Solutions Homework 4 March 5, 2018

Similar documents
MATH 281A: Homework #6

Using T.O.M to Estimate Parameter of distributions that have not Single Exponential Family

1 Matrix representations of canonical matrices

Maximum Likelihood Estimation

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

Composite Hypotheses testing

Maximum Likelihood Estimation of Binary Dependent Variables Models: Probit and Logit. 1. General Formulation of Binary Dependent Variables Models

ANSWERS. Problem 1. and the moment generating function (mgf) by. defined for any real t. Use this to show that E( U) var( U)

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Estimation: Part 2. Chapter GREG estimation

Bernoulli Numbers and Polynomials

Salmon: Lectures on partial differential equations. Consider the general linear, second-order PDE in the form. ,x 2

2.3 Nilpotent endomorphisms

Differentiating Gaussian Processes

APPENDIX A Some Linear Algebra

Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space.

6 More about likelihood

First Year Examination Department of Statistics, University of Florida

Section 8.3 Polar Form of Complex Numbers

Fall 2012 Analysis of Experimental Measurements B. Eisenstein/rev. S. Errede

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0

Errata to Invariant Theory with Applications January 28, 2017

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4)

Chapter 7 Generalized and Weighted Least Squares Estimation. In this method, the deviation between the observed and expected values of

ψ = i c i u i c i a i b i u i = i b 0 0 b 0 0

Parametric fractional imputation for missing data analysis. Jae Kwang Kim Survey Working Group Seminar March 29, 2010

= z 20 z n. (k 20) + 4 z k = 4

C/CS/Phy191 Problem Set 3 Solutions Out: Oct 1, 2008., where ( 00. ), so the overall state of the system is ) ( ( ( ( 00 ± 11 ), Φ ± = 1

Maximum Likelihood Estimation of Binary Dependent Variables Models: Probit and Logit. 1. General Formulation of Binary Dependent Variables Models

Limited Dependent Variables

Linear Approximation with Regularization and Moving Least Squares

763622S ADVANCED QUANTUM MECHANICS Solution Set 1 Spring c n a n. c n 2 = 1.

F71SM1 STATISTICAL METHODS TUTORIAL ON 7 ESTIMATION SOLUTIONS

SL n (F ) Equals its Own Derived Group

SELECTED PROOFS. DeMorgan s formulas: The first one is clear from Venn diagram, or the following truth table:

4 Analysis of Variance (ANOVA) 5 ANOVA. 5.1 Introduction. 5.2 Fixed Effects ANOVA

NOTES ON SIMPLIFICATION OF MATRICES

Expected Value and Variance

First day August 1, Problems and Solutions

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction

Solution 1 for USTC class Physics of Quantum Information

NUMERICAL DIFFERENTIATION

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity

Stat260: Bayesian Modeling and Inference Lecture Date: February 22, Reference Priors

BOUNDEDNESS OF THE RIESZ TRANSFORM WITH MATRIX A 2 WEIGHTS

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens

MATH 5707 HOMEWORK 4 SOLUTIONS 2. 2 i 2p i E(X i ) + E(Xi 2 ) ä i=1. i=1

Expectation propagation

Lecture 6/7 (February 10/12, 2014) DIRAC EQUATION. The non-relativistic Schrödinger equation was obtained by noting that the Hamiltonian 2

However, since P is a symmetric idempotent matrix, of P are either 0 or 1 [Eigen-values

Strong Markov property: Same assertion holds for stopping times τ.

MATH 241B FUNCTIONAL ANALYSIS - NOTES EXAMPLES OF C ALGEBRAS

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016

Foundations of Arithmetic

8.6 The Complex Number System

The EM Algorithm (Dempster, Laird, Rubin 1977) The missing data or incomplete data setting: ODL(φ;Y ) = [Y;φ] = [Y X,φ][X φ] = X

Explaining the Stein Paradox

Math 702 Midterm Exam Solutions

FINITELY-GENERATED MODULES OVER A PRINCIPAL IDEAL DOMAIN

Lecture 4: Universal Hash Functions/Streaming Cont d

Primer on High-Order Moment Estimators

MATH 829: Introduction to Data Mining and Analysis The EM algorithm (part 2)

Digital Signal Processing

Causal Diamonds. M. Aghili, L. Bombelli, B. Pilgrim

5 The Rational Canonical Form

Ballot Paths Avoiding Depth Zero Patterns

Notes on Frequency Estimation in Data Streams

Lecture 21: Numerical methods for pricing American type derivatives

3.1 ML and Empirical Distribution

Modelli Clamfim Equazione del Calore Lezione ottobre 2014

Example: (13320, 22140) =? Solution #1: The divisors of are 1, 2, 3, 4, 5, 6, 9, 10, 12, 15, 18, 20, 27, 30, 36, 41,

Math 217 Fall 2013 Homework 2 Solutions

xp(x µ) = 0 p(x = 0 µ) + 1 p(x = 1 µ) = µ

Lecture 5 Decoding Binary BCH Codes

SIO 224. m(r) =(ρ(r),k s (r),µ(r))

10-701/ Machine Learning, Fall 2005 Homework 3

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Georgia Tech PHYS 6124 Mathematical Methods of Physics I

See Book Chapter 11 2 nd Edition (Chapter 10 1 st Edition)

Topic 5: Non-Linear Regression

The Prncpal Component Transform The Prncpal Component Transform s also called Karhunen-Loeve Transform (KLT, Hotellng Transform, oregenvector Transfor

STAT 3008 Applied Regression Analysis

1 Generating functions, continued

Hidden Markov Models & The Multivariate Gaussian (10/26/04)

Complex Numbers. x = B B 2 4AC 2A. or x = x = 2 ± 4 4 (1) (5) 2 (1)

P exp(tx) = 1 + t 2k M 2k. k N

k t+1 + c t A t k t, t=0

Conjugacy and the Exponential Family

Systems of Equations (SUR, GMM, and 3SLS)

CIS526: Machine Learning Lecture 3 (Sept 16, 2003) Linear Regression. Preparation help: Xiaoying Huang. x 1 θ 1 output... θ M x M

Linear Feature Engineering 11

Chat eld, C. and A.J.Collins, Introduction to multivariate analysis. Chapman & Hall, 1980

Formulas for the Determinant

1 (1 + ( )) = 1 8 ( ) = (c) Carrying out the Taylor expansion, in this case, the series truncates at second order:

Supplementary material: Margin based PU Learning. Matrix Concentration Inequalities

Feb 14: Spatial analysis of data fields

Complex Numbers Alpha, Round 1 Test #123

Random Partitions of Samples

COMPLEX NUMBERS AND QUADRATIC EQUATIONS

Transcription:

1 Solutons Homework 4 March 5, 018 Soluton to Exercse 5.1.8: Let a IR be a translaton and c > 0 be a re-scalng. ˆb1 (cx + a) cx n + a (cx 1 + a) c x n x 1 cˆb 1 (x), whch shows ˆb 1 s locaton nvarant and scale equvarant. It s not clear why anyone would want to compute ths. Turnng to the others: ˆb (cx + a) (cx (n) + a) (cx (1) + a) c(x (n) x (1) ) cˆb (x). Note that snce c > 0, f we defne y cx+a then the order statstcs (sorted values) of y satsfy y () cx () + a, a fact that was used n the frst step of the computaton above. Note that the statstc analyzed here s the sample range, whch s used often for varous reasons. ˆb3 (cx + a) 1 n(n 1) cˆb 3 (x). ˆb4 (cx + a) n c (cx () + a) 1 n n c c x () + a c 1 1 n c c x () 1 cˆb 4 (x). (cx + b) c(x j + b) j In the last verfcaton, we used the comment about the relaton between the order statstcs of y and x at the frst step, and the assumpton that c 0 n the second to last step.

Soluton to Exercse 5..9: These are all exponental famles, so we wll apply Proposton 5..3. For part (a), f θ (x) 1 x! θx e θ exp[x log θ θ]h(x). Cleary η(θ) log θ s dfferentable and the dervatve 1/θ has rank 1 for all θ (.e., s nonzero). The formula for the Fsher Informaton n equaton (5.40) s I(θ) (1/θ) Var θ (X) 1/θ. Note as an asde that X s an unbased estmator of θ and Var θ (X) θ 1/I(θ), so X s UMVUE for θ. Turnng to part (b), n f θ (x) θ x (1 θ) n x exp[x log(θ/(1 θ)) + n log(1 θ)]h(x). x The dervatve of the natural parameter functon s η (θ) d dθ log(θ/(1 θ)) 1 θ + 1 1 θ 1 θ(1 θ). As ths clearly exsts for 0 < θ < 1, we conclude that Proposton 5..3 apples. Now, the Fsher Informaton s I(θ) η (θ) Var θ [X] nθ(1 θ) [θ(1 θ)] n θ(1 θ). Agan, as an asde, we can easly see that n 1 X s an unbased estmator of θ whose varance equals the lower bound, so t s UMVUE. Fnally, for part (c), the pmf s m + x 1 f θ (x) θ m (1 θ) x exp[x log(1 θ) + m log(θ)]h(x). m 1 Clearly, η (θ) d 1 log(1 θ) dθ 1 θ exsts. One can use Proposton.3.1(b) (on the moments of the suffcent statstc n an exponental famly; we have to substute n θ 1 e η nto m log(θ)) to derve that Var θ (X) m(1 θ) θ.

3 Therefore, I(θ) 1 m(1 θ) 1 θ θ m θ (1 θ). Soluton to Exercse 5..16: If f( x) f(x), then ψ(x) f (x)/f(x) satsfes ψ( x) ψ(x) (snce f ( x) f (x)), so ψ ( x) ψ (x), and therefore I 1 (a, b) 1 a xψ (x)f(x) dx 0, snce x s an odd functon and ψ (x)f(x) s an even functon. nformaton matrx s dagonal: I(a, b) 1 a [ I11 (1, 0) 0 0 I (1, 0) ], Thus, the and thus I 1 (a, b) a [ I 1 11 (1, 0) 0 0 I 1 (1, 0) Now t s clear that a I 11 (1, 0) (respectvely a I (1, 0)) are the Informatons for locaton estmaton when scale s known (respectvely, scale estmaton when locaton s known) for a sngle sample, so ths shows parts (a) and (b). For part (c), smply note that all the p.d.f. s mentoned there are symmetrc, so the result apples to them. Soluton to Exercse 5.3.: For part (a), the p.d.f. w.r.t. countng measure on IN {0, 1,,...} s whch s an exponental famly wth f µ (x) exp [x log µ µ] 1 x!, ]. T (x) x, η(µ) log(µ). Note that x s not a.s. constant (.e., does not satsfy a lnear constrant n one dmenson) and η ranges over IR as µ ranges over (0, ), so the famly s full rank, and X s complete and suffcent for a sngle sample, and T X s complete and suffcent for n..d. observatons. We know from e.g. a m.g.f. argument that T s P osson(nµ).

4 g(µ) s U-estmable f and only f there s a δ : IN IR such that g(µ) k0 δ(k) (nµ)k e nµ, µ > 0. k! Multplyng through by e nµ we see that e nµ g(µ) has a power seres that converges for all µ > 0, so t must n fact converge for all real µ (ndeed, for all complex µ), and I hope that you at least learned n calculus that f a power seres centered at 0 (a.k.a. a McLaurn seres) converges for some x, then t converges for any y wth y < x. Furthermore, such power seres representatons are unque and are gven by the classcal Taylor formula: e nµ g(µ) δ(k)n k 1 k0 k! µk µ f and only f k dk δ(k) n dµ k [enµ g(µ)] µ0. In concluson, g(µ) s U-estmable f and only f t has a power seres expanson (centered at 0) vald for all real numbers (otherwse sad, s an entre analytc functon), and then the UMVUE s δ(t ) where δ s gven by formula above. For part (c), t s sometmes easer to use ad hoc methods rather than the formula above to fnd the UMVUE. () g(µ) µ k s already gven as a power seres expanson, and t s easy to see that power seres multply, so the power seres we want s e nµ µ k µ k n j µ j. n k 1 j!! n µ! ( k)! j0 k Matchng the coeffcents, t s clear that the UMVUE s δ(t ) where In partcular, δ(t) { 0 f t < k, n k t!/(t k)! f t k. UMVUE for µ n 1 T X

5 and UMVUE for µ { 0 f T 0 or T 1 T (T 1)/n f T. T (T 1)/n Note that f we wrte t!/(t k)! k 1 0 (t ), then the UMVUE for µ k may be wrtten more smply as k 1 δ(t) n k (t ). 0 () Of course, Thus, g(µ) P µ [X 1 k] µk k! e µ. e nµ g(µ) µk k! e(n 1)µ µk k! k j0 1 j! (n 1)j µ j! ( k)!k! n (n 1) k n 1! µ, and we see that the desred UMVUE s t δ(t) n t (n 1) t k. k Note that the bnomal coeffcent our UMVUE of P µ [X 1 0] s ( t k δ(t) n t (n 1) t (1 1/n) t. ) s 0 unless k t. In partcular,

6 () Note that g(µ) log µ does not have the requred Taylor seres expanson (n partcular, t doesn t have a fnte value and dervates at µ 0). No unbased estmate exsts. (v) Pluggng n our formula, (I am changng the estmand to g(µ) e aµ ) e nµ e aµ e (n+a)µ n + a k n k 1 k0 n k! µk, from whch we can read off the UMVUE as δ(t) n + a t (1 + a/n) t. n (v) Pluggng n agan, e nµ e µ + µ k0(nµ ) k 1 k! 1 k k n j µ j+(k j) k! j k0 k0 k j0 (by the bnomal formula) 1 k k n j µ k j k0 k! j j0 k 1 k n k µ k! k 0 k /! ( k)!(k!) n(k ) 1! n µ. In the above, / denotes the smallest nteger /,.e. / f s even and ( + 1)/ f s odd. From ths, we can read off the UMVUE as T T! δ(t ) (T k)!(k T!) n(k T ) k T/ (v) Ths one s easy: e µ s at µ 0, so t can t have the requste Taylor seres expanson.

Soluton to Exercse 5.3.6: We have from the fact that σ n 1 (X X) has a χ n 1 dstrbuton that E[ˆσ (a)] (n 1)aσ Var[ˆσ (a)] (n 1)a σ 4. Usng the formula MSE Bas + Varance, we obtan σ 4 MSE[ˆσ (a)] [a(n 1) 1] + (n 1)a. Thus, wth a lttle algebra we obtan σ 4 { MSE[ˆσ (a)] MSE[ˆσ (1/(n 1))] } (n 1)a (n 1)a + 1 /(n 1) (n 1)[a 1/(n 1)][a 1/(n + 3)]. Note that the r.h.s. s a quadratc functon of a whose graph s an upward openng parabola wth roots at a 1/(n 1) and a 1/(n+3). Thus, ˆσ (a) has smaller MSE than the UMVUE ˆσ (1/(n 1)) for (n+3) 1 < a < (n 1) 1. Soluton to Exercse 5.3.8: (a) The lkelhood s [ f(y; a, b, σ ) (πσ ) n/ exp 1 ] n (y σ (ax + b)) 1 [ (π) n/ exp 1 y σ + a x σ y + b σ 1 (ax σ + b) n ] log σ As long as we have at least two dstnct values of the x, then the famly wll be dentfable, as dfferent values of a and b wll gve dfferent means. Ths assumpton was apparently forgotten. Thus, the suffcent statstc vector T ( x Y, Y, Y ) does not satsfy any lnear constrants. As (a, b, σ ) range over IR IR (0, ), the natural parameter vector ranges over (, 0) IR IR, whch s an open set, so has nonempty nteror. Thus, the famly s full rank, hence T s complete and suffcent. (b) Clearly Ȳ n 1 T s a functon of T. Also, y 7 n x Ỹ 1 x Y x Y Ȳ x

8 s a functon of T. Thus, we only need to check that â s unbased for a. Note that E[Ȳ ] a x + b. Thus, E (a,b,σ )[â] 1 x x E[Ỹ] 1 x x [(ax + b) (a x + b)] a. 1 x a x (x x ) Thus, â s a functon of the complete and suffcent statstcs whch s unbased for a, so t s the UMVUE of a. We see mmedately that ˆb s also a functon of T (snce Ȳ and â are), so we need only check that t s expectaton s always b. Thus, (c) Clearly ˆσ E (a,b,σ )[ˆb] E[Ȳ ] E[â] x a x + b a x b. 1 [ Y n + (âx + ˆb) â x Y ˆb ] Y, whch s a functon of T. Now we want to check that ts expectaton of σ. It s almost always easer to do these types of calculatons f one subtracts and adds E[Y ] n the squared quantty: (n )E[ˆσ ] [ (Y E E[Y ] + E[Y ] âx ˆb ) ] [ (ɛ E + (a â)x + (b ˆb) ) ] (snce Y E[Y ] + ɛ and E[Y ] ax + b) â a Ỹ x x a Ỹ Y Ȳ ax + b + ɛ (a x + b + ɛ) a x + ɛ (where ɛ ɛ ɛ)

9 â a (a x + ɛ ) x x a Put ɛ x x ˆb b Ȳ â x b a x + b + ɛ â x b (a â) x + ɛ (â a)x + (ˆb b) (â a) x + ɛ j ɛ j x j x j x + ɛ j Note that w x ( j x j ) 1/. w 1, w 0, where the latter follows because x (x x) 0. Thus, we have (n )E[ˆσ ] E ɛ w w j ɛ j ɛ j E ɛ w w j ɛ j j E ɛ ɛ w w j ɛ j + w w j ɛ j j j E ɛ w j ɛ j. j Now [ ] E ɛ [ ] E (ɛ ɛ) (n 1)σ.

10 Also, E j w j ɛ j E j E j w j (ɛ j ɛ) w j ɛ j (snce wj 0) σ j w j σ. Thus, we get n the end that E[ˆσ ] σ, as desred.