-
Notifications
You must be signed in to change notification settings - Fork 17
/
lecture7.tex
268 lines (172 loc) · 8.92 KB
/
lecture7.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
\documentclass[11pt]{article}
%%%%%%%%% options for the file macros.tex
\def\showauthornotes{1}
\def\showkeys{0}
\def\showdraftbox{0}
% \allowdisplaybreaks[1]
\input{macros}
\allowdisplaybreaks
\usepackage{tikz}
\usepackage[
backend=biber,
% giveninits=true,
% natbib=true,
style=alphabetic,
url=false,
% doi=true,
hyperref,
backref=true,
backrefstyle=none,
maxbibnames=10,
sortcites
]{biblatex}
\addbibresource{papers.bib}
%%%%%%%%% Authornotes
\newcommand{\Snote}{\Authornote{S}}
\newenvironment{tight_enumerate}{
\begin{enumerate}
\setlength{\itemsep}{2pt}
\setlength{\parskip}{1pt}
}{\end{enumerate}}
\newenvironment{tight_itemize}{
\begin{itemize}
\setlength{\itemsep}{2pt}
\setlength{\parskip}{1pt}
}{\end{itemize}}
\addbibresource{refs.bib}
%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{document}
\newcommand{\coursenum}{{CSC 2421H}}
\newcommand{\coursename}{{Graphs, Matrices, and Optimization}}
\newcommand{\courseprof}{Sushant Sachdeva}
\lecturetitle{7}{Concentration Bounds}{Fengwei Sun}{29 Oct 2018}
\section{Scalar Chernoff Bound}
\begin{definition}
Let $X_1, \dots, X_t$ be independent random variables such that
\[
0 \le X_i \le R, \mathbb{E} \sum_{i} X_i = \sum_{i} \mathbb{E} X_i = \mu
\]
Then for all $0<\epsilon<1$, we have
\[
P[\sum_{i} X_i \ge (1+\epsilon)\mu] \le e^{-\frac{\epsilon^2 \mu}{3R}}, P[\sum_{i} X_i \le (1-\epsilon)\mu] \le e^{-\frac{\epsilon^2 \mu}{2R}}
\]
\end{definition}
\paragraph{Example:}
Suppose we conduct $t$ independent tosses of a fair coin. Let $X_i=\begin{cases} 1 \quad if \; heads \\ 0 \qquad o/w \end{cases}$. Then the number of heads in this trial is $\sum_{i=1}^{t} X_i$, and $\mathbb{E}(\# \; of \; heads)=\mathbb{E}\sum_{i=1}^{t} X_i = \frac{t}{2}$.\\
To obtain a good estimate of the probability that we see at least 600 heads out of 1000 tosses, we can apply the Chernoff bound with the parameters $\epsilon=0.2, R=1, \mu=500$, and get
\[
P(at \; least \; 600 \; heads \; out \; of \; 1000 \; tosses) = P(\sum_{i=1}^{1000} X_i \ge (1+\epsilon) 500) \le e^{-\frac{0.2^2 * 500}{3}}= e^{-\frac{20}{3}} \approx e^{-7}
\]
\paragraph{Question:}
You can a coin with a bias in $\{\frac{1}{2}+\alpha, \frac{1}{2}-\alpha\}$. How many tosses do you need to decide which bias with probability of at least $1-\delta$?
\paragraph{Algorithm:}
\begin{enumerate}
\item Toss the coin $t$ times independently;
\item If there are at least $\frac{t}{2}$ heads, output $\frac{1}{2}+\alpha$; otherwise output $\frac{1}{2}-\alpha$.
\end{enumerate}
We would like to bound $P(failure) \le \delta$. \\
\paragraph{Case 1:}
The coin has a bias $\frac{1}{2}+\alpha$. Then $P(failure) = P(\sum X_i \le \frac{t}{2})$. \\
To apply Chernoff, we find that $R=1, \mu=t(\frac{1}{2}+\alpha)$. Since we want $(1-\epsilon)\mu = \frac{t}{2}$, we have $\epsilon = 1-\frac{t}{2\mu} = 1-\frac{1}{2\alpha}$. Therefore,
\[
P(\sum X_i \le \frac{t}{2}) = P(\sum X_i \le (1-\epsilon)\mu) \le e^{-\frac{\epsilon^2\mu}{2}} \le \delta
\]
Since $\epsilon^2\mu \approx \Theta(t \alpha^2)$, we have
\[
t \ge \frac{\Theta(1)}{\alpha^2}\log \frac{1}{\delta}
\]
\paragraph{Case 2:}
The coin has a bias $\frac{1}{2}-\alpha$. Then $P(failure) = P(\sum X_i \ge \frac{t}{2})$. \\
To apply Chernoff, we find that $R=1, \mu=t(\frac{1}{2}-\alpha)$. Since we want $(1+\epsilon)\mu = \frac{t}{2}$, we have $\epsilon = \frac{t}{2\mu}-1 = \frac{1}{2\alpha}-1$. Therefore,
\[
P(\sum X_i \ge \frac{t}{2}) = P(\sum X_i \ge (1+\epsilon)\mu) \le e^{-\frac{\epsilon^2\mu}{3}} \le \delta
\]
Since $\epsilon^2\mu \approx \Theta(t \alpha^2)$, we have
\[
t \ge \frac{\Theta(1)}{\alpha^2}\log \frac{1}{\delta}
\]
Both cases indicate that the $t$ is bounded by the logarithm of $\frac{1}{\delta}$, which means that $t$ would be relatively small even for very small $\delta$.
\section{Matrix Chernoff Bound}
\begin{definition}
Let $X_1, \dots, X_t \in \mathbb{R}^{d \times d}$ be symmetric independent random variables such that
\[
0 \preceq X_i \preceq RI, \; \mu_{min} I \preceq \mathbb{E}\sum X_i \preceq \mu_{max} I
\]
Then we have
\[
P[\lambda_{max}(\sum X_i) \ge (1+\epsilon) \mu_{max}] \le d e^{-\frac{\epsilon^2 \mu_{max}}{3R}},
\]
\[
P[\lambda_{min}(\sum X_i) \le (1-\epsilon) \mu_{min}] \le d e^{-\frac{\epsilon^2 \mu_{min}}{2R}}
\]
\end{definition}
\paragraph{Note:}
\begin{enumerate}
\item The condition that $0 \preceq X_i \preceq RI$ is equivalent to $||X_i|| \le R$, or $\lambda_{max}(X_i) \le R$, where $||A|| = \max_{x \ne 0} \frac{||Ax||}{||x||}$, and when $A$ is symmetric, $||A||=max\{\lambda_{max}, -\lambda_{min}\}$.
\item $\mu_{min}=\lambda_{min}(\mathbb{E}\sum X_i), \mu_{max}=\lambda_{max}(\mathbb{E}\sum X_i)$.
\end{enumerate}
\paragraph{Example:}
(Construction of random expander graphs) Suppose we would like to generate an expander graph with $n$ vertices (assuming $n$ is even).
Define a matching as a graph where $d_v=1$ for all vertices $v$. Let $H=\frac{1}{t}$(union of $t$ independent perfect matching). Notice that for all vertices $u$ in the graph $H$, $d_u = 1$.
Using the matrix Chernoff bound, we can show that $H$ is an expander.\\
The Laplacian of $H$ is $L_H = \sum_{i} \frac{1}{t} L_i$, where $L_i$ is the Laplacian of the $i^{th}$ matching.
Let $X_i = \frac{1}{t} L_i$. We know that $X_i \succeq 0$ and $\lambda_{max}(X_i) = \frac{1}{t} \lambda_{max}(L_i)=\frac{2}{t}$.
Also, if we look at a specific vertex $u$ in each matching, it is connected to all other vertices with equal probablity $\frac{1}{n-1}$. This indicates that $\mathbb{E}L_H = \mathbb{E}L_1 = \frac{1}{n-1} L_{K_n}$. \\
Before we apply Chernoff bound on $X_i's$, one issue we notice is that $\lambda_{min}(X_i)=0$.
To fix that, we let $X_i = \frac{1}{t} L_i + \frac{1}{t(n-1)} \mathbbm{1}\mathbbm{1}^{\top}$. Now we have $\mathbb{E}\sum X_i = \frac{1}{n-1} L_{K_n} + \frac{1}{n-1} \mathbbm{1}\mathbbm{1}^{\top} = \frac{n}{n-1} I_n$, and thus $\mu_{max}=\mu_{min}=\frac{n}{n-1}$.\\
We can also show that, $\lambda_{max}(X_i) \le \frac{2}{t}$ after the change of variable.
Let $y = \hat{y} + c \frac{\mathbbm{1}}{\sqrt{n}}$ where $\hat{y}^{\top}\mathbbm{1}=0$. Then
\[
y^{\top} X_i y = \hat{y}^{\top} (\frac{1}{t} L_i) \hat{y} + \frac{c^2n}{t(n-1)} \le (\frac{2}{t})\hat{y}^{\top}\hat{y} + \frac{n}{(n-1)t} c^2 \le \frac{2}{t} (\hat{y}^{\top}\hat{y}+c^2) \le \frac{2}{t} ||y||^2 = \frac{2}{t}
\]
\hfill \break
Now we apply the Chernoff bound, and get the following:
\[
P[\lambda_{max}(\sum X_i) \ge (1+\epsilon) \frac{n}{n-1}] \le n e^\frac{-\epsilon^2 \frac{n}{n-1}}{3 \cdot \frac{2}{t}} = n e^{-\frac{\epsilon^2 t}{6}(\frac{n}{n-1})}
\]
If we pick $t \ge \frac{12}{\epsilon^2} \log n$, we have $P[\lambda_{max}(\sum X_i) \ge (1+\epsilon)\frac{n}{n-1}] \le n \cdot \frac{1}{n^2} = \frac{1}{n}$. \\
Similarly, we have $P[\lambda_{min}(\sum X_i) \le (1-\epsilon) \frac{n}{n-1}] \le \frac{1}{n}$. \\
Therefore, we can conclude that, with probability of at least $1 - \frac{2}{n}$,
\[
\lambda_{max}(\sum X_i) \le (1+\epsilon)\frac{n}{n-1}, \lambda_{min}(\sum X_i) \ge (1-\epsilon)\frac{n}{n-1}
\]
or
\[
(1-\epsilon)\frac{n}{n-1} I \preceq \sum X_i \preceq (1+\epsilon)\frac{n}{n-1}
\]
\hfill \break
To see that $H$ is a good approximation of a complete graph, let $ \Pi = I - \frac{1}{n} \mathbbm{1}\mathbbm{1}^{\top} = \frac{1}{n} L_{K_n}$. Notice that $\Pi^2 = \Pi$.
Consider $\Pi^{\top} (\sum X_i) \Pi$. We have
\[
(1-\epsilon)\frac{n}{n-1} \frac{1}{n} L_{K_n} \preceq \Pi^{\top} (\sum X_i) \Pi \preceq (1+\epsilon)\frac{n}{n-1} \frac{1}{n} L_{K_n}
\]
Since
\[
\Pi^{\top} (\sum X_i) \Pi = \Pi^{\top} (L_H + \frac{1}{n-1} \mathbbm{1}\mathbbm{1}^{\top})(I - \frac{1}{n} \mathbbm{1}\mathbbm{1}^{\top})
\]
\[
= \Pi^{\top} (L_H + \frac{1}{n-1} \mathbbm{1}\mathbbm{1}^{\top} - \frac{n}{n(n-1)} \mathbbm{1}\mathbbm{1}^{\top}) = \Pi^{\top} L_H = L_H
\]
Therefore,
\[
(1-\epsilon)\frac{1}{n-1} L_{K_n} \preceq L_H \preceq (1+\epsilon)\frac{1}{n-1} L_{K_n}
\]
\hfill \break
Now we get an $\epsilon$-expander $H$ with $t \cdot \frac{n}{2} = \Theta (\frac{n\log n}{\epsilon^2})$ edges, and $L_H \approx_\epsilon L_{K_n}$.\\
\hfill \break
In general, we would like to write the Chernoff bound as the following: \\
With probability of at least $1 - 2de^{-\frac{\epsilon^2 \mu_{min}}{2R}}$, $(1-\epsilon) \mu_{min} I \preceq \sum X_i \preceq (1+\epsilon)\mu_{max} I$. \\
\begin{definition}
$H=(V, E')$ is an $\epsilon$-spectral sparsifier of $G=(V, E)$ if $\frac{1}{1+\epsilon} L_G \preceq L_H \preceq (1+\epsilon) L_G$, denoted as $L_H \approx_\epsilon L_G$.
Equivalently, $\forall x \in \mathbb{R}^V, \frac{1}{1+\epsilon} x^{\top} L_G x \le x^{\top} L_H x \le (1+\epsilon) x^{\top} L_G x$.
\end{definition}
\paragraph{Note:}
Let $x = \mathbbm{1}_S$ where $S \subset V$. Then $x^{\top} L_H x = \sum_{(u,v) \in E} w(u, v) (x(u)-x(v))^2 = |E(S, \bar{S})|$.
\paragraph{Theorem:}
For all $G=(V, E)$, there exists $H=(V, E')$ such that $L_H \approx_\epsilon L_G$ and $|E'| \le \Theta(\frac{n\log n}{\epsilon^2})$
\printbibliography
\end{document}
%%% Local Variables:
%%% mode: latex
%%% TeX-master: t
%%% End: