TECHNICAL APPENDIX The proofs for the panel data error component SARAR(R,S) framework are given in full length to the benefit of the reader. They build on and generalize analogous proofs by Kelejian and Prucha (2010) for a cross-sectional SARAR(1,1) and Badinger and Egger (2011) for a cross-section SARAR(R,S) model, as well as analogous proofs for a panel SARAR(0,1) model with nonstochastic regressors by Kapoor, Kelejian, and Prucha (2007). APPENDIX A Notation We adopt the standard convention to refer to matrices and vectors with acronyms in boldface. Let A N denote some matrix. Its elements are referred to as aij, N ; ai., N and a.i, N denote the ith row and the i-th column of A N respectively. If A N is a square matrix, A N1 denotes its inverse; if A N is singular, A N denotes its generalized inverse. If A N is a square, symmetric and positive definite matrix, A1N/ 2 denotes the unique positive definite square root of A N and A N1 / 2 denotes ( A N1 )1 / 2 . The (submultiplicative) matrix norm is defined as A N [Tr (AN A N )]1 / 2 . Finally, unless stated otherwise, for expressions involving sums over elements of vectors or matrices that are stacked over all time periods, we adopt the convention to use single indexation i, running from i 1,..., NT , to denote elements of the stacked vectors or matrices. 1 Remark A.1 i) Definition of row and column sum boundedness (compare Kapoor, Kelejian, and Prucha, 2007, p. 99): Let B N , N 1 , be some sequence of NT NT matrices with T some fixed positive integer. We will then say that the row and column sums of the (sequence of) matrices B N are bounded uniformly in absolute value, if there exists a constant c , which does not depend on N, such that NT max 1 i NT b j 1 ij , N c and max 1 j NT NT b i 1 ij , N c for all N 1. The following results will be repeatedly used in the subsequent proofs. 1 Take the vector u N [uN (1),..., uN (T )] , for example. Using indexation i 1,..., NT , the elements ui , N , i 1,..., N refer to t 1, elements ui , N , i N 1,...,2 N refer to t 2 , etc., and elements ui , N , i (T 1) N 1,..., NT refer to t T . The major advantage of this notation is that it avoids the use of double indexation for the cross-section and time dimension. Moreover, it allows us the invoke several results referring to the case of a single cross-section, which still apply to the case of T stacked cross-sections. 1 ii) Let R N be a (sequence of) N N matrices whose row and column sums are bounded uniformly in absolute value, and let S be some T T matrix (with T 1 fixed). Then the row and column sums of the matrix S R N are bounded uniformly in absolute value (compare Kapoor, Kelejian, and Prucha, 2007, p. 118). iii) If A N and B N are (sequences of) NT NT matrices (with T 1 fixed), whose row and column sums are bounded uniformly in absolute value, then so are the row and column sums of A N B N and A N B N . If Z N is a (sequence of) NT P matrices whose elements are bounded uniformly in absolute value, then so are the elements of AN ZN and ( NT ) 1 ZN A N Z N . Of course, this also covers the case ( NT ) 1 ZN Z N for A N I NT (compare Kapoor, Kelejian, and Prucha, 2007, p. 119). iv) Suppose that the row and columns sums of the NT NT matrices A N (aij, N ) are NT bounded uniformly in absolute value by some finite constant c A ; then a i 1 ij , N q c Aq for q 1 (see Kelejian and Prucha, 2008, Remark C.1). v) Let ξ N and η N be NT 1 random vectors (with T 1 fixed), where, for each N, the elements are independently distributed with zero mean and finite variances. Then the elements of ( NT ) 1 / 2 ZN ξ N are O p (1) and ( NT ) 1 ξN A N η N is O p (1) (compare Kelejian and Prucha, 2004, Remark A.1). 2 vi) Let ζ N be a NT 1 random vector (with T 1 fixed), where, for each N, the elements are distributed with zero mean and finite fourth moments. Let π N be some nonstochastic NT 1 vector, whose elements are bounded uniformly in absolute value and let Π N be a NT NT nonstochastic matrix whose row and column sums are bounded uniformly in absolute value. Define the column vector d N π N Π N ζ N . It follows that the elements of d N have finite fourth moments. (Compare Kelejian and Prucha, 2008, Lemma C.2, for the case T 1 and 3 independent elements of ζ N .) 2 Kelejian and Prucha (2004) consider the case T 1 and where the elements of ξ N and η N are identically distributed. Obviously, the results also holds for (fixed) T 1 and under heteroskedasticity, as long as the variances of the elements of ξ N and η N are bounded uniformly in absolute value. 3 The extension to (fixed) T 1 is obvious. Independence of the elements of ζ N is not required for the result to hold. The fourth moments of the elements of d N π N Π N ζ N are NT NT j 1 j 1 given by E ( i , N ij, N j , N )4 24 E[ i4, N ( ij, N j , N )4 ] 2 Remark A2. The matrices Q 0, N and Q1, N have the following properties (see Kapoor, Kelejian, and Prucha, 2007, p. 101): tr (Q0, N ) N (T 1) , tr(Q1, N ) N , Q 0, N (eT I N ) 0 , Q1, N (eT I N ) (eT I N ) , Q 0, N ε N Q 0, N v N , Q1, N ε N (eT I N )μ N Q1, N v N , (IT D N )Q 0, N Q 0, N (IT D N ) , (IT D N )Q1, N Q1, N (IT D N ) , tr[(IT D N )Q0, N ] (T 1)tr(D N ) , tr[(IT D N )Q1, N ] tr (D N ) , where D N is an arbitrary N N matrix. Obviously, the row and column sums of Q 0, N and Q1, N are bounded uniformly in absolute value. APPENDIX B The following lemma will be repeatedly used in the subsequent proofs. 4 Lemma B.1 Let A N be some nonstochastic NT NT matrix (with T fixed), whose row and column sums ~ be a predictor for are bounded uniformly in absolute value. Let u be defined by (2c) and u N N u N . Suppose that Assumptions 1 to 4 hold. Then (a) N 1E uN A N u N O(1) , Var ( N 1uN A N u N ) o(1) , and 1 ~ A u ~ N 1 (u N N N ) N E (u N A N u N ) o p (1) . (b) N 1E d. j , N A N u N O(1) , j 1,..., P , where d. j , N is the j-th column of the NT P matrix ~ N 1E(D A u ) o (1) . D N , and N 1DN A N u N N N N p (c) If furthermore Assumption 6 holds, then 1 / 2 ~ A u ~ N 1 / 2u uN A N u N αN N 1 / 2Δ N op (1) with α N N 1E[DN ( A N AN )u N ] . N N N N NT NT NT NT 24 [ i4,N ij,N ik ,N il ,N im,N E j ,N k , N l ,N m,N ] K , by Hölder’s j 1 k 1 l 1 m1 inequality as long as the fourth moments of the elements of ζ N are bounded uniformly. 4 Compare Lemma C.1 in Kelejian and Prucha (2008) for the case of a cross-sectional SARAR(1,1) model and Lemma C.1 in Badinger and Egger (2011) for the case of a crosssectional SARAR(R,S) model. 3 ~ α o (1) . In light of (b), we have α N O(1) and N 1DN (A N AN )u N N p Proof of part (a) Let ~ ~ A u ~ N N 1uN A N u N and N N 1u N N N, (B.1) then given (4a), we have N N 1εN B N ε N , with the symmetric NT NT matrix B N defined as S S m 1 m 1 B N (1 / 2)[IT (I N m , N Mm , N ) 1 ]( A N AN )[IT (I N m , N M m , N ) 1 ] (B.2) By Assumptions 1-3 and Remark A.1 in Appendix A, the row and column sums of the matrices B N and Ω , N are bounded uniformly in absolute value. It follows that the row and column sums of the matrices B N Ω , N B N Ω , N are bounded uniformly in absolute value. In the following let K be a common bound for the row and column sums of the absolute value of the elements of B N , Ω , N , and B N Ω , N B N Ω , N and of the absolute value of their respective elements. Then NT NT E N E N 1 bij, N i , N j ,N (B.3) i 1 j 1 NT NT N 1 bij, N E i , N j , N i 1 j 1 NT NT N 1 bij,N ,i , j i 1 j 1 TK , 3 where we used Hölder’s inequality in the last step. This proves that E N is O(1). Now consider Var(N ) , invoking Lemma A.1 in KP (2008): Var (N ) Cov( N 1εN B N ε N , N 1εN B N ε N ) (B.4a) NT 2 N 2Tr (B N Ω , N B N Ω , N ) N 2 bii2,* [ (i4,)N 3] , i 1 2N Tr (B N Ω , N B N Ω , N ) N Tr{diag i 1,..., NT (bii2*, N )diag i 1,..., NT [ E(i4,N ) 3]} 2 2 (B.4b) 4 where bii*, N is the i-th diagonal element of B*N (bij,*, N ) SN BN S N and ( 4i ,)N is the fourth moment of the i-th element of the NT 1 vector η N S N1ε N , i.e., ( 4i ,)N E (i4,N ) . In light of Assumption 1 and the properties and Q 0, N and Q1, N , the row and column sums (and the elements) of S N v Q0, N 1 Q1, N are bounded uniformly in absolute value by some finite constant, say K * . W.o.l.o.g. we can choose the bound K used above such that K * K . Moreover, the row and column sums (and the elements) of SN1 v1Q0, N 11Q1, N are also bounded uniformly in absolute value by some constant K ** . W.o.l.o.g. we can choose K such that K ** K . In light of Remark A.1 and Assumption 1 it follows that the elements of η N S 1ε N have finite fourth moments. Denote their bound by K *** . W.o.l.o.g. we assume that K *** K . Hence, we have Var(N ) 2N 2Tr ( KI NT ) N 2Tr[diagi 1,..., NT ( K 2 K )] . 2 N 1TK N 1TK 3 N 1 (2TK TK 3 ) o(1) . The claim in part (a) of Lemma B.1 that N 1 (uN A N u N ) N 1E(uN A N u N ) o p (1) now follows from Chebychev’s inequality (see, for example, White, 2001, p. 35). 1 ~ A u ~ We now prove the second part of (a), i.e., N 1 (u N N N ) N E (u N A N u N ) o p (1) . Since ~ N E (N ) o p (1) , it suffices to show that N N o p (1) . By Assumption 4, we have ~ u D Δ , where D (d ,..., d ~ u D Δ into the ) . Substituting u u N N N N N 1., N NT ., N N N N N ~ expression for N in (B.1), we obtain ~ N N N 1 (uN ΔN DN )A N (u N DN ΔN ) N 1uN A N u N (B.5) N 1 (uN A N u N ΔN DN A N u N uN A N D N Δ N ΔN DN A N D N Δ N uN A N u N ) N 1 (ΔN DN A N u N uN A N D N Δ N ΔN DN A N D N Δ N ) N 1 (ΔN DN A N u N ΔN DN AN u N ΔN DN A N D N Δ N ) N 1[ΔN DN ( A N AN )u N ΔN DN A N D N Δ N ] N N , 5 where S N N 1[ΔN DN ( A N AN )u N ] N 1{ΔN DN ( A N AN )[IT (I N m, N M m, N ) 1 ]ε N } , (B.6) m 1 N (ΔN DN C N ε N ) , with 1 S C N ( A N AN )[IT (I N m , N M m, N ) 1 ] (c1 ., N ,.., cNT ., N ) , m 1 N N ΔN DN A N D N Δ N . 1 (B.7) By Assumption 3 and Remark A.1, the row and column sums of C N are bounded uniformly in absolute value. We next prove that N o p (1) and N o p (1) . Proof that N o p (1) : N N 1 ΔN DN CN ε N N 1 NT Δ d N i 1 (B.8) c εN i ., N i ., N NT N 1 ΔN di., N ci., N ε N i 1 N 1 ΔN N 1 ΔN NT i 1 NT i 1 NT NT c di., N di., N NT c N 1 ΔN N 1 1 ΔN N N N NT j 1 NT 1 / 2 1 / p 1 / 2 j, N j 1 NN N 1/ p 1/ 2 NT Note that cij, N i 1 bound for the row j, N c j 1 NT j ,N j, N NT i 1 ij , N j 1 N 1 ΔN di., N ij , N j 1 i 1 j, N ij , N di., N cij,N NT di., N i 1 N ΔN 1/ 2 ΔN 1/ p p 1/ q q NT cij, N i 1 1 NT N j, N j 1 1 NT N j, N j 1 1 NT N di., N i 1 1 NT N di., N i 1 1/ p p 1/ p p 1/ q q NT cij, N i 1 1/ q q NT cij, N i 1 K by Assumption. In the following we denote by K the uniform and column sums of the absolute value of the elements of A N and C N . 6 From Remark C.1 in KP (2008) (see Remark A.1 in Appendix A) it follows that 1/ q q q NT NT q and thus c K ij, N cij, N i 1 i 1 N K N 1 / p 1 / 2 N 1/ 2 ΔN K . Factoring K out of the sum yields 1 NT N j, N j 1 NT T ( NT ) 1 di., N i 1 1/ p p . This holds for p 2 for some 0 as in Assumption 4 and 1 / p 1 / q 1 . By Assumption N 4, 1/ 2 Δ N Op (1) . Assumption 4 also implies that 1/ p NT p ( NT )1 di., N Op (1) for p 2 and some 0 . i 1 NT Moreover, E j, N K , which implies that N 1 j , N j 1 Op (1) . Since N 1 / p 1 / 2 0 as N it follows that N o p (1) . Next consider NT NT N N 1 ΔN DN A N DN Δ N = N 1 ΔN di., N aij,N d j., N Δ N (B.9) i 1 j 1 N 1 Δ N 2 1 2 N ΔN N 1/ p N 1 / p 1 / 2 NT NT d d i ., N i 1 NT d i ., N i 1 K ΔN N 2 j 1 aij,N j ., N NT d j ., N j 1 1/ p p NT aij, N j 1 NT 1 NT N di., N N 1 d j ., N i 1 j 1 1 / 2 K N 1/ 2 ΔN 2 1 NT N d j ., N j 1 1/ q q 1/ p p p 2/ p o p (1) . From the last inequality we can also see that N 1 / 2 N op (1) . Summing up, we have proved ~ that N N N N o p (1) . Proof of part (b) 7 Denote by s*, N the s-th element of N 1DN A N u N . By Assumptions 3 and 4 and Remark A.1 in Appendix A there exists a constant K such that E (ui2, N ) K and E dij, N p K with p 2 for some 0 . W.o.l.o.g. we assume that the row and column sums of the matrices A N are bounded uniformly by K . Notice first that E ui , N d js, N Eui2, N 1/ 2 Eui2, N 1/ 2 Ed 1/ 2 2 js, N 1/ p p E d js, N K 1 / 2 K 1 / p K 1 / 21 / p with p as before. It follows that NT NT E s*, N N 1 aij, N E ui , N d js, N (B.10) i 1 j 1 NT NT K 1 / 2 1 / p N 1 aij, N K 1 / 2 1 / p N 1NTK TK 3 / 2 1 / p , i 1 j 1 which shows that indeed E N 1d.s , N A N u N O(1) . Of course, the argument also shows that α N N 1E[DN ( A N AN )u N ] O(1) . It is readily verified that Var (s* ) o(1) , such that we have s* E(s* ) op (1) . Next observe that ~ N 1D A u * , N 1DN A N u N N N N N where N* N 1DN A N D N Δ N . By (B.11) arguments analogous to the proof that ~ N N 1[ΔN DN (A N AN )u N ] op (1) , it follows that N* op (1) . Hence s* s* o p (1) , ~ ~ α o (1) . and thus s* E (s* ) o p (1) , which also shows that N 1DN (A N AN )u N N p Proof of part (c) In light of the proof of part (a) 1 / 2 ~ A u ~ uN A N u N [ N 1uN ( A N AN ) DN ]N 1 / 2Δ N N 1 / 2 N , N 1 / 2u N N N N (B.12) 8 where N 1 / 2 N op (1) as shown above, and in light of (b) and since N 1 / 2Δ N Op (1) by Assumption 4, we have 1 / 2 ~ A u ~ N 1 / 2u uN A N uN N 1 / 2αN Δ N op (1) . N N N N (B.13) Proof of Theorem 1a. Consistency of Initial GM Estimator θˆ 0N The objective function of the nonlinear least squares estimator in (17a) and its nonstochastic counterpart are given by ~ ~ 0 R0N (, θ ) (~ γ N0 Γ0Nb0N )(~ γ N0 Γ0Nb0N ) and (B.14a) 0 0 0 R 0N (θ ) ( γ 0N Γ0N b )( γ 0N Γ0N b ) . (B.14b) 0 Since γ 0N Γ 0N b0N 0 , we have RN0 (θ0N ) 0 , i.e., RN0 (θ ) 0 at the true parameter vector θ0N ( 1 ,..., S , v2 ) . Hence, 0 0 0 R 0N (θ ) R 0N (θ0 ) (b b0N )Γ 0N Γ 0N (bN b0N ) . (B.15) In light of Rao (1973, p. 62) and Assumption 5, it follows that: 0 0 0 RN0 (θ ) RN0 (θ0 ) min (Γ 0N Γ 0N )(b b0N )(b b0N ) and RN0 (θ ) RN0 (θ0 ) * (b b0N )(b b0N ) . 0 0 0 2 By the properties of the norm A [tr(AA)]1 / 2 , we have θ θ0 (b b0N )(b b0N ) such 0 0 0 2 that RN0 (θ ) RN0 (θ0 ) * θ θ0 . Hence, for every 0 0 lim 0 inf 0 0 [ RN0 (θ ) RN0 (θ0 )] 0 N {θ 0 : θ θ } inf 0 0 {θ 0 : θ θ } 2 * θ0 θ0 * 2 0 , (B.16) which proves that the true parameter θ 0 is identifiable unique. ~ Next, let FN0 (~ γ N0 ,Γ0N ) and Φ 0N ( γ 0N , Γ 0N ) . The objective function and its nonstochastic counterpart can then be written as 9 0 0 0 RN0 (θ ) (1, b )FN0 FN0 (1, b ) and 0 0 0 RN0 (θ ) (1,b )Φ0N Φ0N (1, b ) . Hence for ρ [a , a ] and v [0, b] it holds that 2 5 0 0 0 0 RN0 (θ ) RN0 (θ ) (1, b )(FN0 FN0 Φ0N Φ0N )(1, b ) . Moreover, since the norm is submultiplicative, i.e., AB A B , we have 0 R (θ ) R (θ ) F FN0 Φ 0N Φ0N (1, bN ) 0 N 0 0 N 0 2 0 N 2S S (S 1) 4 FN0 FN0 Φ0N Φ0N [1 S (a )2 (a ) b 2 ] . 2 It is readily observed from (16), that the elements of the matrices γ 0N and Γ 0N are all of the form uN A N u N , where A N are nonstochastic NT NT matrices (with T fixed), whose row and column sums are bounded uniformly in absolute value. In light of Lemma B.1, the p p elements of Φ 0N are O(1) and it follows that FN0 Φ 0N 0 and FN0 FN0 Φ0N Φ0N 0 as N . As a consequence, we have (for finite S) S (S 1) 4 2 p 0 0 RN0 (θ ) RN0 (θ ) [FN0 FN0 Φ0N Φ0N ] [1 S (a ) 2 (a ) b ] 0 as N 2 ρ[ a ,a ], v2 [ 0,b ] sup (B.17) ~0 Together with identifiable uniqueness, the consistency of θN ( ~1, N ,..., ~S , N ,~v2, N ) now follows directly from Lemma 3.1 in Pötscher and Prucha (1997). Having proved that the estimators ~1, N ..., ~S , N ,~v2, N are consistent for 1,N ..., S ,N , v2 , we now show that 12 can be estimated consistently from the last line (4S 2) of equation system (12), using ~12,N ~4S 2, N ~4S 2,1, N ~1, N ... ~4 S 2, S , N ~S , N ~4S 2, S 1, N ~12,N ~4 S 2, 2S , N ~S2, N ~4 S 2, 2S 1, N ~1, N ~2, N ... ~4S 2, 2 S S ( S 1) / 2, N ~S 1, N ~S , N . 5 (B.18) This should be read as ρ s [a , a ] for all s 1,..., S . 10 Since γ N Γ N b0N 0 , we have ~12 12 (~4 S 2, N 4 S 2, N ) (~4 S 2,1, N 4 S 2,1, N ) ~1, N ... (~4 S 2, S , N 4 S 2, S , N ) ~S , N (~4 S 2, S 1, N 4 S 2, S 1, N ) ~12,N ... (~4 S 2, 2 S , N 4 S 2, 2 S , N ) ~S2, N (~ ) ~ ~ ... (~ ) ~ ~ 4 S 2 , 2 S 1, N 4 S 2,1, N ( ~ 1, N 4 S 2 , 2 S 1, N 1, N 4 S 2 , 2 S S ( S 1) / 2 , N 2, N 1, N ) ... 4 S 2, S , N ( ~ S,N S , N ) 4 S 2, S 1, N ( ~12,N 12, N )... 4 S 2, 2 S , N ( ~S2, N S2, N ) ( ~ ~ 2 )... 4 S 2 , 2 S 1, N 1, N 2, N 1, N 2, N 4 S 2 , 2 S S ( S 1) / 2 , N 4 S 2 , 2 S S ( S 1) / 2 , N S 1, N S,N (B.19) ( ~S 1, N ~S , N S 1, N S , N ). Since FN0 Φ 0N 0 as N and the elements of Φ N are O(1) it follows from the p p consistency of ~1, N ,..., ~S , N that ~12, N 12 0 as N . Proof of Theorem 1b. Consistency of the Weighted GM Estimator The objective function of the weighted GM estimator and its nonstochastic counterpart are given by ~ ~ RN (θ) (~ γ N Γ N b)Θ N (~ γ N Γ N b) and (B.20a) RN (θ) ( γ N Γ N b)Θ N ( γ N Γ N b) (B.20b) First, in order to ensure identifiable uniqueness, we show that Assumption 5 also implies that the smallest eigenvalue of ΓN Θ N Γ N is bounded away from zero, i.e., min (ΓN Θ N Γ N ) 0 for some 0 0. (B.21) Let A (aij ) Γ0N Γ0N and B (bij ) Γ1N Γ1N . Note that Γ 0N and Γ1N are of dimension ( 2 S 1) [2S S ( S 1) / 2 1] (i.e., they have half the rows and one column less than than ΓN ). A and B are of order [2S S ( S 1) / 2 1] [2S S ( S 1) / 2 1] (i.e., they have one row and column less than ΓN ΓN ). Next define Γ N (Γ0N , Γ1N ) , which differs from ΓN only by the ordering of the rows. Γ 0N corresponds to Γ 0N with a zero column appended as last column, i.e., Γ 0N (Γ 0N ,0) , such that 11 0 0 0 0 Γ ΓN ΓN N ΓN 0 ( Γ 0N Γ 0N is of the a1,1 a1, 2 S S ( S 1) / 2 1 . 0 a2 S S ( S 1) / 2 1, 2 S S ( S 1) / 2 1 0 a2 S S ( S 1) / 2 1,1 0 0 0 same dimension as ΓN Γ N , i.e., 0 0 . 0 0 (B.22a) [2 S S ( S 1) / 2 2] [2 S S ( S 1) / 2 2] .) Γ1N is a modified version of Γ1N , with a zero column included as second last column, such that b1,1 0 b1, 2 S S ( S 1) / 2 1 . 0 . 1 1 . ΓN ΓN 0 0 0 0 b2 S S ( S 1) / 2 1,1 . 0 b2 S S ( S 1) / 2 1, 2 S S ( S 1) / 2 1 ( Γ1N Γ1N is of the same dimension as ΓN Γ N , i.e., (B.22b) [2 S S ( S 1) / 2 2] [2 S S ( S 1) / 2 2] .) Since Γ N (Γ0N , Γ1N ) differs from ΓN only by the ordering of the rows, it follows that 0 1 Γ0N 0 0 1 1 Γ N Γ N = Γ N Γ N Γ N Γ N 1 Γ N Γ N Γ N Γ N , i.e., Γ N Γ N Γ N a1,1 . a 2 S S ( S 1) / 2 1,1 0 a1, 2 S S ( S 1) / 2 1 a 2 S S ( S 1) / 2 1, 2 S S ( S 1) / 2 1 0 b1,1 . 0 b2 S S ( S 1) / 2 1,1 0 (B.23) 0 0 0 0 0 . . 0 0 0 . 0 b2 S S ( S 1) / 2 1, 2 S S ( S 1) / 2 1 0 b1, 2 S S ( S 1) / 2 1 We can thus write xΓ N Γ N x xΓ0N Γ0N x xΓ1N Γ1N x xA A N x A xB B N x B . (B.24) 12 The vector x is of dimension [2S S ( S 1) / 2 2] 1 (corresponding to the number of columns of Γ N ), wheras x A and x B are of dimension [2S S ( S 1) / 2 1] , i.e. both have one row less: x A excludes the last element of x, i.e., x 2 S S ( S 1) / 2 2 , x B excludes the secondlast element of x , i.e., x 2 S S ( S 1) / 2 1 . Again, we invoke Rao (1973, p. 62) for each quadratic form. It follows xA A N x A xB B N x B min ( A N )xAx A min (B N )xB x B * (xAx A xB x B ) *xx (B.25) for any x ( x1 , x2 ,..., x2 S 2 ) . Hence, we have shown that xΓN Γ N x *xx , or, equivalently, xΓN Γ N x * for x 0 . xx (B.26) Next, note that in light of Rao (1973, p. 62), min (ΓN Γ N ) inf x xΓN Γ N x * 0 . xx (B.27) Using Mittelhammer (1996, p. 254) we have min (ΓN Θ N Γ N ) inf x xΓN Θ N Γ N x xΓN Γ N x min ( Ξ N1 ) inf x xx xx min (Θ N )min (ΓN Γ N ) 0 0 , (B.28) with 0 ** since min (Θ N ) * 0 by assumption (see Theorem 2). This ensures that the true parameter vector θ N ( 1,N ,..., S ,N , v2 ,12 ) is identifiable unique. Next note that in light of the assumptions in Theorem 2, Θ N is O(1) by the equivalence of matrix norms. Analogous to the prove of Theorem 1, observe that RN (θ) 0 , i.e., RN (θ) 0 at the true parameter vector θ N ( 1,N ,..., S ,N , v2 ,12 ) . It follows that 13 RN (θ) RN (θ N ) (b bN )Γ N Θ N Γ N (b bN ) . (B.29) ~ Moreover, let FN (~ γ N ,Γ N ) and Φ N ( γ N ,Γ N ) , then, RN (θ) (1, b )FN ΘN FN (1, b ) and (B.30a) RN (θ) (1, b )ΦN Θ N Φ N (1, b ) . (B.30b) The remainder of the proof is now analogous to that of Theorem 1a. ~ Proof of Theorem 2. Asymptotic Normality of θN To derive the asymptotic distribution of the vector qN , defined in (30) we invoke the central limit theorem for vectors of linear quadratic forms given by Kelejian and Prucha (2010, Theorem A.1), which is an extension of the central limit theorem for a single linear quadratic form by Kelejian and Prucha (2001, Theorem 1). The vector of quadratic forms in the present context, to which the Theorem is applied is q*N . The variance-covariance matrix of qN was derived above and is denoted as Ψ N . Accordingly, the variance-covariance matrix of q*N N 1 / 2qN is given by Ψ*N NΨ N and (Ψ *N ) 1/ 2 N 1/ 2 Ψ N1/ 2 . Note that in light of Assumptions 1, 2 and 7 (and Lemma B.1), the stacked innovations ξ N , the matrices A1, s , N ,..., A 4, s , N , s 1,..., S , A a, N , and A b, N , and the vectors a1, s, N , …, a 4, s, N , s 1,..., S , a a, N , and ab, N satisfy the assumptions of central limit theorem by Kelejian and Prucha (2010, Theorem A.1). In the application of Theorem A.1, note that the sample size is given by NT N N (T 1) rather than N . As Kelejian and Prucha (2001, p. 227, fn. 13) point out, Theorem A.1 “also holds if the sample size is taken to be kn rather than n (with k n as N ).” In the present case we have K N (T 1) N , with T 1 and fixed, which ensures that K n as N . Consequently, Theorem A.1 still applies to each quadratic form in q*N . Moreover, as can be observed from the proof of Theorem A.1 in Kelejian and Prucha (2010), the extension of the Theorem from a scalar to a vector of vector of quadratic forms holds up under by this alternative definition of the sample size. It follows that d (Ψ*N ) 1 / 2 q*N N 1 / 2 Ψ N1 / 2q*N Ψ N1 / 2qN (0, I 4 S 2 ) , (B.31) 14 since N 1min (Ψ *N ) N 1min ( NΨ N ) min (Ψ N ) 0 by assumption as required in Theorem A.1. Since the row and column sums of the matrices A1, s , N ,..., A 4, s , N , s 1,..., S , A a, N , and A b, N , and the vectors a1, s, N , …, a 4, s, N , s 1,..., S , a a, N , and ab, N , and the variances v2,N and 2 ,N are bounded uniformly in absolute value, it follows in light of (38) that the elements of Ψ N and also those of Ψ1N/ 2 are bounded uniformly in absolute value. ~ We next turn to the derivation of the limiting distribution of the GM estimator θN . In ~ Theorem 1 we showed that the GM estimator θN defined by (18) is consistent. It follows that – apart from a set of the sample space whose probability tends to zero – the estimator satisfies the following first order condition: 6 ~ ~ ~ ~ ~ q N (θN , Δ N ) ~ q N (θN , Δ N )Θ N q N (θN , Δ N ) Θ N q N ( θN , Δ N ) 0 , ρ θ (B.32) which is a ( S 2) 1 vector, the rows corresponding the partial derivatives of the criterion function with respect to s, N , s 1,..., S , v2 , and 12 . Substituting the mean value theorem expression ~ q (θ , Δ ) ~ q N (θN , Δ N ) q N (θ N , Δ N ) N N N (θN θ N ) , θ (B.33) where θN is some between value, into the first-order condition yields ~ ~ q N (θN , Δ N ) ~ q N (θN , Δ N ) 1 / 2 ~ q N (θN , Δ N ) ~ 1 / 2 ΘN N ( θN θ N ) Θ N N q N (θ N , Δ N ) . (B.34) θ θ θ Observe that q N (θ, Δ N ) ~ ΓNB θ N and consider the two ( S 2) ( S 2) matrices ~ q N (θN , Δ N ) ~ q N (θN , Δ N ) ~ ~ ~ ~ ~ ΞN ΘN B N ΓN Θ N Γ N B N , θ θ (B.35) 6 The leading two and the negative sign are ignored without further consequences for the proof. 15 ΞN B N Γ N Θ N Γ N B N , ~ where B N and B N correspond to B (B.36) N ~ as defined above with θN and θN substituted for θ N . Notice that Ξ N is positive definite, since Γ N and Θ N are positive definite by assumption and the [2S S ( S 1) / 2 2] ( S 2) matrix B N has full column rank. ~ p 0 and In the proof of Theorem 1 (and Lemma B.1) we have demonstrated that Γ N Γ N ~ that the elements of Γ N and Γ N are O(1) and O p (1) , respectively. By Assumption 5, ~ ~ ~ ~ and ρ (and thus also B Θ N Θ N o p (1) , Θ N O(1) and Θ N Op (1) . Since ρ N N N and ~ B N ) are consistent and bounded uniformly in probability, if follows that ΞN ΞN o p (1) , ~ ΞN O p (1) , and ΞN O(1) . Moreover, Ξ N is positive definite and thus invertible, and its inverse ΞN1 is also O(1) . ~ ~ Denote ΞN as the generalized inverse of Ξ N . It then follows as a special case of Lemma F1 in ~ Pötscher and Prucha (1997) that Ξ N is non-singular with probability approaching 1 as ~ ~ N , that Ξ N is O p (1) , and that ΞN ΞN1 o p (1) . ~ Pre-multiplying (B.34) with ΞN we obtain, after rearranging terms, N 1/ 2 ~ ~ ~ ~ 1/ 2 ~ 1 / 2 ~ q N ( θ N , Δ N ) ~ ( θ N θ N ) (I S 2 Ξ N Ξ N ) N ( θ N θ N ) N Ξ N Θ N q N (θ N , Δ N ) .(B.37) θ In light of the discussion above, the first term on the right-hand side is zero on -sets of probability approaching 1 (compare Pötscher and Prucha, 1997, pp. 228). This yields N 1/ 2 ~ ~ ~ q N (θN , Δ N ) ~ 1 / 2 (θN θ N ) ΞN Θ N N q N (θ N , Δ N ) o p (1) . θ (B.38) Next observe that ~ ~ q N (θN , Δ N ) ~ ΞN Θ N ΞN1B N ΓN Θ N o p (1) , θ (B.39) ~ q N (θN , Δ N ) ~ 1 B N ΓN o p (1) . since Ξ N Ξ N o p (1) and θ 16 As we showed in section III, the elements of N 1 / 2q N (θ N , Δ N ) can be expressed as N 1 / 2q N (θ N , Δ N ) N 1 / 2q*N op (1) qN op (1) . (B.40) where q*N is defined in (27), and that d (Ψ*N ) 1 / 2 q*N N 1 / 2 Ψ N1 / 2q*N Ψ N1 / 2qN (0, I 4 S 2 ) . (B.41) It now follows from (B.38), (B.39), and (B.40) that ~ N 1 / 2 (θN θ N ) ΞN1JN Θ N Ψ1N/ 2 (Ψ N1 / 2qN ) o p (1) . (B.42) Since all nonstochastic terms on the right hand side from (B.42) are O(1) it follows that ~ ~ N 1 / 2 (θN θ N ) is O p (1) . To derive the asymptotic distribution of N 1 / 2 (θN θ N ) , we invoke Corollary F4 in Pötscher and Prucha (1997). In the present context, we have d ζ N Ψ N1 / 2qN ζ ~ N (0, I 4 S 2 ) , ~ N 1 / 2 (θN θ N ) A N ζ N o p (1) , with A N ΞN1JN Θ N Ψ1N/ 2 . ~ Furthermore, N 1 / 2 (θN θ N ) Op (1) and its variance-covariance matrix is Ω ~θ (Θ N ) (JN Θ N J N ) 1 JN Θ N Ψ N Θ N J N (JN Θ N J N ) 1 , N where Ω ~θ is positive definite. N As a final point it has to be shown that lim inf N min (A N A N ) 0 as required in Corollary F4 in Pötscher and Prucha (1997). Observe that min (A N A N ) min (ΞN1JN Θ N Ψ N ΘN J N ΞN1 ) (B.43) min (Ψ N )min (Θ N ΘN )min (ΞN1ΞN1 )min (ΓN Γ N )min (B N B N ) 0 , since the matrices involved are all positive definite. 17 Consistency Proof for Estimates of Third and Fourth Moments of the Error Components Consistent estimates for the second moments of vit , N and i, N are delivered by the GM estimators defined in (17) and (18), respectively (See Theorems 1a and 1b). In the following we prove that the estimators for the third and fourth moments of vit , N and it, N , defined in (39) and (40) are also consistent. The proof draws on Gilbert (2002), who considers the estimation of third and fourth moments in error component models without spatial lags and without spatial regressive disturbances. For reasons that will become clear below, we depart from the convention adopted so far to use indexation i 1,..., NT for the stacked series. In the subsequent proof we use the double indexation it , with i 1,..., N , t 1,..., T . Preliminary Remarks Note first that ~ε ε η , where N N N (B.44) S S η N [ ( m , N ~m , N )IT M m , N ][IT (I N m , N M m , N ) 1 ]ε N m 1 m 1 S S m 1 m 1 [IT (I N m , N M m, N )]D N Δ N [ ( m, N ~m , N )(IT M m, N )]D N Δ N . This can also be written as ηN R N gN , (B.45) where R N (R1, N , R 2, N , R 3, N ) with S R 1, N IT (I N m , N M m , N )D N , m 1 S S m1 m1 R 2, N {( I T M1, N )[IT (I N m, N M m, N ) 1 ]ε N ,..., (IT M S , N )[I T (I N m, N M m, N ) 1 ]ε N } , R 3, N [(IT M1, N )D N ,..., (IT M S , N )D N ] , ΔN ~ gN (ρ N ρN ) . ~ )Δ (ρ N ρ N N 18 In light of Assumption 3 and since the elements of D N (d1 ., N ,..., dN ., N ) have bounded fourth moments (by Assumption in Theorem 3), each column of the matrix R N is of the form π N Π N ξ N , where the elements of the NT 1 vector π N are bounded uniformly in absolute value by some finite constant, the row and column sums of the NT NT matrix Π N are bounded uniformly in absolute value by some finite constant, and the fourth moments of the elements of ξ N are also bounded by some finite constant. It follows that the fourth moments of the elements of R N are also bounded by some finite constant by Lemma C.2 in Kelejian and Prucha (2010). 7 As a consequence, ηN R N gN , or for the i-th element of the NT 1 vector η N , i , N gN ri., N N i , N , (B.46) 4 where ri., N denotes the i-th row of R N , N gN , and i , N ri., N with E i , N K . Without loss of generality we can select K such that E (i, N ) K for 4 . By Assumption 1 there is also some K such that E i, N K for 4 . In the following we use K to denote the larger bound, i.e., K max( K , K ) . Also note that N 1 / 2 N Op (1) . Estimation of Third Moments 1 N T ~3 We first consider (3) E ( it3, N ) and its estimate ~( 3, N) it, N . NT i 1 t 1 Using (B.44) we have ~(3, N) 1 N T ( it, N it, N )3 NT i 1 t 1 (B.47) 1 N T 3 ( it , N 3 it2, Nit , N 3it2, N it , N it3, N ) NT i 1 t 1 1 N T 3 3 N T 2 3 N T 2 1 N T 3 it , N it, N NT it, N it, N NT it , N it , N NT i 1 t 1 NT i 1 t 1 i 1 t 1 i 1 t 1 1, N 2, N 3, N 4, N . By the weak law of large numbers for i.i.d. random variables, 1, N converges to ( 3) as N . 7 See also Remark A.1 in Appendix A. 19 Next consider 1 NT 2, N N T 3 i 1 t 1 it , N 2 it , N (B.48) N T 3 N it2, N it , N NT i 1 t 1 1/ 2 1/ 2 4 2 N T N T 3 N it ,N it ,N NT i1 t 1 i1 t 1 1/ 2 3( N N ) N 1/ 2 1 / 2 1/ 2 N T N T 4 2 ( NT )1 it , N ( NT )1 it , N i 1 t 1 i 1 t 1 o p (1) , 1/ 2 since E it , N 4 N T 4 K and thus ( NT )1 it , N i 1 t 1 2 Op (1) , E it , N K and thus 1/ 2 N T 2 ( NT ) 1 it , N i 1 t 1 Op (1) , N 1 / 2 N Op (1) , and N 1 / 2 o(1) . Observe further that 1 NT 3, N N 3 NT N T 3 i 1 t 1 T i 1 t 1 2 N 2 it , N it , N i2,N it , N (B.49) 3 2 NT 2 N i , N it , N NT i 1 1/ 2 1/ 2 N T N T 4 2 3( N N ) N ( NT )1 it , N ( NT )1 i , N i 1 t 1 i 1 t 1 1/ 2 1 2 o p (1) , 1/ 2 since E it , N 2 N T 2 K and thus ( NT ) 1 it ,N i 1 t 1 4 Op (1) , E i , N K and thus 1/ 2 N T 4 ( NT )1 i , N i 1 t 1 Op (1) , N 1 / 2 N Op (1) , and N 1 o(1) . Finally, 4, N 1 NT 1 NT N N T i 1 t 1 T i 1 t 1 3 N 3 it , N i,N (B.50) 3 20 N T 3 1 ( N 1/ 2 N )3 N 3 / 2 N N 1 i , N NT i 1 t 1 NT 3 ( N 1 / 2 N )3 N 3 / 2 ( NT ) 1 i , N o p (1) , i 1 since N 3 / 2 NT 3 1 ( NT ) i ,N O p (1) , N 1 / 2 N Op (1) , and E ( i , N ) K i 1 o(1) . As a consequence, we have ~(3, N) (3) o p (1) , (3) O(1) by Assumption 1, 3 and thus and ~(3, N) Op (1) . Next consider the third moments of the unit-specific error component (3) and its estimate ~(3) , which can be expressed (compare Gilbert, 2002, p. 48ff.) as (3) E( is, N it2, N ) for any given i and s t , ~(3, )N N T T 1 ~is, N ~it2, N . NT (T 1) i 1 s 1 t 1 (B.50a) (B.50b) ts Notice that by Assumption 1, (3) is invariant to the choice of i, s and t. Using (B.44), we have ~(3, )N N T T 1 ( is, N is, N )( it , N it , N ) 2 NT (T 1) i 1 s 1 t 1 (B.51) ts N T T 1 ( is, N it2, N 2it, N is, N it2, N is, N is, N it2, N 2is, Nit, N is, Nit2, N ) NT (T 1) i 1 s 1 t 1 ts 1, N 2, N 3, N 4, N 5, N 6, N Observe that N T T i 1 s ts N T T 1, N is, N it2, N ( is, N vis, N )( it , N vit , N ) 2 N T (B.52) i 1 s 1 t 1 ts T ( i3, N 2i2, N vit , N i , N vit2, N vis, N i2, N 2i , N vit , N vis, N vit2, N vis, N ) i 1 s 1 t 1 ts 11, N 12, N 13, N 14, N 15, N 16, N . 21 By the weak law of large numbers 11, N converges in probability to (3) . Notice further that, by the properties of vit , N and it, N (see Assumption 1), 12, N ,13, N ,14, N ,15, N , and 16, N are all o p (1) . As a consequence, 1, N converges in probability to (3) . Next observe that 2 , N N T T 1 2it, N is, N NT (T 1) i 1 s 1 t 1 (B.53) ts N T T 2 N is, N it , N NT (T 1) i 1 s 1 t 1 ts N T T 1/ 2 1 / 2 2 N N N [ NT (T 1)]1 is, N i 1 s 1 t 1 t s 3, N 1/ 2 2 1/ 2 N T T 1 2 [ NT (T 1)] it ,N i 1 s 1 t 1 t s N T T 1 it2, N is, N NT (T 1) i 1 s 1 t 1 o p (1) , (B.54) ts N T T 1/ 2 2 1 ( N N ) N [ NT (T 1)]1 is,N i 1 s 1 t 1 t s 4 , N 1/ 2 2 1/ 2 N T T 1 4 [ NT (T 1)] it ,N i 1 s 1 t 1 t s o p (1) , N T T 1 is, N it2, N NT (T 1) i 1 s 1 t 1 (B.55) ts 1/ 2 N N N 1/ 2 5, N 1 / 2 1/ 2 N T T N T T 4 [ NT (T 1)]1 is, N [ NT (T 1)]1 it2, N i 1 s t s i 1 s t s N T T 1 2is,Nit,N NT (T 1) i1 s t s o p (1) , (B.56) N T T 2 ( N 1 / 2 N ) 2 N 1 is, N it , N NT (T 1) i 1 s 1 t 1 ts 1/ 2 N T T 1/ 2 2 1 1 2 2( N N ) N [ NT (T 1)] is,N i 1 s 1 t 1 t s 1/ 2 N T T 1 2 [ NT (T 1)] it ,N i 1 s 1 t 1 t s o p (1) , 22 6, N N T T 1 is, Nit2, N NT (T 1) i 1 s 1 t 1 (B.57) ts 1/ 2 N T T 1/ 2 3 3 / 2 1 2 2( N N ) N [ NT (T 1)] is, N i 1 s 1 t 1 t s 1/ 2 N T T 1 4 [ NT (T 1)] it ,N i 1 s 1 t 1 t s o p (1) , because N 1 / 2 N is O p (1) and the terms in brackets expressions are all O p (1) , since E is, N K and E it , N K for 4 and all N. It follows that ~(3, )N (3) op (1) , (3) O(1) by Assumption 1, and that ~(3, )N Op (1) . Obviously, we then also have that (~(3, N) ~(3, )N ) v(3) ~v(,3N) v(3) op (1) . Estimation of Fourth Moments Consider the fourth moment of i, N and its estimate, which can be expressed as (compare Gilbert, 2002, p. 48ff.): 4 E( is it3 ) 3E( is it )[ E( it2 ) E( is it )] for any given i and s t , ~( 4, N) (B.58a) N T T 1 ~is, N ~it3, N NT (T 1) i 1 s 1 t 1 (B.58b) ts N T T N T N T T 3 1 ~ ~ ( 1 ~2 ~is, N ~it , N ) is , N it , N it , N NT (T 1) i 1 s 1 t 1 NT i 1 t 1 NT (T 1) i 1 s 1 t 1 ts ts 1, N 2, N (~2, N 2, N ) . Observe that 1, N N T T 1 ( is, N is, N )( it , N it , N )3 NT (T 1) i 1 s 1 t 1 (B.59) ts N T T 1 ( is, N it3, N 3 is, N it2, Nit , N 3it2, N is, N it , N is, Nit3, N NT (T 1) i 1 s 1 t 1 ts is, N it3, N 3 it2, Nit, Nis, N 3it2, Nis, N it, N it3, Nis, N ) 11, N 12, N 13, N 14, N 15, N 16, N 17, N 18, N . The first term 11, N can also be written as 23 11, N is, N it3, N (is, N vis, N )( it, N vit, N )3 (B.60) (is, N vis, N )( it3, N 3it2, N vit, N 3it, N vit2, N vit3, N ) (is, N it3, N 3is, N it2, N vit, N 3is, N it, N vit2, N is, N vit3, N vis, N it3, N 3it2, N vis, N vit, N 3it, N vis, N vit2, N vis, N vit3, N ) (i4,N 3i3, N vit, N 3i2,N vit2, N is, N vit3, N vis, N it3, N 3it2, N vis, N vit, N 3it, N vis, N vit2, N vis, N vit3, N ) . By the properties of vit , N and it, N (see Assumption 1), 11, N converges in probability to ( 4) 3 2 v2 . Moreover, it follows from the properties of v N and μ N (see Assumption 1), that the terms 12, N , 13, N , 14, N , 15, N , 16, N , 17, N , 18, N are o p (1) . It follows that 1, N converges in probability to ( 4) 3 2 v2 . Next consider 2, N which E[ N T T 3 ( is, N it, N )( it, N it, N ) NT (T 1) i 1 s 1 t s (B.61) N T T 3 ( is, N is, N )( it , N it , N ) NT (T 1) i 1 s 1 t s N T T 3 ( is, N it, N is, N it, N is, Nit, N is, Nit, N ) , NT (T 1) i 1 s 1 t s converges 2 to by the weak law of large numbers, since T 1 T 1 is, N it , N ] 2 for s t by the properties of vit , N and it, N and the sum T s 1 (T 1) t s over the remainder terms appearing in 2, N are o p (1) by arguments analogous to those for 2, N and 5, N (see B.53 and B.56). Finally, by arguments similar to the proof for the 1 N T ~2 consistency of the estimate of the third moment, it follows that ~2, N it, N NT i 1 t 1 converges in probability to 2 . As a consequence, ~( 4) ( 4) o (1) , ( 4) O(1) by Assumption 1, and that ~ ( 4) ,N ,N p Op (1) . We next consider the fourth moment of vit , N and its estimate, which can be written as (compare Gilbert, 2002, p. 48ff.): 24 v( 4) E ( it4 ) E ( is it3 ) 3E ( is it )[ E ( it2 ) E ( is it )] for any given i and s t , (B.62a) ~v(,4N) N T T 1 N T ~4 1 it , N ~is, N ~it3, N NT i 1 s 1 NT (T 1) i 1 s 1 t 1 (B.62b) ts N T T N T N T T 3 1 ~ ~ ( 1 ~2 ~is, N ~it , N ) is , N it , N it , N NT (T 1) i 1 s 1 t 1 NT i 1 t 1 NT (T 1) i 1 s 1 t 1 ts ts 1, N 1, N 2, N (~2, N 2, N ) . We have already shown that 1, N converges in probability to ( 4) 3 2 v2 and that 3 2, N (~2, N 2, N ) converges in probability to 3 2 v2 . Next, expanding 1, N , we obtain 1, N 1 N T ( it , N it , N ) 4 NT i 1 t 1 (B.63) 1 N T 2 ( it , N 2 it , Nit , N it2, N )( it2, N 2 it , Nit , N it2, N ) NT i 1 t 1 1 N T 4 ( it, N 2 it3, Nit, N it2, Nit2, N ) NT i 1 t 1 1 N T (2 it3, Nit , N 4 it2, Nit2, N 2 it , Nit3, N ) NT i 1 t 1 1 N T 2 2 ( it , Nit , N 2 it , Nit3, N it4, N ) NT i 1 t 1 1 N T 4 ( it, N 4 it3, Nit, N 6 it2, Nit2, N 4 it, Nit3, N it4, N ) NT i 1 t 1 1 N T 4 it , N converges in probability to ( 4) v( 4) 6 2 v2 and that the NT i 1 s 1 remainder terms of 1, N are all o p (1) , it follows that ~v(,4N) v( 4) o p (1) , v( 4 ) O (1) by Observing that Assumption 1, and that ~v(,4N) Op (1) . III. Proof of Theorem 3 (Variance-Covariance Estimation) Lemma B.2 S Suppose Assumptions 1-4 hold. Furthermore, assume that sup N m , N 1 , and that the row m 1 and column sums of M m, N , m 1,..., S are bounded uniformly in absolute value by 1 and ρs , N , s 1,..., S , be estimators, satisfying some finite constant respectively. Let ~v2 , ~12 , and ~ 25 ~v2 v2 op (1) , ~12 12 o p (1) , ~ρs , N ρs , N o p (1) , s 1,..., S . Let the NT NT or N N matrix FN be of the form (compare Lemmata 1 and 2): S (a) Fv, N [I T (I N m , N Mm , N ) 1 ]H N , m 1 S F , N (eT I N )[IT (I N m , N Mm , N ) 1 ]H N , m 1 S (b) * Fv*,N ( v 2Q0, N 1 2Q1, N )H*N Ωε,1N [IT (I N m, N M m, N )]H N , m 1 ** ,N F S [ (eT I N )]H*N [ 1 2 (eT I N )][ IT (I N m, N M m , N )]H N , 2 1 m 1 where H N is a N P* matrix whose elements are bounded uniformly in absolute value by ~ ~ ~ ~ some constant c . The corresponding estimates Fv, N , F , N , Fv*,*N , and F**, N are defined ~ analogously, replacing v2 , 12 , Ωε, N , and m, N , m 1,..., S , with ~v2 ,~12 , Ωε , N , and ~m, N , m 1,..., S , respectively. ~ ~ (i) Then, N 1FN FN N 1FN FN o p (1) and N 1FN FN O (1) . (ii) Let a N be some NT 1 or N 1 vector, whose elements are bounded uniformly in ~ absolute value. Then N 1FN a N N 1FN a N o p (1) and N 1Fv, N a N O(1) . Proof of part (i) of Lemma B.2 S Under the maintained assumptions there exists a * with sup m , N * 1 . It follows m 1 immediately by the properties of the matrices M m, N that the row and column sums of *M m, N , m 1,..., S are bounded uniformly in absolute value by 1 and some finite constant respectively. For later reference, note that the elements of the vector *k MkN h.s , N are also bounded uniformly in absolute value by c for some finite integer k. S Next define FN G N K N with K N [IT (I N m , N Mm , N ) 1 ]H N . Denote the (r,s)-th m 1 ~ ~ element of the difference N FN FN N 1FN FN as νN , which is given by 1 ~ ~ N N 1 ( f.r, N f.s , N f.r , N f.s , N ) , r, s 1,..., P* , (B.64a) or ~ ~ ~ ~ N N 1 (k.r , N GN G N k .s , N k.r , N GN G N k .s , N ) . (B.64b) 26 ~ ~ ~ Define further E N GN G N , such that N N 1 (k.r ,N E N k .s ,N k.r , N E N k .s ,N ) . Proof under Assumption (a) ~ 8 Consider first the case FN F , N ; then it holds that E N E N TQ1, N . (The subsequent proof ~ also covers the case with Hence, FN Fv, N E N EN I NT .) ~ 3 ~ N N 1 (k.r ,N E N k .s ,N k.r ,N E N k .s ,N ) , which can be written as N i , N with i 1 ~ ~ 1, N N 1 (k .r , N k .r , N )E N (k .s , N k .s , N ) (B.65) ~ 2, N N 1 (k .r , N k .r , N )E N k .s , N ~ 3, N N 1k.r , N E N (k .s , N k .s , N ) Note that S S ~ k .s , N k .s , N [IT (I N ~m , N Mm, N ) h.s , N [IT (I N m , N Mm, N ) 1 ]h.s , N m 1 (B.66) m 1 We next show that i , N o p (1) , i 1,...,3 , invoking the following theorem (see, e.g., Resnik, p X if and 1999, p. 171): Let ( X , X N , N 1 ) be real valued random variables. Then, X N only if each subsequence X N a contains a further subsequence X N a that converges almost surely to X . As we show below we will be confronted with terms of the form: (Nk ,l ) N 1 p*l k h.r , N (IT MNl )EN (IT MNk ) h.s,N , (B.67) where M N is a matrix, whose row and column sums are bounded uniformly in absolute value by some constant c M . By the properties of E N TQ1, N , the row and column sums of the matrix (IT M Nl )E N (IT M Nk ) are bounded uniformly in absolute value, and (Nk ,l ) O(1) (compare Remark A.1 in the Appendix). 8 For FN F , N , we have G N (eT I N ) . Hence GN G N (eT I N ) (eT I N ) (eT eT I N ) (JT I N ) TQ1, N . 27 Now, let the index N a denote some subsequence. In light of the aforementioned equivalence, there exists a subsequence of this subsequence ( N a ) such that for events A , with P( AC ) 0 , it holds that ~ ρm , N a ( ) ρm , N a 0 , m 1,..., S (B.68) and that for some N a N , (Nka,l ) ( ) K for some K with 0 K , and (B.69) S S ~ρ m , N a m 1 ( ) p** , where p** sup N m, N p* m1 2 In the following, assume that N a N . Since 1. S ~ m 1 m , N a (B.70) ( ) 1 , it follows from Horn and S Johnson (1985, p. 301) that (I N ~m, N a ( )M m, N a ) is invertible, such that m 1 S S ~ k .s , N a k .s , N a [IT (I N ~m, N a ( )Mm, N a ) 1 IT (I N m , N a Mm, N a ) 1 ]h.s , N a (B.71) m 1 m 1 l l S S {IT [I N ~m, N a ( )Mm, N a ] IT [I N m, N a Mm, N a ]}h.s , N a l 1 m 1 l 1 m 1 l l S S ~ IT m, N a ( )Mm, N a m, N a Mm, N a h.s , N a . l 1 m 1 m 1 Substituting into 1, N a , we have ~ ~ 1, N N a 1 (k .r , N k .r , N )E N (k .s , N k .s , N ) a a a a a a (B.72) l l S S ~ N a h.r , N a IT m, N a ( )M m, N a m, N a M m, N a l 1 m 1 m 1 k k S ~ S E N a I T m, N a ( )M m, N a m, N a M m, N a h .s , N a k 1 m 1 m1 1 l l S S ~ N a h.r , N a IT m, N a ( )M m, N a m, N a M m, N a k 1 l 1 m 1 m 1 1 28 k k S S ~ E N a I T m, N a ( )M m, N a m, N a M m, N a h .s , N a , m1 m1 where E N a TQ1, N a for FN F , N (and E N I NT for FN Fv , N ). A single element with index (k,l) of this infinite double sum over k and l is given by l l S S ~ N a h.r , N a IT m, N a ( )M m, N a m, N a M m, N a . m 1 m 1 1 k k S S ~ E N a IT m, N a ( )M m, N a m, N a M m, N a h.s , N a . m 1 m 1 (B.73) ρ N a ( ) there exist matrices M N a and M N a , Next note that for any values of ρ N a and any ~ whose row and column sums are bounded uniformly in absolute value, such that: S m 1 m , N a S Mm , N a m, N a M N a and m 1 S ~ m 1 m , N a S ( )Mm , N a ~m, N a ( )M N a . (B.74) m 1 M N a and M N a can thus be factored out of the sum, yielding l l S l S l ~ m, N a ( ) M N a m, N a M N a . m 1 m 1 (B.75) S S By the same reasoning, for any values of ~m, N a ( ) and m, N a , there exists a matrix m 1 m1 MN a , whose row and column sums are bounded uniformly in absolute value, such that: l l l l S l S l S ~ S l ~ m, N a ( ) M N a m, N a M N a m, N a ( ) m, N a M N a . m 1 m 1 m 1 m 1 (B.76) Substituting MN a into 1, N a , we obtain 1, N a l l S S l ~ N a h.r , Na IT m, Na ( ) m, Na MNa k 1 l 1 m1 m1 1 29 k k S S k ~ E N a I T m, N a ( ) m, N a M N a h .s , N a . m1 m1 Hence, we can write 1, N X N( k,l ) , a k 1 l 1 (B.77) a where X N( ka,l ) aN( ka,l )(Nka,l ) with aN( ka,l ) l l S S ~ m, N a m, N a m 1 m 1 *l k k S S ~ m, N a m, N a m 1 m 1 *k (Nka,l ) Na1 p*l k h.r , N a (IT MNl a )EN a (IT MNk )h.s, N a . and (B.78) (B.79) Note that aN( ka,l ) 0 in light of the aforementioned results and since (Nka,l ) K , it follows that X N( ka ,l ) 0 . Moreover, l aN a l l S ~ S S ~ S m , N a m , N a m , N a m , N a m 1 m 1 m 1 m 1 l k * 2 ** * l k * 2 ** 4 ** * * k (B.80) lk . For N a N , (Nka,l ) ( ) K , such that we have X ( k ,l ) N a B (l , k ) 4 K ** * l k . (B.81) Hence, there exists a dominating function B ( l ,k ) for all values of k,l. Moreover, since ** 1 by construction, we also have that * 30 B (l , k ) B (l , k ) , k 1 l 1 (B.82) k 1 l 1 i.e., the dominating function is integrable (summable). It follows from dominated convergence that lim N a 1, N a 0 . (B.83) The same holds for the i , N a , i 2,...,3 . It follows that i , N a 0 as N a and in light of Resnik, 1999, p. 171) that N o p (1) . ~ ~ Thus, N 1FN FN N 1FN FN o p (1) . That N 1FN E N FN O(1) follows from the properties of FN and E N (compare Remark A.1). As already mentioned above, an analogous proof applies ~ to the case FN Fv, N with E N EN I NT . Proof under Assumption (b) We first consider the case * . Then, FN Fv*,N FN G N K N with G N Ωε,1N and S K N [IT (I N m , N M m , N )]H N , such that E N Ωε,1N Ωε,1N Ωε,2N ( v4Q0, N 14Q1, N ) m 1 ~ ~ and EN Ωε,2N (~v4Q0, N ~14Q1, N ) . Notice that the subsequent proof also covers the case FN F**, N . 9 7 ~ ~ ~ First, we rewrite N N 1 (k.r ,N E N k .s ,N k.r , N E N k .s ,N ) as N i , N , with i 1 ~ ~ ~ ~ ~ 1, N N 1 (k .r , N k .r , N )(E N E N )(k .s , N k .s , N ) , (B.84) 2, N N 1 (k .r , N k .r , N )(E N E N )k .s , N , ~ ~ 3, N N 1k.r , N (E N E N )(k .s , N k .s , N ) , ~ 4, N N 1k.r , N (E N E N )k .s , N , ~ ~ 5, N N 1 (k .r , N k .r , N )E N (k .s , N k .s , N ) , 9 S In case FN F**, N [ 1 2 (eT I N )][ IT (I N m, N M m, N )]H N , we have m 1 G N [ 12 (eT I N )] , such that EN 12 (eT I N )12 (eT I N ) 14 (eT eT I N ) 14Q1, N . 31 ~ 6, N N 1 (k .r , N k .r , N )E N k .s , N , ~ 7, N N 1k.r , N E N (k .s , N k .s , N ) . For the sake of simplicity, we define E N E0, N E1, N , where E0, N v4Q0, N and E1, N 14Q1, N and consider only E1, N in the following; it is obvious that an analogous proof applies to E 0, N and thus also E N . Next, consider S S ~ k .s , N k .s , N [IT (I N ~m, N M m, N )]h.s , N [IT (I N m , N M m , N )]h.s , N . m 1 (B.85) m 1 We next show that i , N o p (1) , i 1,...,7 . Consider S S S m 1 m 1 m 1 I T [I N ~m, N M m, N ] I T (I N m, N M m, N ) I T ( ~m, N m , N )M m, N (B.86) and note that for any values of ~m, N and m, N , there exists a matrix M N , whose row and column sums are uniformly bounded in absolute value, such that S S m 1 m 1 IT ( ~m, N m, N )M m, N IT ( ~m, N m, N )M N . (B.87) Substituting MN a into (the first part of) the expression for 1, N , we obtain S S ~ 1, N N 1h.r , N [IT ( ~m, N m, N )M N ]E1, N [IT ( ~m, N m, N )M N ]h.s ,N m1 (B.88) m1 S S m1 m1 N 1h.r , N [IT ( ~m, N m, N )M N ]E1, N [I T ( ~m, N m, N )M N ]h.s , N . 2 S ~ N ( ~m,N m,N ) h.r ,N (IT MN )E1,N (IT M N )h.s ,N m1 1 2 S N ( ~m,N m,N ) h.r ,N (IT MN )E1, N (IT M N )h.s ,N . m1 1 2 S ~14 N 1 ( ~m, N m, N ) h.r , N (I T MN )Q1, N (IT M N )h.s , N m1 32 2 S N ( ~m, N m, N ) h.r , N (IT MN )Q1, N (IT M N )h.s , N . m 1 4 1 1 2 S (~14 14 ) N 1 ( ~m,N m,N ) h.r ,N (IT MN )Q1,N (IT M N )h.s ,N m1 o p (1)o p (1)O(1) o p (1) . The same holds for the i , N a , i 2,...,7 . Obviously, an analogous proof applies for E 0 , N and E N E0, N E1, N . N o p (1) . thus also for It follows that Thus, ~ ~ ~ N 1FN E N FN N 1FN E N FN o p (1) . That N 1FN E N FN O(1) follows from the properties the FN elements of E1, N ( E 0, N ). Proof of part (ii) of Lemma B.2 ~ Denote the r-th element of the difference N 1FN a N N 1FN a N as wN , which is given by ~ wN N 1 ( f.r, N a N f.r , N a N ) , r 1,..., P* , (B.89) or ~ ~ wN N 1 (k.r , N GN a N k.r , N GN a N ) . Proof under Assumption (a) S Consider first the case FN F , N (eT I N )[IT (I N m, N Mm, N ) 1 ]H N . We then have m 1 FN G N K N S with K N [IT (I N m , N Mm , N ) 1 ]H N m 1 and G N (eT I N ) . Hence, GN (eT I N ) . Obviously, the subsequent proof also covers the case FN Fv, N . It follows that ~ wN N 1 (k.r , N k.r , N ) aN ) , (B.90) where the elements of the vector aN (eT I N )a N are uniformly bounded in absolute value since the row and columns sums of (eT I N ) are uniformly bounded in absolute value and since the elements of a N are uniformly bounded in absolute value. (See Remark A.1 in the Appendix.) 33 In light of the properties of Q 0, N , Q1, N , and aN , it follows directly from the proof of part (i) ~ of Lemma B.2, where we showed that 2, N N 1 (k .r , N k .r , N )E N k .s , N o p (1) , that ~ wN o p (1) and thus N 1FN a N N 1FN a N o p (1) . That N 1FN a N O(1) follows immediately from the properties of FN and a N by Remark A.1. Proof under Assumption (b) Consider first the case FN Fv*,*N . Then, we have FN G N K N with G N Ωε,1N and S K N [IT (I N m , N M m , N )]H N . Notice that the subsequent proof also covers the case m 1 FN F**, N . In that case, ~ ~ wN N 1 (k.r , N GN a N k.r , N GN a N ) ~ ~ ~ N 1 (k .r , N k .r , N )Ωε,1N a N N 1k .r , N (Ωε,1N Ωε,1N )a N . (B.91) ~ Substituting Ωε,1N ~v2Q 0, N ~2Q1, N , we have wN wNv w1N , where ~ wNv N 1k .r , N ~v,2N Q 0, N a N N 1k .r , N v2 Q 0, N a N , ~ w1N ~1, 2N N 1k.r , N Q1, N a N 12k.r , N Q1, N a N . Considering wNv , we have wNv w1v, N w2v, N , where ~ w1v, N ~v2 N 1 (k.r , N k.r , N )Q0, N a N , (B.92) w2v, N (~v2 v2 ) N 1k.r , N Q0, N a N . The first term w1v, N o p (1) by arguments analogous to the proof of Part (i) under Assumption (b) by the consistency of ~s, N , s 1,..., S . In light of the properties of k .r , N , Q 0, N , and a N , N 1k Q a O(1) and since (~ 2 2 ) o (1) , it follows that wv o (1) by .r , N 0, N N v, N v p 2, N p ~ substituting Ωε,1N Ωε,1N (~v2 v2 )Q0, N (~2 2 )Q1, N . Analogous arguments apply to wN such that wN o p (1) . Since this holds for all r 1,..., P* , we have that ~ N 1Fv*,*N a N N 1Fv*,*N a N o p (1) . That N 1Fv*,*N a N O(1) follows from immediately from the properties of Fv*,*N and a N by Remark A.1. Analogous proof can be used to show that ~ N 1Fv*,*N a N N 1Fv*,*N a N o p (1) and that N 1Fv, N a N O(1) , which completes the proof. Proof of Theorem 3 34 ~ As part of proving Theorem 3 it has to be shown that Ψ N Ψ N o p (1) . Observe that in light ~ of (15), each element of Ψ N and the corresponding element of Ψ N can be written as ~ ~ ~ ~ ~ Erp,,sq, N Erp,,sq,,*N Erp,,sq,,*N* Erp,,sq,,*N** Erp,,sq,,*N*** , (B.93) Erp,,sq, N Erp,,sq,,*N Erp,,sq,,*N* Erp,,sq,,*N** Erp,,sq,,*N*** , for p, q 1,...,4, (a, b) , r , s 1,..., S 1 , where 10 ~ Erp,,sq,,*N 2 N 1~v4,NTr (Avp,r ,N Avq,s,N ) 2 N 1~4,NTr (Ap,r ,N Aq,s,N ) 4~2~v2 N 1Tr[( Avp,,r ,N )Avq,,s,N ] , ~ Erp,,sq,,*N* ~v2, N N 1~ a pv , r, N ~ aqv, s , N ~2, N N 1~ a p, r, N ~ aq, s , N N ~ p ,q ,*** ~ (3) 1 NT ~ v v v v ( 3) 1 ~ ~ E r ,s , N v , N N (a p ,r ,i , N aq ,s ,ii, N a p ,r ,ii, N aq ,s ,i , N ) , N N (a~p,r ,i , N aq,s ,ii, N a~p,r ,ii, N aq,s ,i , N ) , i 1 i 1 NT N ~ Erp,,sq,,*N*** (~v(,4N) 3~v4, N ) N 1 a vp , r ,ii, N a~qv, s ,ii, N (~( 4, N) 3~4, N ) N 1 a~p, r ,ii, N aq, s ,ii, N , i 1 i 1 and Erp,,sq,,*N 2N 1 v4Tr (Avp,r ,N Avq,s,N ) 2N 1 4Tr (Ap,r ,N Aq,s,N ) 4 2 v2 N 1Tr[( Avp,,r ,N )Avq,,s,N ] , Erp,,sq,,*N* v2 N 1avp , r, N avq , s , N 2 N 1a p , r, N a q, s , N , NT N E rp,,sq,,*N** v(3) N 1 (a vp ,r ,i , N aqv,s ,ii, N a vp ,r ,ii, N aqv,s ,i , N ) (3) N 1 (a p ,r ,i , N aq,s ,ii, N a p ,r ,ii, N aq,s ,i , N ) , i 1 E p , q ,**** r , s, N ( ( 4) v i 1 3 ) N 4 v 1 NT a i 1 v p , r , ii , N v q , s , ii , N a ( 3 ) N ( 4) 4 1 N a i 1 p , r , ii , N aq, s ,ii, N . ~ ~ ~ To proof that Ψ N Ψ N o p (1) , we show that Erp,,sq,,*N Erp,,sq,,*N o p (1) , Erp,,sq,,*N* Erp,,sq,,*N* o p (1) , ~ ~ Erp,,sq,,*N** Erp,,sq,,*N** o p (1) , and Erp,,sq,,*N*** Erp,,sq,,*N*** o p (1) . ~ i) Proof that Erp,,sq,,*N Erp,,sq,,*N o p (1) Consider ~ Erp,,sq,,*N Erp,,sq,,*N (~v4, N v4 )2 N 1Tr (Avp , r , N Avq, s , N ) (~4, N 4 )2 N 1Tr (A p , r , N A q, s , N ) , 10 See Eqs. (34) in the main text for the structure of the matrix Ψ N and the proper indexation of its elements. 35 and note that the row and column sums of the matrices A vp , r , N , Avq, s , N , A p , r , N , Aq, s , N , A vp,,r , N , and Avq,,s, N are uniformly bounded in absolute value by some constant, say K A . It follows that 2 N 1Tr (Avp, r , N Avq, s , N ) 2 N 1NTK A2 2TK A2 O(1) , 2 N 1Tr (Ap, r , N Aq, s , N ) 2Ka2 2 N 1NK A2 O(1) , 4 N 1Tr[( Avp,,r , N )Avq,,s , N )] 4 N 1NTK A2 4TK A2 O(1) , ~ and thus Erp,,sq,,*N Erp,,sq,,*N o p (1) . ~ ii) Proof that Erp,,sq,,*N* Erp,,sq,,*N* o p (1) Note that ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ 1~ ~ P ~2 Erp,,sq,,*N* Erp,,sq,,*N* ~v2, N N 1α p , r , N N Fv , N Fv , N PN α q , s , N , N N α p , r , N PN F , N F , N PN α q , s , N v2 N 1αp, r , N PN Fv, N Fv, N PN αq, s, N 2 N 1αp, r , N PN F , N F , N PN αq, s, N v p , q ,**, , where pr ,,qs,,**, N r , s , N ~ ~ ~ ~ ~ v 2 1 ~ P pr ,,qs ,,**, ~v2, N N 1α N p ,r , N N Fv , N Fv , N PN α q ,s , N v N α p ,r , N PN Fv , N Fv , N PN α q ,s , N , and ~ ~ ~ ~ ~ 2 1 ~ P pr ,,qs ,,**, ~2,N N 1α N p ,r , N N F , N F , N PN α q ,s , N N α p ,r , N PN F , N F , N PN α q ,s , N . v Consider pr ,,qs ,,**, N , which can be written as ~ ~ ~ ~ ~ v 2 1 ~ P pr ,,qs ,,**, ~v2, N N 1α N p ,r , N N Fv , N Fv , N PN α q ,s , N v N α p ,r , N PN Fv , N Fv , N PN α q ,s , N v pr ,,qs ,,**, (~v2,N v2 ) N 1αp,r ,N PN Fv,N Fv,N PN α q,s ,N N ~ ~ ~ ~ ~ 1 ~ P ~v2, N ( N 1α p , r , N N Fv , N Fv , N PN α q , s , N N α p , r , N PN Fv , N Fv , N PN α q , s , N ) . ~ Note that (~v2, N v2 ) o p (1) and that ~v2, N Op (1) . By assumption PN PN o p (1) , ~ PN O(1) and thus PN Op (1) , where the dimension of PN is P* P ; by Lemma B.1 ~ O (1) , where the dimension of α is P 1 . ~ α o (1) , α O(1) and thus α α N N p N N p N ~ ~ Moreover, N 1Fv, N Fv, N N 1Fv, N Fv, N is o p (1) and N 1Fv, N Fv, N O(1) by Lemma B.2. It ~ ~ v follows that pr ,,qs ,,**, o p (1) . By Lemma B.2 it also holds that N 1F , N F , N N 1F , N F , N and N ~ accordingly, pr ,,qs ,,**, o p (1) . It follows that Erp,,sq,,*N* Erp,,sq,,*N* o p (1) , and in light of Lemma B.2 N 36 it is readily observed that this also holds when F**, N and Fv*,*N are used instead of F , N and Fv, N . ~ iii) Proof that Erp,,sq,,*N** Erp,,sq,,*N** o p (1) Observe that ~ Erp,,sq,,*N** Erp,,sq,,*N** pr ,,qs ,,*N**,v pr ,,qs ,,*N**, , where NT NT i 1 i 1 N N i 1 i 1 pr ,,qs,,*N**,v ~v(,3N) N 1 (a~pv,r ,i , N aqv, s ,ii, N a vp ,r ,ii, N a~qv, s ,i , N ) v(3, N) N 1 (a vp , r ,i , N aqv, s ,ii, N a vp , r ,ii, N aqv, s ,i , N ) , pr ,,qs ,,*N**, ~(3, N) N 1 (a~p,r ,i , N aq, s ,ii, N a p ,r ,ii, N a~q, s ,i , N ) (3, )N N 1 (a p , r ,i , N aq, s ,ii, N a p , r ,ii, N aq, s ,i , N ) . ~ ~ ~ v a pv , r , N Fv , N PN α Next consider pr ,, qs ,,*N**,v and note that ~ p , r , N ; moreover, define a p , r , N as NT 1 vector made up of the main diagonal elements a vp , r ,ii, N of matrix A vp , r , N . It follows that a pv , r,i , N avq , s , N avp , r, N avq , s , N ) ~v(,3N) N 1 (a vp , r, N ~ aqv, s , N a vp , r, N avq , s , N ) pr ,,qs,,*N**,v ~v(,3N) N 1 (~ (~v(,3N) v(3, N) ) N 1 (avp , r, N aqv, s , N a vp , r, N avq , s , N ) . Note that v(3) O(1) , (~v(,3N) v(3) ) o p (1) , ~v(,3N) Op (1) . Moreover, by the properties of avp , r, N , a vp , r , N , N 1 (avp ,r , N aqv, s , N a vp ,r , N avq , s , N ) O(1) such that the term in the second line is ~ ~ v v ~ P v v o p (1) . Next note that ~ a pv , r , N a vp , r , N α p , r , N N Fv , N a p , r , N , and a p , r , N a p , r , N α p , r , N PN Fv , N a p , r , N , such that the first term in the first line is given by ~ 1~ ~ ~ P v 1 v α p , r , N N N Fv , N a p , r , N α p , r , N PN N Fv , N a p , r , N . By assumption PN PN o p (1) , PN O(1) ~ ~ α o (1) , and thus PN Op (1) , where the dimension of PN is P* P . Moreover, α N N p ~ O (1) , where the dimension of α is P 1 . By Lemma B.2, part α O(1) and thus α N N p N ~ (ii), we have that N 1Fv, N a vp , r , N N 1Fv, N a vp, r , N o p (1) and N 1Fv, N a vp, r , N O(1) . It follows ~ 1~ ~ P v 1 v that α p , r , N N N Fv , N a p , r , N α p , r , N PN N Fv , N a p , r , N o p (1) . By analogous arguments the second term in the first line is o p (1) , from which it follows that pr ,,qs ,,*N**,v op (1) . An ~ analogous proof can be used to show that pr ,,qs ,,*N**, o p (1) . Hence, Erp,,sq,,*N** Erp,,sq,,*N** o p (1) . 37 ~ iv) Proof that Erp,,sq,,*N*** Erp,,sq,,*N*** o p (1) Consider ~ Erp,,sq,,*N*** Erp,,sq,,*N*** [(~v( 4) 3~v4 ) ( v( 4) 3 v4 )] N 1a vp , r, N aqv, s , N [(~( 4 ) 3~4 ) ( ( 4 ) 3 4 )] N 1a p, r, N aq, s , N . By the properties of the matrices A vp , r , N , Avq, s , N , A p , r , N , and Aq, s , N , it follows by arguments analogous to above that N 1a vp , r, N aqv, s , N O(1) and that N 1a p, r, N aq, s , N O(1) . Since (~v(,4N) v( 4) ) op (1) and (~( 4, N) ( 4) ) op (1) , and since (~v2, N v2 ) o p (1) , (~2, N 2 ) o p (1) , and thus also (~v4, N v4 ) o p (1) , (~4, N 4 ) o p (1) , it follows that ~ Erp,,sq,,*N*** Erp,,sq,,*N*** o p (1) . ~ ~ ~ We have shown that Erp,,sq,,*N Erp,,sq,,*N o p (1) , Erp,,sq,,*N* Erp,,sq,,*N* o p (1) , Erp,,sq,,*N** Erp,,sq,,*N** o p (1) , ~ and Erp,,sq,,*N*** Erp,,sq,,*N*** o p (1) for all elements of Ψ N , p, q 1,...,4, (a, b) , r , s 1,..., S 1 . It ~ follows that Ψ N Ψ N o p (1) . We are now ready to prove consistency of the estimate of the variance-covariance matrix, i.e., ~ that Ω~θ Ω~θ o p (1) , where Ω ~θ N (Θ N ) (JN Θ N J N ) 1 JN Θ N Ψ N Θ N J N (JN Θ N J N ) 1 . As N N ~ ~ proved above, we have Ψ N Ψ N o p (1) . By Assumption 5, we have Θ N Θ N o p (1) , ~ Θ N O(1) and Θ N O p (1) . Let ΞN JN ΘJ N B N Γ N Θ N Γ N B N as defined in Theorem 2 ~ ~ ~ ~ ~ 11 ~ ~ ~~ ~ and ΞN JN ΘJ N B N ΓN Θ N Γ N B N . In Theorem 2, we showed that J N O p (1) , J N O(1) , ~ ~ ~ and J N J N o p (1) and that ΞN O p (1) , ΞN1 Op (1) and that ΞN ΞN1 o p (1) . It now ~ follows that Ω~θ Ω~θ o p (1) . N N ~ ~ There is a slight discrepancy to the definition of Ξ N in Theorem 2: Here B N is used rather ρN and ρN are than B N , which does not affect the proof, however, noting that both ~ 11 consistent. 38 ~ and Other Model Parameters) III. Proof of Theorem 4 (Joint Distribution of ρ N The subsequent proof will focus on the case Fv, N and Fv, N ; this also convers the case for Fv*,*N and F**, N . The first line in Theorem 4 holds in light of Assumption 7 (for N 1/ 2 Δ N ), bearing in ~ mind that Tv, N Fv, N PN , T , N F , N PN , and Theorem 2 (for N 1/ 2 (θN θ N ) ). d We next prove that ξ o, N Ψo1,N/ 2[( NT )1 / 2 ξN FN ,qN ] N (0, I P* 4 S 2 ) by verifying that the assumptions of the central limit theorem A.1 by Kelejian and Prucha (2010) are fulfilled. Note that min (Ψo, N ) cΨ* o 0 by assumption. In Theorem 2, we verified that the stacked innovations ξ N , the matrices A1, s , N ,..., A 4, s , N , s 1,..., S , A a, N , and A b, N , and the vectors a1, s, N , …, a 4, s, N , s 1,..., S , a a, N , and ab, N satisfy the assumptions of central limit theorem by Kelejian and Prucha (2010, Theorem A.1). Next, consider the two blocks of FN (Fv, N , F , N ) , which are given by S Fv, N [I T (I N m , N Mm , N ) 1 ]H N , and m 1 S F , N [ 1 2 (eT I N )]Ωε , N [IT (I N m , N Mm , N ) 1 ]H N . m 1 S Since the row and columns sums of [ 12 (eT I N )] , Ωε, N , and [I T (I N m, N Mm , N ) 1 ] m 1 are uniformly bounded in absolute value and since the elements of the matrix H N are uniformly bounded in absolute value, it follows that the elements of FN are also uniformly bounded in absolute value. Hence, the linear form FN ξ N Fv, N v N F , N μ N also fulfils the d assumptions of Theorem A.1. As a consequence, ξ o, N N (0, I P* 4 S 2 ) . ~ In the proofs of Theorems 2 and 3, we showed that Ψ N Ψ N o p (1) , Ψ N O(1) , and ~ Ψ N O p (1) . By analogous arguments, this also holds for the submatrices Ψ , N and Ψ , N . ~ ~ ~ Hence, Ψo, N Ψo, N o p (1) , Ψ o , N O(1) and Ψo, N Ψo, N o p (1) , and thus Ψ o, N Op (1) . ~ ~ ~ PN PN o p (1) , PN O(1) , and PN Op (1) as well as Θ N Θ N o p (1) , ~ ~ Θ N O(1) and Θ N O p (1) . In the proof of Theorem 2 we showed that J N J N o p (1) , ~ ~ ~ ~ J N O(1) , and J N O p (1) , and that (JN Θ N J N ) (JN Θ N J N )1 o p (1) , (JN Θ N J N ) 1 O(1) , By assumption 39 ~ ~ ~ ~ and ( JN Θ N J N ) Op (1) . It now follows that Ωo, N Ωo, N o p (1) and Ω o , N O(1) and thus ~ Ωo, N Op (1) . APPENDIX C. Proof of Lemma 1. In light of Eqs. (4a) and (4b), Assumptions 3 and 8, as well as sup N β N b , it follows that all columns of Z N ( X N , YN ) are of the form N π N Π N ε N , where the elements of the vector π N and the row and column sums of the matrix Π N are bounded uniformly in absolute value (see Remark A.1 in Appendix A). It follows from Lemma C.2 in Kelejian and Prucha (2010) that the fourth moments of the elements of the matrix D N Z N are bounded uniformly by some finite constant and that Assumption 6 holds. Next, note that ~ ~ ~ ( NT )1 / 2 (δN δ N ) PN ( NT )1 / 2 Fv, N v N PN ( NT )1 / 2 F , N μ N , ~ where PN is defined in the Lemma, and S Fv , N [IT (I N m , N Mm, N ) 1 ]H N and m 1 S F , N [ 1 2 (eT I N )]Ω , N [IT (I N m , N Mm , N ) 1 ]H N . m 1 ~ In light of Assumption 10, PN PN o p (1) and PN O(1) , with PN as defined in the Lemma. By Assumptions 2, 3 and 9, the elements of Fv, N and F , N are bounded uniformly in absolute value. By Assumption 2, E ( v N ) 0 , E (μ N ) 0 , and the diagonal variance-covariance matrices of v N and μ N have uniformly bounded elements. Thus, E[( NT )1 / 2 Fv, N v N ] 0 and the elements of the variance-covariance matrix of N 1 / 2Fv, N v N , i.e., ( NT )1 v2Fv,N Fv,N , are bounded uniformly in absolute value (see Remark A.1 in Appendix A). Moreover, E[( NT )1 / 2 F2, N μ N ] 0 , and the elements of the variance-covariance matrix of N 1 / 2F , N μ N , i.e., ( NT )1 2 F ,N F ,N , are bounded uniformly in absolute value. It follows from Chebychev’s consequently that ( NT )1 / 2 Fv, N v N Op (1) , ( NT )1 / 2 F , N μ N Op (1) , ~ ( NT )1 / 2 (δN δ N ) PN ( NT )1 / 2 Fv, N v N PN ( NT )1 / 2 F , N μ N o p (1) inequality and and 40 PN ( NT )1 / 2 Fv, N v N PN ( NT )1 / 2 F , N μ N Op (1) . This completes the proof, recalling that TN (Tv, N , T , N ) (PN Fv, N , PN F , N ) . Proof of Lemma 2. The FGTSLS estimator is given by ˆ ˆ ** ** 1 ˆ ** ** ** ** ** δ N (Z N Z N ) Z N y N , where y N Z N δ N u N with S u*N* Ω1,N/ 2 [I T (I N m , N M m, N )]u N . m 1 ˆ Substituting Z*N* PH** Z*N* H *N* (H *N*H *N* ) 1 H *N*Z*N* , we obtain N ˆ ( NT )1/ 2 [δ N δ N ] ( NT )1/ 2 Δ*N* ( NT ) 1/ 2 PN** H *N*u*N* , with PN** {[( NT ) 1 Z*N*H *N* ][( NT ) 1 H *N*H *N* ]1[( NT ) 1 H *N*Z*N* ]}1[( NT ) 1 Z*N*H *N* ][( NT ) 1 H *N*H *N* ]1 . Next note that S u*N* (θ N ) u*N* (θ N ) ( m , N m, N )Ω1,N/ 2 (IT M N )u N , m 1 S (Ω1,N/ 2 Ω1,N/ 2 )[I T (I N m , N M m , N )]u N m 1 (Ω1,N/ 2 Ω1,N/ 2 )(IT M N )u N S (Ω1,N/ 2 Ω1,N/ 2 ) ( m , N m , N )(I T M N )u N , m 1 where M N is a matrix, whose row and columns sums are bounded uniformly in absolute value, satisfying S ( m 1 m,N m , N )M m , N S ( m 1 m,N m,N )M N . Substituting for u *N* , we obtain 5 ˆ ( NT )1 / 2 (δ N δ N ) ( NT )1/ 2 Δ*N* d1,N d2, N d3, N d4, N d5, N di , N , where i 1 41 d1, N ( NT ) 1/ 2 PN**H*N*Ω1,N/ 2ε N , S S d2, N ( NT ) 1/ 2 PN** (θ N ) H *N* ( m, N m, N )Ω1,N/ 2 (I T M N )[I T (I N m, N M m, N ) 1 ]ε N , m1 d3, N m1 ( NT )1 / 2 PN**H*N* (Ω1,N/ 2 Ω1,N/ 2 ) ε N , S d4, N ( NT ) 1/ 2 PN**H *N* (Ω1,N/ 2 Ω1,N/ 2 )(I T M N )[I T (I N m , N M m , N ) 1 ]ε N , m 1 S S d5, N ( NT ) 1/ 2 PN**H*N* ( m, N m, N )(Ω1,N/ 2 Ω1,N/ 2 )(I T M N )[I T (I N m, N M m, N ) 1 ]ε N . m1 m1 Note that the FGTSLS estimator uses (generated) transformed instruments, which are based on the estimate θ̂ N . Observe that S H*N* H*N* ( m , N m , N )Ω1,N/ 2 (IT M N )H N m 1 S (Ω1,N/ 2 Ω1,N/ 2 )[I T (I N m , N M m , N )]H N m 1 (Ω1,N/ 2 Ω1,N/ 2 )(IT M N )H N S (Ω1,N/ 2 Ω1,N/ 2 ) ( m , N m , N )(I T M N )H N . m 1 5 Substituting for H *N* , we obtain di , N dij, N , i 1,...,5 . Considering d1, N , we have j 1 d11, N ( NT ) 1 / 2 PN**F1*,*N v N ( NT ) 1 / 2 PN**F2*,*N μ N , with F1*,*N H *N [( v2Q 0, N 12Q1, N ) H *N Ω1,N , and F2*,*N H*N 12 (eT I N ) . S d12, N ( NT ) 1 / 2 PN** ( m , N m , N )HN (IT M N )Ω1,N ε N , m 1 S d13, N ( NT ) 1/ 2 PN**HN [I T (I N m, N Mm, N )](Ω1,N/ 2 Ω1,N/ 2 )Ω1,N/ 2ε N , m1 d14, N ( NT ) 1/ 2 PN**HN (I T M N )(Ω1,N/ 2 Ω1,N/ 2 )Ω1,N/ 2ε N , S d15, N ( NT ) 1 / 2 PN** ( m , N m , N )HN (IT M N )(Ω1,N/ 2 Ω1,N/ 2 )Ω1,N/ 2ε N m 1 42 Regarding d2 , N we have S S d21, N ( NT ) 1/ 2 PN**H *N* ( ˆ m, N m, N )Ω1,N/ 2 (I T M N )[I T (I N m, N M m, N ) 1 ]ε N , m1 m1 S d22, N ( NT ) 1/ 2 PN**[ ( m, N m, N )]2 HN (I T M N )Ω1,N (I T M N )[I T (I N m, N M m, N ) 1 ]ε N , S m1 m1 S S d23, N ( NT ) 1 / 2 PN** ( m , N m , N )HN [IT (I N m , N Mm , N )] m1 m1 S (Ω1,N/ 2 Ω1,N/ 2 )Ω1,N/ 2 (I T M N )[I T (I N m, N M m, N ) 1 ]ε N , m 1 d24, N ( NT ) 1 / 2 PN** ( m , N m , N )HN (IT M N ) S m1 S 1/ 2 1 / 2 1 / 2 (Ω , N Ω , N )Ω , N (I T M N )[I T (I N m, N M m, N ) 1 ]ε N , m 1 S d25, N ( NT ) 1 / 2 PN**[ ( m , N m , N )]2 HN (IT M N ) m 1 S (Ω1,N/ 2 Ω1,N/ 2 )Ω1,N/ 2 (I T M N )[I T (I N m, N M m, N ) 1 ]ε N . m 1 Regarding d3, N we have d31, N ( NT )1 / 2 PN**H*N* (Ω1,N/ 2 Ω1,N/ 2 ) ε N , S d32, N ( NT ) 1 / 2 PN** ( m , N m , N )HN (IT M N )Ω1,N/ 2 (Ω1,N/ 2 Ω1,N/ 2 ) ε N , m 1 S d33, N ( NT ) 1/ 2 PN**HN [I T (I N m , N Mm , N )](Ω1,N/ 2 Ω1,N/ 2 )(Ω1,N/ 2 Ω1,N/ 2 ) ε N m 1 ( NT ) 1/ 2 PN**HN [I T (I N m, N Mm , N )](Ω1,N/ 2 Ω1,N/ 2 ) 2 ε N , S m 1 d34, N ( NT ) 1/ 2 PN**HN (I T M N )(Ω1,N/ 2 Ω1,N/ 2 )(Ω1,N/ 2 Ω1,N/ 2 ) ε N ( NT ) 1/ 2 PN**HN (I T M N )(Ω1,N/ 2 Ω1,N/ 2 ) 2 ε N , S d35, N ( NT ) 1/ 2 PN** ( m , N m , N )HN (I T M N )(Ω1,N/ 2 Ω1,N/ 2 )(Ω1,N/ 2 Ω1,N/ 2 ) ε N m 1 S ( NT ) 1/ 2 PN** ( m , N m , N )HN (I T M N )(Ω1,N/ 2 Ω1,N/ 2 ) 2 ε N . m 1 Regarding d4 , N we have 43 S d41, N ( NT ) 1 / 2 PN**H*N* (Ω1,N/ 2 Ω1,N/ 2 )(IT M N )[IT (I N m , N M m , N ) 1 ]ε N , m 1 S d42, N ( NT ) 1 / 2 PN** ( m , N m , N )HN (IT M N )Ω1,N/ 2 m 1 S (Ω1,N/ 2 Ω1,N/ 2 )(I T M N )[I T (I N m , N M m , N ) 1 ]ε N , m 1 d43, N ( NT ) 1/ 2 PN**HN [I T (I N m , N Mm , N )] S m1 S (Ω1,N/ 2 Ω1,N/ 2 ) 2 (I T M N )[I T (I N m, N M m, N ) 1 ]ε N m1 S d44, N ( NT ) 1/ 2 PN** HN (I T M N )(Ω1,N/ 2 Ω1,N/ 2 ) 2 (I T M N )[I T (I N m, N M m, N ) 1 ]ε N m 1 d45, N S ( NT ) 1/ 2 PN** ( m, N m, N )HN (I T M N ) m1 S (Ω1,N/ 2 Ω1,N/ 2 ) 2 (I T M N )[I T (I N m, N M m, N ) 1 ]ε N m1 Regarding d5, N we have S S d51, N ( NT ) 1/ 2 PN**H *N* ( m, N m, N )(Ω1,N/ 2 Ω1,N/ 2 )(I T M N )[I T (I N m, N M m, N ) 1 ]ε N , m 1 m 1 S d52, N ( NT ) 1/ 2 PN**[ ( m, N m, N )]2 HN (I T M N )Ω1,N/ 2 m1 S ˆ 1/ 2 Ω 1/ 2 )(I M )[I (I M ) 1 ]ε , (Ω m,N m,N N ,N ,N T N T N m 1 S d53, N ( NT ) 1/ 2 PN** ( m, N m , N )HN [I T (I N m , N Mm , N )] S m 1 m 1 S (Ω1,N/ 2 Ω1,N/ 2 ) 2 (I T M N )[I T (I N m, N M m, N ) 1 ]ε N , m1 d54, N ( NT ) 1/ 2 PN** ( m, N m, N )HN (I T M N ) S m 1 S (Ω1,N/ 2 Ω1,N/ 2 ) 2 (I T M N )[I T (I N m, N M m, N ) 1 ]ε N , m1 S d55, N ( NT ) 1/ 2 PN**[ ( m, N m , N )]2 HN (I T M N ) m 1 S (Ω1,N/ 2 Ω1,N/ 2 ) 2 (I T M N )[I T (I N m , N M m, N ) 1 ]ε N . m 1 44 Summing up we have 5 5 ( NT )1/ 2 Δ*N* dij, N . i 1 j 1 Next note that, in light of Assumption 10 and since θ N is N 1 / 2 -consistent, it follows that ˆ ** ** 1 ( NT ) 1 Z N Z N Q H**Z** Q H**H**Q H**Z** o p (1) . By Assumption 10b we also QH**Z** QH1**H**QH**Z** O(1) have and thus ( QH**Z** QH1**H**QH**Z** )1 O(1) . It follows as a special case of Pötscher and Prucha (1997, Lemma F1) that ˆ ** ** 1 1 1 [ ( NT ) 1Z o p (1) . N Z N ] (Q H**Z** Q H**H**Q H**Z** ) It follows further that PN** PN** o p (1) and PN** O(1) with PN** defined in the Lemma. Next observe that (ρ N ρ N ) o p (1) and that (Ω , N Ω , N ) o p (1)I NT . Note further that all terms dij, N except for d11, N are of the form o p (1)PN** ( NT )-1 / 2 D N ε N , where D N are NT P* S matrices involving products of (I T M N ) , [I T (I N m, N M m , N ) 1 ] , H N , Ω , N ( Ω1,N ), m 1 and Ω 1/ 2 ,N 1/ 2 ,N (Ω elements of ). By the maintained assumptions regarding these matrices it follows that the DN are bounded uniformly in absolute value. As a consequence, E[( NT ) 1/ 2 D N ε N ] 0 and the elements of the variance-covariance matrix of ( NT ) 1/ 2 D N ε N , i.e., ( NT ) 1 D N Ω ε,N D N , are bounded uniformly in absolute value (see Remark A.1 in Appendix A). It follows from Chebychev’s inequality that ( NT ) 1/ 2 D N ε N Op (1) . As a consequence, all terms dij, N except for d11, N are o p (1) , and d11, N O p (1) . Finally, observe that d11, N ( NT ) 1 / 2 PN**Fv*,*N v N ( NT ) 1 / 2 PN**F**, N μ N , with S Fv*,*N ( v 2Q 0, N 1 2Q1, N )H*N Ω1,N [IT (I N m, N M m , N )]H N , and m 1 S F**, N [ 1 2 (eT I N )]H*N [ 1 2 (eT I N )][ IT (I N m, N M m , N )]H N , m 1 which completes the proof, recalling that T F P . ** N ** N ** N 45