在这样的理解下(用GPT的话说,就是This works as long as the number of firms G→∞ , and no single firm becomes too large relative to the whole sample.),β^1 是一致的。
Famous Discrete Probability Distributions
摘一段Lecture Notes中的原话:
In this subsection, we will introduce some famous discrete distributions that are ubiquitous in probability and statistics. Each distribution has an underlying story that explains why they are so often used. It is important to understand these distributions in terms of their stories. These stories capture some generic structure that frequently appears in many problems, and they also help us better understand how the distributions are connected to one another. We will describe the distributions mainly using PMF. But you should be able to get their CDF.
我没做到“understand these distributions in terms of their stories”,惭愧,因此写在本篇。
Bernoulli Distribution
Too simple, pass.
Binomial Distribution
n 个独立同分布的 Bernoulli trials,每次成功概率都是 p,X 代表成功的总次数。
P(X=k)=(kn)pk(1−p)n−k,k=0,1,…,n
X=X1+⋯+Xn
For the sum to be valid, these random variables need to be defined on the same sample space. Here we can think of the sample space S as the set of all possible combined outcomes of the n Bernoulli trials. Each outcome can be written as a length-n string of 1's and 0's, with a 1 in the ith position standing for a success of the ith Bernoulli trial.
Hypergeometric Distribution
Consider an urn with w white balls and b black balls. We draw n balls out of the urn at random without replacement. Let X be the number of white balls in the sample. Then X is said to have the Hypergeometric distribution with parameters w, b, and n; we denote this by X∼HGeom(w,b,n).
P(X=k)=(nw+b)(kw)(n−kb)
There are two ways to deduce the pmf. One with considering the order, one without. They reach the same destination with different sample sapces.
Poisson Distribution
There is no simple experiment on which the Poisson distribution is based, but we will shortly describe how it can be obtained from the binomial distribution by certain limiting operations.
P(X=k)=e−λk!λk,k=0,1,2,…
Poisson ApproximationBin(n,p) with n→∞, p→0, and np→λ converges to the Poisson distribution with parameter λ.
A poisson dist can approximately characterize counts of rare events in a large number of trials. The parameter λ can be interpreted as the rate of occurence of these rare events.
Geometric Distribution
Consider a sequence of independent Bernoulli trials, each with the same success probability p∈(0,1), with trials performed until a success occurs. Let X be the number of failures before the first successful trial. Then X has the Geometric distribution with parameter p; we denote this by X∼Geom(p).
P(X=k)=(1−p)kp,k=0,1,2,…
Memoryless property
P(X≥m+n∣X≥m)=P(X≥n),m,n≥0
The only discrete distribution with memoryless property. It characterizes the notion of waiting time.
Negative Binomial Distribution
In a sequence of independent Bernoulli trials with success probability p, if X is the number of failures before the rth success, then X is said to have the Negative Binomial distribution with parameters r and p, denoted X∼NBin(r,p).
P(X=n)=(r−1n+r−1)pr(1−p)n,n=0,1,2,…
X=X1+⋯+Xr
where the Xi are i.i.d. Geom(p)
One convenient sample space is the set of all finite 0-1 strings whose last entry is the rth success. More precisely, an outcome with n failures is a string of length n+r containing exactly r ones, with the last symbol equal to 1. Then X(ω) is just the number of 0's in ω, namely the number of failures before the rth success.
Famous Continuous Probability Distributions
Uniform Distribution
U∼Unif(a,b) if it pdf is:
f(x)={b−a10if a<x<bo.w.
Normal Distribution
X∼N(μ,σ2) if its pdf is:
f(x)=2πσ1e−2σ2(x−μ)2,x∈R
Let φ(z) be the pdf of standard normal distribution. If X∼N(μ,σ2) with pdf f(x) , then f(x)=σ1φ(σx−μ).
truncated first moment
E[Z1{Z>t}]=∫t∞zφ(z)dz=φ(t)
Exponential Distribution
X∼Exp(λ) if its pdf is:
f(x)={λe−λx0if x>0o.w.
The exponential distribution is closely related to the Poisson distribution. To explain the connection, we need to introduce a concept called the Poisson process.
Poisson Process
A process of arrivals in continuous time is called a Poisson process with rate λ if the following two conditions hold:
The number of arrivals in any interval of length t has a Poisson(λt) distribution.
The numbers of arrivals in disjoint intervals are independent.
We may be interested in the waiting time until the first arrival in a Poisson process. It turns out that this waiting time has an Exp(λ) distribution. To see this, let T1 be the waiting time until the first arrival and Nt denote is the number of arrivals at or before time t. Then
P(T1>t)=P(Nt=0)=e−λt
So T1 has the Exp(λ) distribution.
Memoryless Property
Let X∼Exp(λ), then for any s,t≥0,
P(X>s+t∣X>s)=P(X>t)
▶
An example
Joe moves to a new city with inferior urban planning and where buses are much more erratic. Now, when any bus arrives, the time until the next bus arrives is a random variable with the Exponential distribution Expo(1/10). Joe arrives at the bus stop at a random time, not knowing how long ago the previous bus came. What is the distribution of Joe’s waiting time for the next bus?
▶
Solution
Let I be the length of the interval when Joe arrives at the bus stop, X be the interval between consecutive buses, T be the waiting time. It is easy to see that T∣I∼Unif(0,I) . So if we know the distribution of I , it's all done.
Firstly I thought I shares the same distribution as X , but the intuition is wrong. Let's rectify it.
Imagine a time axis divided by intervals whose lengths follow Expo(1/10). Now we pick a point on the time axis uniformly at random. The probability that this point falls in an interval of length t can be understood as the proportion of the time axis covered by intervals of length t, which is tfX(t) (not only we need to consider the number proportion, which is fX(t), but also the length proportion, which is t). So we can roughly write down fI(t)=tfX(t), but remember the basic property of pdf, we need to normalize it by dividing the constant ∫0∞tfX(t)dt,which is E(X) .
The following compution is passed over, but the final result is I∼Gamma(2,1/10) and T∼Expo(1/10).
Or we can directly use the memoryless property of the exponential distribution to get the same result.
Gamma Distribution
It is a generalization of the exponential distribution. A Gamma r.v. represents the total waiting time for multiple arrivals.
Definition:
An r.v. Y is said to have the Gamma distribution with parameters a and $\lambda $, where a>0 and λ>0, if its PDF is
f(y)=Γ(a)λaya−1e−λy
where Γ(a)=∫0∞xa−1e−xdx is the Gamma function.
Theorem:
Let Y1,…,Yn be i.i.d. Exp(λ) random variables, then Y1+⋯+Yn∼Gamma(n,λ).
The Gamma distribution is the continuous analogue of the Negative Binomial distribution: the negative binomial is a sum of geometric waiting times, and the gamma is a sum of exponential waiting times.
Expectation
An example of non-existing expectatin
Let X be a random variable with the following distribution:
f(x)=π11+x21,x∈R
Then X has a Cauchy distribution, and r.v. ∣X∣ does not have a finite expectation.
Law of the Unconcious Statistician (LOTUS) Let X be a random variable and g be a function. Then
E[g(X)]={∑xg(x)P(X=x)∫−∞∞g(x)f(x)dxif X is discreteif X is continuous
A Summary of These Distributions
Distribution
Mean
Variance
MGF MX(t)
PMF / PDF
CDF
Bern(p)
p
p(1−p)
1−p+pet
P(X=x)=px(1−p)1−x,x∈{0,1}
F(x)=⎩⎨⎧0,1−p,1,x<00≤x<1x≥1
Bin(n,p)
np
np(1−p)
(1−p+pet)n
P(X=k)=(kn)pk(1−p)n−k
F(x)=∑k=0⌊x⌋(kn)pk(1−p)n−k
HGeom(w,b,n)
w+bnw
(w+b)2(w+b−1)nwb(w+b−n)
no simple closed form
P(X=k)=(nw+b)(kw)(n−kb)
F(x)=∑k≤⌊x⌋(nw+b)(kw)(n−kb)
Pois(λ)
λ
λ
exp(λ(et−1))
P(X=k)=e−λk!λk
F(x)=e−λ∑k=0⌊x⌋k!λk
Geom(p)
p1−p
p21−p
1−(1−p)etp,t<−log(1−p)
P(X=k)=(1−p)kp,k=0,1,2,…
F(x)={0,1−(1−p)⌊x⌋+1,x<0x≥0
NBin(r,p)
pr(1−p)
p2r(1−p)
(1−(1−p)etp)r,t<−log(1−p)
P(X=n)=(r−1n+r−1)pr(1−p)n
F(x)=∑n=0⌊x⌋(r−1n+r−1)pr(1−p)n
Unif(a,b)
2a+b
12(b−a)2
t(b−a)etb−eta(t=0), 1(t=0)
f(x)={b−a1,0,a<x<bo.w.
F(x)=⎩⎨⎧0,b−ax−a,1,x≤aa<x<bx≥b
N(μ,σ2)
μ
σ2
exp(μt+2σ2t2)
f(x)=2πσ1e−2σ2(x−μ)2
F(x)=Φ(σx−μ)
Exp(λ)
λ1
λ21
λ−tλ,t<λ
f(x)={λe−λx,0,x>0o.w.
F(x)={0,1−e−λx,x≤0x>0
Gamma(a,λ)
λa
λ2a
(λ−tλ)a,t<λ
f(x)={Γ(a)λaxa−1e−λx,0,x>0o.w.
F(x)=Γ(a)1∫0λxua−1e−udu,x>0
Mutiple Random Variables
Change of two variables
Let X=(X1,X2) be a continuous random vector with joint PDF fX(x1,x2) and support set A={(x1,x2):fX(x1,x2)>0} .Let g:A→R2 be a continuously differentiable and invertible function, and let Y=g(X) . Then the joint PDF of Y is given by
fY(y)=fX(x)⋅∣det∂y∂x∣
f(X,Y) and g(X,Y) may be independent
Let U∼Unif(0,2π), and let T∼Expo(1) be independent of U. Define
Suppose that X1,...,Xn are i.i.d. with finite mean μ and finite variance σ2 .
Strong law of large numbers: P(Xˉn→μ)=1 (almost sure convergence Xˉn→a.s.μ )
Weak law of large numbers: ∀ϵ>0,P(∣Xˉn−μ∣>ϵ)=0 as n→∞ (convergence in probability Xˉn→pμ)
almost sure convergence implies convergence in probability.
Tail Probability Inequalities
Markov's For any r.v. X and constant a>0 ,
P(∣X∣>a)≤aE∣X∣
Proof: Note that
∣X∣=∣X∣I(∣X∣>a)+∣X∣I(∣X∣<a)≥aI(∣X∣>a)
take expectations on both side, then one can get the proof.
Chebyshev's
Let X have mean μ and variance σ2 . Then for any a>0,
P(∣X−μ∣>a)≤a2σ2
CLT
Central Limit Theorem
Suppose that X1,...,Xn are i.i.d. with finite mean μ and finite variance σ2 . Then
n→∞limP(σ/nXˉn−μ≤z)=Φ(z),for any z∈R
(Convergence in distribution →d )
Lyapunov's condition Suppose that X1,...,Xn are independent with means μi and variances σi2 . Let sn2=∑i=1nσi2 . Lyapunov's condition is that there exists δ>0 such that