STAT 3690 Lecture 05
Hello, dear friend, you can consult us at any time if you have any questions, add WeChat: daixieit
STAT 3690 Lecture 05
2022
Block/partitioned matrix
● A partition of matrix: Suppose A11 is of p x r , A12 is of p x s, A21 is of q x r and A22 is of q x s. Make a new (p + q) x (r + s)-matrix by organizing Aij ’s in a 2 by 2 way:
A = ┐
e.g.,
A = - 0(1) 1(0) 3(2) ┐
if
A11 = ┌ 0(1) 1(0) ┐ , A12 = ┌ 3(2) ┐ , A21 = ┌ 4 5 ┐ , and A22 = ┌ 6 ┐ .
● Operations with block matrices
- Working with partitioned matrices just like ordinary matrices
- Matrix addition: if dimensions of Aij and Bij are quite the same, then
A + B = ┐ + ┐ = ┐
- Matrix multiplication: if Aij Bj亿 makes sense for each i, j, k, then
AB = ┐ ┌ ┐ = ┐
- Inverse: if A, A11 and A22 are all invertible, then
A二1 = ┌ -A22(二)1 A2(A二111)1(2)A 11(二)11(二)1(1)2 A22(二)1
* A11 .2 = A11 - A12 A22(二)1 A21
* A22 .1 = A22 - A21 A 11(二)1 A12
┐
options(digits = 4)
(Sigma = matrix(c ( 1 , .5 , .5 ,
.5 , 3 , .5 ,
.5 , .5 , 7),
nrow = 3 , ncol = 3))
# Uer乞jy t右e 乞Xuerse pj 夕αrt乞t乞pX mαtr乞g
## Method 1: by the above formula
(Sigma11 = Sigma[1 :2 , 1:2])
(Sigma12 = as.matrix(Sigma[1 :2 , 3]))
(Sigma21 = t(Sigma12))
(Sigma22 = as.matrix(Sigma[3 , 3]))
(Sigma11.2 = Sigma11 - Sigma12 %*% solve(Sigma22) %*% Sigma21)
(Sigma22.1 = Sigma22 - Sigma21 %*% solve(Sigma11) %*% Sigma12)
(SigmaInv = rbind(
cbind(solve(Sigma11.2), -solve(Sigma11.2) %*% Sigma12 %*% solve(Sigma22)),
cbind(-solve(Sigma22) %*% Sigma21 %*% solve(Sigma11.2), solve(Sigma22.1)) ))
## Method 2: solve()
solve(Sigma)
● Conditional mean vectors and covariance matrices: If X ~ (u, Σ) and
X = ┌ X(X)2(1) ┐ , u = ┌ u(u)2(1) ┐ and Σ = ┐ > 0,
where E(Xi ) = ui and cov(Xi , Xj ) = Σij , then
- E(Xi | Xj = mj ) = ui + Σij Σjj(二)1 (mj - uj ) for i j and Σjj > 0
- cov(Xi | Xj = mj ) = Σii - Σij Σjj(二)1 Σji for i j and Σjj > 0
Multivariate normal (MVN) distribution
● Standard normal random vector
φZ (5) = (2π)二夕/2 exp(-5T 5/2), 5 = [z1 , . . . , z夕]T e R夕
● (General) normal random vector
- Def: The distribution of X is MVN iff there exists q e Z+ , u e Rζ , A e Rζx夕 and Z ~ MVN夕 (0, I) such that X = AZ + u
* Limit the discussion to non-degenerate cases, i.e., rk(A) = q
* X ~ MVNζ (u, Σ), i.e.,
fx (m) = exp{-(m - u)T Σ 二1 (m - u)/2}, m e Rζ
' Σ = var(X) = AAT > 0
● Exercise:
1. Σ = AAT > 0 兮 rk(A) = q (Hint: SVD of A);
2. Σ > 0 ÷ there exists a q x q positive definite matrix, say Σ 1/2, such that Σ = Σ 1/2Σ 1/2 and Σ 二1 = Σ 二1/2Σ 二1/2 (Hint: spectral decomposition of Σ).
2022-03-17