Supporting information for: Functional Mixed Effects Model for Small Area Estimation

Σχετικά έγγραφα
Πανεπιστήµιο Κρήτης - Τµήµα Επιστήµης Υπολογιστών. ΗΥ-570: Στατιστική Επεξεργασία Σήµατος. ιδάσκων : Α. Μουχτάρης. εύτερη Σειρά Ασκήσεων.

Multi-dimensional Central Limit Theorem

One and two particle density matrices for single determinant HF wavefunctions. (1) = φ 2. )β(1) ( ) ) + β(1)β * β. (1)ρ RHF

Multi-dimensional Central Limit Theorem

α & β spatial orbitals in

Ordinal Arithmetic: Addition, Multiplication, Exponentiation and Limit

Variance of Trait in an Inbred Population. Variance of Trait in an Inbred Population

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

ST5224: Advanced Statistical Theory II

1 Complete Set of Grassmann States

2 Composition. Invertible Mappings

Finite Field Problems: Solutions

8.324 Relativistic Quantum Field Theory II

Statistical Inference I Locally most powerful tests

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

Uniform Convergence of Fourier Series Michael Taylor

C.S. 430 Assignment 6, Sample Solutions

ORDINAL ARITHMETIC JULIAN J. SCHLÖDER

EE512: Error Control Coding

Other Test Constructions: Likelihood Ratio & Bayes Tests

A Note on Intuitionistic Fuzzy. Equivalence Relation

Example Sheet 3 Solutions

CHAPTER 25 SOLVING EQUATIONS BY ITERATIVE METHODS

Supplementary materials for Statistical Estimation and Testing via the Sorted l 1 Norm

Symplecticity of the Störmer-Verlet algorithm for coupling between the shallow water equations and horizontal vehicle motion

Neutralino contributions to Dark Matter, LHC and future Linear Collider searches

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)

Problem Set 3: Solutions

LECTURE 4 : ARMA PROCESSES

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

SOLUTIONS TO MATH38181 EXTREME VALUES AND FINANCIAL RISK EXAM

6.1. Dirac Equation. Hamiltonian. Dirac Eq.

Nowhere-zero flows Let be a digraph, Abelian group. A Γ-circulation in is a mapping : such that, where, and : tail in X, head in

Generalized Fibonacci-Like Polynomial and its. Determinantal Identities

THE SECOND WEIGHTED MOMENT OF ζ. S. Bettin & J.B. Conrey

Section 8.3 Trigonometric Equations

SOLUTIONS TO MATH38181 EXTREME VALUES AND FINANCIAL RISK EXAM

Tridiagonal matrices. Gérard MEURANT. October, 2008

derivation of the Laplacian from rectangular to spherical coordinates

Chapter 6: Systems of Linear Differential. be continuous functions on the interval

Homework 3 Solutions

SCHOOL OF MATHEMATICAL SCIENCES G11LMA Linear Mathematics Examination Solutions

Econ 2110: Fall 2008 Suggested Solutions to Problem Set 8 questions or comments to Dan Fetter 1

Matrices and Determinants

Reminders: linear functions

SCITECH Volume 13, Issue 2 RESEARCH ORGANISATION Published online: March 29, 2018

Math221: HW# 1 solutions

Fourier Series. MATH 211, Calculus II. J. Robert Buchanan. Spring Department of Mathematics

8.1 The Nature of Heteroskedasticity 8.2 Using the Least Squares Estimator 8.3 The Generalized Least Squares Estimator 8.

Derivation for Input of Factor Graph Representation

Concrete Mathematics Exercises from 30 September 2016

Lecture 34 Bootstrap confidence intervals

2 Lagrangian and Green functions in d dimensions

A Class of Orthohomological Triangles

Fractional Colorings and Zykov Products of graphs

Lecture 15 - Root System Axiomatics

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

Areas and Lengths in Polar Coordinates

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

Exercises 10. Find a fundamental matrix of the given system of equations. Also find the fundamental matrix Φ(t) satisfying Φ(0) = I. 1.

6.3 Forecasting ARMA processes

Math 6 SL Probability Distributions Practice Test Mark Scheme

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

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

Partial Differential Equations in Biology The boundary element method. March 26, 2013

4.6 Autoregressive Moving Average Model ARMA(1,1)

Lecture 21: Properties and robustness of LSE

Areas and Lengths in Polar Coordinates

Appendix. Appendix I. Details used in M-step of Section 4. and expect ultimately it will close to zero. αi =α (r 1) [δq(α i ; α (r 1)

Inverse trigonometric functions & General Solution of Trigonometric Equations

Commutative Monoids in Intuitionistic Fuzzy Sets

Mean-Variance Analysis

The challenges of non-stable predicates

Solutions to Exercise Sheet 5

Lecture 13 - Root Space Decomposition II

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

SOME PROPERTIES OF FUZZY REAL NUMBERS

ΚΥΠΡΙΑΚΗ ΕΤΑΙΡΕΙΑ ΠΛΗΡΟΦΟΡΙΚΗΣ CYPRUS COMPUTER SOCIETY ΠΑΓΚΥΠΡΙΟΣ ΜΑΘΗΤΙΚΟΣ ΔΙΑΓΩΝΙΣΜΟΣ ΠΛΗΡΟΦΟΡΙΚΗΣ 6/5/2006

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

Solutions for Mathematical Physics 1 (Dated: April 19, 2015)

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

Duals of the QCQP and SDP Sparse SVM. Antoni B. Chan, Nuno Vasconcelos, and Gert R. G. Lanckriet

1. Introduction and Preliminaries.

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

DESIGN OF MACHINERY SOLUTION MANUAL h in h 4 0.

Congruence Classes of Invertible Matrices of Order 3 over F 2

Figure A.2: MPC and MPCP Age Profiles (estimating ρ, ρ = 2, φ = 0.03)..

Chapter 6: Systems of Linear Differential. be continuous functions on the interval

The one-dimensional periodic Schrödinger equation

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

CRASH COURSE IN PRECALCULUS

Estimators when the Correlation Coefficient. is Negative

A General Note on δ-quasi Monotone and Increasing Sequence

Notes on the Open Economy

Lecture 2. Soundness and completeness of propositional logic

SUPPLEMENT TO SEQUENTIAL ESTIMATION OF STRUCTURAL MODELS WITH A FIXED POINT CONSTRAINT (Econometrica, Vol. 80, No. 5, September 2012, )

1 String with massive end-points

12. Radon-Nikodym Theorem

Section 9.2 Polar Equations and Graphs

Local Approximation with Kernels

Transcript:

Supportng nformaton for: Functonal Mxed Effects Model for Small Area Estmaton Tapabrata Mat 1, Samran Snha 2 and Png-Shou Zhong 1 1 Department of Statstcs & Probablty, Mchgan State Unversty, East Lansng, MI 48824 mat@stt.msu.edu, pszhong@stt.msu.edu and 2 Department of Statstcs, Texas A&M Unversty, College Staton, TX 77843 snha@stat.tamu.edu Ths document contans some lemmas and ther proof that are ey n provng Theorems 1 and 2 stated n the man paper. Frst, we re-state some notatons that we have already ntroduced n the man paper. Model (3.3) n the paper s Y = Z F b F + W ν + Υ, where W = (Z R, M 0,, M p ) and ν = (b T R, U T 0, U T 1,, U T p ) T. Denote G = var(ν) = dag(cov(b R ), I n Σ u0,, I n Σ up ) and cov(y ) = Σ. Also, Σ = Σ (δ) = Z F cov(b R )Z T F + cov(u 0 ) + p =1 Dag(X ) cov(u )Dag(X ) + Ω for = 1,..., n. The estmator of b F s bf = (Z T F Σ 1 Z F ) 1 Z T F Σ 1 Y and the predctor of ν s ν = GW T (Y Z F bf ). The covarance matrx Σ(δ) s nvolved wth parameters δ = (σ 2 b 0,, σ 2 b p, ψ 0,, ψ p, ρ 0,, ρ p ) T. Let s = (p + 1)(q + 2) be the number of parameters n δ. These parameters are estmated through restrcted maxmum lelhood method by maxmzng l(δ) = 1 2 log ZT F Z F 1 2 log Σ 1 2 (Y Z F b F ) T (Y Z F bf ). =1 Then there exst some T such that T T Z F = 0 and ran(t ) = mn L 1 (p + 1), and defne P = T (T T ΣT ) 1 T T = Z F (Z T F Z F ) 1 Z T F such that the lelhood can be wrtten as (S1) l(δ) = 1 2 log T T ΣT 1 2 Y T P Y. (S2) and defne the REML estmator δ as the soluton to the score equaton l(δ)/ δ = 0. Now we shall state the lemmas. Lemma 1 s needed for Lemma 2 whch s needed for Theorem 2. Lemmas 3 and 4 are needed for provng Theorems 1 and 2. Lemma 1 Let Σ = Σ (δ ) and Σ = Σ (δ). If max, X <, then tr{(σ Σ ) 2 } C δ δ δ 2, where C δ s a constant and δ δ 2 = (δ δ) T (δ δ). 1

Proof: We wrte Σ Σ = J 1 + J 2 + J 3 + J 4 + J 5, where Therefore, J 1 = Z F Dag{(σ b 0 2 σ 2 b 0 )Dag(λ 1 L 1 ),, (σ b p 2 σ 2 b p )Dag(λ 1 L 1 )}Z T F, J 2 = (ψ 0 ψ 0 )A m (ρ 0 ), J 3 = ψ0a m (ρ 0) ψ0a m (ρ 0 ), p J 4 = Dag(X )[ψ{a m (ρ ) A m (ρ )}]Dag(X ), J 5 = =1 p Dag(X ){(ψ ψ )A m (ρ )}Dag(X ). =1 tr{(σ Σ ) T (Σ Σ )} C[tr(J T 1 J 1 ) + tr(j T 2 J 2 ) + tr(j T 3 J 3 ) + tr(j T 4 J 4 ) + tr(j T 5 J 5 )]. We see that tr(j T 1 J 1 ) C p m (σb 2 σb 2 ) 2 =1 {Z ()T F j,l=1 (t j )Dag(λ 2 L 1 )Z () F (t l )} 2. Also, notce that tr(j T 2 J 2 ) = tr{(ψ 0 ψ 0 )A m (ρ 0 )} 2 C(ψ 0 ψ 0 ) 2 tr{a m (ρ 0 )} 2 and tr[{ψ 0A m (ρ 0) ψ 0A m (ρ 0 )} 2 ] = tr { [(ψ 0 ψ 0 ){A m (ρ 0) A m (ρ 0 )} + ψ 0 {A m (ρ 0) A m (ρ 0 )}] 2} Hence, tr(j T 3 J 3 ) C δ δ 2. Next, tr(j T 4 J 4 ) C C 2{(ψ 0 ψ 0 ) 2 + ψ 2 0}tr{A m (ρ 0) A m (ρ 0 )} 2 C{(ψ 0 ψ 0 ) 2 + ψ 2 0} ρ 0 ρ 0 2. p tr { Dag(X )[ψ{a m (ρ ) A m (ρ )}]Dag(X ) } 2 =1 p =1 [ tr { Dag(X )[(ψ ψ ){A m (ρ ) A m (ρ )}]Dag(X ) } 2 +tr { Dag(X )[ψ {A m (ρ ) A m (ρ )}]Dag(X ) } ] 2 p C [(ψ ψ ) 2 tr{a m (ρ ) A m (ρ )} 2 tr{dag(x )} 4 =1 +ψ 2 tr{a m (ρ ) A m (ρ )} 2 tr{dag(x )} 4 ]. If A m (ρ ) has bounded second dervatves wth respect to ρ, then tr{a m (ρ ) A m(ρ )} 2 C ρ ρ. And tr(j T 5 J 5 ) C p =1 (ψ ψ ) 2 tr { Dag(X )A m (ρ )Dag(X ) } 2. In summary, tr{(σ Σ ) 2 } C δ δ δ 2. 2

Lemma 2 Let C δ be the constant defned n Lemma 1 and assume that the smallest egenvalue of Σ s bounded below by c 0 > 0. Suppose that δ δ 2 such that 2C δ /c 2 0 < 1 when n s large enough. Then tr{(σ 1 ) 2 } [ ] (1 2c 4 0 Cδ 2 2 ) 1 2tr{( ) 2 } + 2C δ /c 4 0. Proof: By the matrx nverse formula, Σ 1 Σ ), we then have tr{(σ 1 ) 2 } 2tr{( +2tr[{ = Because (Σ Σ )Σ 2 (Σ Σ ) s non-negatve defnte, tr[{ (Σ Σ )Σ 1 (Σ Σ ) (Σ Σ ) + ) 2 } + 2tr[{ (Σ Σ ) } 2 ] (Σ Σ )Σ 1 (Σ (Σ Σ )Σ 1 (Σ Σ ) } 2 ]. (S3) } 2 ] tr 2 { (Σ Σ )Σ 1 (Σ Σ ) } tr{(σ 1 ) 2 }tr[{(σ Σ )Σ 2 (Σ Σ )} 2 ] tr 2 {(Σ Σ )Σ 2 (Σ Σ )} tr{(σ 1 ) 2 }λ 4 mn (Σ )tr 2 {(Σ Σ ) 2 } tr{(σ 1 ) 2 }c 4 0 Cδ 2 2, where the last nequalty follows from Lemma 1 and the assumpton n ths Lemma. In addton, tr{( (Σ Σ ) ) 2 } λ 4 mn (Σ )tr{(σ Σ ) 2 }. Hence, from (S3), tr{(σ 1 ) 2 } (1 2c 4 0 Cδ 2 2 ) 1 (2tr{( ) 2 } + 2C δ /c 4 0). Ths completes the proof of Lemma 2. Lemma 3 Let d 2 = max j, { ( tr(p V P V ), tr P V P V ) (, tr P 2 V P 2 V )} δ δ and d = mn d. Then there exsts δ such that for any 0 < q 0 < 1 and large n, δ δ = A 1 a + o p (d 2q 0 ), where a = l(δ)/ δ and A = E{ 2 l(δ)/ δ 2 }, on the set B wth P (B) convergng to 1. Proof: We wll apply Theorem 2.1 of Das et al. (2004). Let us frst verfy the followng condtons. 3

The gth moment of the followng quanttes are bounded for some d and d = mn d, 1 l(δ) 1 2 l(δ), E( δ δ0 d d j δ j δ0 2 l(δ) ) δ j δ0, d M j, d d j d d where M j = sup δ Sδ (δ 0 ) 3 l(δ)/( δ ) wth S γ (δ 0 ) = {δ : δ δ 0 γd /d 1 s}. Usng the lelhood gven n (S2), we obtan the frst dervatve of l(δ) wth respect to δ l(δ) = 1 2 {ɛt P V P ɛ tr(p V )}, (S4) where ɛ = Y Z F b F and V = dag(v 1,, V n ). Let ɛ = Σ 1/2 u and u N(0, I mn ). Then for any g 2, E l(δ) g = 2 g E u T Σ 1/2 P V P Σ 1/2 u E(u T Σ 1/2 P V P Σ 1/2 u) c Σ 1/2 P V P Σ 1/2 g 2 = ctr(v P V P ) g/2. Thus, f we tae d = tr(v P V P ) 1/2, the gth moment of (1/d ) l(δ 0 )/ δ s bounded. Because P / = T (T T ΣT ) 1 T T ( Σ/ )T (T T ΣT ) 1 T T = P ( Σ/ )P = P V j P, we have 2 l(δ) = 1 { ɛ T Q j ɛ + tr(p V j P V ) tr(p V } ), (S5) 2 where Q j = P {V j P V + V P V j ( V / )}P := P K j P. Then we have where E 2 l(δ) E( 2 l(δ) ) g = 2 g E u T Σ 1/2 Q j Σ 1/2 u E(u T Σ 1/2 Q j Σ 1/2 u) c Σ 1/2 Q j Σ 1/2 g 2 = ctr(k j P K j P ) g/2, tr(k j P K j P ) = tr((v j P V + V P V j V )P (V j P V + V P V j V )P ) = 2tr(V j P V P V j P V P ) + 2tr(V P V P V j P V j P ) 2tr(V j P V P V P ) 2tr(V P V j P V P ) + tr( V P V P ) and applyng Lemma 5.2 of Das et al. (2004), we have tr(v j P V P V P ) tr(v P V P V j P V j P ) 1/2 tr( V P V P ) 1/2 ; tr(v P V j P V P ) tr(v P V P V j P V j P ) 1/2 tr( V P V P ) 1/2 ; tr(v j P V P V j P V P ) tr(v P V P V j P V j P ). g g 4

Therefore, tr(k j P K j P ) { 2tr(V P V P V j P V j P ) 1/2 + tr( V P V P ) 1/2} 2. Notce that tr(a 2 ) tr 2 (A) for any non-negatve matrx A. Snce P 1/2 V P V P 1/2 s a nonnegatve defnte matrx, we have tr(v P V P V j P V j P ) tr(v P V P V P V P ) 1/2 tr(v j P V j P V j P V j P ) 1/2 tr(p 1/2 V P V P 1/2 )tr(p 1/2 V j P V j P 1/2 ) = tr(v P V P )tr(v j P V j P ). Hence f we tae d = max j [tr(v P V P ) 1/2, tr{ V / )P ( V / )P } 1/2 ] the gth moment of 1 2 l(δ) E( d d j δ j δ0 2 l(δ) ) δ j δ0 s bounded for any g 2. Next, we compute the thrd dervatves, 3 l(δ) δ { = 2 1 ɛ T P V P (V j P V + V P V j V )P + P ( V j δ P V V j P V P V V j P V δ + V δ P V j V P V P V j + V P V j P (V j P V + V P V j + V )P V P } ɛ E(ɛ T R j ɛ) δ = ɛ T (P V P V j P V P + P V P V P V j P + P V P V P V j P )ɛ + 2 1 ɛ T P 2 V δ P ɛ 2 V δ )P ɛ T (P V P V P + P V P V j P + P V j P V P )ɛ + E{ɛ T P (R j + Rj 2 1 )P ɛ}, δ δ δ where R j = V P V j P V + V P V P V j + V P V P V j and R j = V P ( V / ) + V P ( V j / δ ) + V j P ( V / δ ). Consder the frst term n the thrd dervatves. Denote Σ for Σ evaluated at δ and smlarly for Ṽj. Then t can be shown that (T T ΣT ) 1 = (T T ΣT ) 1 + (T T ΣT ) 1 T T (Σ Σ)T (T T ΣT ) 1 and T T Ṽ j T = T T V j T + T T (Ṽj V j )T. For convenence, denote H = (T T ΣT ) 1 and G = T T V T. Further 1 = HT T (Σ Σ)T H, 2j = T T (Ṽj V j )T. It can be seen that 2 V H = H + (δ δ )HG H + ψ0 HT T ( A m (ρ0) I n )T H + 2(p+1) =1 5 p =1 ψ HT T D{ A m (ρ )}T H,

where D( A m ) = dag{dag(x 1 ) A m dag(x 1 ),, dag(x n ) A m dag(x n )}, A m (ρ ) = A m (ρ ) A m ( ρ ).For 1 j (p + 1), Ṽ j = V j ; f j = p + 2, V j Ṽj = A m (ρ 0 ) I n ; f p + 3 j 2(p + 2), V j Ṽj = D{ A m (ρ )}; f 2(p + 1) + 1 j 2(p + 1) + q, V j Ṽj = ( A m (ρ 0 )/ ρ 0,j T ) I n wth j = j 2(p + 1) and f 2(p + 1) + ( 1)q + 1 j 2(p + 1) + q, V j Ṽj = D{ ( A m (ρ )/ ρ,j )} wth 2 (p + 1) and j = j 2(p + 1) + ( 1)q. Snce H, A m (ρ ) and A T m(ρ ) are postve defnte, f γ n S γ s small enough such that, (1/2)H H 2H, (1/2)A m (ρ ) A m ( ρ ) 2A m (ρ ) and (1/2)A T m(ρ ) A T m( ρ ) 2A T m(ρ ). Then f p + 1 H 1/2 G HT T ɛ = H 1/2 G HT T ɛ 2 H 1/2 G HT T ɛ 2 H 1/2 G HT T ɛ + 2(p+1) δ j δ j H 1/2 G HG j HT T ɛ + ψ 0 H 1/2 G HT T ( A m (ρ 0 ) I n )T HT T ɛ p + ψ j H 1/2 G HT T D( A m (ρ j ))T HT T ɛ. It can be shown that there exsts some constant C(γ) such that =1 H 1/2 G HT T ( A m (ρ 0 ) I n )T HT T ɛ C(γ) H 1/2 G H G 2p+3 HT T ɛ C(γ) H 1/2 G H 1/2 H 1/2 G2p+3 HT T ɛ, and H 1/2 G HT T D( A m (ρ j ))T HT T ɛ C(γ) H 1/2 G H 1/2 H 1/2 G HT T ɛ for = j+2(p+1)+1. Therefore, for p + 1, H 1/2 G HT T ɛ 2 H 1/2 G HT T ɛ + 2(p+1) 2 δ j δ j H 1/2 G H 1/2 H 1/2 G j HT T ɛ + 2 3(p+1) =2(p+1)+1 C(γ) ψ 2p 3 H 1/2 G H 1/2 H 1/2 G HT T ɛ 2(p+1) 2 H 1/2 G HT T ɛ + 2γd H 1/2 G H 1/2 d 1 j H 1/2 Gj HT T ɛ +2C(γ) H 1/2 G H 1/2 3(p+1) =2(p+1)+1 ψ 2p 3 H 1/2 G HT T ɛ, 6

where H 1/2 G H 1/2 = tr(p V P V ) 1/2. For (q + 2)(p + 1) > p + 1, H 1/2 G HT T ɛ 2{1 + C (γ)} H 1/2 G HT T ɛ In summary, defne then It follows that j = 2(p+1) +2γ{1 + C (γ)}d H 1/2 G H 1/2 d 1 j H 1/2 Gj HT T ɛ +2{1 + C (γ)}c(γ) H 1/2 G H 1/2 3(p+1) =2(p+1)+1 { 2γ{1 + C (γ)}d 1 j d H 1/2 G H 1/2 for j p + 1 2C{1 + C (γ)}c(γ) H 1/2 G H 1/2 for p + 1 j s, H 1/2 G HT T ɛ 2(1 + C (γ)) H 1/2 G HT T ɛ + sup H 1/2 G HT T ɛ 2{1 + C (γ)} H 1/2 G HT T ɛ + δ S γ If we tae γ smaller enough such that s j < 1, then ψ 2p 3 H 1/2 G HT T ɛ. s j H 1/2 Gj HT T ɛ. s sup δ S γ H 1/2 G HT T ɛ 2{1 + C (γ)} H 1/2 G HT T ɛ For any g > 4 and some constant C, + 2{1 + C (γ)}(1 s j ) 1 j sup δ S γ H 1/2 Gj HT T ɛ. s j H 1/2 G j HT T ɛ. E H 1/2 G j HT T ɛ g = E ɛ T T HG j HG j HT T ɛ g/2 = E ɛ T P V j P V j P ɛ g/2 Ctr g/2 (P V j P V j ). Hence the frst term n 3 l(δ)/( δ ) can be bounded by (S6) (S7) (S8) ɛ T P Ṽ P Ṽ j P Ṽ P ɛ = ɛ T T H G H Gj H G HT T ɛ λ max ( H 1/2 Gj H1/2 ) H 1/2 G HT T ɛ H 1/2 G HT T ɛ C 1 (γ)λ max (H 1/2 G j H 1/2 ) H 1/2 G HT T ɛ H 1/2 G HT T ɛ. Combnng (S8) and the above two nequalty, t can be seen that ( d ) g E sup ɛ T P Ṽ P Ṽ j P Ṽ P ɛ d d j d δ S γ C g 1(γ)λ g max(h 1/2 G j H 1/2 )E( 1 d d sup δ S γ H 1/2 G HT T ɛ sup δ S γ H 1/2 G HT T ɛ ) g C g 1(γ)λ g max(h 1/2 G j H 1/2 ). 7

We choose γ small enough such that C 1 (γ)λ max (H 1/2 G j H 1/2 ) <. The other terms n 3 l(δ)/( δ ) can be bounded smlarly. For example, ɛ T P Ṽ P Ṽ P ɛ H1/2 G HT T ɛ H 1/2 G HT T ɛ, where the bound for the rght hand sde can be obtaned smlarly as of (S8). condton (v) n Theorem 2.1 of Das et al. (2004) holds. Notce that from (S5), (A) j = E( 2 l(δ) { ) = 2 1 tr(q j Σ) tr(p V j P V ) tr(p V } ) = 2 1 tr(p V P V j ). Therefore, Then the (, j)th component of D1 1 AD1 1, where D 1 = Dag(d 1,, d n ), s tr(p V P V j )/(d d j ). Condton () n Das et al. (2004) s equvalent to requre that the smallest egenvalue of (D 1 AD 1 ) must be bounded away from 0 and. (D 1 AD 1 ) s λ mn. Snce λ x T ( D 1 )( A)( D 1 )x mn = nf x 0 x T x we requre that x T D 2 x λ max ( A) nf x 0 x T x Suppose the smallest egenvalue of λ max ( A)(mn (d )) 2 <, λ max ( A) = O(mn d 2 ). (S9) Under condton (S9), condton () of Das et al. (2004) holds. Therefore, condtons ()-(v) n Theorem 2.1 of Das et al. (2004) hold and g can be any nteger greater than 4. Ths fnshes the proof of Lemma 3. Lemma 4 Defne t(δ) = l T bf + m T ν as the BLUP estmator of Y 0 (t m ; δ) for some specfc and δ be the REML estmator of δ. If condtons (a)-(d) hold, then { t(δ) } 2 E{t( δ) t(δ)} 2 = E δ ( δ δ) + o(n 1 ). Proof: For convenence, let us defne ũ := (ũ T 1,, ũ T n) T = Y Z F bf, u = Y Z F b F and ζ T (δ) := m T GW T = (ζ1 T (δ),, ζn T (δ)) where ζ T (δ) = Z T R 0 (t m )cov(b R )Z T R f Z T R 0 (t m )cov(b R )Z T R 0 +Σ (m) u0 + p q=1 X (t m )[Σ uq Dag(X q0 )] (m) f =. where s the area we are nterested n predctng (n the man text, we used nstead of. In ths supplemental, we used ), Σ (m) u0 s the mth row of Σ u0 and [Σ uq Dag(X q0 )] (m) s the 8

mth row of Σ uq Dag(X q0 ). Let C 1 and C 2 be constants whch may tae dfferent values n each appearance. By the Taylor expanson of t( δ) around δ, we have where δ δ δ δ. Then t( δ) t(δ) = t(δ) δ ( δ δ) + 1 2 ( δ δ) T 2 t(δ ) δ 2 ( δ δ) E{t( δ) t(δ)} 2 = E{ t(δ) { t(δ) δ ( δ δ)} 2 + E δ ( δ δ)( δ δ) T 2 t(δ ) } ( δ δ) δ { 2 +(1/4)E ( δ δ) T 2 t(δ ) } 2 { t(δ) } 2 ( δ δ) := E δ 2 δ ( δ δ) + R1 + R 2. Frst, we would le to show R 2 = o(n 1 ). Notce that { E ( δ δ) T 2 t(δ ) } 2 { ( δ δ) = E[ tr 2 2 t(δ ) }] ( δ δ)( δ δ) T δ 2 δ ( { 2 E tr ( 2 t(δ ) } ]) ) 2 tr [{( δ δ)( δ δ) T } 2 δ ( { 2 = E tr ( 2 t(δ ) }{ } 2 ) ) 2 ( δ δ) T ( δ δ) δ 2 = s =1 s [ E Because s s a fxed number, we only need to show that [ E The frst dervatve of t(δ) s ( 2 t(δ ) δ δ j t(δ) = l T b F ( 2 t(δ ) δ δ j } 2 ] ) {( δ 2 δ) T ( δ δ). } 2 ] ) {( δ 2 δ) T ( δ δ) = o(n 1 ). (S10) + bt (δ) ũ b T b F (δ)z F, where b F / = (Z T F Σ 1 Z F ) 1 Z T F Σ 1 ( Σ/ ) ũ and the second dervatves of t(δ) s 2 t(δ) T 2 bf = l + 2 ζ T (δ) ũ ζt (δ) b F Z F := J 1 (δ) + J 2 (δ) + J 3 (δ) + J 4 (δ) + J 5 (δ), ζt (δ) b F Z F ζ T 2 bf (δ)z F 9

where 2 bf = (ZF T Z F ) 1 Z T 1 Σ F Σ Z F (ZF T Z F ) 1 Z T 1 Σ F Σ Z F ũ (ZF T Z F ) 1 ZF T 1 Σ Σ Z F (ZF T Z F ) 1 Z T 1 Σ F Σ Z F ũ +(ZF T Z F ) 1 ZF T 1 Σ Σ Σ ũ +(ZF T Z F ) 1 ZF T 1 Σ Σ Σ ũ 2 Σ (ZF T Z F ) 1 ZF T ũ := I 1 (δ) + I 2 (δ) + I 3 (δ) + I 4 (δ) + I 5 (δ). Let us loo at J 1 (δ). We can wrte J 1 (δ) = l T {I 1 (δ) + I 2 (δ) + I 3 (δ) + I 4 (δ) + I 5 (δ)}. Snce lt I 1 (δ ) and l T I 2 (δ ) are smlar, we only show that l T I 1 (δ ) s bounded by l T I 1 (δ ) C 1 n 1/2 u T u 1/2 (1 + C 2 δ δ ), where C 1 and C 2 are some constants. By the Cauchy- Schwarz nequalty and /2 Z T F (ZT F Σ 1 Z F ) 1 Z F /2 beng an dempotent matrx, we have l T I 1 (δ) = l T (ZF T Z F ) 1 ZF T 1 Σ Σ Z F (ZF T Z F ) 1 Z T 1 Σ F Σ Z F ũ l T (ZF T Z F ) 1 ZF T 1 Σ Σ Σ Z F (ZF T Z F ) 1 l 1/2 ũ T 1 Σ Σ Σ ũ 1/2. Denote d T = l T (Z T F Σ 1 Z F ) 1 Z T F /2 and P 1 (δ) = l T (ZF T Z F ) 1 ZF T 1 Σ Σ Σ Z F (ZF T Z F ) 1 l. Then we can wrte P 1 (δ) = = = d T Σ Σ δ d j =1 [tr(d d T d d T )] 1/2 [tr(/2 Σ =1 (d T d )[tr(/2 Σ =1 (d T d )tr 2 (/2 Σ /2 ) =1 =1 (d T d )tr 2 ( Σ ). Σ /2 )] 1/2 Σ /2 )] 1/2 10

Smlar to the proof of Lemma 2, we have tr 2 {Σ 1 ( Σ / δ j )} C 1 tr 2 { ( Σ / )}(1 + C 2 δ δ ) and assumng that max 1n tr 2 { ( Σ / )} <, we have P 1 (δ ) C 1 max 1n tr2 ( Σ )(1 + C 2 δ δ ) = C 1 max 1n tr2 ( Σ )(1 + C 2 δ δ ) d T d = C 1 max 1n tr2 ( Σ )(1 + C 2 δ δ ) l T (ZF T Σ 1 Z F ) 1 l. =1 lt (ZF T Σ 1 Z F ) 1 ZF T Σ 1 Z F (ZF T Σ 1 Z F ) 1 l It can be shown that l T (Z T F Σ 1 Z F ) 1 l = C1 lt (Z T F Σ 1 Z F ) 1 l(1 + C2 δ δ ). Snce ZF T Σ 1 Z F = O(n 1 ) and max 1n tr 2 { ( Σ / } <, P 1 (δ ) C 1 n 1 (1 + C 2 δ δ ). Next, ũ T 1 Σ Σ Σ ũ = Hence ũ T Σ 1 Σ δ =1 =1 [tr(/2 ũ ũ T ũũ T /2 )] 1/2 [tr( Σ Σ ) 2 ] 1/2 ũ T 1 Σ ũtr(σ Σ ) ũ T ũtr 2 ( Σ ). =1 =1 Σ 1 Σ Σ 1 ũ C δ 1 max 1n tr2 ( Σ )(1 + C 2 δ δ ) ũ T Σ 1 ũ C 1 u T u (1 + C 2 δ δ ). (S11) Note that we used the fact that ũ T ũ u T u. It follows that l T I 1 C 1 n 1/2 u T u 1/2 (1+ C 2 δ δ ). Smlarly, l T I 2 C 1 n 1/2 u T u 1/2 (1 + C 2 δ δ ). The thrd term n J 1 s l T I 3 = l T (ZF T Z F ) 1 ZF T 1 Σ Σ Σ ũ l T (ZF T Z F ) 1 l 1/2 ũ T 1 Σ Σ Σ Σ Σ ũ 1/2. Applyng the nequalty tr(a 2 ) tr 2 (A) for any nonnegatve matrx A, we have ũ T 1 Σ Σ Σ Σ Σ ũ = ũ T Σ Σ Σ Σ =1 [tr(/2 ũ ũ T ũũ T /2 )] 1/2 [tr( Σ Σ Σ Σ δ )2 ] 1/2 =1 ũ T ũtr 2 ( )2 tr( Σ ) 2 tr( Σ ) 2. =1 11 ũ

Hence, ũ T Σ 1 Σ δ C 1 max C 1 max tr 2 ( tr 2 ( Σ 1 Σ δ j Σ 1 Σ δ j )2 max tr( Σ )2 max tr( Σ ) 2 max ) 2 max Σ 1 Σ Σ 1 ũ δ =1 ũ T Σ 1 tr( Σ ) 2 (1 + C 2 δ δ ) tr( Σ ) 2 (1 + C 2 δ δ ) ũ tr 2 (Σ 1 ) 2 tr( Σ =1 =1 ũ T Σ 1 ũ u T u, δ j ) 2 tr( Σ ) 2 δ and t s easy to see that l T (Z T F Σ 1 Z F ) 1 l C1 l T (Z T F Σ 1 Z F ) 1 l (1 + C2 δ δ ). Therefore, l T I 3 (δ ) C 1 n 1/2 u T u 1/2 (1 + C 2 δ δ ). Smlarly, we can bound l T I 3 (δ ) and l T I 5 (δ ). So, n summary, J 1 (δ ) C 1 n 1/2 u T u 1/2 (1 + C 2 δ δ ). Next, for J 3, we have J 3 = ζt (δ) b F Z F Notng that for = ζt (δ) ζt (δ) Z F (ZF T Z F ) 1 ZF T ζ T (δ) = ZR T 0 (t m )cov(b R )ZR T Σ j Z F (ZF T Z F ) 1 ZF T 1 Σ Σ ũ ζ(δ) 1/2 ũ T 1 Σ Σ Σ ũ 1/2. + ZR T 0 (t m ) cov(b R) Z T R = O(n 1/2 ), where cov(b R )/ = 0 f δ j σb 2 and cov(b R )/ = Dag(0,, Dag(λ 1 L 1 ),..., 0) f δ j = σb 2, and ζ T 0 (δ)/ = O(1) for all δ j. Hence, { ζ T (δ)/ }Z F s of order O(1) for each component. It follows that { ζ T (δ)/ }Z F (Z T F Σ 1 Z F ) 1 Z T F { ζ(δ)/} = O(n 1 ). We have already shown n (S11) that ũ T Σ 1 Σ δ Σ 1 Σ Σ 1 ũ C δ 1 u T u (1 + C 2 δ δ ). Therefore, J 3 (δ ) C 1 n 1/2 u T u 1/2 (1 + C 2 δ δ ). Smlarly, we can show the same bound for J 4 (δ ). Now let us chec J 5, the proof s almost the same as J 1, where we replace l by ζ T (δ)z F. Notce that each component of ζ T (δ)z F s O(1). Then ζ T (δ)z F (Z T F Σ 1 Z F ) 1 Z T F ζ(δ) = O(n 1 ). Hence, as we have shown for J 1, t can also be shown that J 5 (δ ) C 1 n 1/2 u T u 1/2 (1 + C 2 δ δ ). It remans to show J 2 = O p (1). Notce that for, ζ T (δ) = ZR T 0 (t m ) cov(b R) j ZR T 1 Σ 1 + ZR T 0 (t m )cov(b R )ZR T Σ + Z T R 0 (t m )cov(b R )Z T R + ZR T 0 (t m ) cov(b R) Σ 2 Σ Σ = O(n 1/2 ), 12 ZR T 1 Σ 1 + Z T R 0 (t m )cov(b R )Z T R 1 Σ Σ

and { ζ T 0 (δ)/ } = O(1). It follows that J 2 = ζt (δ) ζ T ũ = (δ) ũ ζt (δ) ũ =1 =1 =1 ( ζt (δ) Σ ζ T (δ) ) 1/2 (ũ T ũ) 1/2, where ũ = (ũ 1,, ũ m ) T. It s easy to see that ũ T Σ 1 ũ = O p (1). Hence J 2 (δ ) C 1 n 1/2 u T u 1/2 (1 + C 2 δ δ ). In summary, from J 1 (δ ) J 5 (δ ), 2 t(δ ) δ C 1 n 1/2 (u T u) 1/2 (1 + C 2 δ δ ), δ j where C s some constant. Applyng the Cauchy-Schwarz nequalty, R 2 = E(( 2 t(δ ) 2) ) [( δ 2 δ) T ( δ δ)] 2Cn 1 {E(u T u δ δ 4 ) + E(u T u δ δ 6 )} δ δ j [ 2Cn 1 {E(u T u) 2 } 1/2 {E( δ δ 8 )} 1/2 +{E(u T u) 2 } 1/2 {E( δ δ 12 )} 1/2 ]. Because E(u T u) 2 = O(n 2 ) and E( δ δ 8 ) = O(n 4 ), we have R 2 = o(n 1 ). To show the order of R 1, we would le to now the order of E{ t(δ) s δ ( δ δ)} 2 C {E( t(δ) ) 4 } 1/2 {E( δ δ ) 4 )} 1/2. Now we can rewrte t(δ)/ n the followng form ( t(δ) = f j (δ) ζ(δ)z F (ZF T Z F ) 1 Z T Σ F =1 D + b(δ) ) D ɛ := h j (δ) T ɛ, (S12) where f j (δ) = l T (ZF T Σ 1 Z F ) 1 ZF T ( Σ/)D, D = I Z F (ZF T Σ 1 Z F ) 1 ZF T Σ 1. Defne h (2) j (δ) T = ζ T (δ)z F (ZF T Σ 1 Z F ) 1 ZF T ( Σ/)D and h (3) j (δ) = { ζ T (δ)/ }D. Snce ɛ N(0, Σ), E( t(δ) ) 4 = 3(h j (δ) T Σh j (δ)) 2 6{(f T j (δ)σf j (δ)) 2 + (h (2) j (δ) T Σh (2) j (δ)) 2 + (h (3) j (δ) T Σh (3) j (δ)) 2 }. Defne B = Z F (Z T F Σ 1 Z F ) 1 Z T F. Notce that Σ 1/2 B/2 s an dempotent matrx. Then the frst term on the rght hand sde of (S12) s fj T (δ)σf j (δ) = l T (ZF T Z F ) 1 ZF T Σ DΣD Σ Z F (ZF T Z F ) 1 l = l T (ZF T Z F ) 1 ZF T Σ (Σ B) Σ Z F (ZF T Z F ) 1 l = l T (Z T F Z F ) 1 Z T F =1 Σ Σ Σ Z F (Z T F Z F ) 1 l lt (Z T F Z F ) 1 Z T F Σ Σ Σ Z F (Z T F Z F ) 1 l λ max ( Σ Σ Σ ) lt (ZF T Z F ) 1 ZF T Z F (ZF T Z F ) 1 l, =1 13

whch s of order O(n 1 ). Smlarly, the second term of the rght hand sde of (S12) s h (2) j (δ) T Σh (2) j (δ) = ζ(δ)b Σ (Σ B) Σ Bζ T (δ) ζ T (δ)b Σ Σ Σ Bζ(δ) ζ T (δ)σ Σ Σ Σ Σζ(δ) T ζ T Σ Σ (δ)σ Σ Σ ζ (δ). =1 { } If λ 1 L 1 = O(n 1/2 (2) ) and λ max ( Σ / )Σ ( Σ / )Σ <. Then h j (δ) T Σh (2) j (δ) = O(1). Then the thrd term on the rght hand sde of (S12) s h (3) j (δ) T Σh (3) j (δ) = ζt (δ) = ζ(δ) = =1 DΣD T ζ(δ) = ζ(δ) {Σ Z F (ZF T Z F ) 1 ZF T } ζ(δ)t Σ 1/2 {I /2 Z F (ZF T Z F ) 1 ZF T /2 1/2 ζ(δ)t }Σ ζ (δ) ζ (δ) T Σ λ max (Σ ) =1 ζ (δ) ζ (δ) T. ζ(δ) Σ ζ(δ)t If λ max (Σ ) < and λ 1 L 1 = O(n 1/2 ), then ζ (δ)/ = O(n 1/2 ) and hence h (3) j (δ) T Σh (3) j (δ) = O(1). Hence, E {( t(δ)/ ) 4 } = O(1). It follows that E{( t(δ)/ δ)( δ δ)} 2 = O(n 1 ). Agan by the Cauchy-Schwarz nequalty, t s easy to see that R 1 = o(n 1 ). Therefore, we have Ths completes the proof of Lemma 4. E{t( δ) t(δ)} 2 = E{ t(δ) δ ( δ δ)} 2 + o(n 1 ). 14

Some addtonal detals n the proof of Theorem 2: For K 1, because m T G m = Z T R 0 (t m )cov(b R )Z R0 + e T m,0(i n Σ u0 )e m,0 + =1 et,x (I n Σ u )e,x, m T ( 2 G/ ) m s a summaton of fxed number functons of varance components δ. Therefore, t can be shown that K1 K 1 C δ δ. For K 2, notce that { 2 γ T (δ) 0 both δ and δ j are σ b 2 s; γ(δ) = δ 2 γ T (δ) 0 j γ (δ) f one of δ and δ j s not σb 2 s, where γ 0 (δ) s the th m-dmensonal subvector of γ T (δ) = (γ T 1 (δ),, γ T m(δ)). Therefore, K2 K 2 tr{γ 0 (δ ) 2 γ T 0 (δ ) Σ 1 } tr{γ 0 (δ) 2 γ T 0 (δ) } tr{(γ 0 (δ ) 2 γ T (δ ) γ 0 (δ) 2 γ T 0 (δ) } δ ) j + tr{γ 0 (δ) 2 γ T 0 (δ) (Σ 1 δ )} j + tr{(γ 0 (δ ) 2 γ T (δ ) γ 0 (δ) 2 γ T 0 (δ) )(Σ 1 )} := K(1) 2 + K (2) 2 + K (3) From Lemma 2 we now tr(σ 1 ) C δ δ, hence to show that K2 K 2 C δ δ, t s enough to show that ( 2 γ (l) (δ )/ )γ () (δ ) ( 2 γ (l) (δ)/ )γ () (δ) C δ δ where subscrpt (l) denotes the lth component. Notce that 2 γ (l) (δ ) γ () (δ ) 2 γ (l) (δ) (δ) C 2 γ (l) (δ ) 2 γ (l) (δ) + C 2 γ (l) (δ ) 2 γ (l) (δ) γ () + C γ () γ () 2. (δ ) γ () (δ) (δ ) γ () (δ). Clearly, γ () (δ ) γ () (δ) C δ δ from the expresson of γ () (δ) and t also easy to show that 2 γ (l) (δ )/ 2 γ (l) (δ)/ C δ δ. It follows that K2 K 2 C δ δ. The dervaton of K 3 to K 7 are smlar, here we only gve the detals for K 4. We frst wrte K 4 K 4 C C =1 =1 γt (δ ) Σ 1 Σ Σ 1 (K () 41 + K () 42 + K () 41 K () 42 ), γ (δ ) γt (δ) Σ δ γ (δ) j where K () 41 = tr[{γ (δ )( γ T (δ )/ ) γ (δ)( γ T (δ)/)} ( Σ / ) ] and K() tr[{σ 1 ( Σ / δ j )Σ 1 ( Σ / ) }γ (δ)( γ T (δ)/)]. It can be seen that K () 41 = tr{(γ (δ ) γ (δ)) γt (δ) + tr{(γ (δ ) γ (δ))( γt (δ ) Σ δ } + tr{γ (δ)( γt (δ ) j γt (δ) 15 ) Σ δ }. j γt (δ) 42 = ) Σ δ } j

For, γ (l) (δ ) γ (l) (δ) C λ L 1 1 p =0 σ2 b σb 2 and each element of γ T (δ)/ s of the same order of λ L1 1. Hence, tr{(γ (δ ) γ (δ))( γ T (δ)/) ( Σ / ) } C λ L1 2 δ δ. Smlarly, we can show the other terms are also bounded by C λ L1 2 δ δ. It follows that K () 41 C λ L1 2 δ δ f. By notng that tr(σ 1 )2 C δ δ, tr{( Σ / δ j ) ( Σ / )} 2 C δ δ, γ (δ) and γ T (δ)/ are both O( λ L1 1 ) for, t can be shown that K () 42 C λ L1 2 δ δ for. For =, K () 41 C δ δ and K () 42 C δ δ. In summary, usng the assumpton λ L1 = O(n 1/2 ), we have K 4 K 4 C( λ L1 2 + 1) δ δ C δ δ. Here we show that g 4 (δ)/ = o(n 1/2 ). Observe that g jl 4 (δ) δ = ηt j δ ΣP V j P V l P Ση l + η T j By the Cauchy-Schwarz nequalty, Σ δ P V j P V l P Ση l + η T j ΣP V P P V j P V l P Ση l +η T j ΣP V j δ P V l P Ση l + η T j ΣP V j P V P V l P Ση l + η T j ΣP V j P V l δ P Ση l +η T j ΣP V j P V l P V P Ση l + η T j ΣP V j P V l P Σ δ η l + η j ΣP V j P V l P Σ ηt l δ. η T j ΣP V j δ P V l P Ση l (η T j ΣP V j δ P V j δ P Ση j ) 1/2 (η T l ΣP V l P V l P Ση l ) 1/2 By the defnton of η and h(δ) T Σh(δ) = o(n 3/2 ), we can see that η T j ΣP ( V j / δ )P V l P Ση l = o(n 1/2 ). The order of the other terms of g 4 (δ)/ can be derved smlarly. 16