일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
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 |
- 언리얼엔진
- Font Generation
- cv
- 폰트생성
- 생성모델
- ue5.4
- Stat110
- Diffusion
- RNN
- ddpm
- NLP
- Few-shot generation
- BERT
- Unreal Engine
- WinAPI
- GAN
- 모션매칭
- multimodal
- 오블완
- CNN
- WBP
- Generative Model
- dl
- motion matching
- deep learning
- animation retargeting
- UE5
- 딥러닝
- userwidget
- 디퓨전모델
- Today
- Total
Deeper Learning
[Statistics 110] Lecture 24: Gamma distribution and Poisson process 본문
[Statistics 110] Lecture 24: Gamma distribution and Poisson process
Dlaiml 2023. 1. 8. 00:19Gamma Function(감마 함수)
Γ(a)=∫0inf
성질
\Gamma(n) = (n-1)!\\ \Gamma(x+1) = x\Gamma(x)\\ \Gamma(1/2) =\sqrt \pi
Gamma Distribution(감마 분포)
모든 가능한 구간에 대해 적분값이 1이 되도록 정규화 하기 위해 간단하게 감마 함수를 감마함수로 나누면
\int_0^{\inf}\frac1{\Gamma(a)}x^ae^{-x}\frac{dx}{x} = 1\\p.d.f\space f_X(x) = \frac1{\Gamma(a)}x^ae^{-x}\frac{1}{x}
X \sim Gamma(a,1)의 pdf식을 위와 같이 구할 수 있다.
Gamma(a,\lambda)의 PDF
Y = \frac X\lambda, X \sim Gamma(a,1)\\Y \sim Gamma(a, \lambda)\\f_Y(y) = f_X(x)\frac{dx}{dy}\\ y=\frac x\lambda, x=\lambda y, \frac{dx}{dy}=y\\ \therefore f_Y(y) = \frac1{\Gamma(a)}(\lambda y)^ae^{-\lambda y}\frac{1}{y}
Poisson Process
감마분포와 지수분포의 관계
T_n은 n번째 이메일이 도착한 시점
t시점까지 도착한 이메일의 수를 N_t라고 하면 N_t \sim Pois(\lambda t).
- 가정: 서로 다른 시간 구간에서 이메일의 도착 확률은 독립적
P(T_1>t) = P(N_t = 0) = e^{-\lambda t}
지수분포와의 관계를 볼 수 있음
지수함수의 무기억성에 의해 1 번째 메일 ~ 2 번째 메일이 오는 사이 시간의 분포도 지수분포를 따름.
도착 간격 시간은 i.i.d Expo(\lambda).
시간 간격이 아닌 절대적인 시간을 구해보자.
T_n은 n번째 이메일이 도착한 시점으로 0번째 ~ 1번째 메일 도착시간 간격 + 그 후 다음 메일 도착 시간 간격을 쭉 더해서 구할 수 있다.
T_n = \sum_{j=1}^nX_j, X_j = Expo(\lambda),i.i.d \\ T_n \sim Gamma(n,\lambda)
적률생성함수를 통한 증명
X_j가 i.i.d Expo(1)을 따를 때 T = \sum_{j=1}^nX_j가 Gamma(n,1) 을 따르는 것을 증명하라
X_1의 MGF는 \frac1{1-t}(t<1).
X_j끼리는 i.i.d로 모두 독립이기 때문에 T의 MGF는 X_j의 MGF를 모두 곱한것과 같다.
MGF of T_n = (\frac1{1-t})^n.
이제 감마분포의 MGF가 위와 같음을 증명하면 된다.
Y \sim Gamma(n,1)이라고 하자. LOTUS에 의해
E(e^{tY}) = \frac1{\Gamma(n)}\int_0^{\inf}y^ne^{-ty}e^{-y}\frac{dy}{y} = \frac1{\Gamma(n)}\int_0^{\inf}y^ne^{-(1-t)y}\frac{dy}{y}\\x = (1-t)y, dx=(1-t)dy\\\frac1{\Gamma(n)}\int_0^{\inf}y^ne^{-(1-t)y}\frac{dy}{y} = \frac{(1-t)^{-n}}{\Gamma(n)}\int_0^{\inf}x^ne^{-x}\frac{dx}{x} = (\frac{1}{1-t})^n
n차 적률
X \sim Gamma(a,1)일 때, E(X^c) 구하라
LOTUS에 의해
E(X^c) = \frac{1}{\Gamma(a)}\int_0^{\inf}x^cx^ae^{-x}\frac{dx}{x} = \frac{1}{\Gamma(a)}\int_0^{\inf}x^{a+c}e^{-x}\frac{dx}{x}
감마함수의 식의 형태와 같기 때문에 치환하면
\Gamma (a) = \int_0^{\inf}x^ae^{-x}\frac{dx}{x}, (a>0,\R) \\\frac{1}{\Gamma(a)}\int_0^{\inf}x^{a+c}e^{-x}\frac{dx}{x} = \frac{\Gamma(a+c)}{\Gamma(a)},(a+c>0)
이를 이용해서 1,2차 적률과 분산을 구해보면
E(X^1) = \frac{\Gamma(a+1)}{\Gamma(a)} = \frac{a\Gamma(a)}{\Gamma(a)} = a\\E(X^2) = \frac{\Gamma(a+2)}{\Gamma(a)} = \frac{(a+1)\Gamma(a+1)}{\Gamma(a)} =\frac{a(a+1)\Gamma(a)}{\Gamma(a)} = a(a+1)\\Var(X) = a(a+1) - a^2 = a
\lambda가 1일 때, 평균과 분산이 모두 a.
Gamma(a,\lambda)의 평균은 a/\lambda, 분산은 a/\lambda^2.
Reference
[0] https://www.youtube.com/playlist?list=PL2SOU6wwxB0uwwH80KTQ6ht66KWxbzTIo
Statistics 110: Probability
Statistics 110 (Probability) has been taught at Harvard University by Joe Blitzstein (Professor of the Practice in Statistics, Harvard University) each year ...
www.youtube.com
'Statistics & Math > Statistics 110: Probability' 카테고리의 다른 글
[Statistics 110] Lecture 25: Order Statistics and Conditional Expectations (0) | 2023.01.28 |
---|---|
[Statistics 110] Lecture 26: Conditional Expectation Continued (0) | 2023.01.26 |
[Statistics 110] Lecture 23: Beta distribution (0) | 2023.01.08 |
[Statistics 110] Lecture 22: Transformations and Convolutions (0) | 2022.12.25 |
[Statistics 110] Lecture 21: Covariance and Correlation (0) | 2022.12.25 |