1 Additional lemmas. Supplementary Material APPENDIX. that N 1 N } E } E { N } E + O(N 3 ), Proof. The results follow by straightforward calculation.

Σχετικά έγγραφα
Analytical Expression for Hessian

Matrix Hartree-Fock Equations for a Closed Shell System

e t e r Cylindrical and Spherical Coordinate Representation of grad, div, curl and 2

(a,b) Let s review the general definitions of trig functions first. (See back cover of your book) sin θ = b/r cos θ = a/r tan θ = b/a, a 0

Laplace s Equation in Spherical Polar Coördinates

Uniform Convergence of Fourier Series Michael Taylor

Space Physics (I) [AP-3044] Lecture 1 by Ling-Hsiao Lyu Oct Lecture 1. Dipole Magnetic Field and Equations of Magnetic Field Lines

Curvilinear Systems of Coordinates

Areas and Lengths in Polar Coordinates

ANTENNAS and WAVE PROPAGATION. Solution Manual

Areas and Lengths in Polar Coordinates

ST5224: Advanced Statistical Theory II

4.2 Differential Equations in Polar Coordinates

Math 446 Homework 3 Solutions. (1). (i): Reverse triangle inequality for metrics: Let (X, d) be a metric space and let x, y, z X.

2 Composition. Invertible Mappings

MATH 38061/MATH48061/MATH68061: MULTIVARIATE STATISTICS Solutions to Problems on Matrix Algebra

М. I. Parolya, М. М. Sheremeta ESTIMATES FROM BELOW FOR CHARACTERISTIC FUNCTIONS OF PROBABILITY LAWS

Exercise, May 23, 2016: Inflation stabilization with noisy data 1

EE512: Error Control Coding

Phys460.nb Solution for the t-dependent Schrodinger s equation How did we find the solution? (not required)

Other Test Constructions: Likelihood Ratio & Bayes Tests

Solutions to Exercise Sheet 5

Every set of first-order formulas is equivalent to an independent set

A Note on Intuitionistic Fuzzy. Equivalence Relation

Tutorial Note - Week 09 - Solution

Fractional Colorings and Zykov Products of graphs

A Two-Sided Laplace Inversion Algorithm with Computable Error Bounds and Its Applications in Financial Engineering

TL-Moments and L-Moments Estimation for the Generalized Pareto Distribution

Approximation of distance between locations on earth given by latitude and longitude

SOLUTIONS TO MATH38181 EXTREME VALUES AND FINANCIAL RISK EXAM

ORDINAL ARITHMETIC JULIAN J. SCHLÖDER

Example 1: THE ELECTRIC DIPOLE

Oscillating dipole system Suppose we have two small spheres separated by a distance s. The charge on one sphere changes with time and is described by

Problem Set 3: Solutions

Solution Series 9. i=1 x i and i=1 x i.

D Alembert s Solution to the Wave Equation

6.3 Forecasting ARMA processes

SOLUTIONS TO MATH38181 EXTREME VALUES AND FINANCIAL RISK EXAM

Lecture 34 Bootstrap confidence intervals

Limit theorems under sublinear expectations and probabilities

The semiclassical Garding inequality

Example Sheet 3 Solutions

Ordinal Arithmetic: Addition, Multiplication, Exponentiation and Limit

Pg The perimeter is P = 3x The area of a triangle is. where b is the base, h is the height. In our case b = x, then the area is

Section 7.6 Double and Half Angle Formulas

6. MAXIMUM LIKELIHOOD ESTIMATION

forms This gives Remark 1. How to remember the above formulas: Substituting these into the equation we obtain with

Numerical Analysis FMN011

F19MC2 Solutions 9 Complex Analysis

5. Choice under Uncertainty

Homework 8 Model Solution Section

Finite Field Problems: Solutions

ω ω ω ω ω ω+2 ω ω+2 + ω ω ω ω+2 + ω ω+1 ω ω+2 2 ω ω ω ω ω ω ω ω+1 ω ω2 ω ω2 + ω ω ω2 + ω ω ω ω2 + ω ω+1 ω ω2 + ω ω+1 + ω ω ω ω2 + ω

( y) Partial Differential Equations

Lecture 15 - Root System Axiomatics

Second Order Partial Differential Equations

= λ 1 1 e. = λ 1 =12. has the properties e 1. e 3,V(Y

k A = [k, k]( )[a 1, a 2 ] = [ka 1,ka 2 ] 4For the division of two intervals of confidence in R +

Απόκριση σε Μοναδιαία Ωστική Δύναμη (Unit Impulse) Απόκριση σε Δυνάμεις Αυθαίρετα Μεταβαλλόμενες με το Χρόνο. Απόστολος Σ.

21. Stresses Around a Hole (I) 21. Stresses Around a Hole (I) I Main Topics

dx x ψ, we should find a similar expression for rθφ L ψ. From L = R P and our knowledge of momentum operators, it follows that + e y z d

Tridiagonal matrices. Gérard MEURANT. October, 2008

Overview. Transition Semantics. Configurations and the transition relation. Executions and computation

Lecture 13 - Root Space Decomposition II

HOMEWORK 4 = G. In order to plot the stress versus the stretch we define a normalized stretch:

Homework 3 Solutions

2x 2 y x 4 +y 2 J (x, y) (0, 0) 0 J (x, y) = (0, 0) I ϕ(t) = (t, at), ψ(t) = (t, t 2 ), a ÑL<ÝÉ b, ½-? A? 2t 2 at t 4 +a 2 t 2 = lim

New bounds for spherical two-distance sets and equiangular lines

CHAPTER 101 FOURIER SERIES FOR PERIODIC FUNCTIONS OF PERIOD

4.6 Autoregressive Moving Average Model ARMA(1,1)

Lecture 2: Dirac notation and a review of linear algebra Read Sakurai chapter 1, Baym chatper 3

2. Let H 1 and H 2 be Hilbert spaces and let T : H 1 H 2 be a bounded linear operator. Prove that [T (H 1 )] = N (T ). (6p)

w o = R 1 p. (1) R = p =. = 1

1 3D Helmholtz Equation

A Bonus-Malus System as a Markov Set-Chain. Małgorzata Niemiec Warsaw School of Economics Institute of Econometrics

Arithmetical applications of lagrangian interpolation. Tanguy Rivoal. Institut Fourier CNRS and Université de Grenoble 1

q-analogues of Triple Series Reduction Formulas due to Srivastava and Panda with General Terms

Inverse trigonometric functions & General Solution of Trigonometric Equations

On mixing generalized poison with Generalized Gamma distribution

A NEW CLASS OF MODULAR EQUATIONS IN RAMANUJAN S ALTERNATIVE THEORY OF ELLIPTIC FUNCTIONS OF SIGNATURE 4 AND SOME NEW P-Q ETA-FUNCTION IDENTITIES

Srednicki Chapter 55

CHAPTER 25 SOLVING EQUATIONS BY ITERATIVE METHODS

6.1. Dirac Equation. Hamiltonian. Dirac Eq.

Practice Exam 2. Conceptual Questions. 1. State a Basic identity and then verify it. (a) Identity: Solution: One identity is csc(θ) = 1

Problem Set 9 Solutions. θ + 1. θ 2 + cotθ ( ) sinθ e iφ is an eigenfunction of the ˆ L 2 operator. / θ 2. φ 2. sin 2 θ φ 2. ( ) = e iφ. = e iφ cosθ.

Lecture 21: Properties and robustness of LSE

Statistical Inference I Locally most powerful tests

C.S. 430 Assignment 6, Sample Solutions

Empirical best prediction under area-level Poisson mixed models

derivation of the Laplacian from rectangular to spherical coordinates

Μηχανική Μάθηση Hypothesis Testing

3.4 SUM AND DIFFERENCE FORMULAS. NOTE: cos(α+β) cos α + cos β cos(α-β) cos α -cos β

ANSWERSHEET (TOPIC = DIFFERENTIAL CALCULUS) COLLECTION #2. h 0 h h 0 h h 0 ( ) g k = g 0 + g 1 + g g 2009 =?

Matrices and Determinants

Solutions Ph 236a Week 2

Appendix to On the stability of a compressible axisymmetric rotating flow in a pipe. By Z. Rusak & J. H. Lee

Math 6 SL Probability Distributions Practice Test Mark Scheme

1 String with massive end-points

VEKTORANALYS. CURVILINEAR COORDINATES (kroklinjiga koordinatsytem) Kursvecka 4. Kapitel 10 Sidor

Second Order RLC Filters

Transcript:

1 Additional lemmas Supplementay Mateial APPENDIX Lemma A1. Let (T 1,1, T 2,1, T 3,1, T 4,1 ),..., (T 1,N, T 2,N, T 3,N, T 4,N ) be independent andom vectos of length 4 such that E(T,i ) = 0 (i = 1,..., N, = 1, 2, 3, 4). Assume that N 1 N i=1 E(T,i) 2 = O(1) fo = 1, 2, 3, 4. Then E Ẽ(T1 )Ẽ(T N 2) = N 2 E(T 1,i T 2,i ), i=1 E Ẽ(T1 )Ẽ(T2)Ẽ(T N 3) = N 3 E(T 1,i T 2,i T 3,i ), i=1 E Ẽ(T1 )Ẽ(T2)Ẽ(T3)Ẽ(T 4) = N 4 κ N i=1 E(T κ1,it κ2,i) N =1 E(T κ3,t κ4,) + O(N 3 ), whee κ = (κ 1, κ 2, κ 3, κ 4 ) and κ is taken ove κ = (1, 2, 3, 4), (1, 3, 2, 4), (1, 4, 2, 3). Poof. The esults follow by staightfowad calculation. Lemma A2. Let τ 1, τ 2, τ 3, τ 4 be vaiables taking values τ 1,i, τ 2,i, τ 3,i, τ 4,i (i = 1,..., N) on the finite population. Define T,i = (p 1 i R i 1)τ,i and T,i O = (p 1 i R i 1)τ,i O fo = 1, 2, 3, 4. Assume that (10) holds, lim inf N Ẽp(1 p) > 0, and Ẽ(τ 2 ) = O(1) fo = 1, 2, 3, 4. Then the following esults hold. (i) E Re Ẽ(T1 )Ẽ(T O 2) = E P ois Ẽ(T )Ẽ(T O 1 2 ) + o(n 1 ). (ii) Ẽ(T1 )Ẽ(T 2), Ẽ(T 3 )Ẽ(T O 4) = cov P ois Ẽ(T )Ẽ(T O 1 2 ), Ẽ(T3 O )Ẽ(T 4 O ) + o(n 2 ). (iii) Ẽ(T1 ), Ẽ(T 2 )Ẽ(T O 3) = cov P ois Ẽ(T 1 ), Ẽ(T2 O )Ẽ(T 3 O ) + σ 1 Ẽp(1 p)τ1 O O E P ois Ẽ(T )Ẽ(T O 2 3 ) + o(n 2 ). (iv) Ẽ(T1 ), Ẽ(T 2 )Ẽ(T3)Ẽ(T O 4) = cov P ois Ẽ(T 1 ), Ẽ(T2 O )Ẽ(T 3 O )Ẽ(T 4 O ) + o(n 2 ). Thoughout, E Re and ae computed unde eective sampling and E P ois and cov P ois ae computed unde Poisson sampling. 1

Poof. We show esult (iii) in detail. The othe esults can be shown similaly. By simple manipulation, [Ẽ(p 1 R 1)τ 1, Ẽ(p 1 R 1)τ 2 Ẽ(p 1 R 1)τ 3, = 1 E N 3 Re (R i π i )(R p )(R k p k ) p 1 i τ 1,i p 1 τ 2, p 1 k i,,k τ 3,k. By Haek (1981, Chapte 14), conside the egession pedicto ˆπ i p i )σ 1 (K n), whee K = N i=1 R i and n = N i=1 p i. Then = p i p i (1 E Re (Ri π i ) 2 = E P ois (Ri ˆπ i ) 2 + o(σ 1 ), E Re (R i π i )(R π ) = E P ois (R i ˆπ i )(R ˆπ ) + o(σ 1 ) (i ), whee σ o(σ 1 ) 0 as σ unifomly in i,. Futhemoe, conside the quadatic egession pedicto ˆπ iq = p i p i (1 p i )σ 1 (K n) p i (1 p i )( p p i )σ 1, whee p is defined as befoe in (18). Then E Re (Ri π i ) 3 = E P ois (Ri ˆπ iq ) 3 + o(σ 1 ) E Re (Ri π i ) 2 (R π ) = E P ois (Ri ˆπ iq ) 2 (R ˆπ Q ) + o(σ 1 ) (i ), E Re (R i π i )(R π )(R k π k ) = E P ois (R i ˆπ iq )(R ˆπ Q )(R k ˆπ kq ) + o(σ 2 ) (i k), whee σ o(σ 1 ) 0 and σ 2 o(σ 2 ) 0 as σ unifomly in i,, k. By using these elations and (18) and appoximating R p R ˆπ = (R ˆπ Q ) + (ˆπ Q ˆπ ), we find that [Ẽ(p 1 R 1)τ 1, Ẽ(p 1 R 1)τ 2 Ẽ(p 1 R 1)τ 3, ( 1 = 1 + o(1) N 3 i,,k [ 1 1 (1 p i )( p p i )τ 1,i Nσ N 2 i E P ois (R i ˆπ i )(R ˆπ )(R k ˆπ k ) p 1 i,k = (R π ) + (π p ) by τ 1,i p 1 τ 2, p 1 k τ 3,k E P ois (R ˆπ )(R k ˆπ k ) p 1 τ 2, p 1 k τ 3,k The esult follows because N 1 (R i ˆπ i )p 1 i τ 1,i = Ẽ(p 1 R 1)τ1 O, etc. The tem Ẽ(1 p)( p p)τ 1 equals Ẽ(1 p)( p p)τ O 1 because Ẽ(1 p)( p p)p = 0 and hence equals Ẽp(1 p)τ O 1 because Ẽ(1 p)τ O 1 = 0. ). 2

Lemma A3. In the setup of Lemma A2, define S,i = (π 1 i R i 1)τ,i fo = 1, 2, 3, 4. Unde the assumptions in Lemma A1, the following esults hold. (i) E Re Ẽ(S1 )Ẽ(S O 2) = E P ois Ẽ(T )Ẽ(T O 1 2 ) + o(n 1 ). (ii) Ẽ(S1 )Ẽ(S 2), Ẽ(S 3 )Ẽ(S O 4) = cov P ois Ẽ(T )Ẽ(T O 1 2 ), Ẽ(T3 O )Ẽ(T 4 O ) + o(n 2 ). (iii) Ẽ(S1 ), Ẽ(S 2 )Ẽ(S O 3) = cov P ois Ẽ(T 1 ), Ẽ(T2 O )Ẽ(T 3 O ) + σ 1 Ẽp(1 p)τ1 O O E P ois Ẽ(T )Ẽ(T O 2 3 ) + σ 1 Ẽp(1 p)τ2 O O E P ois Ẽ(T )Ẽ(T O 1 3 ) + σ 1 Ẽp(1 p)τ3 O O E P ois Ẽ(T )Ẽ(T O 1 2 ) + o(n 2 ). (iv) Ẽ(S1 ), Ẽ(S 2 )Ẽ(S3)Ẽ(S O 4) = cov P ois Ẽ(T 1 ), Ẽ(T2 O )Ẽ(T 3 O )Ẽ(T 4 O ) + o(n 2 ). Poof. We show esult (iii) in detail. By simple manipulation, [Ẽ(π 1 R 1)τ 1, Ẽ(π 1 R 1)τ 2 Ẽ(π 1 R 1)τ 3, = 1 E N 3 Re (R i π i )(R π )(R k π k ) π 1 i τ 1,i π 1 τ 2, π 1 k τ 3,k i,,k = 1 N 3 i,,k E Re (R i π i )(R p )(R k p k ) p 1 i i,,k τ 1,i p 1 τ 2, p 1 k τ 3,k + 1 E N 3 Re (R i π i )(p π )(R k p k ) p 1 i τ 1,i p 1 τ 2, p 1 + 1 E N 3 Re (R i π i )(R π )(p k π k ) p 1 i i,,k τ 1,i p 1 k τ 3,k τ 2, p 1 k τ 3,k + o(n 2 ). Result (ii) follows by simila calculations to those fo Lemma A2 using the appoximation of p i π i by ˆπ i ˆπ iq. 2 Poofs Poof of Poposition 1. (i) Diect calculation shows that ˆα eg = g 1 + g 2 (ˆα eg ) Ẽ(eξT ) Ẽ 1 (ξξ T ) V 1 Ẽ(ξ). 3

Expansion (4) holds because Ẽ 1 (ξξ T ) V 1 = Ẽ 1 (ξξ T )Ẽ(ξξT ) V V 1 = V 1 Ẽ(ξξT ) V V 1 + o p (N 1/2 ) by the assumption E( ξ 4 ) <. the expansion of ˆα egc holds because Similaly, ˆα egc = g 1 + g 2 (ˆα egc ) + Ẽ(e 0)Ẽ(ξT )Ẽ 1 (ξξ T )Ẽ(ξ) Ẽ(e 0ξ T ) Ẽ 1 (ξξ T ) V 1 Ẽ(ξ). (ii) By Owen (2001, Chapte 11), max i ξ i = o p (N 1/2 ) and ˆλ = O p (N 1/2 ) povided that E( ξ 2 ) <. A Taylo expansion of (3) gives [ 0 = Ẽ(ξ) )ˆλ + Ẽ(ξ Ẽ(ξξT ξξ T )ˆλ ˆλ 1 ξ ξ k ξξ T Ẽ 2 (1 + λ ˆλ ˆλk ˆλ, T ξ) 4 whee λ is on the line oining 0 and ˆλ. By the Cauchy-Schwatz and tiangle inequalities, the diffeence between the last tem and k Ẽ(ξ ξ k ξξ T )ˆλ ˆλk ˆλ/2 is bounded in absolute value by [ 1 Ẽ ξ ξ k ξξ T max (1 + 2 λ T ξ i ) 4 1 ˆλ ˆλk ˆλ = o p (N 3/2 ) i k because max i λ T ξ i = o p (1) and E( ξ 4 ) <. By eaanging the expansion of (3) and solving fo ˆλ as in Newey and Smith (2004, Lemma A4), we find ˆλ = h 1 +h 2 (ˆλ)+ h 3 (ˆλ) + o p (N 3/2 ), whee h 1 = V 1 Ẽ(ξ) and h 2 (ˆλ) = V Ẽ(ξξT 1 ) V V 1 Ẽ(ξ) + V Ẽ(ξT 1 )V 1 E(ξ ξξ T ) V 1 Ẽ(ξ), h 3 (ˆλ) = V Ẽ(ξξT 1 ) V V Ẽ(ξξT 1 ) V V 1 Ẽ(ξ) V Ẽ(ξξT 1 ) V V Ẽ(ξT 1 )V 1 E(ξ ξξ T ) V 1 Ẽ(ξ) V [Ẽ(ξT 1 )V Ẽ(ξξT 1 ) V V 1 E(ξ ξξ T ) V 1 Ẽ(ξ) + V [Ẽ(ξT 1 )V Ẽ(ξT 1 )V 1 E(ξ ξξ T ) V 1 E(ξ ξξ T ) V 1 Ẽ(ξ) V Ẽ(ξT 1 )V 1 E(ξ ξξ T ) V Ẽ(ξξT 1 ) V V 1 Ẽ(ξ) + V Ẽ(ξT 1 )V 1 E(ξ ξξ T ) V ẼT 1 (ξ)v 1 E(ξ ξξ T ) V 1 Ẽ(ξ) + V [Ẽ(ξT 1 )V Ẽ(ξ 1 ξξt ) E(ξ ξξ T ) V 1 Ẽ(ξ) [ 1 2 V 1 E(ξ ξ k ξξ T ) V 1 Ẽ(ξ) V 1 Ẽ(ξ) V 1 Ẽ(ξ) k k Thoughout, each expession of the fom ẼT (ϕ T )UE(ψ ξξ T ), enclosed by culy backets, is defined as k Ẽ(ϕ k)u k E(ψ ξξ T ). Similaly, Ẽ T (ϕ T )UẼ(ψ ξξt ) is defined 4 k

as k Ẽ(ϕ k)u k Ẽ(ψ ξξ T ). Futhe, a Taylo expansion of ˆα lik gives [ ˆα lik = Ẽ(η) Ẽ(ηξT )ˆλ + ˆλ T Ẽ(ηξξ T )ˆλ 1 2 ˆλ ηξ ξξ T T Ẽ (1 + λ ˆλ T ξ) 4 ˆλ, whee λ is on the line oining 0 and ˆλ. The last tem is ˆλ T Ẽ(ηξ ξξ T )ˆλ ˆλ/2 + o p (N 3/2 ) similaly as in the expansion of (3). Wite h 2 = h 2 (ˆλ) and h 3 = h 3 (ˆλ). By substituting the expansion of ˆλ into that of ˆα lik, we find ˆα lik = Ẽ(η) E(ηξT )(h 1 + h 2 + h 3 ) Ẽ(ηξT ) E(ηξ T ) (h 1 + h 2 ) + h T 1E(ηξξ T )(h 1 + h 2 ) + h T 2E(ηξξ T )h 1 + h T 1 Ẽ(ηξξT ) E(ηξξ T ) 1 2 ht 1 E(ηξ ξξ T )(h 1 ) h 1 + O p (N 3/2 ). By eaanging the tems, we have g 2 (ˆα lik ) = E(ηξ T )h 2 Ẽ(ηξT ) E(ηξ T ) h 1 + h T 1E(ηξξ T )h 1, g 3 (ˆα lik ) = E(ηξ T )h 3 Ẽ(ηξT ) E(ηξ T ) h 2 + h T 1E(ηξξ T )h 2 + h T 2E(ηξξ T )h 1 + h T 1 Ẽ(ηξξT ) E(ηξξ T ) 1 2 ht 1 E(ηξ ξξ T )(h 1 ) h 1. Finally, by diect calculation, the g 2 tem is simplified to that in Poposition 1, and the g 3 tem is simplified to g 3 (ˆα lik ) = Ẽ(eξT )V Ẽ(ξξT 1 ) V V 1 Ẽ(ξ) Ẽ(eξT )V Ẽ(ξT 1 )V 1 E(ξ ξξ T ) V 1 Ẽ(ξ) Ẽ(ξT )V Ẽ(ξξT 1 ) V V 1 E(eξξ T )V 1 Ẽ(ξ) + Ẽ(ξT )V Ẽ(ξT 1 )V 1 E(ξ ξξ T ) V 1 E(eξξ T )V 1 Ẽ(ξ) Ẽ(ξT )V 1 E(eξξ T )V Ẽ(ξξT 1 ) V V 1 Ẽ(ξ) + Ẽ(ξT )V 1 E(eξξ T )V Ẽ(ξT 1 )V 1 E(ξ ξξ T ) V 1 Ẽ(ξ) + Ẽ(ξT )V Ẽ(eξξT 1 ) E(eξξ T ) V 1 Ẽ(ξ) 1 )V Ẽ(ξT 1 )V 1 E(eξ ξξ T ) V 1 Ẽ(ξ). 2Ẽ(ξT h 1 h 1 5

Poofs of Popositions 2 3. Poposition 2 follows by simple calculation. To show V(ˆα lik ), wite g 2 = g 2 (ˆα lik ), g 3 = g 3 (ˆα lik ), and the 8 tems of g 3 as g 31, g 32 and so on. By diect but tedious calculation using Lemma A1, va(g 2 ) = N 2 t(v 1 Σ eξ:eξ ) t(v 1 Σ eξξ V 1 Σ eξξ ), cov(g 1, g 21 ) = N 2 t(v 1 Σ e0 eξξ), cov(g 1, g 22 ) = N 2 t(v 1 Σ e0 ξξv 1 Σ eξξ ), cov(g 1, g 31 ) = N 2 t(v 1 Σ e0 ξξv 1 Σ eξξ ) + N 2 u, cov(g 1, g 32 ) = N 2 u, cov(g 1, g 33 ) = cov(g 1, g 35 ) = N 2 t(v 1 Σ e0 ξξv 1 Σ eξξ ), cov(g 1, g 37 ) = N 2 t(v 1 Σ e0 eξξ), up to tems of ode O(N 2 ) and cov(g 2, g 3l ) = o(n 2 ) fo l = 4, 6, 8, whee Σ eξ:eξ = E(e 2 ξξ T ), Σ eξξ = E(eξξ T ), and Σ e0 ξξ = E(e 0 ξξ T ) as defined in Popositions 2 3, Σ e0 eξξ = E(e 0 eξξ T ), and u = t E(e 0 eξ T )V 1 E(ξ ξξ T )V 1 = E(e 0 eξ T )V 1 E(ξξ T V 1 ξ). The fomula follows by substituting these expessions into the definition of V(ˆα lik ). To show V(ˆα eg,bc ), decompose the bias coection tem as (η ˆβ T ξ)ξ T Ẽ 1 (ξξ T )ξ N 1 Ẽ = N 1 Ẽ(eξ T V 1 ξ) + N Ẽ(eξT 1 V 1 ξ) E(eξ T V 1 ξ) N 1 Ẽ ( ˆβ β) T ξξ T V 1 ξ + N 1 Ẽ [eξ T Ẽ 1 (ξξ T ) V 1 ξ Substituting the fist-ode expansions fo ˆβ and Ẽ 1 (ξξ T ) above leads to thid-ode expansion (4) fo ˆα eg,bc, whee g 2 (ˆα eg,bc ) = Ẽ(eξT )V 1 Ẽ(ξ) + N 1 E(eξ T V 1 ξ), g 3 (ˆα eg,bc ) = Ẽ(eξT )V Ẽ(ξξT 1 ) V V 1 Ẽ(ξ) + N Ẽ(eξT 1 V 1 ξ) E(eξ T V 1 ξ). N 1 E(ξ T V 1 ξξ T )V 1 Ẽ(eξ) [ N 1 t V Ẽ(ξξT 1 ) V V 1 E(eξξ T ). 6

Wite g 2 = g 2 (ˆα eg,bc ), g 3 = g 3 (ˆα eg,bc ), and the 4 tems of g 3 as g 31, g 32 and so on. The fomula fo V(ˆα eg,bc ) follows by diect calculation: va(g 2 ) = N 2 t(v 1 Σ eξ:eξ ) + t(v 1 Σ eξξ V 1 Σ eξξ ), cov(g 1, g 2 ) = N 2 t(v 1 Σ e0 eξξ), cov(g 1, g 31 ) = N 2 t(v 1 Σ e0 ξξv 1 Σ eξξ ) + N 2 u, cov(g 1, g 32 ) = N 2 t(v 1 Σ e0 eξξ), cov(g 1, g 33 ) = N 2 u, cov(g 1, g 34 ) = N 2 t(v 1 Σ e0 ξξv 1 Σ eξξ ), up to tems of ode O(N 2 ). It is inteesting to mention that without bias coection, V(ˆα eg ) is equal to va(g 2 ) + 2cov(g 1, g 2 ) + 2cov(g 1, g 31 ). V(ˆα egc,bc ) can be poved by simila calculation. Poof of Poposition 4. Finally, the fomula fo (i) The expansion of α eg holds similaly to that of ˆα eg because α eg = g 1 +g 2 ( α eg ) Ẽ(εξT ) Ẽ 1 (ξζ T ) V 1 Ẽ(ξ) and Ẽ 1 (ξζ T ) V 1 = V 1 Ẽ(ξζ T ) V V 1 + o p (N 1/2 ). (ii) By Tan (2013), max i=1,...,n p 1 (X i ) 1X i = o p (N 1/2 ) and λ = O p (N 1/2 ). A Taylo expansion of (8) gives [ 0 = Ẽ(ξ) Ẽ(ζξT ) λ + Ẽ(ζ ξξ T ) λ λ 1 ζ ξ k ξξ T Ẽ 2 (1 + λ λ λk λ, T ξ) 4 k whee λ is on the line oining 0 and λ. The last tem is k Ẽ(ζ ξ k ξξ T )ˆλ ˆλk ˆλ/2+ o p (N 3/2 ) similaly as in the expansion of (3). By solving fo λ fom the expansion of (8), we find λ = h 1 + h 2 ( λ) + h 3 ( λ) + o p (N 3/2 ), whee h 2 ( λ) and h 3 ( λ) ae defined as h 2 (ˆλ) and h 3 (ˆλ) with ξξ T eplaced by ζξ T, ξ ξξ T by ζ ξξ T, and ξ ξ k ξξ T by ζ ξ k ξξ T thoughout. Futhe, a Taylo expansion of α lik gives α lik = Ẽ(η) Ẽ(ηξT ) λ + λ T Ẽ(ηξξ T ) λ 1 2 λ T [ ηξ ξξ T Ẽ (1 + λ λ T ξ) 4 λ, whee λ is on the line oining 0 and λ. The last tem is λ T Ẽ(ηξ ξξ T ) λ λ/2+ o p (N 3/2 ) similaly as in the expansion of (3). By substituting the expansion of λ into that of α lik and by simila calculation as in the poof of Poposition 1, we obtain 7

g 2 ( α lik ) as shown in Poposition 1 and g 3 ( α lik ) = Ẽ(εξT )V Ẽ(ζξT 1 ) V V 1 Ẽ(ξ) Ẽ(εξT )V Ẽ(ξT 1 )V 1 E(ζ ξξ T ) V 1 Ẽ(ξ) Ẽ(ξT )V Ẽ(ζξT 1 ) V V 1 E(εξξ T )V 1 Ẽ(ξ) + Ẽ(ξT )V Ẽ(ξT 1 )V 1 E(ζ ξξ T ) V 1 E(εξξ T )V 1 Ẽ(ξ) Ẽ(ξT )V 1 E(εξξ T )V Ẽ(ζξT 1 ) V V 1 Ẽ(ξ) + Ẽ(ξT )V 1 E(εξξ T )V Ẽ(ξT 1 )V 1 E(ζ ξξ T ) V 1 Ẽ(ξ) + Ẽ(ξT )V Ẽ(εξξT 1 ) E(εξξ T ) V 1 Ẽ(ξ) 1 )V Ẽ(ξT 1 )V 1 E(εξ ξξ T ) V 1 Ẽ(ξ). 2Ẽ(ξT Poofs of Popositions 5 6. Poposition 5 follows by simple calculation. To show V( α lik ), wite g 2 = g 2 ( α lik ), g 3 = g 3 ( α lik ), and the 8 tems of g 3 as g 31, g 32 and so on. By diect but tedious calculation using Lemma A1, va(g 2 ) = N 2 t(v 1 Σ εξ:εξ ) t(v 1 Σ εξξ V 1 Σ εξξ ), cov(g 1, g 21 ) = N 2 t(v 1 Σ e0 εξξ), cov(g 1, g 22 ) = N 2 t(v 1 Σ e0 ξξv 1 Σ εξξ ), cov(g 1, g 31 ) = N 2 t(v 1 Σ e0 ζξv 1 Σ εξξ ) + ν, cov(g 1, g 32 ) = N 2 ν, cov(g 1, g 33 ) = cov(g 1, g 35 ) = N 2 t(v 1 Σ e0 ζξv 1 Σ εξξ ), cov(g 1, g 37 ) = N 2 t(v 1 Σ e0 εξξ), up to tems of ode O(N 2 ) and cov(g 2, g 3l ) = o(n 2 ) fo l = 4, 6, 8, whee Σ εξ:εξ = E(ε 2 ξξ T ) and Σ εξξ = E(εξξ T ) as in Popositions 5 6, Σ e0 ξξ = E(e 0 ξξ T ) as in Poposition 2, Σ e0 ζξ = E(e 0 ζξ T ), Σ e0 εξξ = E(e 0 εξξ T ), and ν = t E(e 0 εξ T )V 1 E(ζ ξξ T )V 1 = E(e 0 εξ T )V 1 E(ζξ T V 1 ξ). The fomula fo V( α lik ) follows because Σ e0 ξξ = Σ e0 ζξ A. To show V( α eg,bc ), we 8

find that α eg,bc admits expansion (4) similaly to ˆα eg,bc, whee g 2 ( α eg,bc ) = Ẽ(εξT )V 1 Ẽ(ξ) + N 1 E(εξ T V 1 ξ), g 3 ( α eg,bc ) = Ẽ(εξT )V Ẽ(ζξT 1 ) V V 1 Ẽ(ξ) + N Ẽ(εξT 1 V 1 ξ) E(εξ T V 1 ξ) N 1 E(ξ T V 1 ξζ T )V 1 Ẽ(εξ) [ N 1 t V Ẽ(ζξT 1 ) V V 1 E(εξξ T ). Wite g 2 = g 2 ( α eg,bc ), g 3 = g 3 ( α eg,bc ), and the 4 tems of g 3 as g 31, g 32 and so on. The fomula fo V( α eg,bc ) follows by diect calculation: va(g 2 ) = N 2 t(v 1 Σ εξ:εξ ) + t(v 1 Σ εξξ V 1 Σ εξξ ), cov(g 1, g 2 ) = N 2 t(v 1 Σ e0 εξξ), cov(g 1, g 31 ) = N 2 t(v 1 Σ e0 ζξv 1 Σ eξξ ) + N 2 ν, cov(g 1, g 32 ) = N 2 t(v 1 Σ e0 εξξ), cov(g 1, g 33 ) = N 2 ν, cov(g 1, g 34 ) = N 2 t(v 1 Σ e0 ζξv 1 Σ eξξ ), up to tems of ode O(N 2 ). It is inteesting to mention that without bias coection, V( α eg ) is equal to va(g 2 ) + 2cov(g 1, g 2 ) + 2cov(g 1, g 31 ). Poof of Coollay 1. Assume that E(Y X) = β T X. Then A = 0 and Σ εξξ = 0. Theefoe, V( α eg,bc ) = V( α lik ) = N 2 t(v 1 Σ εξ:εξ ). By diect calculation, Σ εξ:εξ = E(ε 2 ξξ T ) = E (p 1 1) 2 p 1 (Y β T X) 2 XX T, Σ eξ:eξ = E(e 2 ξξ T ) = E (p 1 1) 2 (Y β T X) 2 XX T [ + E (p 1 1) (p 1 1)(Y β T X)X + (β T X)X 2, whee τ 2 = ττ T fo a column vecto τ. Then V(ˆα lik ) V( α lik ) = N 2 t[v 1 (Σ eξ:eξ Σ εξ:εξ ) Σ eξξ V 1 Σ eξξ. The matix in the culy backet equals E(β T X) 2 ξξ T E(β T X)ξξ T V 1 E(β T X)ξξ T, which is the vaiance matix of the esidual in poecting (β T X)ξ on ξ and hence nonnegative definite. Theefoe, V( α lik ) V(ˆα lik ). Poof of Poposition 7. The poof is simila to those of Popositions 1 and 4. The moment condition Ẽ( X 4 ) = O(1) is imposed fo ˆα eg and α eg to show that 9

Ẽ(ξξ T ) = V + O p (N 1/2 ) in g 3 (ˆα eg ) and Ẽ(ξζ T ) = V + O p (N 1/2 ) in g 3 ( α eg ) by Chebyshevs inequality. Similaly, the moment condition Ẽ( X 6 ) = O(1) is imposed fo ˆα lik and α lik to show that Ẽ(e ξξ T ) = E Ẽ(e ξξ T ) + O p (N 1/2 ) in g 3 (ˆα lik ) and Ẽ(ε ξξ T ) = E Ẽ(ε ξξ T ) + O p (N 1/2 ) in g 3 ( α lik ). Poof of Poposition 8. (i) The bias fomulas follow by simple calculation. To show V ( α lik ), wite g 2 = g 2 ( α lik ), g 3 = g 3 ( α lik ), and the 8 tems of g 3 as g 31, g 32 and so on. Let e = e Ȳ. By diect but tedious calculation using Lemma A1, va (g 2 ) = N 2 t(v 1 Ω ε ξ:ε ξ) t(v 1 Ω ε ξξv 1 cov (g 1, g 21 ) = N 2 t(v 1 Ω e ε ξξ) + N 2 t(v 1 B ), cov (g 1, g 22 ) = N 2 t(v 1 Ω e ξξv 1 Ω ε ξξ), cov (g 1, g 31 ) = N 2 t(v 1 Ω e ζξv 1 Ω εξξ) + N 2 ν, cov (g 1, g 32 ) = N 2 ν, cov (g 1, g 33 ) = cov (g 1, g 35 ) = N 2 t(v 1 cov (g 1, g 37 ) = N 2 t(v 1 Ω e ε ξξ), Ω e ζξv 1 Ω ε ξξ), Ω ε ξξ), up to tems of ode O(N 2 ) and cov (g 2, g 3l )=o(n 2 ) fo l = 4, 6, 8. Hee Ω ε ξ:ε ξ= Nva Ẽ(ε ξ) and Ω ε ξξ = E Ẽ(ε ξξ T ) as in Poposition 8, Ω e ζξ = E Ẽ(e ζξ T ) = Ω εξξ, Ω e ξξ = E Ẽ(e ξξ T ) = Ω εξξ A, Ω e εξξ = E Ẽ(e ε ξξ T ) = Ω εξ:εξ, and [ ν = t E Ẽ(e ε ξ T ) V 1 E Ẽ(ζ ξξt ) V 1 = E Ẽ(e ε ξ T ) V 1 E Ẽ(ζξT V 1 ξ). To show V (ˆα lik ), wite g 2 = g 2 (ˆα lik ), g 3 = g 3 (ˆα lik ), and the 8 tems of g 3 as g 31, g 32 and so on. By diect but tedious calculation using Lemma A1, va (g 2 ) = N 2 t(v 1 Ω e ξ:e ξ) t(v 1 Ω e ξξv 1 cov (g 1, g 21 ) = N 2 t(v 1 Ω e e ξξ) + N 2 t(v 1 B ), cov (g 1, g 22 ) = N 2 t(v 1 Ω e ξξv 1 Ω eξξ), cov (g 1, g 31 ) = N 2 t(v 1 Ω e ξξv 1 Ω e ξξ) + N 2 u, cov (g 1, g 32 ) = N 2 u, cov (g 1, g 33 ) = cov (g 1, g 35 ) = N 2 t(v 1 cov (g 1, g 37 ) = N 2 t(v 1 Ω e e ξξ), Ω e ξξv 1 Ω e ξξ), Ω eξξ), 10

up to tems of ode O(N 2 ) and cov (g 2, g 3l ) = o(n 2 ) fo l = 4, 6, 8. Hee Ω e ξ:e ξ = Nva Ẽ(e ξ) and Ω e ξξ = E Ẽ(e ξξ T ) as in Poposition 8, Ω e e ξξ = E Ẽ( e e ξξ T ), and [ u = t E Ẽ(e e ξ T ) V 1 E Ẽ(ξ ξξt ) V 1 = E Ẽ(e e ξ T ) V 1 E Ẽ(ξξT V 1 ξ). (ii) Expansion (4) can be poved fo ˆα eg,bc and α eg,bc similaly as in the missingdata poblem. The g 1, g 2, and g 3 tems ae defined as in the poofs of Popositions 3 and 6 except that e is eplaced thoughout by e, ε by ε, V by V, and so on. Then the second-ode vaiance can be calculated fo ˆα eg,bc and α eg,bc similaly in the missing-data poblem. The fomulas of va (g 2 ), cov (g 1, g 2 ), etc., ae simila to va(g 2 ), cov(g 1, g 2 ), etc., in the poofs of Popositions 3 and 6, but with subtle changes as in the calculation of V (ˆα lik ) and V ( α lik ). Poof of Coollay 2. All the finite population quantities such as V, β, Ω ε ξξ, and A involve means ove the finite population of polynomials of p 1 (X), Y, and X with degee of Y 2 and that of X 4. By Chow and Teiche (1997, Coollay 8.1.7), each of these finite population means conveges to the coesponding expectation unde L yx. Fo example, V conveges to V yx = E yx (p 1 1)XX T, β to β yx = V 1 yx E yx (p 1 1)XY, Ω ε ξξ to E yx (p 1 1) 2 (Y β T yxx)xx T, and A to E yx (p 1 1)(Y β T yxx)xx T, whee E yx ( ) denotes the expectation of a function of (Y, X) unde L yx. The fomulas of second-ode vaiances emain valid with the finite population means eplaced by the coesponding expectations. Assume that the conditional expectation of Y given X unde L yx, E yx (Y X), is of fom b T X fo a constant vecto b. Then E yx (Y X) = β T yxx and hence Ω εξξ = o(1) and A = o(1). Theefoe, V ( α eg,bc ) = V ( α lik ) = N 2 t(v 1 Ω εξ:εξ) + 2t(V 1 B ) + o(n 2 ). By diect calculation, Ω ε ξ:ε ξ = Nva Ẽ(ε ξ) = Ẽ (p 1 1) 3 (Y β T X) 2 XX T, [ Ω e ξ:e ξ = Nva Ẽ(e ξ) = Ẽ (p 1 1) (p 1 1)(Y β T X)X + (β T X)X 2. Then V (ˆα lik ) V ( α lik ) = N 2 t[v 1 (Ω eξ:eξ Ω εξ:εξ) Ω eξξv 1 Ω eξξ+o(n 2 ). The matix in the culy backet conveges to E yx (p 1 1)(β T yxx) 2 XX T E yx (p 1 1)(β T yxx)xx T V 1 yx E yx (p 1 1)(β T yxx)xx T, which is the expectation of the pod- 11

uct of the esidual and its tanspose in poecting (p 1 1) 1/2 (β T yxx)x on (p 1 1) 1/2 X and hence nonnegative definite. Theefoe, V ( α lik ) V (ˆα lik ). Poofs of Popositions 9 and 10. The poof of Poposition 9 is simila to that of Poposition 7. Fo Poposition 10, the bias fomulas follow by simple calculation using Lemma A2. To show V ( α lik ), wite g 2 = g 2 ( α lik ), g 3 = g 3 ( α lik ), and the 8 tems of g 3 as g 31, g 32, etc. Recall that ξ 0,i = R i p i and U 0 = Ẽp(1 p). By diect but tedious calculation using Lemma A2, va (g 2 ) = N 2 t( 1 Γ εξ 1 :ε ξ 1 ) t( 1 Γ εξ 1 ξ 1 1 Γ εξ 1 ξ 1 ), cov (g 1, g 21 ) = N 2 t( 1 Γ ε ξ 1 :ε ξ 1 ) + N 2 t( 1 D 1 ) N 2 ψt( 1 Γ ε ξ 1 ξ 1 ), cov (g 1, g 22 ) = N 2 t( 1 Γ e ξ 1 ξ 1 1 Γ ε ξ 1 ξ 1 ) + N 2 ψt( 1 Γ ε ξ 1 ξ 1 ), cov (g 1, g 31 ) = N 2 t( 1 Γ e ζ 1 ξ 1 1 Γ ε ξ 1 ξ 1 ) + t( 1 Γ e ζ 1 ξ 0 0 Γ T ε ξ 1 ξ 0 ) + N 2 ν, cov (g 1, g 32 ) = N 2 ν, cov (g 1, g 33 ) = cov (g 1, g 35 ) = N 2 t( 1 Γ e ζ 1 ξ 1 1 Γ ε ξ 1 ξ 1 ) + t( 1 Γ e ζ 1 ξ 0 0 Γ T ε ξ 1 ξ 0 ), cov (g 1, g 37 ) = N 2 t( 1 Γ e :ε ξ 1 ξ 1 ), up to tems of ode O(N 2 ) and cov (g 2, g 3l ) = o(n 2 ) fo l = 4, 6, 8. Hee Γ ε ξ 1 :ε ξ 1 = Nva Ẽ(ε ξ 1 ) = Ẽ[(p 1 1)(p 1 1)(Y β T X)Z O 2 +o(1), Γ ε ξ 1 ξ 1 = E P ois Ẽ(ε ξ 1 ξ T 1 ), and Γ ε ξ 1 ξ 0 = E P ois Ẽ(ε ξ 1 ξ T 0 ) as in Poposition 10, Γ e ζ 1 ξ 1 = E P ois Ẽ(e ζ 1 ξ T 1 ) = Γ ε ξ 1 ξ 1, Γ e ζ 1 ξ 0 = E P ois Ẽ(e ζ 1 ξ 0 ) = Γ ε ξ 1 ξ 0, Γ e ξ 1 ξ 1 = E P ois Ẽ( e ξ 1 ξ T 1 ) = Γ εξ 1 ξ 1 C 1, Γ e :ε ξ 1 ξ 1 = E P ois Ẽ(e ε ξ 1 ξ T 1 ) = Ẽ[(p 1 1) 3 (Y β T X) 2 ZZ T = Γ εξ 1 :ε ξ 1 + Q 1 + o(1), and ν = E P ois Ẽ(e ε ξ1 T ) U1 1 E P ois Ẽ(ζ1 ξ1 T U1 1 ξ 1 ) + E P ois Ẽ(e ε ξ 0 ) 0 E P ois Ẽ(Rξ T 1 1 ξ 1 ) To show V (ˆα lik ), wite g 2 = g 2 (ˆα lik ), g 3 = g 3 (ˆα lik ), and the 8 tems of g 3 as g 31, g 32,. 12

etc. By diect but tedious calculation using Lemma A2, va (g 2 ) = N 2 t( 1 Γ e ξ 1 :e ξ 1 ) t( 1 Γ e ξ 1 ξ 1 1 Γ e ξ 1 ξ 1 ), cov (g 1, g 21 ) = N 2 t( 1 Γ e :e ξ 1 :ξ 1 ) + N 2 t( 1 D 1 ) N 2 ψt( 1 Γ e ξ 1 ξ 1 ), cov (g 1, g 22 ) = N 2 t( 1 Γ e ξ 1 ξ 1 1 Γ e ξ 1 ξ 1 ) + N 2 ψt( 1 Γ e ξ 1 ξ 1 ), cov (g 1, g 31 ) = N 2 t( 1 Γ e ξ 1 ξ 1 1 Γ eξ 1 ξ 1 ) + t( 1 Γ e ξ 1 ξ 0 0 Γ T e ξ 1 ξ 0 ) + N 2 u, cov (g 1, g 32 ) = N 2 u, cov (g 1, g 33 ) = cov (g 1, g 35 ) = N 2 t( 1 Γ e ξ 1 ξ 1 1 Γ e ξ 1 ξ 1 ) + t( 1 Γ e ξ 1 ξ 0 0 Γ T e ξ 1 ξ 0 ), cov (g 1, g 37 ) = N 2 t( 1 Γ e :e ξ 1 ξ 1 ), up to tems of ode O(N 2 ) and cov (g 2, g 3l ) = o(n 2 ) fo l = 4, 6, 8. Hee Γ e ξ 1 :e ξ 1 = Nva Ẽ(e ξ 1 ) = Ẽ[(p 1 1)(p 1 1)(Y β T X)Z+(β T X)Z O 2 +o(1), Γ eξ 1 ξ 1 = E P ois Ẽ(e ξ 1 ξ T 1 ), and Γ eξ 1 ξ 0 = E P ois Ẽ(e ξ 1 ξ 0 ) as in Poposition 10, Γ e ξ 1 ξ 0 = E P ois Ẽ(e ξ 1 ξ 0 ) = Γ εξ 1 ξ 0 C 0, Γ e :e ξ 1 :ξ 1 = Γ εξ 1 :ε ξ 1 + Ẽ[(p 1 1)(p 1 2)(Y β T X)Z O (β T X)Z T O, Γ e :e ξ 1 ξ 1 = E P ois Ẽ(e e ξ 1 ξ T 1 ) = Γ e :ε ξ 1 ξ 1 +Ẽ(p 1 1)(p 1 2)(Y β T X)(β T X)ZZ T = Γ e :e ξ 1 :ξ 1 + Q 1 + K 1 + o(1), and u = E P ois Ẽ(e e ξ1 T ) U1 1 E P ois Ẽ(ξ1 ξ1 T U1 1 ξ 1 ) + E P ois Ẽ(e e ξ 0 ) 0 E P ois Ẽ(ξ0 ξ T 1 1 ξ 1 ) Finally, V (ˆα eg,bc ) and V ( α eg,bc ) can be shown similaly as in Poposition 8, but with subtle changes as in the foegoing calculation of V (ˆα lik ) and V ( α lik ). Poof of Coollay 3. The poof is simila to that of Coollay 2. Let U yx = E yx (p 1 1)ZZ T. Assume that E yx (Y X) is of fom b T X fo a constant vecto b. Then Γ ε ξ 1 ξ 1, Γ ε ξ 1 ξ 0, C 1, C 0, Q 1, K 1, and ψ ae o(1). Theefoe, V ( α eg,bc ) = V ( α lik ) = N 2 t( 1 Γ ε ξ 1 :ε ξ 1 ) + 2t( 1 D 1 ) + o(n 2 ). Moeove, V (ˆα lik ) V ( α lik ) = N 2 t[ 1 (Γ e ξ 1 :e ξ 1 Γ ε ξ 1 :ε ξ 1 ) Γ e ξ 1 ξ 1 1 Γ e ξ 1 ξ 1 + o(n 2 ). The matix in the culy backet conveges to E yx [(p 1 1)(β T yxx)z O 2 E yx (p 1 1)(β T yxx)zz T yx E yx (p 1 1)(β T yxx)zz T, which is the expectation of the poduct of the esidual and its tanspose in poecting (p 1 1) 1/2 (β T yxx)z O on (p 1 1) 1/2 Z and hence nonnegative definite. Theefoe, V ( α lik ) V (ˆα lik ).. 13

Poof of Poposition 11. The poof is simila to that of Poposition 10, by diect calculation using Lemma A3. Fo example, the tem cov (g 1, g 21 ) fo α lik,π involves the additional tem N 2 ϕ T 1 Γ εε ξ 1 by esult (iii) of Lemma A3. REFERENCES Chow, Y.-S. and Teiche, H. (1997) Pobability Theoy (3d edition), New Yok: Spinge. 14