• 검색 결과가 없습니다.

LECTURE NOTE 6 – SUM OF RANDOM VARIABLES For random variables X

N/A
N/A
Protected

Academic year: 2022

Share "LECTURE NOTE 6 – SUM OF RANDOM VARIABLES For random variables X"

Copied!
4
0
0

로드 중.... (전체 텍스트 보기)

전체 글

(1)

Probability and Random Process, Junhee Seok, Korea University Lecture Note 6 – Sum of Random Variables

1

LECTURE NOTE 6 – SUM OF RANDOM VARIABLES

For random variables X1, X2, …, Xn, we will discuss about the properties of their sum ∑𝑛 𝑋𝑖 𝑖=1 .

PDF OF THE SUM OF TWO RANDOM VARAIABLES

When W = X+Y,

𝐹𝑊(𝑤) = Pr[𝑋 + 𝑌 ≤ 𝑤] = ∫𝑥=−∞ �∫𝑦=−∞𝑤−𝑥 𝑓𝑋,𝑌(𝑥, 𝑦)𝑑𝑦� 𝑑𝑥 𝑓𝑊(𝑤) =𝑑𝑑𝑑𝑤𝑊(𝑤)= ∫𝑥=−∞𝑑𝑤𝑑𝑦=−∞𝑤−𝑥 𝑓𝑋,𝑌(𝑥, 𝑦)𝑑𝑦� 𝑑𝑥

= ∫𝑥=−∞𝑑𝑤𝑑𝑑𝑑𝑑𝑑𝑦=−∞𝑑 𝑓𝑋,𝑌(𝑥, 𝑦)𝑑𝑦� 𝑑𝑥 ← 𝑢 = 𝑤 − 𝑥

= ∫ 𝑓−∞ 𝑋,𝑌(𝑥, 𝑢)𝑑𝑥= ∫ 𝑓−∞ 𝑋,𝑌(𝑥, 𝑤 − 𝑥)𝑑𝑥 Finally, 𝒇𝑾(𝒘) = ∫ 𝒇−∞ 𝑿,𝒀(𝒙, 𝒘 − 𝒙)𝒅𝒙.

Example

Especially, when X and Y are independent,

𝒇𝑿+𝒀(𝒘) = ∫ 𝑓−∞ 𝑋,𝑌(𝑥, 𝑤 − 𝑥)𝑑𝑥= ∫ 𝑓−∞ 𝑋(𝑥)𝑓𝑌(𝑤 − 𝑥)𝑑𝑥

= (𝒇𝑿∗ 𝒇𝒀)(𝒘) → 𝐜𝐜𝐜𝐜𝐜𝐜𝐜𝐜𝐜𝐜𝐜 of 𝑓𝑋 and 𝑓𝑌

Convolution

For two function 𝑓() and 𝑔(), their convolution is defined as (𝑓 ∗ 𝑔)(𝑥) = ∫ 𝑓(𝑢)𝑔(𝑥 − 𝑢)𝑑𝑢−∞ .

Example

In general, we can derive the pdf of 𝑊 = ∑𝑛𝑖=1𝑋𝑖 by recursively applying the pdf calculation of the sum of two random variables, which means 𝑊𝑛= 𝑋𝑛+ 𝑊𝑛−1. While it is not easy to show, as special cases,

(1) 𝑋𝑖 ~ Poi(𝜆𝑖) and 𝑋𝑖’s are indep.  ∑𝑛 𝑋𝑖

𝑖=1 ~ Poi(∑𝑛 𝜆𝑖 𝑖=1 ) (2) 𝑋𝑖 ~ N(𝜇𝑖, 𝜎𝑖2) and 𝑋𝑖’s are indep.  ∑𝑛 𝑋𝑖

𝑖=1 ~ N(∑𝑛 𝜇𝑖

𝑖=1 , ∑𝑛𝑖=1𝜎𝑖2)

Note that E[∑𝑛𝑖=1𝑋𝑖] = ∑ E[𝑋𝑛𝑖=1 𝑖] and Var[∑𝑛𝑖=1𝑋𝑖] = ∑ Var[𝑋𝑛𝑖=1 𝑖]. We will see this in the next section.

https://en.wikipedia.

org/wiki/Convolutio n

(2)

Probability and Random Process, Junhee Seok, Korea University Lecture Note 6 – Sum of Random Variables

2

EXPECTATION OF THE SUM OF RANDOM VARIABLES

Let 𝑊𝑛= 𝑋1+ 𝑋2+ ⋯ + 𝑋𝑛= ∑𝑛 𝑋𝑖

𝑖=1 . And then,

E[𝑊𝑛] = E[𝑋1] + E[𝑋2] + ⋯ + E[𝑋𝑛] = ∑ E[𝑋𝑛𝑖=1 𝑖] Var[𝑊𝑛] = E[(∑ 𝑋𝑛 𝑖

𝑖=1 − ∑𝑛 𝜇𝑖

𝑖=1 )2] = E[(∑ (𝑋𝑛𝑖=1 𝑖− 𝜇𝑖))2]

= E�∑𝑛𝑖=1∑ (𝑋𝑛𝑗=1 𝑖− 𝜇𝑖)�𝑋𝑗− 𝜇𝑗��

= ∑𝑛𝑖=1𝑛𝑗=1Cov�𝑋𝑖, 𝑋𝑗�= ∑ Cov�𝑋𝑖=𝑗 𝑖, 𝑋𝑗�+ ∑ Cov�𝑋𝑖≠𝑗 𝑖, 𝑋𝑗

= ∑𝑛𝑖=1Var[𝑋𝑖]+ 2 ∑𝑛𝑖=1𝑛𝑗=𝑖+1Cov�𝑋𝑖, 𝑋𝑗

If Xi’s are independent, Var[𝑊𝑛] = ∑ Var[𝑋𝑛𝑖=1 𝑖]. If Xi’s are iid, Var[𝑊𝑛] = 𝑛Var[𝑋].

Example

Expectation and Variance of the Mean of iid Random Variables

Let 𝑊𝑛= ∑𝑛𝑖=1𝑋𝑛𝑖 when 𝑋𝑖’s are iid random variables such that 𝑋𝑖~𝑋. Then,

E[𝑊𝑛] = ∑ E �𝑛𝑖=1 𝑋𝑛𝑖�= E[𝑋]

Var[𝑊𝑛] = ∑ Var �𝑛𝑖=1 𝑋𝑛𝑖�= ∑𝑛𝑖=1𝑛12Var[𝑋𝑖]=1𝑛Var[𝑋]

The variance of Wn is close to zero when n is large, which means we can estimate the E[X] very accurately through the averaging.

Law of large numbers: for iid Xi’s, lim𝑛→∞1𝑛𝑛𝑖=1𝑋𝑖= E[𝑋].

Example

(1) We have an unfair coin of which head probability is p. But, we don’t know the value of p. Now, we want to estimate p by flipping the coin n times. Is each flipping iid? How much accurately can we estimate p when flipping 20 times compared with 10 times?

(2) Two survey companies report the supporting rate for the president. Company A surveys 100 people, and Company B surveys 500 people. Which company will report the more accurate rate and how much more accurate?

(3)

Probability and Random Process, Junhee Seok, Korea University Lecture Note 6 – Sum of Random Variables

3

CENTRAL LIMIT THEOREM

The CLT makes Gaussian distribution universal, the king of random variables.

𝑋𝑖’s are iid with E[𝑋] = 𝜇 and Var[𝑋] = 𝜎2. And then, E[∑𝑛 𝑋𝑖

𝑖=1 ] = 𝑛𝜇 and Var[∑ 𝑋𝑛 𝑖

𝑖=1 ] = 𝑛𝜎2. Here, For 𝑍𝑛=𝑛𝑖=1�𝑛𝜎𝑋𝑖−𝑛𝑛2 , E[𝑍𝑛] = 0 and Var[𝑍𝑛] = 1.

When n is close to infinity,

lim𝑛→∞𝐹𝑍𝑛(𝑧) = Φ(𝑧) or lim

𝑛→∞𝑍𝑛 ~ 𝑁(0,1) : Central Limit Theorem

The proof is very hard beyond the scope of this course. Actually, every textbook says this comment, lol.

Approximation using the CLT

𝑋𝑖’s are iid with E[𝑋] = 𝜇 and Var[𝑋] = 𝜎2. For 𝑊𝑛= ∑𝑛 𝑋𝑖

𝑖=1 = √𝑛𝜎2𝑍𝑛+ 𝑛𝜇, 𝐹𝑊𝑛(𝑤) = Pr�√𝑛𝜎2𝑍𝑛+ 𝑛𝜇 ≤ 𝑤� = 𝐹𝑍𝑛𝑤−𝑛𝑛�𝑛𝜎2� → Φ �𝑤−𝑛𝑛�𝑛𝜎2� when n  ∞.

When n is large, we can approximate 𝐹𝑊𝑛(𝑤) with Φ �𝑤−𝑛𝑛�𝑛𝜎2�.

Or we can approximate 𝑊𝑛 with a Gaussian random variable, N(𝑛𝜇, 𝑛𝜎2).

Example

(1) Let 𝑊𝑛= ∑𝑛𝑖=1𝑋𝑖 and 𝑋𝑖 ~ Bern(p). Then, Wn ~ B(n,p). When n is large enough, how can we approximate B(n,p)?

(2) One million people vote either candidate A or B equally and randomly. What is the probability that A wins by more than 2,000 votes?

Tips for the Approximation for a Discrete Random Variable

For a discrete random variable X with µ and σ2,

Pr[𝑘1≤ 𝑋 ≤ 𝑘2] = Pr[𝑘1− 0.5 ≤ 𝑋 < 𝑘2+ 0.5] ~Φ �𝑘2+0.5−𝑛𝜎 � − Φ �𝑘1−0.5−𝑛𝜎 � rather than

Pr[𝑘1≤ 𝑋 ≤ 𝑘2] ~ Φ �𝑘2𝜎−𝑛� − Φ �𝑘1𝜎−𝑛

Example

When X ~ B(20,0.4), Pr[X=8] = Pr[8≤X≤8] = ?

(4)

Probability and Random Process, Junhee Seok, Korea University Lecture Note 6 – Sum of Random Variables

4

SUMMARY

When W = X+Y, 𝒇𝑾(𝒘) = ∫ 𝒇−∞ 𝑿,𝒀(𝒙, 𝒘 − 𝒙)𝒅𝒙.

When X and Y are independent, 𝒇𝑿+𝒀(𝒘) = (𝒇𝑿∗ 𝒇𝒀)(𝒘) = ∫ 𝒇−∞ 𝑿(𝒙)𝒇𝒀(𝒘 − 𝒙)𝒅𝒙

When 𝑊𝑛= 𝑋1+ 𝑋2+ ⋯ + 𝑋𝑛= ∑𝑛𝑖=1𝑋𝑖,

E[𝑊𝑛] = ∑ E[𝑋𝑛𝑖=1 𝑖] and Var[𝑊𝑛] = ∑ Var[𝑋𝑛𝑖=1 𝑖]+ 2 ∑𝑛𝑖=1𝑛𝑗=𝑖+1Cov�𝑋𝑖, 𝑋𝑗�. When Xi’s are iid, E[𝑊𝑛] = 𝑛E[𝑋] and Var[𝑊𝑛] = 𝑛Var[𝑋].

Law of large numbers: for iid Xi’s, lim𝑛→∞1 𝑛𝑛 𝑋𝑖

𝑖=1 = E[𝑋].

Central Limit Theorem

When 𝑋𝑖’s are iid with E[𝑋] = 𝜇 and Var[𝑋] = 𝜎2,

𝑛→∞lim

𝑛 𝑋𝑖 𝑖=1 − 𝑛𝜇

√𝑛𝜎2 ~ 𝑁(0,1) When 𝑋𝑖’s are iid and n is large, we can approximate 𝑊𝑛= ∑𝑛 𝑋𝑖

𝑖=1 by N(𝑛𝜇, 𝑛𝜎2).

참조

관련 문서

negatively superadditive dependent random variables, Marcinkie- wicz-Zygmund strong law of large numbers, weighted sums, the three series theorem.. Supported by the National

Key words and phrases: multivariate linear process, linearly positive quadrant dependent random vectors, central limit theorem2. This work was supported by

In section 2, we provide the establish the exponential in- equalities for sum of extended acceptable random variables and in section 3, we obtain a result dealing with