partial \boldsymbol{z}^{\star}} = \boldsymbol{p}$\\
$\frac{\partial\left(\boldsymbol{z}^{\mathrm{H}} \boldsymbol{R} z\right)}{\partial \boldsymbol{z}^{*}}=\boldsymbol{R}\boldsymbol{z}$\\
$\frac{\partial \operatorname{tr}\left(\boldsymbol{S}^{\mathrm{H}} \boldsymbol{B}\right)}{\partial \boldsymbol{S}^{*}} = \boldsymbol{B}$
\end{emphbox}
\end{sectionbox}
\begin{sectionbox}
\subsection{Quadratic minimization with linear equality constraints}
\textbf{Problem:}\\
$\min _{\boldsymbol{z}} \boldsymbol{z}^{\mathrm{H}} \boldsymbol{R} \boldsymbol{z}, \quad \text { such that } \quad \boldsymbol{A}^{\mathrm{H}} \boldsymbol{z}=\boldsymbol{b}, \quad \boldsymbol{R}=\boldsymbol{R}^{\mathrm{H}}>\boldsymbol{0}$\\
\textbf{Corresponding Lagrange-ian function:}\\
$\mathcal{L}=\boldsymbol{z}^{\mathrm{H}} \boldsymbol{R} \boldsymbol{z}+\boldsymbol{\lambda}^{\mathrm{H}}\left(\boldsymbol{A}^{\mathrm{H}} \boldsymbol{z}-\boldsymbol{b}\right)+\left(\boldsymbol{z}^{\mathrm{H}} \boldsymbol{A}-\boldsymbol{b}^{\mathrm{H}}\right) \boldsymbol{\lambda}$\\
\textbf{Resulting dual optimization problem:} $\min _{z} \max _{\lambda} \mathcal{L}$\\
\textbf{Solution:}
\begin{emphbox}
\vspace{-0.2cm}
$$\boldsymbol{z}_{\mathrm{opt}}=\boldsymbol{R}^{-1} \boldsymbol{A}\left(\boldsymbol{A}^{\mathrm{H}} \boldsymbol{R}^{-1} \boldsymbol{A}\right)^{-1} \boldsymbol{b}$$
$$\operatorname{min} \boldsymbol{z}^H\boldsymbol{R}\boldsymbol{z} = \boldsymbol{b}^H(\boldsymbol{A}^H\boldsymbol{R}^{-1}\boldsymbol{A})^{-1}\boldsymbol{b}$$
\end{emphbox}
\end{sectionbox}
\begin{sectionbox}
\subsection{Real-valued quadratic minimization with linear inequality constraints}
\textbf{Problem:}\\
$\min _{\boldsymbol{x}} \boldsymbol{x}^{\mathrm{T}} \boldsymbol{C} \boldsymbol{x}, \quad \text { subject to } \quad \boldsymbol{A} \boldsymbol{x} \leq \boldsymbol{b}$\\
\textbf{Corresponding Lagrange-ian function:}\\
$\mathcal{L}=\boldsymbol{x}^{\mathrm{T}} \boldsymbol{C} \boldsymbol{x}+\boldsymbol{\lambda}^{\mathrm{T}}(\boldsymbol{A} \boldsymbol{x}-\boldsymbol{b})$
\textbf{Resulting dual optimization problem:} $\min \max \mathcal{L} \text { subject to } \lambda \geq 0$\\
\textbf{Algorithm for determining the solution:}
\begin{center}
\includegraphics[width = 6.8cm]{img/quadr-opt.png}
\end{center}
\end{sectionbox}
\section{Linear Algebra}
\begin{sectionbox}
\subsection{Vector Space}
A complex vector space $\mathcal{V}$ is a set with the following properties:
\begin{enumerate}
\item $\forall \boldsymbol{a}, \boldsymbol{b} \in \mathcal{V}: \quad \boldsymbol{a}+\boldsymbol{b} \in \mathcal{V}$
\item $\forall \boldsymbol{a}, \boldsymbol{b} \in \mathcal{V}: \quad \boldsymbol{a}+\boldsymbol{b}=\boldsymbol{b}+\boldsymbol{a}$
\item $\forall \boldsymbol{a}, \boldsymbol{b}, \boldsymbol{c} \in \mathcal{V}: \quad(\boldsymbol{a}+\boldsymbol{b})+\boldsymbol{c}=\boldsymbol{a}+(\boldsymbol{b}+\boldsymbol{c})$
\item $\forall \boldsymbol{a} \in \mathcal{V}: \exists \mathbf{0} \in \mathcal{V}: \quad \boldsymbol{a}+\mathbf{0}=\boldsymbol{a}$
\item $\forall \boldsymbol{a} \in \mathcal{V}: \exists-\boldsymbol{a} \in \mathcal{V}: \quad \boldsymbol{a}+(-\boldsymbol{a})=\mathbf{0}$
\item $\forall \boldsymbol{a} \in \mathcal{V}: \quad 1 \boldsymbol{a}=\boldsymbol{a}$
\item $\forall \boldsymbol{a} \in \mathcal{V}, \forall \lambda, \mu \in \mathbb{C}: \quad \lambda(\mu \boldsymbol{a})=(\lambda \mu) \boldsymbol{a}$
\item $\forall \boldsymbol{a}, \boldsymbol{b} \in \mathcal{V}, \forall \lambda \in \mathbb{C}: \quad \lambda(\boldsymbol{a}+\boldsymbol{b})=\lambda \boldsymbol{a}+\lambda \boldsymbol{b}$
\item $\forall \boldsymbol{a} \in \mathcal{V}, \forall \lambda, \mu \in \mathbb{C}: \quad(\lambda+\mu) \boldsymbol{a}=\lambda \boldsymbol{a}+\mu \boldsymbol{a}$
\end{enumerate}
\end{sectionbox}
\begin{sectionbox}
\subsection{Linear Subspace}
A set $\mathcal{S}$ is called a subspace of a complex vector space $\mathcal{V}$ iff:
\begin{enumerate}
\item $\mathcal{S} \subseteq \mathcal{V}$ \item $\forall \boldsymbol{a}, \boldsymbol{b} \in \mathcal{V}: \quad \boldsymbol{a}+\boldsymbol{b}=\boldsymbol{b}+\boldsymbol{a}$
\item $\forall \boldsymbol{a}, \boldsymbol{b} \in \mathcal{S}: \quad \boldsymbol{a}+\boldsymbol{b} \in \mathcal{S}$
\item $\forall \boldsymbol{a} \in \mathcal{S}, \forall \lambda \in \mathbb{C}: \quad \lambda \boldsymbol{a} \in \mathcal{S}$
\end{enumerate}
\end{sectionbox}
\begin{sectionbox}
\subsection{Linear (In)dependence}
The vectors $\boldsymbol{v}_{1}, \ldots, \boldsymbol{v}_{n} \in \mathcal{V}$ are said to be linearly independent iff:\\
$\sum_{k=1}^{n} a_{k} \boldsymbol{v}_{k}=\mathbf{0} \quad \Longrightarrow \quad a_{1}=\cdots=a_{n}=0$\\
The vectors are linearly dependent iff:\\
$\exists i: \exists b_{1}, \ldots, b_{i-1}, b_{i+1}, \ldots, b_{n} \in \mathbb{C}: \quad \boldsymbol{v}_{i}=\sum\limits_{k=1, k \neq i}^{n} b_{k} \boldsymbol{v}_{k}$\\
\begin{itemize}
\item $\boldsymbol{v}_{1}, \ldots, \boldsymbol{v}_{n} \in \mathcal{V} \text { are } L I, \text { and } \boldsymbol{s} \in \mathcal{V}$ cannot be expressed as a linear combination, then $\boldsymbol{v}_{1}, \ldots, \boldsymbol{v}_{n}, \boldsymbol{s}$ are LI
\item $\operatorname{dim}(\mathcal{S})$ of a subspace is the maximum number of LI vectors that fit in it
\item For every subspace $\mathcal{S}, \text { with } \operatorname{dim}(\mathcal{S})=n$, and any LI vectors $\boldsymbol{v}_{1}, \ldots, \boldsymbol{v}_{n} \in \mathcal{S}$ we have $\mathcal{S}=\operatorname{Sp}\left(\boldsymbol{v}_{1}, \ldots, \boldsymbol{v}_{n}\right)$
\item Orthonormal vectors are LI
\end{itemize}
\end{sectionbox}
\begin{sectionbox}
\subsection{Gram-Schmidt}
$\boldsymbol{u}_{2}=\boldsymbol{v}_{2}-\boldsymbol{u}_{1} \frac{\boldsymbol{u}_{1}^{\mathrm{H}} \boldsymbol{v}_{2}}{\boldsymbol{u}_{1}^{\mathrm{H}} \boldsymbol{u}_{1}} \dots$\\
\end{sectionbox}
\begin{sectionbox}
\subsection{Matrix Cookbook}
$\operatorname{tr}(\boldsymbol{A} \boldsymbol{B})=\operatorname{tr}(\boldsymbol{B} \boldsymbol{A})$\\
$\operatorname{tr}(\boldsymbol{CDE})=\operatorname{tr}(\boldsymbol{ECD})=\operatorname{tr}(\boldsymbol{DEC})$\\
$(\boldsymbol{A} \boldsymbol{B})^{\mathrm{T}}=\boldsymbol{B}^{\mathrm{T}} \boldsymbol{A}^{\mathrm{T}}$\\
$(\boldsymbol{A B})^{\mathrm{H}}=\boldsymbol{B}^{\mathrm{H}} \boldsymbol{A}^{\mathrm{H}}$\\
$(\boldsymbol{A} \boldsymbol{B})^{-1}=\boldsymbol{B}^{-1} \boldsymbol{A}^{-1}$\\
$\left(\boldsymbol{A}^{-1}\right)^{\mathrm{H}}=\left(\boldsymbol{A}^{\mathrm{H}}\right)^{-1}$\\
\end{sectionbox}
\begin{sectionbox}
\subsection{Determinant}
$det: \mathbb{C}^{m \times m} \ni \boldsymbol{A} \mapsto \operatorname{det} \boldsymbol{A} \in \mathbb{C}$ having the following properties:
\begin{enumerate}
\item $\operatorname{det} \mathbf{I}_{m}=1$
\item If $\boldsymbol{A}$ has LD columns, then $\operatorname{det} \boldsymbol{A}=0$.
\item $\operatorname{det} \boldsymbol{A}$ is linear in the columns of $\boldsymbol{A}$.
\end{enumerate}
For the determinant of MxM matrices the following is true:
\begin{enumerate}
\item For $m=1: \operatorname{det} A=A$.
\item $\operatorname{det} \boldsymbol{A}=\sum_{i=1}^{m}(-1)^{i+j} A_{i, j} \operatorname{det} \boldsymbol{A}^{[i, j]}$, for any $j \in\{1, \ldots m\}$, where $\boldsymbol{A}^{[i, j]}$ is the matrix that results from $\boldsymbol{A}$ if the $i$-th row and the $j$-th column are removed.
\item $\operatorname{det}(\boldsymbol{A} \boldsymbol{B})=(\operatorname{det} \boldsymbol{A})(\operatorname{det} \boldsymbol{B})$
\item $\operatorname{det}\left(\boldsymbol{A}^{-1}\right)=(\operatorname{det} \boldsymbol{A})^{-1}$
\item $\operatorname{det}\left(\boldsymbol{A}^{\mathrm{T}}\right)=\operatorname{det} \boldsymbol{A}$
\end{enumerate}
\end{sectionbox}
\begin{sectionbox}
\subsection{Eigenvalue Decomposition (EVD)}
\begin{emphbox}
$\boldsymbol{A}=\boldsymbol{B} \boldsymbol{\Lambda} \boldsymbol{B}^{-1}$
\end{emphbox}
$\operatorname{tr} \boldsymbol{A}=\sum_{i=1}^{m} \lambda_{i}$\\
$\operatorname{det} \boldsymbol{A}=\prod_{i=1}^{m} \lambda_{i}$\\
$\boldsymbol{A}^{k}=\boldsymbol{B} \boldsymbol{\Lambda}^{k} \boldsymbol{B}^{-1}$\\
Every scalar function which has a Taylor series expansion can be generalized to matrices:\\
$\boldsymbol{h}(\boldsymbol{A})=\boldsymbol{B} \boldsymbol{h}(\boldsymbol{\Lambda}) \boldsymbol{B}^{-1}$\\
\end{sectionbox}
\begin{sectionbox}
\subsection{Hermitian Matrices}
\begin{emphbox}
$\boldsymbol{A}=\boldsymbol{A}^H \in \mathbb{C}^{m\times m}$
\end{emphbox}
Properties:
\begin{itemize}
\item m orthogonal (LI) eigenvectors $\rightarrow$ EVD exists
\item real eigenvalues
\item EVD has the form $\boldsymbol{A}=\boldsymbol{B} \boldsymbol{\Lambda} \boldsymbol{B}^{\mathrm{H}}$
\end{itemize}
\end{sectionbox}
\begin{sectionbox}
\subsection{Gramian Matrices}
\begin{emphbox}
$\exists \boldsymbol{C} \in \mathbb{C}^{m \times m}: \quad \boldsymbol{A}=\boldsymbol{C} \boldsymbol{C}^{\mathrm{H}}$
\end{emphbox}
Properties:
\begin{itemize}
\item Hermitian matrix
\item non-negative real eigenvalues
\item EVD exists
\end{itemize}
\end{sectionbox}
\begin{sectionbox}
\subsection{Sherman-Morrison-Woodbury identity}
$$(\boldsymbol{A}+\boldsymbol{B} \boldsymbol{C} \boldsymbol{D})^{-1}=\boldsymbol{A}^{-1}-\boldsymbol{A}^{-1} \boldsymbol{B}\left(\boldsymbol{C}^{-1}+\boldsymbol{D} \boldsymbol{A}^{-1} \boldsymbol{B}\right)^{-1} \boldsymbol{D} \boldsymbol{A}^{-1}$$
$$\boldsymbol{A} \in \mathbb{C}^{M \times M}, \boldsymbol{C} \in \mathbb{C}^{N \times N}, \boldsymbol{B} \in \mathbb{C}^{M \times N}, \boldsymbol{D} \in \mathbb{C}^{N \times M}$$
$$\operatorname{rank} \boldsymbol{A}=M, \operatorname{rank} \boldsymbol{C}=N$$
Tipps for Reformulation:
\begin{itemize}
\item insert $\boldsymbol{C}^{-1}\boldsymbol{C}$
\item ausklammern wos geht
\end{itemize}
\end{sectionbox}
\begin{sectionbox}
\subsection{Singular Value Decomposition (SVD)}
\begin{emphbox}
$\boldsymbol{A}=\boldsymbol{U} \boldsymbol{\Sigma} \boldsymbol{V}^{\mathrm{H}}$\\
$\boldsymbol{U} \in \mathbb{C}^{m \times m}, \quad \text { with } \quad \boldsymbol{U}^{-1}=\boldsymbol{U}^{\mathrm{H}}$\\
$\boldsymbol{V} \in \mathbb{C}^{n \times n}, \quad \text { with } \quad \boldsymbol{V}^{-1}=\boldsymbol{V}^{\mathrm{H}}$\\
$\boldsymbol{\Sigma} \in \mathbb{R}^{m \times n}, \quad \text { with } \quad \boldsymbol{\Sigma}_{i, j \neq i}=0, \quad \boldsymbol{\Sigma}_{i, i} \geq 0$\\
\end{emphbox}
\begin{itemize}
\item exists for every matrix
\end{itemize}
Relation to SVD:\\
\begin{itemize}
\item $\boldsymbol{V}$: SVD of $\boldsymbol{A}^{\mathrm{H}} \boldsymbol{A}$
\item $\boldsymbol{U}$: SVD of $\boldsymbol{A} \boldsymbol{A}^{\mathrm{H}}$
\item $\boldsymbol{\Sigma}\boldsymbol{\Sigma}^{\mathrm{T}} = \text{diag}(\lambda_1 \dots \lambda_m)$
\item $s_i=\sqrt{\lambda_{i}}, \quad 1 \leq i \leq \min (m, n)$
\end{itemize}
\begin{emphbox}
$\boldsymbol{A}=\boldsymbol{U}_{1} \boldsymbol{\Sigma}_{1} \boldsymbol{V}_{1}^{\mathrm{H}}=\left[\begin{array}{ll}
\boldsymbol{U}_{1} & \boldsymbol{U}_{2}
\end{array}\right]\left[\begin{array}{cc}
\boldsymbol{\Sigma}_{1} & \mathbf{O} \\
\mathbf{O} & \mathbf{O}
\end{array}\right]\left[\begin{array}{c}
\boldsymbol{V}_{1}^{\mathrm{H}} \\
\boldsymbol{V}_{2}^{\mathrm{H}}
\end{array}\right]$
\end{emphbox}
Properties:
\begin{itemize}
\item $\boldsymbol{U}_{1}^{\mathrm{H}} \boldsymbol{U}_{1}=\boldsymbol{V}_{1}^{\mathrm{H}} \boldsymbol{V}_{1}=\mathbf{I}_{r}$\\
\item $\boldsymbol{U}_{2}^{\mathrm{H}} \boldsymbol{U}_{2}=\mathbf{I}_{m-r}$, $\mathbf{V}_{2}^{\mathrm{H}} \mathbf{V}_{2}=\mathbf{I}_{n-r}$\\
\item $\boldsymbol{U}_{1}^{\mathrm{H}} \boldsymbol{U}_{2}=\mathbf{O}_{r,(m-r)}$, $\boldsymbol{U}_{2}^{\mathrm{H}} \boldsymbol{U}_{1}=\mathbf{O}_{(m-r), r}$\\
\item $\boldsymbol{V}_{1}^{\mathrm{H}} \boldsymbol{V}_{2}=\mathbf{O}_{r,(n-r)}$, $\boldsymbol{V}_{2}^{\mathrm{H}} \boldsymbol{V}_{1}=\mathbf{O}_{(n-r), r}$\\
\item $\boldsymbol{U}_{1} \boldsymbol{U}_{1}^{\mathrm{H}}+\boldsymbol{U}_{2} \boldsymbol{U}_{2}^{\mathrm{H}}=\mathbf{I}_{m}$\\
\item $\boldsymbol{V}_{1} \boldsymbol{V}_{1}^{\mathrm{H}}+\boldsymbol{V}_{2} \boldsymbol{V}_{2}^{\mathrm{H}}=\mathbf{I}_{n}$\\
\item $\operatorname{im}\boldsymbol{V}_{1}$ and $\operatorname{im}\boldsymbol{V}_{2}$ are complementary subspaces
\item $\operatorname{im}\boldsymbol{U}_{1}$ and $\operatorname{im}\boldsymbol{U}_{2}$ are complementary subspaces
\end{itemize}
\textbf{Four elementary subspaces of a matrix:}\\
Image/column space:
\begin{itemize}
\item $\operatorname{im} \boldsymbol{A}=\operatorname{im} \boldsymbol{A} \boldsymbol{A}^{\mathrm{H}}=\operatorname{im} \boldsymbol{U}_{1}$
\item $\operatorname{dim}(\operatorname{im} \boldsymbol{A})= \operatorname{rank}(\boldsymbol{A}) = r$
\item $\operatorname{rank} \boldsymbol{A}=\operatorname{rank} \boldsymbol{A}^{\mathrm{H}}=\operatorname{rank} \boldsymbol{A}^{\mathrm{T}}=\operatorname{rank} \boldsymbol{A} \boldsymbol{A}^{\mathrm{H}}=\operatorname{rank} \boldsymbol{A}^{\mathrm{H}} \boldsymbol{A}$
\item $\boldsymbol{P}_{\mathrm{im} \boldsymbol{A}}=\boldsymbol{U}_{1} \boldsymbol{U}_{1}^{\mathrm{H}}$
\end{itemize}
Null space:
\begin{itemize}
\item $\text { null } \boldsymbol{A}=\operatorname{im} V_{2}$
\item $\operatorname{dim}(\text { null } \boldsymbol{A})=n-r$
\item if $r=n$, the nullspace is $\{\mathbf{0}\}$
\item $\boldsymbol{P}_{\text {null } \boldsymbol{A}}=\boldsymbol{V}_{2} \boldsymbol{V}_{2}^{\mathrm{H}}$
\end{itemize}
Left null space:\\
Right image space:\\
\textbf{Complementary subspaces:}\\
$\mathcal{S}_{1} \bigcap \mathcal{S}_{2}=\{\boldsymbol{0}\} \quad \text { and } \quad \mathcal{S}_{1} \bigcup \mathcal{S}_{2}=\mathbb{C}^{m \times 1}$
\end{sectionbox}
\begin{sectionbox}
\subsection{Projectors}
\begin{itemize}
\item $\operatorname{im} \boldsymbol{P}_{\mathcal{S}}=\mathcal{S}$
\item $\boldsymbol{P}_{\mathcal{S}}=\boldsymbol{P}_{\mathcal{S}}^{\mathrm{H}}$
\item $\boldsymbol{P}_{\mathcal{S}} \boldsymbol{P}_{\mathcal{S}}=\boldsymbol{P}_{\mathcal{S}}$
\end{itemize}
\begin{emphbox}
$\boldsymbol{z} \in \mathcal{S} \Longleftrightarrow \boldsymbol{P}_{\mathcal{S}} \boldsymbol{z}=\boldsymbol{z}$
\end{emphbox}
\begin{itemize}
\item Projectors are unique for their subspace
\end{itemize}
\end{sectionbox}
\begin{sectionbox}
\subsection{Eckart-Young}
$\arg \min _{\boldsymbol{B}}\|\boldsymbol{A}-\boldsymbol{B}\|_{\mathrm{F}}^{2}, \quad \text { s.t. } \quad \operatorname{rank} \boldsymbol{B}=k<r=\operatorname{rank} \boldsymbol{A}$
\vspace{-0.2cm}
\begin{emphbox}
$\boldsymbol{B}=\sum_{i=1}^{k} \boldsymbol{u}_{i} s_{i} \boldsymbol{v}_{i}^{\mathrm{H}}$
\end{emphbox}
$$\begin{aligned}\|\boldsymbol{A}-\boldsymbol{B}\|_{\mathrm{F}}^{2}&=\|\boldsymbol{\Sigma}-\boldsymbol{M}\|_{\mathrm{F}}^{2}, \text{ where } \boldsymbol{A}=\boldsymbol{U} \boldsymbol{\Sigma} \boldsymbol{V}^{\mathrm{H}},\boldsymbol{M}=\boldsymbol{U}^{\mathrm{H}} \boldsymbol{B} \boldsymbol{V}\\&=\sum_{i=1}\left|s_{i}-M_{i, i}\right|^{2}+\sum_{i>r}\left|M_{i, i}\right|^{2}+\sum_{i, j \neq i}\left|M_{i, j}\right|^{2}\end{aligned}$$
Minimum: $M_{i,i} = 0, i>r$, $M_{i,j\neq i} = 0$, $M_{i,i} = s_i, i=1\dots k$
\end{sectionbox}
\begin{sectionbox}
\subsection{Frobenius Norm}
$\|\boldsymbol{C}\|_{\mathrm{F}}^{2}=\operatorname{tr}\left(\boldsymbol{C}^{\mathrm{H}} \boldsymbol{C}\right)=\operatorname{tr}\left(\boldsymbol{C}\boldsymbol{C}^{\mathrm{H}}\right)$\\
$\|\boldsymbol{C}\|_{\mathrm{F}}^{2}=\left\|\boldsymbol{U}^{\mathrm{H}} \boldsymbol{C} \boldsymbol{V}\right\|_{\mathrm{F}}^{2}$\\
With $\boldsymbol{M}=\boldsymbol{U}^{\mathrm{H}} \boldsymbol{B} \boldsymbol{V}$: \\
$\|\boldsymbol{A}-\boldsymbol{B}\|_{\mathrm{F}}^{2}=\|\boldsymbol{\Sigma}-\boldsymbol{M}\|_{\mathrm{F}}^{2}=\sum_{i=1}^{r}\left|s_{i}-\boldsymbol{M}_{i, i}\right|^{2}+\sum_{i>r}\left|\boldsymbol{M}_{i, i}\right|^{2}+\sum_{i, j \neq i}\left|\boldsymbol{M}_{i, j}\right|^{2}$\\
\end{sectionbox}
\begin{sectionbox}
\subsection{Linear System of Equations}
\textbf{Exakt solution:}\\
If and only if $b \in \operatorname{im} \boldsymbol{A}$, the system $\boldsymbol{A} \boldsymbol{w}=\boldsymbol{b}$, with $\boldsymbol{A} \in \mathbb{C}^{m \times n}$, has the following exact solution(s):
\begin{emphbox}
$\boldsymbol{w}=\boldsymbol{V}_{1} \boldsymbol{\Sigma}_{1}^{-1} \boldsymbol{U}_{1}^{\mathrm{H}} \boldsymbol{b}+\boldsymbol{V}_{2} \boldsymbol{z}, \quad \text { for any } \quad \boldsymbol{z} \in \mathbb{C}^{(n-r) \times 1}, \quad r=\operatorname{rank} \boldsymbol{A}$
\end{emphbox}
The solution is unique iff. $\boldsymbol{A}$ has full column rank.
\textbf{Minimum Norm Solution:}\\
If $b \in \operatorname{im} \boldsymbol{A}$ and $\boldsymbol{A}$ has full row rank, the solution $w_{\mathrm{MN}} \text {, of } \boldsymbol{A} w_{\mathrm{MN}}=b$, which has the smallest euclidian norm is given by:
\begin{emphbox}
$\boldsymbol{w}_{\mathrm{MN}}=\boldsymbol{A}^{\mathrm{H}}\left(\boldsymbol{A} \boldsymbol{A}^{\mathrm{H}}\right)^{-1} \boldsymbol{b}$ ( full row rank)\\
\end{emphbox}
\textbf{Least Squares Solution:}\\
While for $b \notin \operatorname{im} \boldsymbol{A}$, with $\boldsymbol{A} \in \mathbb{C}^{m \times n}$, there is no exact solution for the system $\boldsymbol{A} \boldsymbol{w}=\boldsymbol{b}$, an approximate solution, $\boldsymbol{w}_{\mathrm{LS}}$, can be defined as:
\begin{emphbox}
$\boldsymbol{w}_{\mathrm{LS}}=\arg \min _{\boldsymbol{w}}\|\boldsymbol{A} \boldsymbol{w}-\boldsymbol{b}\|_{2}^{2}$\\
$\boldsymbol{w}_{\mathrm{LS}}=\left(\boldsymbol{A}^{\mathrm{H}} \boldsymbol{A}\right)^{-1} \boldsymbol{A}^{\mathrm{H}} \boldsymbol{b}$ (full column rank)
\end{emphbox}
\end{sectionbox}
\begin{sectionbox}
\subsection{Pseudoinverse}
$\boldsymbol{A}^{+}= \begin{cases}\boldsymbol{V}_{1} \boldsymbol{\Sigma}_{1}^{-1} \boldsymbol{U}_{1}^{\mathrm{H}} & \text { for } \boldsymbol{A} \neq \mathbf{O} \\ \boldsymbol{A}^{\mathrm{T}} & \text { else. }\end{cases}$\\
$\boldsymbol{A}^{+}= \begin{cases}\boldsymbol{A}^{\mathrm{H}}\left(\boldsymbol{A} \boldsymbol{A}^{\mathrm{H}}\right)^{-1} & \text { for full row-rank } \boldsymbol{A} \\ \left(\boldsymbol{A}^{\mathrm{H}} \boldsymbol{A}\right)^{-1} \boldsymbol{A}^{\mathrm{H}} & \text { for full column-rank } \boldsymbol{A} \\ \lim _{\epsilon \rightarrow 0} \boldsymbol{A}^{\mathrm{H}}\left(\boldsymbol{A} \boldsymbol{A}^{\mathrm{H}}+\epsilon \mathbf{I}\right)^{-1} & \\ \lim _{\epsilon \rightarrow 0}\left(\boldsymbol{A}^{\mathrm{H}} \boldsymbol{A}+\epsilon \mathbf{I}\right)^{-1} \boldsymbol{A}^{\mathrm{H}} & \end{cases}$\\
Relation to the projectors:
\begin{itemize}
\item $\boldsymbol{P}_{\mathrm{im} \boldsymbol{A}}=\boldsymbol{A} \boldsymbol{A}^{+}=\boldsymbol{U}_{1} \boldsymbol{U}_{1}^{\mathrm{H}}$
\item $\boldsymbol{P}_{\text {null } \boldsymbol{A}}=\boldsymbol{I}-\boldsymbol{A}^{+} \boldsymbol{A} =\boldsymbol{V}_{2} \boldsymbol{V}_{2}^{\mathrm{H}}$
\end{itemize}
\end{sectionbox}
\section{Random Processes}
\begin{sectionbox}
\subsection{Discrete random processes}
Definition: function $x[n]$ which is selected from an ensemble of possible functions by random.
Properties obtained by averaging over the ensemble:
\begin{emphbox}
Expectation function:\\
$\mu_{x}[n]=\mathrm{E}[x[n]]$\\
Autocorrelation function:\\
$r_{x}[n, k]=\mathrm{E}[x[n] x[n-k]]$\\
Autocovariance function:\\
$\begin{aligned}
c_{x}[n, k] &=\mathrm{E}\left[\left(x[n]-\mu_{x}[n]\right)\left(x[n-k]-\mu_{x}[n-k]\right)\right] \\
&=r_{x}[n, k]-\mu_{x}[n] \mu_{x}[n-k]
\end{aligned}$\\
Cross-correlation function:\\
$r_{x, y}[n, k]=\mathrm{E}[x[n] y[n-k]]$\\
Cross-covariance function:\\
$\begin{aligned}
c_{x, y}[n, k] &=\mathrm{E}\left[\left(x[n]-\mu_{x}[n]\right)\left(y[n-k]-\mu_{y}[n-k]\right)\right] \\
&=r_{x, y}[n, k]-\mu_{x}[n] \mu_{y}[n-k]
\end{aligned}$\\
\end{emphbox}
\end{sectionbox}
\begin{sectionbox}
\subsection{Approximations obtained by averaging over time:}
\begin{emphbox}
Expectation function:\\
$\hat{\mu}_{x}^{[N]}[n]=\frac{1}{2 N+1} \sum_{i=-N}^{N} x[n+i]$\\
Cross-correlation function:\\
$\hat{r}_{x, y}^{[N]}[n, k]=\frac{1}{2 N+1} \sum_{i=-N}^{N} x[n+i] y[n+i-k]$\\
\end{emphbox}
The approximations are random processes themselves.
\end{sectionbox}
\begin{sectionbox}
\subsection{Wide-sense-stationary (WSS) processes}
$\mu_{x}[n], r_{x}[n, k], c_{x}[n, k], r_{x, y}[n, k]$ and $c_{x, y}[n, k]$ are independent of the time index $n$.
\begin{emphbox}
$\mu_{x}=\mathrm{E}[x[n]]$\\
$r_{x}[k]=\mathrm{E}[x[n] x[n-k]]$, $r_{x}[k]=r_{x}[-k]$\\
$r_{x, y}[k]=\mathrm{E}[x[n] y[n-k]]$, $r_{x, y}[k]=r_{y, x}[-k]$\\
$\mathrm{E}\left[\hat{\mu}_{x}^{[N]}[n]\right]=\mu_{x}, \quad \mathrm{E}\left[\hat{r}_{x}^{[N]}[n, k]\right]=r_{x}[k]$
\end{emphbox}
For zero-mean processes: $\mu_{x}=0, \quad \mu_{y}=0, \quad \ldots$: correlation and covariance are identical
\end{sectionbox}
\begin{sectionbox}
\subsection{Ergodic processes}
Averages over ensembles can be purely obtained from averages over time. Ergodicity implies WSS.
\begin{emphbox}
$\lim _{N \rightarrow \infty} \hat{\mu}_{x}^{[N]}[n]=\mu_{x}$\\
$\lim _{N \rightarrow \infty} \hat{r}_{x}^{[N]}[n, k]=r_{x}[k]$
\end{emphbox}
\end{sectionbox}
\begin{sectionbox}
\subsection{Complex processes}
$u[n]=x[n]+\mathrm{j} y[n]$\\
Complex autocorrelation function:\\
$r_{u}[k]=\mathrm{E}\left[u[n] u^{*}[n-k]\right]$\\
$r_{u}[k]=r_{x}[k]+r_{y}[k]+\mathrm{j}\left(r_{x, y}[-k]-r_{x, y}[k]\right)$\\
Adjunct complex autocorrelation function:\\
$\tilde{r}_{u}[k]=\mathrm{E}[u[n] u[n-k]]$\\
$\tilde{r}_{u}[k]=r_{x}[k]-r_{y}[k]+\mathrm{j}\left(r_{x, y}[k]+r_{x, y}[-k]\right)$
Reformulation:\\
$r_{x}[k] =\frac{1}{2} \operatorname{Re}\left\{r_{u}[k]+\tilde{r}_{u}[k]\right\}$ \\
$r_{y}[k] =\frac{1}{2} \operatorname{Re}\left\{r_{u}[k]-\tilde{r}_{u}[k]\right\}$ \\
$r_{x, y}[k] =-\frac{1}{2} \operatorname{Im}\left\{r_{u}[k]-\tilde{r}_{u}[k]\right\}$
\end{sectionbox}
\begin{sectionbox}
\subsection{Proper WSS processes}
Equivalent definitions (iff):\\
$\forall k: \quad \tilde{r}_{u}[k]=0$\\
$\forall k: \quad r_{x}[k]=r_{y}[k], \quad \text { and } \quad \forall k: \quad r_{x, y}[k]=-r_{x, y}[-k]$\\
$\mathrm{E}\left[\boldsymbol{u}[n] \boldsymbol{u}^{\mathrm{T}}[n]\right]=\left[\begin{array}{cccc}
\tilde{r}_{u}[0] & \tilde{r}_{u}[1] & \tilde{r}_{u}[2] & \cdots \\
\tilde{r}_{u}[1] & \tilde{r}_{u}[0] & \tilde{r}_{u}[1] & \cdots \\
\vdots & \ddots & \ddots & \vdots
\end{array}\right]=\mathbf{O}$
$\Rightarrow$ The autocorrelation function is completely described by the complex autocorrelation function.
\end{sectionbox}
\section{Overview}
\begin{sectionbox}
\subsection{Linear estimation for matrices}
\textbf{Problem:} Find $\hat{\boldsymbol{S}}$ from $\boldsymbol{X}$
$$\boldsymbol{X}=\boldsymbol{A} \boldsymbol{S}+\boldsymbol{\Upsilon}$$
Four different cases:
\begin{enumerate}
\item nothing is known (except structure) $\rightarrow$ MUSIC
\item only $\boldsymbol{A}$ is known $\rightarrow$ Least Squares
\item $\boldsymbol{A}$ and $\mathrm{E}\left[\boldsymbol{\Upsilon} \boldsymbol{\Upsilon}^{\mathrm{H}}\right]$ are known $\rightarrow$ BLUE
\item $\boldsymbol{A}$, $E\left[\boldsymbol{\Upsilon} \boldsymbol{\Upsilon}^{\mathrm{H}}\right]$ and $\mathrm{E}\left[\boldsymbol{S} \boldsymbol{S}^{\mathrm{H}}\right]$ are known
\end{enumerate}
\end{sectionbox}
\section{Kolmogorov-Wiener Filters}
\begin{sectionbox}
\subsection{Linear Filters}
$$u = h * s$$
If $h$ has $K+1$ coefficients, its memory is $K$.\\
If $M$ output samples are required: $\boldsymbol{H} \in \mathbb{C}^{M\times(M+K)}$
\end{sectionbox}
\begin{sectionbox}
\subsection{Kolmogorov-Wiener filter for SISO/$\quad\quad\quad\quad\quad\quad\quad$ time domain equalizer for the linear multipath channel}
% $h[n]=\sum_{k=0}^{K} h[k] \delta[n-k]$\\
% $u[n]=\sum_{k=0}^{K} h[k] s[n-k]+v[n]$\\
% $y[n]=\sum_{m=0}^{M-1} w^{*}[m] u[n-m]$\\
% K: channel memory\\
\textbf{Optimization problem:}
\begin{emphbox}
$\mathbf{M S E}=\mathrm{E}[|\underbrace{y[n]-d[n]}_{e[n]}|^{2}]$\\
$(w[0], \ldots, w[M-1])_{\mathrm{opt}}=\arg \min _{(w[0], \ldots, w[M-1])} \operatorname{MSE}$
\end{emphbox}
$\boldsymbol{u}[n] \in \mathbb{C}^{M \times 1}$, $\boldsymbol{s}[n] \in \mathbb{C}^{N \times 1}$, $\boldsymbol{H} \in \mathbb{C}^{M \times N}$, $N=M+K$\\
\begin{emphbox}
$y[n]=\boldsymbol{w}^{\mathrm{H}} \boldsymbol{u}[n]=\boldsymbol{w}^{\mathrm{H}}(\boldsymbol{H s}[n]+\boldsymbol{v}[n])$\\
\end{emphbox}
$\boldsymbol{R} = \mathrm{E}\left[\boldsymbol{u}[n] \boldsymbol{u}^{\mathrm{H}}[n]\right] = \boldsymbol{H R}_{s} \boldsymbol{H}^{\mathrm{H}}+\boldsymbol{H R}_{s, v}+\boldsymbol{R}_{v, s} \boldsymbol{H}^{\mathrm{H}}+\boldsymbol{R}_{v}$\\
$\boldsymbol{p} = \mathrm{E}\left[\boldsymbol{u}[n] d^{*}[n]\right]$\\
$\sigma_{s}^{2} = \mathrm{E}\left[d[n]d^*[n]\right]$
\textbf{Solutions in terms of u:}
\begin{emphbox}
$\mathrm{E}\left[\boldsymbol{u}[n] e^{*}[n]\right]=0$ (principle of orthogonality)\\
$\boldsymbol{w}_{\mathrm{opt}}=\boldsymbol{R}^{-1} \boldsymbol{p}$, or: $\boldsymbol{w}_{\mathrm{opt}}^{\mathrm{H}}=\boldsymbol{p}^{\mathrm{H}} \boldsymbol{R}^{-1}$\\
$\mathrm{MSE}_{\min }=\sigma_{s}^{2}-\boldsymbol{p}^{\mathrm{H}} \boldsymbol{R}^{-1} \boldsymbol{p}$\\
\end{emphbox}
$\boldsymbol{w}_{\mathrm{opt}}$ minimizes MSE if: $\boldsymbol{R}>0 \Leftrightarrow \boldsymbol{R}$ invertible ($ \boldsymbol{R}$ is Gramian)\\
In general: $\mathbf{M S E}=\boldsymbol{w}^{\mathrm{H}} \boldsymbol{R} \boldsymbol{w}-\boldsymbol{w}^{\mathrm{H}} \boldsymbol{p}-\boldsymbol{p}^{\mathrm{H}} \boldsymbol{w}+\sigma_{s}^{2}$\\
Noise variance: $E[(\boldsymbol{w}_{\mathrm{opt}}^H\boldsymbol{v}[n])(\boldsymbol{w}_{\mathrm{opt}}^H\boldsymbol{v}[n])^H]$\\
Overall impulse response: coefficients given by $\boldsymbol{w}_{\mathrm{opt}}^H\boldsymbol{H}$\\
\textbf{Special case: }$d[n] = s[n-l]$, $s$ and $v$ are uncorrelated
\begin{emphbox}
$\begin{aligned}\boldsymbol{w}_{\mathrm{opt}}&=\left(\boldsymbol{H} \boldsymbol{R}_{\boldsymbol{s}} \boldsymbol{H}^{\mathrm{H}}+\boldsymbol{R}_{\boldsymbol{v}}\right)^{-1} \boldsymbol{H} \boldsymbol{R}_{\boldsymbol{s}} \mathbf{e}_{l+1}\\&=\boldsymbol{R}_{\boldsymbol{v}}^{-1} \boldsymbol{H}\left(\boldsymbol{R}_{\boldsymbol{s}}^{-1}+\boldsymbol{H}^{\mathrm{H}} \boldsymbol{R}_{\boldsymbol{v}}^{-1} \boldsymbol{H}\right)^{-1} \mathbf{e}_{l+1}\end{aligned}$
\end{emphbox}
$\boldsymbol{R} = \boldsymbol{H R}_{s} \boldsymbol{H}^{\mathrm{H}} + \boldsymbol{R}_{v}$\\
$\boldsymbol{p} = \boldsymbol{H} \boldsymbol{R}_{s} \mathbf{e}_{l+1}$\\
$\boldsymbol{R}_s = \mathrm{E}\left[\boldsymbol{s}[n] \boldsymbol{s}^{\mathrm{H}}[n]\right]$\\
$\boldsymbol{R}_{s,v} = \mathrm{E}\left[\boldsymbol{s}[n] \boldsymbol{v}^{\mathrm{H}}[n]\right] = \boldsymbol{O}$\\
$\boldsymbol{R}_v = \mathrm{E}\left[\boldsymbol{v}[\boldsymbol{n}] \boldsymbol{v}^{\mathrm{H}}[\boldsymbol{n}]\right]$\\
Optimization of $l_{\text{opt}}$:\\
$\begin{aligned}l_{\mathrm{opt}}&=\arg \min _{l \in\{0,1, \ldots N-1\}} \sigma_{s}^{2}-\boldsymbol{p}^{\mathrm{H}} \boldsymbol{R}^{-1} \boldsymbol{p}\\&=\arg \max _{l \in\{0,1, \ldots N-1\}} \boldsymbol{p}^{\mathrm{H}} \boldsymbol{R}^{-1} \boldsymbol{p}\\&=\text{argmin} \sigma_{s}^{2}-\mathbf{e}_{l+1}^{\mathrm{T}} \boldsymbol{R}_{\boldsymbol{s}} \boldsymbol{H}^{\mathrm{H}}\left(\boldsymbol{H} \boldsymbol{R}_{\boldsymbol{s}} \boldsymbol{H}^{\mathrm{H}}+\boldsymbol{R}_{v}\right)^{-1} \boldsymbol{H} \boldsymbol{R}_{\boldsymbol{s}} \mathbf{e}_{l+1}\end{aligned}$\\
This is done by trying a few neighbors of $N/2 = (M+K)/2$.\\
Noise variance: $\sigma_n^2 = E[(\boldsymbol{w}_{\mathrm{opt}}^H\boldsymbol{v}[n])(\boldsymbol{w}_{\mathrm{opt}}^H\boldsymbol{v}[n])^H]$\\
Overall impulse response: coefficients given by $\boldsymbol{h} \leftarrow \boldsymbol{w}_{\mathrm{opt}}^H\boldsymbol{H}$\\
Signal-to-interference ratio: $$\text{SINR} = \frac{\sigma_s^{\prime 2}|h[l]|^2}{\sigma_s^{\prime 2}\sum_{i, i\neq l} |h[i]|^2 + \sigma_n^2}, \quad\text{assuming}\, \boldsymbol{R}_s = \sigma_s^2 \boldsymbol{I}$$.
\end{sectionbox}
\begin{sectionbox}
\subsection{Kolmogorov-Wiener filter for diversity reception}
L multipath channels with just \textbf{one transmit signal}:\\
$\boldsymbol{u}_{i}[n]=\boldsymbol{H}_{i} \boldsymbol{s}[n]+\boldsymbol{v}_{i}[n], i \in\{1,2, \ldots L\}$, $\boldsymbol{H}_{i} \in \mathbb{C}^{M\times(M+K)}$
$$\underbrace{\left[\begin{array}{c}
u_{1}[n] \\
u_{2}[n] \\
\vdots \\
u_{L}[n]
\end{array}\right]}_{\boldsymbol{u}[n]}=\underbrace{\left[\begin{array}{c}
H_{1} \\
H_{2} \\
\vdots \\
H_{L}
\end{array}\right]}_{\boldsymbol{H}\in\mathbb{C}^{LM\times(M+K)}} s[n]+\underbrace{\left[\begin{array}{c}
v_{1}[n] \\
v_{2}[n] \\
\vdots \\
v_{L}[n]
\end{array}\right]}_{\boldsymbol{v}[n]}$$
Same formula as for SISO but with different matrices:\\
$\boldsymbol{w}_{\mathrm{opt}}=\left(\boldsymbol{H} \boldsymbol{R}_{\boldsymbol{s}} \boldsymbol{H}^{\mathrm{H}}+\boldsymbol{R}_{\boldsymbol{v}}\right)^{-1} \boldsymbol{H} \boldsymbol{R}_{\boldsymbol{s}} \mathbf{e}_{l+1}$\\
$\boldsymbol{R}_{\boldsymbol{s}}=\mathrm{E}\left[\boldsymbol{s}[n] \boldsymbol{s}[n]^{\mathrm{H}}\right] \in \mathbb{C}^{(M+K) \times(M+K)}$\\
$\boldsymbol{R}_{\boldsymbol{v}}=\mathrm{E}\left[\boldsymbol{v}[n] \boldsymbol{v}[n]^{\mathrm{H}}\right] \in \mathbb{C}^{L M \times L M}$\\
$y[n]=\boldsymbol{w}_{\mathrm{opt}}^{\mathrm{H}} \boldsymbol{u}[n]$\\
usually $l_{\text{opt}}=\lfloor(M+K) / 2\rfloor$\\
Diversity reception
\begin{itemize}
\item one signal is transmitted over multiple channels
\item others can still be used if one breaks (redundancy)
\item space-diversity, frequency-diversity, time-diversity, or combinations
\end{itemize}
\end{sectionbox}
\begin{sectionbox}
\subsection{Kolmogorov-Wiener filter for multi-streaming}
L channels with \textbf{different transmit signals}:\\
\includegraphics[width = 6.8cm]{img/wiener-multi.png}
$\boldsymbol{H}_{i, j} \in \mathbb{C}^{M \times(M+K)}$ connects j-th transmitter to i-th receiver\\
$\mathbf{H} \in \mathbb{C}^{M m_{\mathrm{Rx}} \times(M+K) m_{\mathrm{Tx}}}$\\
$\boldsymbol{s}[n] \in \mathbb{C}^{(M+K) m_{\mathrm{Tx}}}$\\
$\boldsymbol{v}[n] \in \mathbb{C}^{M m_{\mathrm{Rx}}}$\\
$K$ is the maximum channel memory of all channels\\
Main difference to before: Ordering of the elements in $\boldsymbol{s}[n]$ has changed to groups of time series.\\
\textbf{Solution:}\\
$$\boldsymbol{w}_{j, \mathrm{opt}}=\underbrace{\left(\boldsymbol{H} \boldsymbol{R}_{\boldsymbol{s}} \boldsymbol{H}^{\mathrm{H}}+\boldsymbol{R}_{v}\right)^{-1}}_{\boldsymbol{R}^{-1}} \underbrace{\boldsymbol{H} \boldsymbol{R}_{\boldsymbol{s}} \mathbf{e}_{(M+K)(j-1)+l_{j+1}}}_{\boldsymbol{p}}$$
$$y_{j}[n]=\boldsymbol{w}_{j, \mathrm{opt}}^{\mathrm{H}} \boldsymbol{u}[n], \quad j \in\left\{1,2, \ldots, m_{\mathrm{Tx}}\right\}$$
$\boldsymbol{w}_{j, \mathrm{opt}}$ is the optimum filter for recovering the signal of the j-th transmitter with time delay $l_j \in \{0,\dots,M+K-1\}$
\end{sectionbox}
\begin{sectionbox}
\subsection{Kolmogorov-Wiener Filter for SISO}
\textbf{Setup:}
$$\boldsymbol{u} = \boldsymbol{h}s+\boldsymbol{v}, \quad \boldsymbol{s} = \boldsymbol{w}^H\boldsymbol{u}$$
$$\boldsymbol{R}_{\boldsymbol{s}}=1, \quad \boldsymbol{R}_{v}=\mathbf{I},\quad s,v \text{ uncorrelated}$$
\textbf{MSE solution:}
$$\boldsymbol{w} = (\boldsymbol{I} + \boldsymbol{h}\boldsymbol{h}^H)^{-1}\boldsymbol{h} =\frac{\boldsymbol{h}}{1+\boldsymbol{h}^H\boldsymbol{h}}$$
\end{sectionbox}
\begin{sectionbox}
\subsection{General Properties of Kolmogorov-Wiener}
\begin{itemize}
\item Linear filter: can only use first order correlations of $d$ and $u$, $\quad\quad\quad\quad$ $\rightarrow$ $p$ must be $\neq 0$
\item Filtering without noise: MSE drops exponentially when increasing reconstruction filter length (if $\boldsymbol{H}$ has full column rank)
\item Filtering with noise: MSE saturates and cannot drop exponentially
\end{itemize}
\end{sectionbox}
\begin{sectionbox}
\subsection{Kolmogorov-Wiener Filter for SNR $\rightarrow \infty$}
\textbf{Setup:}
$$\boldsymbol{u} = \boldsymbol{H}\boldsymbol{s}+\boldsymbol{v}, \quad \hat{\boldsymbol{s}} = \boldsymbol{w}^H\boldsymbol{u}$$
$$\boldsymbol{R}_{\boldsymbol{s}}=\sigma_{s}^{2} \mathbf{I}, \quad \boldsymbol{R}_{v}=\sigma_{v}^{2} \mathbf{I},\quad s,v \text{ uncorrelated}$$
\textbf{MSE solution:}
$$\begin{aligned}\lim _{\sigma_{s}^{2} / \sigma_{v}^{2} \rightarrow \infty} \boldsymbol{w}_{\mathrm{opt}}^{\mathrm{H}} &= \begin{cases}\mathbf{e}_{l+1}^{\mathrm{T}} \boldsymbol{H}^{\mathrm{H}}\left(\boldsymbol{H} \boldsymbol{H}^{\mathrm{H}}\right)^{-1}, & \text {full row rank } \boldsymbol{H} \\ \mathbf{e}_{l+1}^{\mathrm{T}}\left(\boldsymbol{H}^{\mathrm{H}} \boldsymbol{H}\right)^{-1} \boldsymbol{H}^{\mathrm{H}}, & \text {full column rank}\boldsymbol{H}\end{cases}\\&=\mathbf{e}_{l+1}^{\mathrm{T}} \boldsymbol{H}^{+}\end{aligned}$$
Perfect reconstruction only if $H$ has full column rank:\\
$$\hat{\boldsymbol{s}}_{l+1}=\mathbf{e}_{l+1}^{\mathrm{T}} \boldsymbol{H}^{+} \boldsymbol{H} \boldsymbol{s}=\mathbf{e}_{l+1}^{\mathrm{T}} \underbrace{\left(\boldsymbol{H}^{\mathrm{H}} \boldsymbol{H}\right)^{-1} \boldsymbol{H}^{\mathrm{H}} \boldsymbol{H}}_{\mathbf{I}} \boldsymbol{s}=\boldsymbol{s}_{l+1}$$
Imperfect reconstruction if $H$ has not full column rank:\\
$$\hat{s}_{l+1}=\mathbf{e}_{l+1}^{\mathrm{T}} \underbrace{\boldsymbol{H}^{+} \boldsymbol{H}}_{\neq \mathbf{I}} \boldsymbol{s} \neq \boldsymbol{s}_{l+}$$
Because with $\operatorname{rank} \boldsymbol{H}<N$:\\
$$\boldsymbol{H}^{+} \boldsymbol{H}=\mathbf{I}-\boldsymbol{P}_{\text {null } \boldsymbol{H}}$$
$$\operatorname{dim} \text { null } \boldsymbol{H}=N-\operatorname{rank} \boldsymbol{H}>0$$
$$\boldsymbol{P}_{\text {null } \boldsymbol{H}} \neq \mathbf{O}$$
\end{sectionbox}
\begin{sectionbox}
\subsection{Steepest Descent Algorithm (SDA)}
Kolmogorov-Wiener filters need to solve $\boldsymbol{R}\boldsymbol{w}_{\text{opt}} = \boldsymbol{p}$, where:\\
$\boldsymbol{R}=\mathrm{E}\left[\boldsymbol{u}[n] \boldsymbol{u}^{\mathrm{H}}[n]\right]$ and $\boldsymbol{p}=\mathrm{E}\left[\boldsymbol{u}[n] d^{*}[n]\right]$\\
Problems with solving explicitly
\begin{itemize}
\item Accuracy: computation of $m^2$ complex numbers for $R^{-1}$ and matrix vector multiplication add errors when $m$ is large
\item Computational load: Recomputing $R^{-1}$ at every time step is costly
\item Gaussian elimination needs to restart the entire computation every time
\item Triangular schemes bring no benefit as $R$ also changes
\end{itemize}
\textbf{Gradient of MSE}:\\
$\operatorname{MSE}(w)=F\left(\frac{1}{2}\left(w+w^{\star}\right), \frac{1}{2}\left(w-w^{\star}\right) / \mathrm{j}\right)=G\left(w, w^{*}\right)$\\
$\begin{aligned}\mathrm{dMSE}&=\left(\frac{\partial G}{\partial \boldsymbol{w}}\right)^{\mathrm{T}} \mathrm{d} \boldsymbol{w}+\left(\frac{\partial G}{\partial \boldsymbol{w}^{*}}\right)^{\mathrm{T}} \mathrm{d} \boldsymbol{w}^{*}\\&=\left(\frac{\partial G^{*}}{\partial \boldsymbol{w}^{*}}\right)^{\mathrm{H}} \mathrm{d} \boldsymbol{w}+\left(\left(\frac{\partial G}{\partial \boldsymbol{w}^{*}}\right)^{\mathrm{H}} \mathrm{d} \boldsymbol{w}\right)^{*}\\&=2 \operatorname{Re}\left\{\left(\frac{\partial G}{\partial w^{\star}}\right)^{\mathrm{H}} \mathrm{d} w\right\} = 2 \operatorname{Re}\left\{(\boldsymbol{R} w-\boldsymbol{p})^{\mathrm{H}} \mathrm{d} \boldsymbol{w}\right\}\\&\leq 2\left|(\boldsymbol{R} w-\boldsymbol{p})^{\mathrm{H}} \mathrm{d} \boldsymbol{w}\right|, \text{ with equality for } \mathrm{d} \boldsymbol{w}=(\boldsymbol{R} \boldsymbol{w}-\boldsymbol{p}) \mathrm{d} t\end{aligned}$\\
\textbf{Gradient descent}:\\
$$\boldsymbol{w}_{n+1}=\boldsymbol{w}_{n}-\left(\boldsymbol{R} \boldsymbol{w}_{n}-\boldsymbol{p}\right) \mu, \quad \mu>0$$
$$\boldsymbol{w}_{n+1}=(\mathbf{I}-\mu \boldsymbol{R}) \boldsymbol{w}_{n}+\mu \boldsymbol{p}$$
\textbf{Convergence}:\\
$$\boldsymbol{c}_{n}=\boldsymbol{w}_{n}-\boldsymbol{w}_{\text {opt }}, \boldsymbol{c}_{n+1}=(\mathbf{I}-\mu \mathbf{R}) \boldsymbol{c}_{n}$$
$$\boldsymbol{R}=\boldsymbol{Q}\boldsymbol{\Lambda}\boldsymbol{Q}^H, \boldsymbol{z}_{n}=\mathbf{Q}^{\mathrm{H}} \boldsymbol{c}_{n}, \boldsymbol{z}_{n+1}=(\mathbf{I}-\boldsymbol{\mu} \boldsymbol{\Lambda}) \boldsymbol{z}_{n}$$
$$\Rightarrow \forall i \in\{1,2, \ldots, m\}: \left|1-\mu \lambda_{i}\right|<1 \Leftrightarrow 0< \mu < 2/\lambda_{\text{max}}$$
sufficient (not necessary): $0<\mu<\frac{2}{\operatorname{tr} R}$\\
\textbf{Steepest Descent Algorithm}:\\
%$$\boldsymbol{w}_{n+1}=\boldsymbol{w}_{n}+\mu \boldsymbol{r}_{n}, \boldsymbol{r}_{n}=\boldsymbol{p}-\boldsymbol{R} \boldsymbol{w}_{n}, \mu_{\mathrm{opt}}=\frac{\boldsymbol{r}_{n}^{\mathrm{H}} \boldsymbol{r}_{n}}{\boldsymbol{r}_{n}^{\mathrm{H}} \boldsymbol{R} \boldsymbol{r}_{n}}$$
$\mu_{\mathrm{opt}}$ is obtained by minimizing $\operatorname{MSE}_{n+1}$ wrt. $\mu$, i.e. $\frac{\partial \mathrm{MSE}_{n+1}}{\partial \mu^{*}} = 0$
$\mathbf{M S E}_{n}=\boldsymbol{w}_{n}^{\mathrm{H}} \boldsymbol{R} \boldsymbol{w}_{n}-\boldsymbol{w}_{n}^{\mathrm{H}} \boldsymbol{p}-\boldsymbol{p}^{\mathrm{H}} \boldsymbol{w}_{n}+\sigma_{d}^{2}$\\
$\boldsymbol{w}_{n+1}=\boldsymbol{w}_{n}+\mu \boldsymbol{r}_{n},\quad \boldsymbol{r}_{n} = \boldsymbol{p}-\boldsymbol{R}\boldsymbol{w}_{n}$\\
\end{sectionbox}
\begin{sectionbox}
\includegraphics[width = 6.8cm]{img/sda.png}
Complexity: $M^{2} \log M$ scalar arithmetic operations/series time-steps\\
Complexity with full parallelization: $(\log M)^{2}$ time-steps\\
Steps 2-5 require $4m^2+5n-1$ scalar arithmetic operations per iteration\\
Number of iterations: $\approx 15(\operatorname{log}(M)-1), M > 15$
\end{sectionbox}
\begin{sectionbox}
\subsection{SDA for constrained optimization}
\textbf{Problem:}\\
$$\min _{w} \boldsymbol{w}^{\mathrm{H}} \boldsymbol{A} \boldsymbol{w}, \quad \text { s.t. } \quad \boldsymbol{B}^{\mathrm{H}} \boldsymbol{w}=\boldsymbol{c}$$
$$\mathbb{C}^{M \times M} \ni \boldsymbol{A}=\boldsymbol{A}^{\mathrm{H}}>\mathbf{0}, \boldsymbol{\boldsymbol{B}} \in \mathbb{C}^{M \times L}, \boldsymbol{c} \in \mathbb{C}^{L \times 1}, \text { rank } \boldsymbol{B}=L$$
\textbf{Ansatz:}\\
Decompose solution into fixed term $\boldsymbol{w}_{q}$ and variable term $\boldsymbol{z}$\\
$$\boldsymbol{B}^{\mathrm{H}} \boldsymbol{w}_{\mathrm{q}}=\boldsymbol{c}, \quad \boldsymbol{z} \in \text { null } \boldsymbol{B}^{\mathrm{H}}$$
$$\boldsymbol{w}=\boldsymbol{w}_{q}-\boldsymbol{z}$$
Parameterize $\boldsymbol{z}$ by $\boldsymbol{w}_a$:
$$\boldsymbol{B}=\underbrace{\left[\begin{array}{ll}
\boldsymbol{U}_{1} & \boldsymbol{U}_{2}
\end{array}\right]}_{\boldsymbol{U}}\left[\begin{array}{ll}
\boldsymbol{\Sigma}_{1} & \mathbf{O} \\
\mathbf{O} & \mathbf{O}
\end{array}\right]\left[\begin{array}{l}
\boldsymbol{V}_{1}^{\mathrm{H}} \\
\boldsymbol{V}_{2}^{\mathrm{H}}
\end{array}\right]$$
$$\boldsymbol{z}=\boldsymbol{U}_{2} \boldsymbol{w}_{\mathrm{a}}, \quad \boldsymbol{w}_{\mathrm{a}} \in \mathbb{C}(M-L) \times 1 \Rightarrow \boldsymbol{z} \in \text{null}\boldsymbol{B}^H \forall \boldsymbol{w}_a$$
Obtaining $\boldsymbol{U}_{2}$ without SVD:
\begin{enumerate}
\item Init: $\boldsymbol{U} \leftarrow\left[\begin{array}{ll}\boldsymbol{B} & \boldsymbol{F}\end{array}\right]$, where $\boldsymbol{F} \in \mathbb{C}^{M \times(M-L)}$ has i.i.d. random components.
\item Orthogonalize with all yet orthogonalized columns.\\For $i=0$ to $M-1$:
$$\boldsymbol{u}_{i} \leftarrow \boldsymbol{u}_{i}-\sum_{j=1}^{i-1} \boldsymbol{u}_{j}\left(\boldsymbol{u}_{j}^{\mathrm{H}} \boldsymbol{u}_{i}\right) /\left(\boldsymbol{u}_{j}^{\mathrm{H}} \boldsymbol{u}_{j}\right)$$
\item Normalize: For $i \in\{L+1, L+2, \ldots M\}$ do $\boldsymbol{u}_{i} \leftarrow \boldsymbol{u}_{i} / \sqrt{\boldsymbol{u}_{i}^{\mathrm{H}} \boldsymbol{u}_{i}}$
\item Output: $\boldsymbol{U}_{2} \leftarrow\left[\boldsymbol{u}_{L+1} \boldsymbol{u}_{L+2} \cdots \boldsymbol{u}_{M}\right] \in \mathbb{C}^{M \times(M-L)}$
\end{enumerate}
\vspace{0.2cm}
Finding $w_q$ by SDA:\\
$$w_{q}=\boldsymbol{B} \underbrace{\left(\boldsymbol{B}^{\mathrm{H}} \boldsymbol{B}\right)^{-1} c}_{q}$$
$$\left(\boldsymbol{B}^{\mathrm{H}} \boldsymbol{B}\right) q=c$$
$$w_{q}=\boldsymbol{B} q$$
Reformulate:\\
$$\min _{\boldsymbol{w}_{\mathrm{a}}}\left(\boldsymbol{w}_{\mathrm{q}}^{\mathrm{H}}-\boldsymbol{w}_{\mathrm{a}}^{\mathrm{H}} \boldsymbol{U}_{2}^{\mathrm{H}}\right) \boldsymbol{A}\left(\boldsymbol{w}_{\mathrm{q}}-\boldsymbol{U}_{2} \boldsymbol{w}_{\mathrm{a}}\right)$$
\textbf{Solution:} run SDA with:\\
$$\boldsymbol{R}=\boldsymbol{U}_{2}^{\mathrm{H}} \boldsymbol{A} \boldsymbol{U}_{2} \in \mathbb{C}^{(M-L) \times(M-L)}, \quad \text { and } \quad \boldsymbol{p}=\boldsymbol{U}_{2}^{\mathrm{H}} \boldsymbol{A} \boldsymbol{w}_{\mathrm{q}}$$
\end{sectionbox}
\begin{sectionbox}
\subsection{Steepest Descent Procedure (SDP)}
Problem: $R$ and $p$ are unknown and need to be estimated.\\
Drop assumption that $u[n]$ is WSS (e.g. channel changes) $\rightarrow$ $\boldsymbol{R}[n]$\\
Estimation with exponential weighting:\\
$$\widehat{\boldsymbol{R}}[n]=\frac{\sum_{k=0}^{\infty} \boldsymbol{u}[n-k] \boldsymbol{u}^{\mathrm{H}}[n-k] \alpha[k]}{\sum_{k=0}^{\infty} \alpha[k]}$$
$$\alpha[k]= \begin{cases}1 & \text { for } k=0 \\ \eta^{k} & \text { for } k>0 \\ 0 & \text { else }\end{cases}, \sum_{k=0}^{\infty} \eta^{k}=\frac{1}{1-\eta}$$
$$\widehat{\boldsymbol{R}}[n]=\eta \widehat{\boldsymbol{R}}[n-1]+(1-\eta) \boldsymbol{u}[n] \boldsymbol{u}^{\mathrm{H}}[n]$$
$$\widehat{\boldsymbol{p}}[n]=\eta \widehat{\boldsymbol{p}}[n-1]+(1-\eta) \boldsymbol{u}[n] d^{*}[n]$$
\includegraphics[width = 6.8cm]{img/sdp.png}
\end{sectionbox}
\begin{sectionbox}
\subsection{SDP: Determining $\eta$}
If $R[n]$ and $p[n]$ are changing slowly $\rightarrow$ choose $\eta$ close to 1\\
If $R[n]$ and $p[n]$ remain const for $N_1$ time slots, but have completely changed after $N_2 > N_1$:\\
$\eta^{N_{2}}=\gamma \ll 1$ and $\eta^{N_{1}}=1-\gamma$\\
$\rightarrow$ $\eta=x^{1 / N_{1}}, \quad \text { where } \quad x^{N_{2} / N_{1}}+x-1=0, \quad 0<x<1$\\
$N_{2} / N_{1}=30$ is a reasonable choice in practice $\rightarrow$ $\eta=\exp \left(-\frac{2.5}{N_{2}}\right)$\\
\textbf{Radio Communications:}\\
$N_2$ is the number of samples in which the receiver has moved by $\lambda$.
For sample rate (bandwidth) $B$, wavelength $\lambda$, and speed $v$: $$\eta=\exp \left(-2.5 \frac{v}{B \lambda}\right)$$
\end{sectionbox}
\begin{sectionbox}
\subsection{Least Mean Square (LMS)}
Special Case of SDP for $\eta = 0$:
$$\widehat{\boldsymbol{R}}[n]=\boldsymbol{u}[n] \boldsymbol{u}^{\mathrm{H}}[n], \quad \widehat{\boldsymbol{p}}[n]=\boldsymbol{u}[n] d^{*}[n]$$
$$\mu=\frac{\boldsymbol{r}^{\mathrm{H}} \boldsymbol{r}}{\boldsymbol{r}^{\mathrm{H}} \widehat{\boldsymbol{R}} \boldsymbol{r}}=\frac{1}{\|\boldsymbol{u}[n]\|_{2}^{2}}$$
$$\boldsymbol{r}=\widehat{\boldsymbol{p}}-\widehat{\boldsymbol{R}} \boldsymbol{w}=\boldsymbol{u}[n] (\underbrace{d^{*}[n]-\overbrace{\boldsymbol{u}^{\mathrm{H}}[n] \boldsymbol{w}}^{y^{*}[n]}}_{e^{*}[n]}) =\boldsymbol{u}[n] e^{*}[n]$$
\includegraphics[width = 6.8cm]{img/lms.png}
$\alpha > 0$: avoids very large step sizes, if $\boldsymbol{u}[n]$ is small\\
$\beta > 0$: is for fine-tuning (found experimentally as well as $\alpha$)\\
\end{sectionbox}
\begin{sectionbox}
\subsection{Block diagrams}
Standard SDP and LMS:\\
\includegraphics[width = 6.8cm]{img/blk-diagrams.png}\\
Constrained SDP and LMS:
$$z[n]=\boldsymbol{w}^{\mathrm{H}}[n] \boldsymbol{x}[n]$$
$$\min _{w[n]} \mathrm{E}\left[|z[n]|^{2}\right], \quad \text { s.t. } \quad \boldsymbol{B}^{\mathrm{H}} \boldsymbol{w}[n]=\boldsymbol{c}$$
\includegraphics[width = 6.8cm]{img/blk-diagrams2.png}
$$\boldsymbol{x}[n]=s_{i}[n] \boldsymbol{b}_{i} \quad \Longrightarrow \quad z[n]=s_{i}[n] c_{i}^{*}$$
$$\boldsymbol{x}[n]\notin \operatorname{im} \boldsymbol{B} \quad \Longrightarrow \quad z[n]=0$$
\end{sectionbox}
\pagebreak
\section{Least Squares}
\begin{sectionbox}
\subsection{Least Squares}
Only $\boldsymbol{A}$ is known, no information about $\boldsymbol{\Upsilon}$ $\rightarrow$ ignore it all together\\
\textbf{Least squares problem:}\\
$$\boldsymbol{X} \approx \boldsymbol{A} \boldsymbol{S}$$
$$\hat{\boldsymbol{S}}_{\mathrm{LS}}=\arg \min _{\boldsymbol{S}}\|\boldsymbol{X}-\boldsymbol{A S}\|_{\mathrm{F}}^{2}$$
\textbf{Derivation:}
$$\mathcal{E}=\|\boldsymbol{X}-\boldsymbol{A} \boldsymbol{S}\|_{\mathrm{F}}^{2}=\operatorname{tr}\left((\boldsymbol{X}-\boldsymbol{A} \boldsymbol{S})^{\mathrm{H}}(\boldsymbol{X}-\boldsymbol{A} \boldsymbol{S})\right)$$
$$\frac{\partial \mathcal{E}}{\partial \boldsymbol{S}^{*}}=-\boldsymbol{A}^{\mathrm{H}} \boldsymbol{X}+\boldsymbol{A}^{\mathrm{H}} \boldsymbol{A} \boldsymbol{S} \leftarrow \frac{\partial \operatorname{tr}\left(\boldsymbol{S}^{\mathrm{H}} \boldsymbol{B}\right)}{\partial \boldsymbol{\boldsymbol{S}}^{*}} = \boldsymbol{B}$$
\begin{emphbox}
$\begin{aligned}
\hat{\boldsymbol{S}}_{\mathrm{LS}} &=\left(\boldsymbol{A}^{\mathrm{H}} \boldsymbol{A}\right)^{-1} \boldsymbol{A}^{\mathrm{H}} \boldsymbol{X} \text{ (full column rank)}\\
&=\boldsymbol{A}^{+} \boldsymbol{X} = \boldsymbol{A}^{+} \boldsymbol{A} \boldsymbol{S}+\boldsymbol{A}^{+} \boldsymbol{\Upsilon}\text{ (general)}
\end{aligned}$
\end{emphbox}
$$\boldsymbol{A} \text { has full column rank } \Longrightarrow \hat{\boldsymbol{S}}_{\mathrm{LS}}=\boldsymbol{S}+\boldsymbol{A}^{+} \boldsymbol{\Upsilon}$$
$$\mathrm{E}[\boldsymbol{\Upsilon}]=\mathbf{0}\Longrightarrow \text{ unbiased}$$
Estimation noise:\\
$$\mathrm{E}\left[\left\|\boldsymbol{A}^{+} \boldsymbol{\Upsilon}\right\|_{\mathrm{F}}^{2}\right] = \operatorname{tr}\left(\boldsymbol{A}^{+} \mathrm{E}\left[\boldsymbol{\Upsilon} \boldsymbol{\Upsilon}^{\mathrm{H}}\right] \boldsymbol{A}^{+\mathrm{H}}\right)$$
For white noise $\mathrm{E}\left[\boldsymbol{\Upsilon} \boldsymbol{\Upsilon}^{\mathrm{H}}\right]=\mathbf{I}$ and full column rank $\boldsymbol{A}$:\\
$$\mathrm{E}\left[\left\|\boldsymbol{A}^{+} \boldsymbol{\Upsilon}\right\|_{\mathrm{F}}^{2}\right] = \operatorname{tr}\left(\left(\boldsymbol{A}^{\mathrm{H}} \boldsymbol{A}\right)^{-1}\right) = \sum_{i} \frac{1}{\lambda_{i}},$$
where $\lambda_i$ are the eigenvalues of $\boldsymbol{A}^H\boldsymbol{A}$ $\rightarrow$ minimal if all are the same\\
For white observation noise the smallest possible variance is achieved iff:\\
$$\boldsymbol{A}^{\mathrm{H}} \boldsymbol{A}=c \mathbf{I}, \quad \text { for any } c>0$$
among all matrices $\boldsymbol{A} \in \mathbb{C}^{M \times d} \text { with }\|\boldsymbol{A}\|_{\mathrm{F}}^{2}=c$.
\begin{itemize}
\item $\boldsymbol{A}^{\mathrm{H}} \boldsymbol{A}$ has $\lambda_i = c/d$
\item $\boldsymbol{A}$ must have orthogonal columns with the same euclidean norm
\item use orthogonal pilot sequences
\item purely deterministic approach, no need to know statistical properties
\end{itemize}
\end{sectionbox}
\begin{sectionbox}
\subsection{Pilot Sequence}
\textbf{Setup:}
$$\boldsymbol{u}=\boldsymbol{H}\boldsymbol{p}+\boldsymbol{v}\quad\Longleftrightarrow \quad\boldsymbol{u}=\boldsymbol{A} \boldsymbol{h}+\boldsymbol{v}$$
$$\boldsymbol{A}=\left[\begin{array}{cccc}
P_{0} & P_{1} & \cdots & P_{K} \\
P_{1} & P_{2} & \cdots & P_{K+1} \\
\vdots & \vdots & \vdots & \vdots \\
P_{q-K-1} & P_{q-K} & \cdots & P_{q-1}
\end{array}\right]$$
$$\boldsymbol{h}=\left[\begin{array}{llll}
h_{0} & h_{1} & \cdots & h_{K}
\end{array}\right]^{\mathrm{T}}$$
Necessary for full column rank: $q \geq 2 K+1$
\end{sectionbox}
\begin{sectionbox}
\subsection{LS curve fitting}
$$\hat{y}(x)=\frac{a}{x}+b+c x+d x^{2}+e x^{3}$$
$$\underbrace{\left[\begin{array}{ccccc}
1 / x_{1} & 1 & x_{1} & x_{1}^{2} & x_{1}^{3} \\
1 / x_{2} & 1 & x_{2} & x_{2}^{2} & x_{2}^{3} \\
\vdots & \vdots & \vdots & \vdots & \vdots \\
1 / x_{N} & 1 & x_{N} & x_{N}^{2} & x_{N}^{3}
\end{array}\right]}_{\boldsymbol{A}} \underbrace{\left[\begin{array}{c}
a \\
b \\
c \\
d \\
e
\end{array}\right]}_{\boldsymbol{w}} \approx \underbrace{\left[\begin{array}{c}
y_{1} \\
y_{2} \\
\vdots \\
y_{N}
\end{array}\right]}_{\boldsymbol{y}}$$
$$\boldsymbol{w}_{\mathrm{LS}}=\boldsymbol{A}^{+} \boldsymbol{y}$$
\end{sectionbox}
\begin{sectionbox}
\subsection{Numerical integration with LS}
Within every 3x3 window:
$$\hat{f}(x, y)=a+b x^{2}+c y^{2}+d x^{2} y^{2}$$
$$\left[\begin{array}{cccc}
1 & h_{x}^{2} & h_{y}^{2} & h_{x}^{2} h_{y}^{2} \\
1 & 0 & h_{y}^{2} & 0 \\
\vdots & \vdots&\vdots&\vdots\\
1 & h_{x}^{2} & h_{y}^{2} & h_{x}^{2} h_{y}^{2}
\end{array}\right] \underbrace{\left[\begin{array}{c}
a \\
b \\
c \\
d
\end{array}\right]}_{w} \approx\left[\begin{array}{c}
f\left(-h_{x}, h_{y}\right) \\
f\left(0, h_{y}\right) \\
\vdots\\
f\left(h_{x},-h_{y}\right)
\end{array}\right]$$
$$\hat{F}=\frac{4}{9} h_{x} h_{y}\left(9 a+3 b h_{x}^{2}+\left(3 c+d h_{x}^{2}\right) h_{y}^{2}\right)$$
And without solving $w_{\mathrm{LS}}=\boldsymbol{A}^{+} \boldsymbol{f}$:
$$\begin{aligned}
\hat{F}=(& f\left(-h_{x}, h_{y}\right)+f\left(h_{x}, h_{y}\right)+f\left(-h_{x},-h_{y}\right)+f\left(h_{x},-h_{y}\right) \\
&+4\left(f\left(0, h_{y}\right)+f\left(-h_{x}, 0\right)+f\left(h_{x}, 0\right)+f\left(0,-h_{y}\right)\right) \\
&+16 f(0,0)) \frac{h_{x} h_{y}}{9}
\end{aligned}$$
3x3 window: quadratic convergence when increasing $M$\\
5x5 window: cubic convergence when increasing $M$\\
Procedure:\\
$$F=\int_{y=y_{\min }}^{y_{\max }} \int_{x=x_{\min }}^{x_{\max }} f(x, y) \mathrm{d} x \mathrm{~d} y$$
$$h_{x}=\frac{x_{\max }-x_{\min }}{M-1}, \quad h_{y}=\frac{y_{\max }-y_{\min }}{M-1}$$
3x3 window: $ M \in\{3,5,7, \ldots\}$, 5x5 window: $M \in \{5,9,13, \ldots\}$\\
Divide the integration region into $M$x$M$ parts. Move the local approximation by $3-1$, or $5-1$ grid points and sum up all.
\begin{center}\includegraphics[width = 4cm]{img/grid2.jpeg}\end{center}
\end{sectionbox}
\begin{sectionbox}
\subsection{Least squares as a projection}
\textbf{Setup:}
$$\boldsymbol{A} \boldsymbol{x}=\boldsymbol{b}, \quad \text { where } \quad \boldsymbol{b} \notin \operatorname{im} \boldsymbol{A}$$
$$\boldsymbol{A} \boldsymbol{x}=\boldsymbol{b}+ \Delta\boldsymbol{b}, \quad \text { where } \quad \boldsymbol{b}+\Delta\boldsymbol{b} \in \operatorname{im} \boldsymbol{A}$$
$$\Delta \boldsymbol{b}_{\mathrm{opt}}=\arg \min _{\Delta \boldsymbol{b}}\|\Delta \boldsymbol{b}\|_{2}^{2}, \quad \text { s.t. } \quad \boldsymbol{b}+\Delta \boldsymbol{b} \in \operatorname{im} \boldsymbol{A}$$
Derivation:
$$\Delta \boldsymbol{b}_{\mathrm{opt}}=\arg \min _{\Delta \boldsymbol{b}}\|\Delta \boldsymbol{b}\|_{2}^{2}, \quad \text { s.t. } \quad \boldsymbol{P}(\boldsymbol{b}+\boldsymbol{\Delta b})=\boldsymbol{b+\Delta b}$$
$$\Delta \boldsymbol{b}_{\mathrm{opt}}=\arg \min _{\Delta \boldsymbol{b}}\|\Delta \boldsymbol{b}\|_{2}^{2}, \quad \text { s.t. } \quad(\mathbf{I}-\boldsymbol{P})(\boldsymbol{b}+\Delta \boldsymbol{b})=\mathbf{0}$$
Lagrangian optimization yields:
$$\Delta \boldsymbol{b}=-(\mathbf{I}-\boldsymbol{P}) \boldsymbol{b}$$
\begin{emphbox}
The least squares solution $\boldsymbol{x} = \boldsymbol{A}^+\boldsymbol{b}$ is the exact solution of $\boldsymbol{A}\boldsymbol{x} = \boldsymbol{P}\boldsymbol{b}$. I.e. the least squares estimation projects the measurements on $\operatorname{im}\boldsymbol{A}$\\
$\boldsymbol{P} = \boldsymbol{A}\boldsymbol{A}^+$
\end{emphbox}
\end{sectionbox}
\begin{sectionbox}
\subsection{Total Least Squares}
\textbf{Setup:}
$$\min \left\|\left[\begin{array}{ll}
\Delta \boldsymbol{A} & \Delta \boldsymbol{b}
\end{array}\right]\right\|_{F}^{2} \quad\text{s.t.} \quad (\boldsymbol{A}+\Delta \boldsymbol{A}) \boldsymbol{x}=\boldsymbol{b}+\Delta \boldsymbol{b}$$
Rewrite:
$$\left[\begin{array}{ll}
\boldsymbol{A} & b
\end{array}\right]\left[\begin{array}{c}
x \\
-1
\end{array}\right]=0, \quad \boldsymbol{b} \notin \operatorname{im} \boldsymbol{A}$$
$\boldsymbol{b} \notin \operatorname{im} \boldsymbol{A}$, thus: $\operatorname{rank} [\boldsymbol{Ab}] = N + 1$ and $\operatorname{null} [\boldsymbol{Ab}] = \{0\}$\\
$$\left[\begin{array}{ll}
\boldsymbol{A} & \boldsymbol{b}
\end{array}\right]=\boldsymbol{U} \boldsymbol{\Sigma} \boldsymbol{V}^{\mathrm{H}}=\sum_{k=1}^{N+1} s_{k} \boldsymbol{u}_{k} \boldsymbol{v}_{k}^{\mathrm{H}}$$