-
Notifications
You must be signed in to change notification settings - Fork 5
/
lec01.rnw
900 lines (788 loc) · 27.1 KB
/
lec01.rnw
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
\documentclass[10pt]{beamer}
\mode<presentation>{
\usetheme{Goettingen}
%\usetheme{Madrid}
\usecolortheme{default}
}
\usepackage{CJK}
\usepackage{graphicx}
\usepackage{amsmath, amsopn}
\usepackage[english]{babel}
\usepackage[T1]{fontenc}
\usepackage[latin1]{inputenc}
\usepackage{enumerate}
\usepackage{animate}
\usepackage{url}
\ifx\hypersetup\undefined
\AtbBeginDocument{%
\hypersetup{unicode=true,pdfusetitle,
bookmarks=true,bookmarksnumbered=false,bookmarksopen=false,
breaklinks=false,pdfborder={0 0 0},pdfborderstyle={},backref=false,colorlinks=false}
}
\else
\hypersetup{unicode=true,pdfusetitle,
bookmarks=true,bookmarksnumbered=false,bookmarksopen=false,
breaklinks=false,pdfborder={0 0 0},pdfborderstyle={},backref=false,colorlinks=false}
\fi
\usepackage{breakurl}
\usepackage{color}
\usepackage{times}
\usepackage{xcolor}
\usepackage{listings}
\lstset{
language=R,
keywordstyle=\color{blue!70}\bfseries,
basicstyle=\ttfamily,
commentstyle=\ttfamily,
showspaces=false,
showtabs=false,
frame=shadowbox,
rulesepcolor=\color{red!20!green!20!blue!20},
breaklines=true}
\title[BI476]{BI476: Biostatistics - Case Studies}
\author[Maoying Wu]{Maoying, Wu\\{\small ricket.woo@gmail.com}}
\institute[CBB] % (optional, but mostly needed)
{
\inst{}
Dept. of Bioinformatics \& Biostatistics\\
Shanghai Jiao Tong University
}
\date{Spring, 2018}
\AtBeginSection[]
{
\begin{frame}<beamer>{Next Section ...}
\tableofcontents[currentsection]
\end{frame}
}
\begin{document}
\begin{CJK*}{UTF8}{gbsn}
<<setup, include=FALSE>>=
library(knitr)
opts_chunk$set(fig.path='figure/beamer-', fig.align='center',
fig.show='hold',size='footnotesize', out.width='0.60\\textwidth')
@
\frame{\titlepage}
\begin{frame}
\tableofcontents
\end{frame}
\section{Syllabus}
\begin{frame}
\frametitle{BI476: Syllabus}
\begin{itemize}
\item This is a one-semester course for undergraduate students majored in biostatistics
or bioinformatics.
\item Topics will cover experiment design, intuitive hypothesis testing, (generalized
linear models (including generalized estimating equations), survival analysis and
multivariate statistics methods.
\item Advanced topics will be included, such as penalized regression or hierarchical
/mixed-effects linear models or Bayesian statistics, if time permits.
\item Estimation, interpretation, and diagnostic approaches will be discussed.
\item Software instruction will be provided in class in R.
\item Performance will be evaluated based on homeworks (35\%), two exams (30\%),
lab assignments (20\%) and projects (15\%).
\item This is a two-credit course.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{BI476: Syllabus}
\begin{itemize}
\item Textbooks: No textbook is required. We will provide the readings and
related materials on the website.
\item Prerequisites: Linear algebra, Probability, Biostatistics
\item Course Objectives:
Upon successful completion of the course, the student will be able to
\begin{itemize}
\item Design the experiment and analyze the experiment data
\item Apply, interpret and diagnose linear regression models
\item Apply, interpret and diagnose logistic, poisson and Cox regresssion models
\item Apply and interpret the multivariate analysis methods.
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}[t]
\frametitle{BI476: Syllabus}
\begin{figure}[ht]
\begin{tabular}{ll}
\textbf{Instructor}: & Maoying Wu (ricket.woo@gmail.com)\\
\textbf{Website}: & \url{http://cbb.sjtu.edu.cn/~mywu/bi476}\\
\textbf{Github}: & \url{https://github.com/ricket.sjtu/bi476}\\
\textbf{Office}: & Biopharmaceutics Building, Rm 4A-223\\
\textbf{Time}: & Mondays, 14:00-15:40\\
\textbf{Location}: & East Lower Hall, Rm 403\\
\textbf{Office Hours}: & Monday-Friday, 8:30 – 17:00\\
\end{tabular}
\end{figure}
\end{frame}
\begin{frame}
\frametitle{BI476: Schedule}
\begin{itemize}
\item Lecture 1: Recaps of Mathematical Knowledge for Biostatistics (1-2)
\item Lecture 2: Observational Studies and Analysis (3-4)
\item Lecture 3: Randomized Clinical Trials and Analysis (5-6)
\item Lecture 4: Linear Regression Models and Extensions (7-8)
\item Midterm (9)
\item Lecture 5: Panel Data Analysis (10-11)
\item Lecture 6: Survival Analysis and Competing Risks (12-13)
\item Lecture 7: Multivariate Statistcal Analysis (14-15)
\item Projects (16)
\item Final (17-18)
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{List of Symbols}
\begin{itemize}
\item $\mathcal{X}$: A set
\item $\mathbb{R}$: Real set
\item $X$: A random variable
\item $\mathbf{X} \in \mathbb{R}^{m \times n}$: A $m$-by-$n$ real matrix
\item $\mathbf{x} \in \mathbb{R}^{n}$: A real vector of length $n$
\item $x \in \mathbb{R}$: A real number
\item $\Phi(x) = P(X \le x)$: Cumulative distribution function
\item $f(x) = f(X=x)$: Probability density function
\end{itemize}
\end{frame}
\begin{frame}[t]
\frametitle{Lecture 1: Mathematics for Biostatistics}
\begin{itemize}
\item Calculus (微积分)
\begin{itemize}
\item Limits (极限)
\item Derivatives (导数): First-order and second-order
\item Integration (积分)
\item Gradient (梯度): Jacobian, Hessian
\item Convex function (凸函数) and Jensen's inequality (简森不等式)
\item Taylor's expansion (泰勒展开)
\end{itemize}
\item Linear algebra (线性代数)
\begin{itemize}
\item Vector (向量), matrix (矩阵)
\item Norm (范数)
\item Rank (秩), determinant (行列式), trace (迹)
\item Matrix multiplication (矩阵乘积)
\item Eigendecomposition (正交分解)
\item Singular value decomposition (SVD,奇异值分解)
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}[t]
\frametitle{Lecture 1: Mathematics for Biostatistics}
\begin{itemize}
\item Probability (概率论)
\begin{itemize}
\item Probability density function (pdf, 概率密度函数)
\item Probability mass function (pmf, 概率质量函数)
\item Cumulative distribution function (cdf, 累积分布函数)
\item Moment generating function (mgf, 矩母函数)
\item Joint probability distribution (联合概率分布)
\item Conditional probability distribution (条件概率分布)
\item Marginal distribution (边缘概率分布)
\item Bayes' Equation/Theorem (贝叶斯公式/定理)
\item Continuous distributions (连续概率分布)
\item Discrete distribution (离散概率分布)
\end{itemize}
\item Numerical Optimization (数值优化方法)
\begin{itemize}
\item Convex set (凸集)
\item Convex function (凸函数), Concave function (凹函数)
\item Gradient descent (梯度下降), gradient ascent (梯度上升)
\item Newton's method (牛顿法)
\item Quasi-Newton's methods (拟牛顿法)
\item Method of multiplers (乘子法)
\item Laglangian method of multiplier (拉格朗日乘子法)
\end{itemize}
\end{itemize}
\end{frame}
\section{Calculus}
\begin{frame}[t]
\frametitle{Limit (极限)}
Compute the limit
$$
\lim_{x \to 0} \frac{3 \sin^2 x}{4 x^2}
$$
\end{frame}
\begin{frame}[t]
\frametitle{Derivative (导数)}
$$
f^\prime(x)=\frac{d}{dx} f(x)=\lim_{h\rightarrow 0} \frac{f(x+h)-f(x)}{h}
$$
The derivative is the slope of the tangent line to the graph of $f(x)$,
assuming the tangent line exists.
\end{frame}
\begin{frame}[t]
\frametitle{Common Derivatives and Rules}
\begin{itemize}
\item $\frac{d}{dx}x^n = n x^{n-1}$
\item $\frac{d}{dx}\ln = \frac{1}{x}$
\item $\frac{d}{dx}a^x = a^x \ln a$
\item $\frac{d}{dx}e^x = e^x$
\end{itemize}
\begin{block}{\center Product rule}
$$
[f(x)g(x)]^{\prime} = f^{\prime}(x)g(x) + f(x)g^{\prime}(x)
$$
\end{block}
\begin{block}{\center Quotient rule}
$$
\left[ \frac{f(x)}{g(x)}\right]^{\prime} = \frac{g(x)f^{\prime}(x) - f(x)g^{\prime}(x)}{g^2(x)}
$$
\end{block}
\begin{block}{\center Chain rule}
Let $y = f(g(x))$,
$$
y^{\prime} = f^{\prime}(g(x)) g^{\prime}(x)
$$
\end{block}
\end{frame}
\begin{frame}[t]
\frametitle{Integrals (积分)}
For a function $f(x)$, its indefinite integral is:
$$
\int f(x)dx = F(x) + C, \mbox{where } F'(x) = f(x)
$$
\end{frame}
\begin{frame}[t]
\frametitle{Common integrals and rules}
\begin{enumerate}[(a)]
\item $\int_a^a f(x)dx = 0$
\item $\int_a^b f(x)dx = -\int_b^a f(x)dx$
\item $\int x^r dx = \frac{1}{r+1}x^{r+1} + C$
\item $\int_a^b f(x)dx = F(b) - F(a)$
\item $\int x^n dx = \frac{1}{n+1} x^{n+1} + C, n \neq -1$
\item $\int \frac{1}{x} dx = \ln |x| + C$
\item $\int e^x dx = e^x + C$
\item $\int_a^b f(g(x)) g'(x) = \int_{g(a)}^{g(b)} f(u)du, \mbox{where } u= g(x)$
\item Integration by parts (分部积分): $\int_a^b u dv = uv|_a^b - \int_a^b v du$
\end{enumerate}
\end{frame}
\begin{frame}
\frametitle{Integrals: Exercises}
\begin{itemize}
\item $\int_0^5 x^2 e^{-x} dx$
\item $\int x \ln x$
\item $\int_0^t \frac{2}{1000^2} x e^{-(x/1000)^2}$
\item $\int_{-\infty}^{\infty} \frac{1}{2} e^{-|y| + ty}$
\end{itemize}
\end{frame}
\begin{frame}[t]
\frametitle{Taylor's Expansion (泰勒展开)}
$$
f(x) = f(x_0) + \frac{f'(x_0)}{1!}(x-x_0) + \frac{f''(x_0)}{2!}(x-x_0)^2 + \dots = \sum_{k=0}^{\infty} \frac{f^{(k)}(x_0)}{k!}(x-x_0)^k
$$
\begin{block}{Exercise}
$$
1 + \frac{1}{2!} + \frac{1}{3!} + \dots
$$
\end{block}
\end{frame}
\begin{frame}
\frametitle{Multivariate Calculus}
\begin{itemize}
\item $f_x(x, y) = \frac{\partial}{\partial x} f(x, y) = \frac{\partial f}{\partial x}$
\item $\int_a^b \left[ \int_c^d f(x, y)dx\right]dy = \int_c^d \left[ \int_a^b f(x, y)dy\right]dx$
\end{itemize}
\end{frame}
\section{Linear algebra}
\begin{frame}[t]
\frametitle{Linear vector space (线性向量空间)}
A linear vector space $\mathcal{X}$ is a collection of elements
satisfying the following properties:
\begin{itemize}
\item Rule of addition (加法律): $\forall x, y, z \in \mathcal{X}$,
\begin{enumerate}
\item $x + y \in \mathcal{X}$
\item $x + y = y + x$
\item $(x + y) + z = x + (y + z)$
\item $\exists 0 \in \mathcal{X}$, such that $x + 0 = x$
\item $\forall x \in \mathcal{X}$, $\exists -x \in \mathcal{X}$ such that $x + (-x) = 0$
\end{enumerate}
\item Rule of multiplication (乘法律): $\forall x, y \in \mathcal{X}$ and $a, b \in R$,
\begin{enumerate}
\item $ax \in \mathcal{X}$
\item $a(bx) = (ab) x$
\item $1x = x, 0x = 0$
\item $a(x + y) = ax + ay$
\end{enumerate}
\end{itemize}
\begin{block}{Example: $\mathbb{R}^n$}
The $n$-dimensional Euclidean $\mathbb{R}^n$, is a linear vector space.
\end{block}
\end{frame}
\begin{frame}[t]
\frametitle{Inner product (向量内积)}
An \textbf{inner product} is a mapping: $\mathcal{X} \times \mathcal{X} \mapsto \mathbb{R}$.
The inner product between any $x, y \in \mathcal{X}$ is denoted by $\langle x, y \rangle$ and it satisfies
the following properties for all $x, y, z \in \mathcal{X}$:
\begin{enumerate}[(1)]
\item $\langle x, y \rangle = \langle y, x \rangle$
\item $\langle ax, y \rangle = a \langle x, y \rangle$ for all scalars $a$
\item $\langle x + y, z \rangle = \langle x, z \rangle + \langle y, z \rangle$
\item $\langle x, x \rangle \ge 0$ and $\langle x, x \rangle = 0 \Rightarrow x = 0$
\end{enumerate}
A space $\mathcal{X}$ equipped with an inner product is called an \textbf{inner product space}.
\end{frame}
\begin{frame}[t]
\frametitle{Orthongonal Vectors (正交向量)}
$\mathbf{x}$ and $\mathbf{y}$ are \textbf{orthogonal vectors} if:
$$
\langle \mathbf{x}, \mathbf{y} \rangle = 0
$$
Let $\mathcal{X} = \mathbb{R}^n$, then
$$
\langle \mathbf{x}, \mathbf{y} \rangle := \mathbf{x}^T \mathbf{y} = \sum_{i=1}^n x_i y_i
$$
\end{frame}
\begin{frame}
\frametitle{Norms (向量范数)}
The inner product induces the defintion of $\ell_2$-norm:
$$
\lVert \mathbf{x} \rVert_2 = \sqrt{\langle \mathbf{x}, \mathbf{x} \rangle}
$$
here the norm measure the size (length) of $\mathbf{x}$.
The inner product can be written into the following form with norms:
$$
\langle \mathbf{x}, \mathbf{y} \rangle = \Vert \mathbf{x} \Vert \Vert \mathbf{y} \Vert \cos \theta
$$
where $\theta$ is the angle between vectors $\mathbf{x}$ and $\mathbf{y}$.
The general $\ell_p$-norm for $\mathbf{x}$ is:
$$
\lVert \mathbf{x} \rVert_p = (\sum_i x_i^p)^{1/p}, p=0,1,2,\dots, \infty
$$
We have $\ell_0$ and $\ell_1$ norms:
$$
\lVert \mathbf{x} \rVert_0 = \sum_i I(x_i \neq 0), \lVert \mathbf{x} \rVert_1 = \sum_i |x_i|
$$
\end{frame}
\begin{frame}[t]
\frametitle{Cauchy-Schwartz inequality (柯西-施瓦茨不等式)}
$$
\langle x, y \rangle \le \Vert x \Vert \Vert y \Vert
$$
\textcolor{red}{Q: When does the equation hold?}
\end{frame}
\begin{frame}[t]
\frametitle{Triangle inequality (三角不等式)}
$$
\Vert x - y \Vert \le \Vert x \Vert + \Vert y \Vert
$$
\textcolor{red}{Q: When does the equation hold?}
\end{frame}
\begin{frame}[t]
\frametitle{Linearly independent (线性无关)}
For a set of vectors
$$
x_1, x_2, \ldots, x_p \in \mathcal{X},
$$
if there exists a set of scalars $a_1, a_2, \ldots, a_p \in \mathbb{R}$ such that not all $a_i = 0$ and
$$
\sum_{i=1}^p a_i x_i = 0
$$
we say that $x_1, x_2, \ldots, x_p$ are \textbf{linearly dependent(线性相关)}.
If equation only holds in the case $a_1=a_2=\ldots=a_p=0$, then we say that the vectors are \textbf{linearly independent (线性无关)}.
\end{frame}
\begin{frame}[t]
\frametitle{Basis (基)}
A set of vectors $\{\phi_i\}(i=1, \dots, n)$ is a \textbf{basis (基)} for $\mathcal{X}$ if an
arbitrary vector $x \in \mathcal{X}$ can be expressed as the linear combination of
$\{\phi_i\}(i=1,\dots,n)$. That is, there exists a set of scalars $\{\theta_i\}(i=1,\dots,n)$,
such that
$$
x = \sum_{i=1}^n \theta_i \phi_i
$$
\begin{block}{Orthonormal basis (正交基)}
The bases $\{\phi_i\}_{i=1}^n$ are orthonormal if
$$
\phi_i^T \phi_j = \left\{
\begin{array}{ll}
0, & i \neq j\\
1, & i = j
\end{array}
\right.
$$
\end{block}
\end{frame}
\begin{frame}
\frametitle{Orthobasis of Hilbert space (希尔伯特空间的正交基)}
Every $x \in \mathcal{H}$ can be represented in terms of an orthonormal basis
$\{\phi_i\}_{i \ge 1}$ (or "orthobasis" for short) according to:
$$
x = \sum_{i \ge 1} \langle x, \phi_i \rangle \phi_i
$$
This is easy to see as follows. Suppose x has a representation $\sum_i \theta_i \phi_i$. Then
$$
\theta_i = \langle x, \phi_i \rangle
$$
\begin{block}{Example: Orthonormal basis for $\mathbb{R}^n$}
$$
\phi_k = \left[0, \cdots, 1, \cdots, 0 \right]^{-1}
$$
where
$$
\phi_{k, i} = \left\{
\begin{array}{ll}
0, & i \neq k\\
1, & i = k
\end{array}
\right.
$$
\end{block}
\end{frame}
\begin{frame}
\frametitle{Subspace (子空间)}
Consider a set of vectors $\{x_i\}_{i=1}^p \in \mathcal{X}$. The \textbf{span} of these vectors
is the set of all vectors $x \in \mathcal{X}$ that can be generated from linear combinations of
the set
$$
\operatorname{span}(\left\{xi\right\}_{i=1}^p) := \left\{ y: y = \sum_{i=1}^p a_i x_i, a_1, \ldots, a_p \in \mathbb{R} \right\}
$$
This set is also called a subspace of $\mathcal{X}$.
A subset $\mathcal{M} \subset \mathcal{X}$ is a subspace if $x, y \in \mathcal{M}$, we have
$$
ax + by \subset \mathcal{M}
$$
注:如果$\phi_1, \cdots, \phi_p$是子空间$\mathcal{M} \subset \mathbb{R}^n$的一组正交基,
则该子空间中的任意向量$x \in \mathcal{M}$可以写成:
$$
x = \sum_{i=1}^p \theta_i \phi_i
$$
这样虽然$x \in \mathbb{R}^n$,但由于其是$\mathcal{M}$中的向量,所以可以写成$p$个自由参数的线性组合,也就是说其自由度为$p$。
\end{frame}
\begin{frame}
\frametitle{Orthogonal projection (正交投影)}
Let $\mathcal{H}$ be a \textbf{Hilbert space} and let $\mathcal{M} \subset \mathcal{H}$ be a subspace.
Every $x \in \mathcal{H}$ can be written as
$$
x = y + z
$$
where $y \in \mathcal{M}$ and $z \perp \mathcal{M}$, which is shorthand for $z$ orthogonal to
$\mathcal{M}$; that is
$$
\forall v \in \mathcal{M}, \langle v, z \rangle = 0
$$
The vector $y$ is the optimal approximation to $x$ in terms of vectors in $M$ in the following sense:
$$
y = \operatorname{argmin}_{v \in \mathcal{M}} \| x - v \|
$$
The vector $y$ is called the \textbf{orthogonal projection} of $x$ onto $\mathcal{M}$.
\end{frame}
\begin{frame}
\frametitle{Orthogonal subspace projection (正交子空间投影)}
Let $\mathcal{M} \subset \mathcal{H}$ and let $\{\phi_i\}_{i=1}^r$ be an orthobasis for $\mathcal{M}$. For any
$x \in \mathcal{H}$, the projection of $x$ onto $\mathcal{M}$ is given by
$$
y = \sum_{i=1}^r \langle \phi_i, x \rangle \phi_i
$$
and this projection can be viewed as a sort of filter that \textbf{removes all components of the signal $x$
that are orthogonal to $\mathcal{M}$}.
\begin{block}{Example}
Let $\mathcal{H} = \mathbb{R}^2$. Consider the canonical coordinate system $\phi_1 = \left[\begin{array}{c}
1\\
0
\end{array}\right]$ and $\phi_2 = \left[\begin{array}{c}
0\\
1
\end{array}\right]$. Let $\mathcal{M}$ be the subspace spanned by $\phi_1$. The projection of any
$x = [x_1\; x_2]^T \in \mathbb{R}^2$ onto $\mathcal{M}$ is
$$
\begin{array}{lcl}
P_1x &=& \langle x, \phi_1 \rangle \phi_1\\
& = & \left[ \begin{array}{cc}
x_1 & x_2
\end{array}\right] \left[ \begin{array}{c}
1\\
0
\end{array}\right] \left[ \begin{array}{c}
1\\
0
\end{array}\right] \\
& = & \left[ \begin{array}{c}
x_1\\
0
\end{array}\right]
\end{array}
$$
The projection operator $P_1$ is just a matrix and it is given by
$$
\begin{array}{lcl}
P_1 &:=& \phi_1 \phi_1^T \\
& = & \left[ \begin{array}{c}
1\\
0
\end{array}\right] \left[ \begin{array}{cc}
1 & 0
\end{array}\right]\\
& = & \left[ \begin{array}{cc}
1 & 0\\
0 & 0
\end{array}\right]
\end{array}
$$
\end{block}
\end{frame}
\begin{frame}
\frametitle{Orthogonal projections in Euclidean subspaces}
More generally suppose we are considering $\mathbb{R}^n$ and we have a orthonormal
basis $\{\phi_i\}_{i=1}^r$ for some $r$-dimensional ($r < n$) subspace $\mathcal{M}$
of $\mathbb{R}^n$. Then the projection matrix is given by
$$
P_{\mathcal{M}} = \sum_{i=1}^r \phi_i \phi_i^T
$$
Moreover, if $\{\phi_i\}_{i=1}^r$ is a basis for $\mathcal{M}$, but not necessarily orthonormal, then
$$
P_{\mathcal{M}} = \Phi(\Phi^T \Phi)^{-1}\Phi^T
$$
where $\Phi = [\phi_1, \ldots, \phi_r]$, a matrix whose columns are the basis vectors.
注:这被用在线性回归模型$y=X\beta$的求解上,其最小二乘解析解就是$y$到$X$张成的$p$-维子空间的正交投影:
$$
\hat{\beta} = (X^T X)^{-1} X^T y
$$
\end{frame}
\begin{frame}
\frametitle{Eigendecomposition of a symmetric matrix (对称阵的特征分解)}
Let $C \in \mathbb{R}^{n \times n}$ is a real, symmetric matrix ($C^T = C$).
$v \in \mathbb{R}^n$ is the \textbf{eigenvector (特征向量)} of $C$ such that:
$$
C v = \lambda v
$$
where $\lambda$ is the eigenvalue (特征值) of $C$ corresponding to $v$.
There are $n$ orthonormal eigenvectors for $C$ such that
$$
\langle v_i, v_j \rangle = \delta_{ij}
$$
Let $V = [v_1, \ldots, v_n]$, then
$$
C = V \Lambda V^T
$$
where $\Lambda = \operatorname{diag}(\lambda_1, \ldots, \lambda_n)$.
\end{frame}
\begin{frame}
\frametitle{Singular value decomposition (SVD, 奇异值分解)}
The SVD of an $n \times p$ matrix $H$ is written as
$$
H = \underbrace{U}_{n \times p}\quad \underbrace{\Sigma}_{p \times p}\quad \underbrace{V^T}_{p \times p}
$$
\begin{itemize}
\item $U=[u_1, \cdots, u_p]$ where $\{u_i\}_{i=1}^p$ are real $n$-dimensional
vectors, and called the \textbf{left singular vectors} of $H$. $U^TU = I_p$.
\item $\Sigma = \operatorname{diag}(\sigma_1, \cdots, \sigma_p), \sigma_1 \ge
\sigma_2 \ge \cdots \ge \sigma_p \ge 0$. And $\{\sigma_i\}_{i=1}^p$ are
called the \textbf{singular values} of $H$.
\item $V=[v_1, \cdots, v_p]$ where $\{v_i\}_{i=1}^p$ are $p$-dimensional vectors,
and called the \textbf{right singular vectors} of $H$. $V^T V = I_p$.
\end{itemize}
\end{frame}
\begin{frame}[t]
\frametitle{SVD}
Also note that:
$$
\begin{array}{lcl}
H^T H & = & (U \Sigma V^T)^T U \Sigma V^T \\
& = & V \Sigma U^T U \Sigma V^T \\
& = & V \Sigma^2 V^T\\
HH^T & = & U \Sigma^2 U^T
\end{array}
$$
Therefore,
\begin{itemize}
\item $\{\sigma_1^2, \cdots, \sigma_p^2\}$ are the eigenvalues of $H^TH$ and
$\{v_1, \cdots, v_p\}$ are the corresponding eigenvectors.
\item $\{\sigma_1^2, \cdots, \sigma_p^2\}$ are the $p$-first eigenvalues of
$HH^T$ (the remaining $n-p$ eigenvalues are all zeros) and $\{u_1, \cdots, u_p\}$
are the associated eigenvectors.
\end{itemize}
\end{frame}
\begin{frame}
\frametitle{Application of SVD}
Say we want to solve an \textbf{over-determined} linear equations:
$$
\underbrace{y}_{n \times 1} = \underbrace{X}_{n \times p}\; \underbrace{\beta}_{p \times 1}
$$
\begin{itemize}
\item If $n=p$ and $X=U \Sigma V^T$ with $\sigma_1 \ge \cdots \ge \sigma_p > 0$,
we say $X$ is \textbf{square and non-singular}, $\beta = X^{-1}y$
\item If $n>p$ and $X=U \Sigma V^T$ with $\sigma_1 \ge \cdots \ge \sigma_p > 0$,
we say $X$ is \textbf{non-square and non-singular}, $\beta = (X^TX)^{-1}X^Ty$.
This is called the least squares solution to the over-determined linear equations.
\item When $n<p$, this is an \textbf{under-determined} linear equations, and can be
solved using \textbf{penalized regression}.
\end{itemize}
\end{frame}
\begin{frame}[t]
\frametitle{Assignment 1}
\begin{enumerate}
\item Can you extend the derivatives to vector/matrix form, say, $\frac{d x^T y}{dx}$,
$\frac{d x^T y}{d x^T}$, $\frac{d x^T A y}{dx}$?
\item What is least squares fitting of $\mathbf{X\beta = y}$? Can you use the above matrix
derivatives to reach the normal equation?
\item What is QR-decomposition and Cholesky decomposition? Can you give some comments
on the application of the two decomposition techniques?
\item \texttt{base::qr()} and \texttt{chol()} can be used to compute the two kinds of
decompositions. Give an example to illustrate the usage of the decompositions.
\item What kinds of matrices are positive-definite? positive semi-definite?
\end{enumerate}
\end{frame}
\section{Optimization}
\begin{frame}[t]
\frametitle{Mathematical Optimization}
\begin{itemize}
\item Optimization uses a rigorous mathematical model to determine the
most efficient solution to a described problem.
\item You should identify an objective function
\begin{itemize}
\item Objective is a quantitative measure of the performance
\item Objective is usually a single number
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}[t]
\frametitle{Classification of optimization problems}
Optimization problem can be constrained or unconstrained.
\begin{block}{Common groups}
\begin{itemize}
\item Linear programming (LP)
\begin{itemize}
\item Objective function and constraints are both linear
\item $\min_{\bf x} {\bf c^T x}$ s.t. ${\bf Ax \le b}$ and ${\bf x} \ge 0$
\end{itemize}
\item Quadratic programming (QP)
\begin{itemize}
\item Objective function is quadratic and constraints are linear
\item $\min_{\bf x} {\bf x^TQx + c^T x}$ s.t. ${\bf Ax \le b}$ and ${\bf x} \ge 0$
\end{itemize}
\item Nonlinear programming (NLP)
\begin{itemize}
\item Objective function or at least one constraint is nonlinear
\end{itemize}
\end{itemize}
\end{block}
\end{frame}
\begin{frame}[t]
\frametitle{Nonlinear optimization}
\scriptsize
\begin{figure}[]
\renewcommand\arraystretch{2.5}
\begin{tabular}{l|p{1.5cm}|p{1.5cm}p{1.5cm}p{1.5cm}}
\hline
\textbf{Dimensionality} & \multicolumn{1}{|c|}{One-dimensional} & \multicolumn{3}{c}{Multi-dimensional}\\
\hline
\textbf{Category} & Non-gradient based & Gradient-based & Hessian-based & Non-gradient based \\
\hline
\textbf{Algorithms} & Golden Section Search & Gradient descent & Newton/Quasi-Newton (L-BFGS, BFGS) & Nelder-Mead\\
\hline
\end{tabular}
\end{figure}
\end{frame}
\begin{frame}[t,fragile,containsverbatim]
\frametitle{One-dimensional nonlinear programming}
\begin{itemize}
\item Golden section search
\item Basic steps:
\begin{enumerate}
\item Golden ratio: $\phi = (\sqrt{5} - 1)/2 = .618$
\item Pick an interval $[a,b]$ containing the optimum
\item Evaluate $f(x_1)$ at $x_1 = a + (1-\phi)(b-a)$ and compare with
$f(x_2)$ at $x_2 = a + \phi(b-a)$
\item if $f(x_1) < f(x_2)$, continue the search in the interval $[a, x_1]$, else $[x_2, b]$
\end{enumerate}
\item R command \texttt{stats::optimize()}
\begin{lstlisting}[language=R]
optimize(f=, interval=, ...,
tol = .Machine$double.eps^0.25)
\end{lstlisting}
\end{itemize}
\end{frame}
\begin{frame}[t]
\frametitle{Nondifferentiable function}
\begin{itemize}
\item $f(x) = |x - 2| + 2|x - 1|$
\item How to solve it?
\end{itemize}
<<echo=FALSE,fig=TRUE>>=
f <- function(x) abs(x-2) + 2*abs(x-1)
curve(f, 0, 3)
@
\end{frame}
\begin{frame}[t]
\frametitle{Newton-Raphson}
\begin{itemize}
\item Newton method is often used to find the zeros of a function
\item Minima fulfill the conditions $f^{\prime}(x^*)=0$ and $f^{\prime\prime}(x^*)>0$, so
Newton can be used to find the zeros of the first derivative
\item Basic steps
\begin{enumerate}
\item Approximate the function at the starting point with a linear
tangent (e.g., second-order Taylor expansion $t(x) \approx f^{\prime}(x_0) + (x-x_0)f^{\prime\prime}(x_0)$)
\item Find the intersect $t(x_i)=0$ as an approximation to $f^{\prime}(x^*)=0$
\item Use the intersect as the new starting point
\item Finally, the algorithm $x_{n+1} = x_n - \frac{f^{\prime}(x_n)}{f^{\prime\prime}(x_n)}$
is repeated until $f^{\prime}(x_n)$ is close enough to 0.
\end{enumerate}
\end{itemize}
\end{frame}
\begin{frame}[t]
\frametitle{Hessian-based: BFGS and L-BFGS-B}
\begin{itemize}
\item Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm builds on
the idea of Newton method to take gradient information into
account
\item Gradient information comes from an approximation of the Hessian
matrix
\item No guaranteed conversion; especially problematic if Taylor
expansion does not fit well
\item L-BFGS-B stands for limited-memory-BFGS-box:
\begin{itemize}
\item Extension of BFGS
\item Memory-efficient implementation
\item Additional handles box constraints
\end{itemize}
\end{itemize}
\end{frame}
\begin{frame}[t]
\frametitle{Himmelblau's function}
$$
f(x, y) = (x^2 + y - 11)^2 + (x + y^2 - 7)^2
$$
\begin{itemize}
\item Himmelblau's function is a popular multi-modal function to
benchmark optimization algorithms.
\item For equivalent minima are located at $f(-3.78,-3.28)=0$,
$f(-2.80,3.13)=0$, $f(3,2)=0$, $f(3.58,-1.85)=0$
\end{itemize}
<<echo=FALSE,fig=TRUE>>=
fn <- function(para) {
matrix.A <- matrix(para, ncol=2)
x <- matrix.A[,1]
y <- matrix.A[,2]
f.x <- (x^2+y-11)^2 + (x+y^2-7)^2
return(f.x)
}
par <- c(1,1)
xy <- as.matrix(expand.grid(seq(-5,5, length=101),
seq(-5,5,length=101)))
colnames(xy) <- c('x', 'y')
df <- data.frame(fnxy = fn(xy), xy)
library(lattice)
wireframe(fnxy ~ x*y, data=df, shade=TRUE, drape=FALSE,
scales=list(arrows=FALSE),
screen=list(z=-240, x=-70, y=0))
@
\end{frame}
\begin{frame}[t]
\frametitle{Gradient-free method: Nelder-Mead}
\begin{itemize}
\item Nelder-Mead sovles multi-dimensional equations using function values
\item Works with non-differentiable functions
\item Basic steps:
\begin{enumerate}
\item Choose a simplex consisting of $n+1$ points $p_1, p-2, \dots, p_{n+1}$
are chosen with $n$ being the number of variables
\item Calculate $f(p_i)$ and sort by value, e.g. $f(p_1) \le f(p_2) \le \dots
\le f(p_{n+1})$
\item Check if the best value is good enough, if so, stop
\item Drop the point with hightest $f(p_i)$ from the simplex
\item Choose a new point to be added to the simplex
\item Continue with step 2
\end{enumerate}
\end{itemize}
\end{frame}
\end{CJK*}
\end{document}