of Statistics and Probability Asymptotic Distribution of One Order Statistic 21 3 Asymptotic Theory of Likelihood Ratio Test Statistics
Statistical asymptotics draws from a variety of sources including (but not restricted to) probability theory, analysis (e g Taylor's theorem), and of
Asymptotic theory (or large sample theory) aims at answering the question: what happens as we gather more and more data? In particular, given random sample,
The asymptotic theory of statistical inference is the study of how well we may succeed in this pursuit, in quantitative terms Any function of the data,
Review of probability theory, probability inequalities • Modes of convergence, stochastic order, laws of large numbers • Results on asymptotic normality
Asymptotic Theory of Statistics and Probability, Springer Serfling, R (1980) Approximation Theorems of Mathematical Statistics, John Wiley, New
To celebrate the 65th birthday of Professor Zhengyan Lin, an Inter- national Conference on Asymptotic Theory in Probability and Statistics
In Chapter 5, we derive exact distributions of several sample statistics based on a random sample of observations • In many situations an exact statistical
22869_6Lecture7_Asymptotics_2019.pdf
Statistics
Asymptotic Theory
Shiu-Sheng Chen
Department of Economics
National Taiwan University
Fall 2019
Shiu-Sheng Chen (NTU Econ)StatisticsFall 20191/28
Asymptotic Theory: Motivation
Asymptotic theory (or large sample theory) aims at answering the question:what happens as we gather more and more data? In particular, given random sample,{X1,X2,X3,...,Xn}, and statistic:
Tn=t(X1,X2,...,Xn),
what is thelimiting behaviorofTnasn?→∞?
Shiu-Sheng Chen (NTU Econ)StatisticsFall 20192/28
Asymptotic Theory: Motivation
Why asking such a question?
For instance, given random sample{Xi}ni=1≂i.i.d.N(μ,σ2), we know that
¯Xn≂N?μ,σ2
n? However, if{Xi}ni=1≂i.i.d.(μ,σ2)without normal assumption, what is the distribution of¯Xn?
We don"t know, indeed.
Is it possible to find a good approximation of the distribution of
¯Xnasn?→∞?
Yes! This is where theasymptotic theorykicks in.
Shiu-Sheng Chen (NTU Econ)StatisticsFall 20193/28
Preliminary Knowledge
Section 1
Preliminary Knowledge
Shiu-Sheng Chen (NTU Econ)StatisticsFall 20194/28
Preliminary Knowledge
Preliminary Knowledge
Limit
Markov Inequality
Chebyshev Inequality
Shiu-Sheng Chen (NTU Econ)StatisticsFall 20195/28
Preliminary Knowledge
Limit of a Real Sequence
Definition (Limit)
If for everyε>0, and an integerN(ε),
?bn-b?<ε,?n>N(ε) then we say that a sequence of real numbers{b1,...,bn}converges to a limitb.
It is denoted by
limn→∞bn=b
Shiu-Sheng Chen (NTU Econ)StatisticsFall 20196/28
Preliminary Knowledge
Markov Inequality
Theorem (Markov Inequality)
Suppose thatXis a random variable such thatP(X≥0)=1. Then for every real numberm>0,
P(X≥m)≤E(X)
m
Shiu-Sheng Chen (NTU Econ)StatisticsFall 20197/28
Preliminary Knowledge
Chebyshev Inequality
Theorem (Chebyshev Inequality)
LetY≂(E(Y),Var(Y)). Then for every numberε>0,
P??Y-E(Y)?≥ε?≤Var(Y)
ε2
Proof: LetX=[Y-E(Y)]2, then
P(X≥0)=1
and
E(X)=Var(Y)
Hence, the result can be derived by applying the Markov
Inequality.
Shiu-Sheng Chen (NTU Econ)StatisticsFall 20198/28
Modes of Convergence
Section 2
Modes of Convergence
Shiu-Sheng Chen (NTU Econ)StatisticsFall 20199/28
Modes of Convergence
Types of Convergence
For a random variable, we consider three modes of convergence:
Converge in Probability
Converge in Distribution
Converge in Mean Square
Shiu-Sheng Chen (NTU Econ)StatisticsFall 201910/28
Modes of Convergence
Converge in Probability
Definition (Converge in Probability)
Let{Yn}be a sequence of random variables and letYbe another random variable. For anyε>0,
P(?Yn-Y?<ε)?→1,asn?→∞
then we say thatYnconverges in probability toY, and denote it by Yn p?→Y
Equivalently,
P(?Yn-Y?≥ε)?→0,asn?→∞
Shiu-Sheng Chen (NTU Econ)StatisticsFall 201911/28
Modes of Convergence
Converge in Probability
{Xi}ni=1≂i.i.d.Bernoulli(0.5)and then computeYn=¯Xn=∑iXi n
In this case,Yn
p?→0.5
02004006008001000
0.2 0.4 0.6 0.8 1.0 toss z Shiu-Sheng Chen (NTU Econ)StatisticsFall 201912/28
Modes of Convergence
Converge in Distribution
Definition (Converge in Distribution)
Let{Yn}be a sequence of random variables with distribution functionFYn(y), (denoted byFn(y)for simplicity). LetYbe another random variable with distribution function,FY(y). If limn→∞Fn(y)=FY(y)at allyfor whichFY(y)is continuous then we say thatYnconverges in distribution toY.
It is denoted by
Yn d?→Y
FY(y)is called thelimiting distributionofYn.
Shiu-Sheng Chen (NTU Econ)StatisticsFall 201913/28
Modes of Convergence
Converge in Mean Square
Definition (Converge in Mean Square)
Let{Yn}be a sequence of random variables and letYbe another random variable. If
E(Yn-Y)2?→0,asn?→∞.
Then we say thatYnconverges in mean square toY.
It is denoted by
Yn ms?→Y
It is also calledconverge in quadratic mean.
Shiu-Sheng Chen (NTU Econ)StatisticsFall 201914/28
Important Theorems
Section 3
Important Theorems
Shiu-Sheng Chen (NTU Econ)StatisticsFall 201915/28
Important Theorems
Theorems
Theorem
Yn ms?→cif and only if limn→∞E(Yn)=c,andlimn→∞Var(Yn)=0.
Proof. It can be shown that
E(Yn-c)2=E([Yn-E(Yn)]2)+[E(Yn)-c]2
Shiu-Sheng Chen (NTU Econ)StatisticsFall 201916/28
Important Theorems
Theorems
Theorem
IfYn ms?→YthenYn p?→Y Proof: Note thatP(?Yn-Y?2≥0)=1, and by Markov Inequality, P(?Yn-Y?≥k)=P(?Yn-Y?2≥k2)≤E(?Yn-Y?2) k2 Shiu-Sheng Chen (NTU Econ)StatisticsFall 201917/28
Important Theorems
Weak Law of Large Numbers, WLLN
Theorem (WLLN)
Given a random sample{Xi}ni=1withσ2=Var(X1)<∞. Let¯Xn denote the sample mean, and note thatE(¯Xn)=E(X1)=μ. Then
¯Xn
p?→μ Proof: (1) By Chebyshev Inequality (2) By Converge in Mean
Square
Sample mean¯Xnis getting closer (in probability sense) to the population meanμas the sample size increases. That is, if we use¯Xnas aguessof unknownμ, we are quite happy that the sample mean makes a good guess. Shiu-Sheng Chen (NTU Econ)StatisticsFall 201918/28
Important Theorems
WLLN for Other Moments
Note that the WLLN can be thought as
∑ni=1Xi n=X1+X2+⋯Xn n p?→E(X1)
LetY=X2, and by the WLLN,
∑ni=1Yi n=Y1+Y2+⋯Yn n p?→E(Y1)
Hence,
∑ni=1X2i n=X21+X22+⋯X2n n p?→E(X21) Shiu-Sheng Chen (NTU Econ)StatisticsFall 201919/28
Important Theorems
Example: An Application of WLLN
AssumeWn≂Binomial(n,μ), and letYn=Wnn. Then Yn p?→μ Why? SinceWn=∑iXi,Xi≂i.i.d.Bernoulli(μ) withE(X1)=μ,
Var(X1)=μ(1-μ), the result follows by WLLN.
Shiu-Sheng Chen (NTU Econ)StatisticsFall 201920/28
Important Theorems
Central Limit Theorem, CLT
Theorem (CLT)
Let{Xi}ni=1be a random sample, whereE(X1)=μ<∞,
Var(X1)=σ2<∞, then
Zn=¯Xn-E(¯Xn)?Var(¯Xn)=
⎷n(¯Xn-μ) σ d?→N(0,1) If a random sample is taken from any distribution with meanμ and varianceσ2, regardless of whether this distribution is discrete or continuous, then the distribution of the random variableZnwill be approximately the standard normal distribution in large sample. Shiu-Sheng Chen (NTU Econ)StatisticsFall 201921/28
Important Theorems
CLT
Using notation of asymptotic distribution,
¯Xn-μ?
σ2 n ≂AN(0,1), Or
¯Xn≂AN?μ,σ2
n?, where≂Arepresents asymptotic distribution, andArepresents
Asymptotically
Shiu-Sheng Chen (NTU Econ)StatisticsFall 201922/28
Important Theorems
An Application of CLT
Example: Assume{Xi}≂i.i.d.Bernoulli(μ), then
¯Xn-μ?μ(1-μ)
n d?→N(0,1). Why?
SinceE(¯Xn)=μ, andVar(¯Xn)=σ2
n=μ(1-μ) n Shiu-Sheng Chen (NTU Econ)StatisticsFall 201923/28
Important Theorems
Continuous Mapping Theorem
Theorem (CMT)
GivenYn
p?→Y, andg(?)is continuous, then g(Yn) p?→g(Y).
Proof: omitted here.
Examples: ifYn
p?→Y, then 1 Yn p?→1 Y Y2n p?→Y2 ⎷Yn p?→⎷Y Shiu-Sheng Chen (NTU Econ)StatisticsFall 201924/28
Important Theorems
Theorem
Theorem
GivenWn
p?→WandYn p?→Y, then Wn+Yn p?→W+Y WnYn p?→WY
Proof: omitted here.
Shiu-Sheng Chen (NTU Econ)StatisticsFall 201925/28
Important Theorems
Slutsky Theorem
Theorem
GivenWn
d?→WandYn p?→c, wherecis a constant. Then Wn+Yn d?→W+c WnYn d?→cW WnYn d?→W cforc≠0
Proof: omitted here.
Shiu-Sheng Chen (NTU Econ)StatisticsFall 201926/28
Important Theorems
The Delta Method
Theorem
Given⎷n(Yn-θ)d?→N(0,σ2).Letg(?)be differentiable, and g′(θ)≠0exists, then ⎷n(g(Yn)-g(θ))d?→N(0,[g′(θ)]2σ2). Proof: (sketch) Given 1st-order Taylor approximation g(Yn)≈g(θ)+g′(θ)(Yn-θ), then⎷n(g(Yn)-g(θ)) g′(θ)≈⎷n(Yn-θ)d?→N(0,σ2) Shiu-Sheng Chen (NTU Econ)StatisticsFall 201927/28
Important Theorems
Example
Given{Xi}ni=1≂i.i.d.(μ,σ2), find the asymptotic distribution of
¯Xn
1-¯Xn.
Note that by CLT,
⎷n(¯Xn-μ)d?→N(0,σ2)
Hence, by the Delta method,
g(¯Xn)=¯Xn
1-¯Xn,g(μ)=μ
1-μ,g′(μ)=1
(1-μ)2 ⎷n?¯Xn
1-¯Xn-μ
1-μ?d?→N?0,1
(1-μ)4σ2? Shiu-Sheng Chen (NTU Econ)StatisticsFall 201928/28