Week 11Spring 2009Lecture 21. Estimation of Large Covariance Matrices: Lowerbound (II)ObserveX1; X2; : : : ; Xni.i.d. from a p-variate Gaussian distribution, N (; pp) .We assume that the covariance matrix pp= (ij)1i;jpis contained in thefollowing parameter space,F (; "; M) =n : jijj M ji jj(+1)for all i 6= j and max() 1="o(1)Theorem 1 Under the assumption (1), we haveinf^supFE^ 2 cn22+1+ clog pn: (2)Last time we have showninf^supFE^ 2 clog pn:In this lecture we will showinf^supFE^ 2 cn22+1by the Assouad’s lemma.We shall now de…ne a parameter space that is appropriate for the minimaxlower bound argument. For given positive integers k and m with 2k p and1 m k, de…ne the p p matrix B(m; k) = (bij)ppwithbij= I fi = m and m + 1 j 2k, or j = m and m + 1 i 2kg :Set k = n12+1and a = k(+1). We then de…ne the collection of 2kcovariancematrices asH =( () : () = Ip+ akXm=1mB(m; k); = (m) 2 f0; 1gk)(3)where Ipis the p p identity matrix and is a constant. It is easy to checkthat as long as 0 < < min fM; (1 ") =2g the collection H F("; M). Wewill showinf^supHE^ 2 cn22+1(4)1A Lower bound by the Assouad’s LemmaWe …rst prove equation (4). Let X1; X2; : : : ; Xnbe i.i.d. N (0; ( )) with () 2 H. Denote the joint distribution by P. We apply Assouad’s Lemma tothe parameter space H,max2H22E^ ()2 minH(;)1 () 02H (; )k2minH(;)=1kP^ PkFrom Lemma 2 we haveminH(;)1 () 02H (; ) cka2and from Lemma 3,minH(;)=1kP^ Pk c > 0thusmax2F1122E^ ()2c22k2a2 c1n22+1.Now we give proofs of auxiliary lemmas.Lemma 2 For () de…ned in (3) we haveminH(;)1 () 02H (; ) cka2:Proof of Lemma 2 : We de…ne v = (1 fk i 2kg). Let () 0 v = (wi) .There are exactly H (; ) number of wisuch that jwij = ka (just consider upperhalf of the matrix), which implies () 0 v22 H (; ) (ka)2and so () 02 H (; ) (ka)2=k cka2.Lemma 3 Let Pbe the joint distribution of n i.i.d. X1; X2; : : : ; Xnwith X1N (0; ()) and () 2 F11. Then for some c1> 0 we haveminH(;)=1kP^ Pk c1.Proof of Lemma 3 : When H; 0= 1, we will showkP0 Pk21 2K (P0jP) = 2n12tr01()12log det01()p2 n cka22for some small c > 0, where K (j) is the Kullback–Leibler divergence andthe …rst inequality follows from the well known Pinsker’s inequality (see, e.g.,Csiszár (1967)). This immediately implies the L1distance between two measuresis bounded away from 1, and then the lemma follows. Write0= D1+ () :Then12tr01()p2=12trD11().Let ibe the eigenvalues of D11(). Since D11() is similar to the sym-metric matrix 1=2() D11=2(), and1=2() D11=2()1=2()kD1k1=2() c1kD1k c1kD1k1 c2ka;then all eigenvalues i’s are real and in the interval [c2ka; c2ka], where ka =k k(+1)= k! 0. Note that the Taylor expansion yieldslog det01()= log detI + D11()= trD11() R3whereR3 c3pXi=12ifor some c3> 0.Write 1=2() = UV1=2UT, where UUT= I and V is a diagonal matrix. Itfollows from the fact that the Frobenius norm of a matrix remains the sameafter an orthogonal transformation thatpXi=12i=1=2() D11=2()2F kV k2UTD1U2F=1()2kD1k2F c4ka2:3Lecture 22. Estimation of Large Covariance Matrices: Discussi onsTopics1. Adaptive estimation2. Estimation und er di¤erent matrix norms3. Estimating fu nctionals of the covariance matrix4. Sparse covariance estimation (graphical models)5. Estimation of covariance function with functional data and its connectionto fu nctional data analysis6. Toeplitz matrix estimation7. all interactions
View Full Document