forked from johnmyleswhite/MLNotes
-
Notifications
You must be signed in to change notification settings - Fork 0
/
mcmc.html
188 lines (143 loc) · 4.82 KB
/
mcmc.html
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
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN"
"http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<!-- saved from url=(0014)about:internet -->
<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
<title></title>
<base target="_blank"/>
<style type="text/css">
body, td {
font-family: sans-serif;
background-color: white;
font-size: 12px;
margin: 8px;
}
tt, code, pre {
font-family: 'DejaVu Sans Mono', 'Droid Sans Mono', 'Lucida Console', Consolas, Monaco, monospace;
}
h1 {
font-size:2.2em;
}
h2 {
font-size:1.8em;
}
h3 {
font-size:1.4em;
}
h4 {
font-size:1.0em;
}
h5 {
font-size:0.9em;
}
h6 {
font-size:0.8em;
}
a:visited {
color: rgb(50%, 0%, 50%);
}
pre {
margin-top: 0;
max-width: 95%;
border: 1px solid #ccc;
}
pre code {
display: block; padding: 0.5em;
}
code.r {
background-color: #F8F8F8;
}
table, td, th {
border: none;
}
blockquote {
color:#666666;
margin:0;
padding-left: 1em;
border-left: 0.5em #EEE solid;
}
hr {
height: 0px;
border-bottom: none;
border-top-width: thin;
border-top-style: dotted;
border-top-color: #999999;
}
@media print {
* {
background: transparent !important;
color: black !important;
filter:none !important;
-ms-filter: none !important;
}
body {
font-size:12pt;
max-width:100%;
}
a, a:visited {
text-decoration: underline;
}
hr {
visibility: hidden;
page-break-before: always;
}
pre, blockquote {
padding-right: 1em;
page-break-inside: avoid;
}
tr, img {
page-break-inside: avoid;
}
img {
max-width: 100% !important;
}
@page :left {
margin: 15mm 20mm 15mm 10mm;
}
@page :right {
margin: 15mm 10mm 15mm 20mm;
}
p, h2, h3 {
orphans: 3; widows: 3;
}
h2, h3 {
page-break-after: avoid;
}
}
</style>
<!-- MathJax scripts -->
<script type="text/javascript" src="https://c328740.ssl.cf1.rackcdn.com/mathjax/2.0-latest/MathJax.js?config=TeX-AMS-MML_HTMLorMML">
</script>
</head>
<body>
<p>In Bayesian statistics, all information about unknown quantities is contained in their posterior distribution after conditioning on observed data. If the quanties are denoted \(\theta\) and the observed data are denoted \(D\), then this quantity is<br/>
\[<br/>
\mathbb{P}(\theta | D) = \frac{\mathbb{P}(D | \theta) \mathbb{P}(\theta)}{\mathbb{P}(D)}<br/>
\]<br/>
While this posterior probability distribution is well-defined mathematically in most cases, it is not computationally usable and is often analytically intractable. There are two problems in particular:</p>
<ul>
<li>The denominator, \[\mathbb{P}(D)\], is often unamenable to mathematical analysis and not computatable because of intractable nested summations. As such, we can learn the density function of the posterior at any given value of \[\theta\] only up to a constant. That constant is \[\mathbb{P}(D)\], because \[\mathbb{P}(D)\] does not vary with \(\theta\).</li>
<li>Even if we knew the denominator perfectly (which we don't), we still would have no computational tools for drawing samples from this complex probability distribution. The simplest method known for generating samples from a probability distribution involves the inverse CDF of the distribution. We do not have access to the inverse CDF of \[\mathbb{P}(\theta | D)\]: if we did, we would be able to produce \[\mathbb{P}(\theta | D)\] – which we just we cannot. So we must employ methods that do not require access to the inverse CDF. The alternative sampling methods we will turn to are universal in that they can be applied anytime we have access to an unnormalized density function (as we do by looking at the numerator in the equation above), but they achieve universality at the expense of substantially increased computation costs.</li>
</ul>
<p>Here I enumerate some methods for working with probability distributions; I will discuss them in detail in other sections of my notes.</p>
<p>Methods for estimating the expectation of an intractable probability distribution for which we have a (possibly unnormalized) density function:</p>
<ul>
<li>Importance sampling</li>
<li>Annealed importance sampling</li>
</ul>
<p>Non-Markov Chain Methods for obtaining samples from a (possibly unnormalized) density function:</p>
<ul>
<li>Rejection sampling</li>
</ul>
<p>Markov Chain methods for obtaining samples from a (possibly unnormalized) density function:</p>
<ul>
<li>Metropolis-Hastings</li>
<li>Gibbs sampling</li>
<li>Slice sampling</li>
<li>Hamiltonian Monte Carlo</li>
<li>Exact sampling</li>
</ul>
<p>Interesting questions remain about the convergence rates, diagnostics, reliability, internal consistency vs. external validity, of these methods.</p>
</body>
</html>