LDA Œ˘l%
Rickjin(@), version 1.0
2013 c 2 8 F
Contents
0.1 m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
0.2 Gamma…Œ . . . . . . . . . . . . . . . . . . . . . . . . . .
0.2.1 Gamma …Œ)P . . . . . . . . . . . . . . . . . . . . .
0.2.2 Gamma …Œ! . . . . . . . . . . . . . . . . . . . . . .
0.2.3 l'Gamma ' . . . . . . . . . . . . . . . . .
0.3 @£Beta/Dirichlet' . . . . . . . . . . . . . . . . . . . . . . .
0.3.1 giZ—@£Beta ' . . . . . . . . . . . . . . . . .
0.3.2 Beta-Binomial . . . . . . . . . . . . . . . . . . . . .
0.3.3 Dirichlet-Multinomial . . . . . . . . . . . . . . . . .
0.3.4 Beta/Dirichlet '5 . . . . . . . . . . . . . . .
0.4 MCMC Gibbs Sampling . . . . . . . . . . . . . . . . . . . . . .
0.4.1 ¯[ . . . . . . . . . . . . . . . . . . . . . . . . . . .
0.4.2 Œ…9†›' . . . . . . . . . . . . . . . . . . . . .
0.4.3 Markov Chain Monte Carlo . . . . . . . . . . . . . . . . .
0.4.4 Gibbs Sampling . . . . . . . . . . . . . . . . . . . . . . . .
0.5 ' . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
0.5.1 Unigram Model . . . . . . . . . . . . . . . . . . . . . . . .
0.5.2 Topic Model PLSA . . . . . . . . . . . . . . . . . . . .
0.6 LDA ' . . . . . . . . . . . . . . . . . . . . . . . . . . . .
0.6.1 iZ5K . . . . . . . . . . . . . . . . . . . . . . . . . . .
0.6.2 nL§') . . . . . . . . . . . . . . . . . . . . . . . . .
0.6.3 Gibbs Sampling . . . . . . . . . . . . . . . . . . . . . . . .
0.6.4 Training and Inference . . . . . . . . . . . . . . . . . . . .
0.7 P . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
2
2
7
10
13
13
17
21
24
24
24
27
31
34
37
38
42
44
44
46
49
52
53
1
0.1 m
3Machine Learning ¥§LDA ·~^.{¡Linear Discrim-
inant Analysis Latent Dirichlet Allocation§3ø'¥•l%
·" LDA ·3'¥Ø˝¶.§aquSVD, PLSA
.§–^uf´',3'´'¥·Øk^." Ø3
·§ø.¥9Œ˘£k:ı§)Gamma …Œ§Dirichlet '§
Dirichlet-Multinomial §Gibbs Sampling, Variational Inference, d'
§PLSA , –9LDA '"
ø'8I§·˚3˘Sn)LDA .¥§I)
›Œ˘£" ·g,?n!¯˘S!Œ
§§ˆø˚§IŒ˘˜:£˜LFWk)
5V˙ŒnO6ø"
'IK!/l%0i§ˇl%¿Xgd!!–UŒ1
§[!?nJk>/¶,•F"l%·ØN·n)§
=Bƒ·’uŒ˘l%" Øu'¥?§Hu&•#L˘
Ærickjin, ‰·ezhihuijin@gmail.com"
0.2 Gamma…Œ
0.2.1 Gamma …Œ)P
∞
˘pŒ˘§•˘SLXek:AGamma…Œ
Γ(x) =
tx−1e−tdt
ˇL'¨'{§–ø…ŒkXe485
0
Γ(x + 1) = xΓ(x)
u·ØN·y†§Γ(x) …Œ–⁄·ƒ3¢Œ8§kXe5
Γ(n) = (n − 1)!
˘SGamma …Œ§ıc–5•kƒfl
1. øøo%…Œ§Œ˘[·Xض
2. ‰ ´Γ … Œ § ƒ ø … Œ ‰ ´ vΓ(n) = n!
·Γ(n) = (n − 1)!
2
C]§uyk'z]0Gamma …Œuy{⁄§
‘§I‰Œ˘§ø·{‘"
1728c§xn3˜ŒflK§ˇ‘·rŒˇœ
“‰´lŒ8¢Œ8§~XŒ1, 4, 9, 16,··· –^ˇœ“n2
g,L§=Bn ¢Œ§øˇœ“·ß—‰´" *‘
·–Ø^†w›ˇLy = x2ˇL⁄kŒ:(n, n2)ø:§
l–r‰´3Œ8œ“¢Œ8" Uxnm'?n
ƒS1, 2, 6, 24, 120, 720,··· , •–O2!, 3!, ·˜–O2.5!Q”•r
—(n, n!):x3I¶§(¢–w§N·x^ˇLø
:†w›
Figure 1: ˇL(n, n!)›
xnˆ{)ßøflK§u·&Z.d.ª|ƒ33ß
Z.ª|§du.ßZ.ª|3‹§ƒˇdøfl
K".u1729 c{)ßøflK§ddΓ …Œ)§
.k22"
fl ¢ ˜ k ) ßn! O fl K · ß Z . ª |§ ƒ u y§ X
Jm, n·Œ§XJm → ∞§k
1 · 2 · 3··· m
(1 + n)(2 + n)··· (m − 1 + n)
(m +
)n−1 → n!
n
2
u·^øˆ¡ƒ¨“–rn!‰´¢Œ8"~X§n = 2.5,
3
m v§˜u“–CqO2.5!"
.,uyn! –^Xeˆ¡ƒ¨L
2
n
3
n
1
4
n
2
3
1
n + 1
2
n + 2
3
n + 3
^4/“§ø“fn–
··· = n!
1 · 2 · 3··· m
(1 + n)(2 + n)··· (m + n)
lim
m→∞
(m + 1)n = n!
>–n
1 · 2 · 3··· m
(1 + n)(2 + n)··· (m + n)
(m + 1)n
=1 · 2 · 3··· n ·
(n + 1)(n + 2)m
(1 + n)(2 + n)··· m
·
(m + 1)n
(m + 1)(m + 2)··· (m + n)
(1)
(2)
(m + 1)n
(m + 1)(m + 2)··· (m + n)
m + 1
m + k
→ n!
(m → ∞)
=n!
=n!
n
k=1
⁄–(1)!(2)“⁄Æ"
.m'}`l{~fm'O§ww·˜k5˘§
.4Œ˘*8B" n = 1/2 §\(1)“O§n
–
1
! =
2 · 4
3 · 3
· 4 · 6
5 · 5
· 6 · 8
7 · 7
· 8 · 10
9 · 9
···
›y = x(1 − x) e¡¨(·»1¡¨)§’
,m>—˝¶Wallis œ“’Ø" Wallis 31665cƒ^{O
2
uπXe(J§
2 · 4
3 · 3
u·§.|^Wallis œ“XeØ⁄(J
· 8 · 10
9 · 9
··· =
π
4
· 6 · 8
· 4 · 6
5 · 5
7 · 7
1
! =
2
√
π
2
.pd·kœŒ˘[§·.pd”»" p
d·Pk§Œ˘~>§uL(J%rgŁ,!§
3e⁄(J§øŒ˘[Øpd1¶.”§†
~ˇL†œ§ƒ'¥ 3eƒXŒ˘Ł
,§'ky>" .˚.dQ‘L0.,ƒ·⁄k<
4
Figure 2: Œ˘[.
P"0¯|3ƒ¶˝5Œ˘6¥Ø.Œ˘8B
“´"
.w( 1
2 )! ¥,kπ, ØŒ˘[§kπ /7,k’
¨'" dd.n! ‰–L,«¨'/“§u·.m'}`rn!
L¨'/“" ,Wallis ¨'vku†5§Wallis ·ƒ^
“O§·Wallis œ“L§˜·3?n¨
2 dx§Wallis Øu§.m'˜Xe/“¨'
' 1
2 (1 − x) 1
0 x 1
1
J(e, n) =
xe(1 − x)ndx
d?n Œ§e ¢Œ"|^'¨'{§N·
0
J(e, n) =
n
e + 1
J(e + 1, n − 1)
›Eƒ^ªSœ“§“–
J(e, n) =
1 · 2··· n
(e + 1)(e + 2)··· (e + n + 1)
u·.Xe›“f
n! = (e + 1)(e + 2)··· (e + n + 1)
1
0
xe(1 − x)ndx
e5§.ƒ^:OE|§e = f /g ¿-f → 1, g → 0, ,Ø
“m>O4(4OL§d?§J§k,˘w¡
º'zj)§u·.Xe{’⁄(J
n! =
(− log t)ndt
1
0
5
.⁄ırn!L¨'/“ XJ•Ct = e−u,–•
~Gamma …Œ/“
∞
0
n! =
une−udu
∞
1
u·,|^“rƒ¢Œ8§•Gamma …Œ/“
Γ(x) =
(− log t)x−1dt =
tx−1e−tdt
0
0
Figure 3: Γ(x)
Gamma …ŒØ§•5ww1flK§Gamma …Œ‰´
Γ(n) = (n − 1)!, øw5˝O" XJ•?e§rGamma …
Œ‰´¥tx−1 Otx
∞
0
Γ(x) =
txe−tdt
ø–ƒΓ(n) = n!" .@Gamma…Œ‰´·X⁄«§
JΓ(n) = n!§·.uoˇ§Y?UGamma …Œ‰
´§ƒΓ(n) = (n − 1)!"V4Œ˘[ØGamma …Œ?\
˜¥§@ø‰´§u·ø‰´⁄Q⁄fl¢" kŒ˘[§
Uˇ·.˜Xe¨'
ø…Œy3¡Beta …Œ"XJGamma …Œ‰´vΓ(n) = (n−1)!,
@ok
1
0
B(m, n) =
xm−1(1 − x)n−1dx
Γ(m)Γ(n)
Γ(m + n)
B(m, n) =
6
~⁄Ø¡/“"·XJΓ(n) = n! ‰´§-
1
E(m, n) =
xm(1 − x)ndx
Kk
0
E(m, n) =
Γ(m)Γ(n)
Γ(m + n + 1)
ø/“w,XB(m, n)‘{§Œ˘[o·Ø3Œ˘œ“{a"
)ıGamma …Œ{⁄§
• Philip J. Davis, Leonhard Euler’s Integral: A Historical Profile of the
Gamma Function
• Jacques Dutka, The Early History of the Factorial Function
• Detlef Gronnau, Why is the gamma function so as it is?
0.2.2 Gamma …Œ!
Each generation has found something of interest to say about the
gamma function. Perhaps the next generation will also.
—Philip J.Davis
Gamma …Œl§)m'NıŒ˘[?1˜§)pd!V4!
%dA.d!7" ø…Œ3yŒ˘'¥\˜§3V˙
¥·ˆ?3§ØıO'ø…Œ’" Gamma …Œƒ
2§˜k§kStirling œ“aq(
√
Γ(x) ∼
2πe−xxx− 1
2
,§Gamma …Œ=–‰´3¢Œ8§–E†¡"
Figure 4: E†¡Gamma …Œ
7