Lecture 27: Optimal Estimators and Functional Delta Method

Similar documents
32 estimating the cumulative distribution function

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.436J/15.085J Fall 2008 Lecture 19 11/17/2008 LAWS OF LARGE NUMBERS II THE STRONG LAW OF LARGE NUMBERS

LECTURE 8: ASYMPTOTICS I

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 21 11/27/2013

Lecture 19: Convergence

7.1 Convergence of sequences of random variables

Convergence of random variables. (telegram style notes) P.J.C. Spreij

1 Convergence in Probability and the Weak Law of Large Numbers

Glivenko-Cantelli Classes

A RANK STATISTIC FOR NON-PARAMETRIC K-SAMPLE AND CHANGE POINT PROBLEMS

Notes 19 : Martingale CLT

Fall 2013 MTH431/531 Real analysis Section Notes

ECONOMETRIC THEORY. MODULE XIII Lecture - 34 Asymptotic Theory and Stochastic Regressors

Notes 27 : Brownian motion: path properties

Singular Continuous Measures by Michael Pejic 5/14/10

Let us give one more example of MLE. Example 3. The uniform distribution U[0, θ] on the interval [0, θ] has p.d.f.

Chapter 8. Uniform Convergence and Differentiation.

7.1 Convergence of sequences of random variables

Seunghee Ye Ma 8: Week 5 Oct 28

LECTURE 14 NOTES. A sequence of α-level tests {ϕ n (x)} is consistent if

CS434a/541a: Pattern Recognition Prof. Olga Veksler. Lecture 5

Chapter 3. Strong convergence. 3.1 Definition of almost sure convergence

Introduction to Extreme Value Theory Laurens de Haan, ISM Japan, Erasmus University Rotterdam, NL University of Lisbon, PT

Lecture 20: Multivariate convergence and the Central Limit Theorem

17. Joint distributions of extreme order statistics Lehmann 5.1; Ferguson 15

Direction: This test is worth 250 points. You are required to complete this test within 50 minutes.

Journal of Multivariate Analysis. Superefficient estimation of the marginals by exploiting knowledge on the copula

Kernel density estimator

Math Solutions to homework 6

sin(n) + 2 cos(2n) n 3/2 3 sin(n) 2cos(2n) n 3/2 a n =

Lecture 33: Bootstrap


MATH301 Real Analysis (2008 Fall) Tutorial Note #7. k=1 f k (x) converges pointwise to S(x) on E if and

Continuous Functions

Sequences and Series of Functions

Introductory statistics

January 25, 2017 INTRODUCTION TO MATHEMATICAL STATISTICS

Apply change-of-basis formula to rewrite x as a linear combination of eigenvectors v j.

Lecture 3 : Random variables and their distributions

MAT1026 Calculus II Basic Convergence Tests for Series

Definition 4.2. (a) A sequence {x n } in a Banach space X is a basis for X if. unique scalars a n (x) such that x = n. a n (x) x n. (4.

Lecture Chapter 6: Convergence of Random Sequences

( θ. sup θ Θ f X (x θ) = L. sup Pr (Λ (X) < c) = α. x : Λ (x) = sup θ H 0. sup θ Θ f X (x θ) = ) < c. NH : θ 1 = θ 2 against AH : θ 1 θ 2

Ma 4121: Introduction to Lebesgue Integration Solutions to Homework Assignment 5

An Introduction to Asymptotic Theory

Resampling Methods. X (1/2), i.e., Pr (X i m) = 1/2. We order the data: X (1) X (2) X (n). Define the sample median: ( n.

MATH 320: Probability and Statistics 9. Estimation and Testing of Parameters. Readings: Pruim, Chapter 4

Lecture 8: Convergence of transformations and law of large numbers

1.3 Convergence Theorems of Fourier Series. k k k k. N N k 1. With this in mind, we state (without proof) the convergence of Fourier series.

It is often useful to approximate complicated functions using simpler ones. We consider the task of approximating a function by a polynomial.

Advanced Stochastic Processes.

Distribution of Random Samples & Limit theorems

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 6 9/23/2013. Brownian motion. Introduction

Entropy Rates and Asymptotic Equipartition

Solutions: Homework 3

d) If the sequence of partial sums converges to a limit L, we say that the series converges and its

4. Partial Sums and the Central Limit Theorem

Empirical Processes: Glivenko Cantelli Theorems

Quiz. Use either the RATIO or ROOT TEST to determine whether the series is convergent or not.

Point Estimation: properties of estimators 1 FINITE-SAMPLE PROPERTIES. finite-sample properties (CB 7.3) large-sample properties (CB 10.

REAL ANALYSIS II: PROBLEM SET 1 - SOLUTIONS

Introduction to Probability. Ariel Yadin

CHAPTER 5 SOME MINIMAX AND SADDLE POINT THEOREMS

TENSOR PRODUCTS AND PARTIAL TRACES

1. (25 points) Use the limit definition of the definite integral and the sum formulas 1 to compute

Analytic Continuation

ECE 901 Lecture 12: Complexity Regularization and the Squared Loss

Math 61CM - Solutions to homework 3

Statistical Theory MT 2008 Problems 1: Solution sketches

Chapter 6 Infinite Series

Statistical Theory MT 2009 Problems 1: Solution sketches

MATH 112: HOMEWORK 6 SOLUTIONS. Problem 1: Rudin, Chapter 3, Problem s k < s k < 2 + s k+1

U8L1: Sec Equations of Lines in R 2

Notes 5 : More on the a.s. convergence of sums

Notes on iteration and Newton s method. Iteration

The Central Limit Theorem

Section 14. Simple linear regression.

On Thresholds for Robust Goodness-of-Fit Tests

Solutions to home assignments (sketches)

Math 113, Calculus II Winter 2007 Final Exam Solutions

Topic 5 [434 marks] (i) Find the range of values of n for which. (ii) Write down the value of x dx in terms of n, when it does exist.

First Year Quantitative Comp Exam Spring, Part I - 203A. f X (x) = 0 otherwise

Taylor expansion: Show that the TE of f(x)= sin(x) around. sin(x) = x - + 3! 5! L 7 & 8: MHD/ZAH

On forward improvement iteration for stopping problems

Math 680 Fall Chebyshev s Estimates. Here we will prove Chebyshev s estimates for the prime counting function π(x). These estimates are

lim za n n = z lim a n n.

EECS564 Estimation, Filtering, and Detection Hwk 2 Solns. Winter p θ (z) = (2θz + 1 θ), 0 z 1

Gamma Distribution and Gamma Approximation

SOME SEQUENCE SPACES DEFINED BY ORLICZ FUNCTIONS

Properties of Fuzzy Length on Fuzzy Set

Definition An infinite sequence of numbers is an ordered set of real numbers.

ECE 330:541, Stochastic Signals and Systems Lecture Notes on Limit Theorems from Probability Fall 2002

Asymptotic distribution of products of sums of independent random variables

Kolmogorov-Smirnov type Tests for Local Gaussianity in High-Frequency Data

Large Sample Theory. Convergence. Central Limit Theorems Asymptotic Distribution Delta Method. Convergence in Probability Convergence in Distribution

Math 25 Solutions to practice problems

for all x ; ;x R. A ifiite sequece fx ; g is said to be ND if every fiite subset X ; ;X is ND. The coditios (.) ad (.3) are equivalet for =, but these

ANSWERS TO MIDTERM EXAM # 2

U8L1: Sec Equations of Lines in R 2

Slide Set 13 Linear Model with Endogenous Regressors and the GMM estimator

Transcription:

Stat210B: Theoretical Statistics Lecture Date: April 19, 2007 Lecture 27: Optimal Estimators ad Fuctioal Delta Method Lecturer: Michael I. Jorda Scribe: Guilherme V. Rocha 1 Achievig Optimal Estimators From the last class, we kow that the best limitig distributio we ca hope for a parameter ψ is N0, ψ I ψ T. The et questio to ask is whether such boud ca be achieved. This is the theme of our et result: Lemma 1. Lemma 8.14 i va der Vaart, 1998 Assume that the eperimet P : Θ is differetiable i quadratic mea at 0 with o-sigular Fisher iformatio matri I. Let ψ be differetiable at 0. Let T be a estimator sequece i the eperimets P : Rk such that: T ψ = 1 ψ I 1 l X i + o P 1, the T is the best regular estimator for ψ at. Coversely, every best regular estimator sequece satisfies this epasio. Proof. Let, := 1 l X i. We kow that, coverges i distributio to a with a N0,I distributio. From Theorem 7.2 i va der Vaart 1998 we kow that: dp + h log = h T 1 2 ht I h + o P 1. dp Usig Slutsky s theorem, we get: T ψ log dp + h dp [ N ψ I 1 h T 1 2 ht I h 0 1 2 ht I h ] [ ψ, I 1 ψ h T ψ T ψ h h T I h ] Usig Le Cam s third lemma we ca coclude that the sequece T ψ uder + h coverges i distributio to N ψ h, ψ I 1 ψ T. Sice ψ is differetiable, we have that ψ + h ψ ψ h as. We coclude that, uder + h, T ψ + h does ot ivolve h, that is, T is regular. 1

2 Lecture 27: Optimal Estimators ad Fuctioal Delta Method To prove the coverse, let T ad S be a two best regular estimator sequeces. Alog subsequeces, it ca be show that: S ψ + h [ ] + h S ψ h T ψ + h T ψ h for a radomized estimator S,T i the limitig eperimet. Because S ad T are best regular, S ad T are best equivariat-i-law. Thus S = T = ψ X almost surely ad, as a result, S T coverges i distributio to S T = 0. As a result, every two best regular estimator sequeces are asymptotically equivalet. To get the result, apply this coclusio to T ad: S = ψ + 1 ψ I 1, Remarks o Theorem 8.14: From theorem 5.39, a Maimum Likelihod Estimator ˆ satisfies: ˆ = 1 I 1 l X i + o P 1 uder regularity coditios. It follows that MLEs are asymptotically efficiet. This result ca be eteded to a trasforms of the MLE ψ for a differetiable ψ by usig the delta method ad observig that ψˆ satisfies the epasio i lemma 8.14. Lemma 8.14 suggests that Rao score fuctios leadig to tests costructed from the scores are asymptotically efficiet. 2 Fuctioal Delta Method The fuctioal delta method aims at etedig the delta method to a oparametric cotet. The high level idea is to iterpret a statistic as a fuctioal φ mappig from the space of probability distributios D to the real lie R ad use a otio of derivative of this fuctioal to obtai the asymptotic distributio of φˆf. We have prove before that: sup ˆF F p 0, Gliveko-Catelli Theorem F ˆF F GF, Dosker Theorem where ˆF is the empirical distributio fuctio based i samples ad G F is the Browia Bridge. Our goal ow is to fid coditios o the fuctioals φ so we ca eted the above modes of covergece to φˆf. As we will see i detail below, cosistecy of the sequece φˆf follows easily from assumig φ to be cotiuous with respect to the supremum orm: this a atural etesio of the cotiuous mappig theorem. The geeralizatio of the delta method to fuctioals is more ivolved as differet otios of differetiability of fuctioals eist. Before we jump ito the cotiuous mappig theorem ad the fuctioal delta method, a few eamples of statistical fuctioals are i order.

Lecture 27: Optimal Estimators ad Fuctioal Delta Method 3 2.1 Eamples of statistical fuctioals The mea: µf = XdFX; The variace: VarF = X µf 2 dfx; Higher order momets: µ k F = X µf k dfx; The Kolmogorov-Smirov statistics: KF = sup F F 0, where F 0 is a fied hypothesized distributio; The Crámer-vo Mises statistics: CF = F F 0 2 df 0, where F 0 is a fied hypothesized distributio; V-statistics: φf = E F TX 1,X 2,...,X p where X 1,X 2,...,X p are idepedet copies of F- distributed radom variables; Quatile fuctioal: φf = F 1 p = if { : F p}; 2.2 Cosistecy of statistical fuctioals Oe possible assumptio to esure that φˆf coverges to φf is cotiuous with respect to the supremum orm. Formally, this is defied as: Defiitio 2. Cotiuity of a fuctioal Let D be the space of distributios ad φ : D R. We say φ is cotiuous with respect to the supremum orm at F if: sup F F 0 φf φf. The et result is a etesio of the cotiuous mappig theorem to fuctioals ad ca be used to establish the cosistecy of statistical fuctioals. Theorem 3. Cotiuous Mappig Theorem for Statistical Fuctioals Let φ : D R be a cotiuous fuctioal at F. It follows that: If F F p 0, the φf φf p 0. Proof. From cotiuity of φ with respect to the sup orm, we have that for every ε > 0, there eists δ > 0 such that: Hece, F F δ φf φf ε. 0 P φf φf > ε P F F > δε 0, where the last covergece is due to F F p 0 by hypothesis.

4 Lecture 27: Optimal Estimators ad Fuctioal Delta Method 2.2.1 Eamples of cotiuous statistical fuctioals The followig two fuctioals are cotiuous with respect to the supremum orm: φf = Fa: The distributio fuctio evaluated at a poit. To establish the cotiuity of this fuctio, otice that for a sequece of distributios such that F F 0: 0 F a Fa sup F F 0 φf = F F 0 2 df 0 : The Crámer-vo Mises fuctioal. Agai, take a sequece of distributios such that F F 0. We have: 0 φf φf = [F 2 F 2 + 2F 0 F F ] df 0 F F 2F 0 F F df 0 }{{} 2 2sup F F df 0 0 2.3 Limitig distributio of statistical fuctioals Recall our goal to determie the limitig distributio of φˆf. Heuristically, we might hope to derive it if we ca fid a liear φ F resultig i a approimatio of the sort: φˆf φf = φ FˆF F + some residual = φ FĜ + some residual As before, we must keep track of the behavior of the residual term as grows. The fact that φ operates o the ifiite dimesioal space of distributio fuctios will require us to be more careful ad resort to some cocepts of fuctioal aalysis. Namely, we will be lookig at the otios of Gateau ad Hadamard derivatives. We start by cosiderig Gateau derivatives. To see how they ca be used, otice that from liearity of φ F, we have that: φ FĜ = 1 φ Fδ Xi F, where δ Xi is the distributio fuctio cocetratig all mass at X i. For each of the terms i the sum, we ca cosider a directioal derivative of φ i the directio δ Xi F. That is the Gateau derivative which i this case is defied as: φ Fδ Xi F = d dt [φ1 tf + tδ X i ]. t=0 The epressio for φ F δ X i F above is kow i the robust statistics literature as the ifluece fuctio: IF φ,f = d dt [φ1 tf + tδ ]. t=0

Lecture 27: Optimal Estimators ad Fuctioal Delta Method 5 To some etet, it measures how much the statistic φf is affected by addig a ew observatio at to the sample. A related cocept is the gross-error sesitivity defied as: For robustess, γ must be bouded. γ = supif φ,f. Goig back to the approimatio of φˆf φf, we ow write: φˆf φf = 1 IF φ,f X i + R. We ow have: E F φ F δ Xi F = φ F δ PdP = 0, If we assume: Var F φ F δ Xi F = IF φ,f 2 df <. ad the residual term R ca be cotrolled somehow more o this i later classes, a cetral limit theorem will hold for the radom variable φ F δ X i F ad we ca epect that: φˆf φf F N0,λ 2. As is the case i multivariate calculus, the eistece of Gateau directioal derivative does ot esure that the residual of the approimatio is well behaved differetiability. I the et classes, we will study the residual term more closely. Comig up et I the et classes, we will: make this heuristic of the fuctioal delta method more precise; look more closely at the otio of Hadamard derivative; use Hadamard derivative to determie coditios that esure the fuctioal delta method works Refereces va der Vaart, A. W. 1998. Asymptotic Statistics. Cambridge Uiversity Press, Cambridge.