Random Vectors and Multivariate Normal Distributions

Transcription

1 Chapter 3 Random Vectors and Multivariate Normal Distributions 3.1 Random vectors Definition Random vector. Random vectors are vectors of random 75

2 variables. For instance, X = X 1 X 2., where each element represent a random variable, is a random vector. Definition Mean and covariance matrix of a random vector. The mean (expectation) and covariance matrix of a random vector X is defined as follows: E [X 1 ] E [X E [X] = 2 ],. E [X n ] and cov(x) = E [{X E (X)} {X E (X)} T] σ1 2 σ σ 1n σ = 21 σ σ 2n,.... X n σ n1 σ n2... σ 2 n (3.1.1) where σ 2 j = var(x j) and σ jk = cov(x j,x k ) for j, k = 1, 2,...,n. Chapter 3 76

3 Properties of Mean and Covariance. 1. If X and Y are random vectors and A,B,C and D are constant matrices, then E [AXB + CY + D] = AE [X]B + CE[Y] + D. (3.1.2) Proof. Left as an excercise. 2. For any random vector X, the covariance matrix cov(x) is symmetric. Proof. Left as an excercise. 3. If X j, j = 1, 2,...,n are independent random variables, then cov(x) = diag(σj 2, j = 1, 2,...,n). Proof. Left as an excercise. 4. cov(x + a) = cov(x) for a constant vector a. Proof. Left as an excercise. Chapter 3 77

4 Properties of Mean and Covariance (cont.) 5. cov(ax) = Acov(X)A T for a constant matrix A. Proof. Left as an excercise. 6. cov(x) is positive semi-definite. Meaningful covariance matrices are positive definite. Proof. Left as an excercise. 7. cov(x) = E[XX T ] E[X] {E[X]} T. Proof. Left as an excercise. Chapter 3 78

5 Definition Correlation Matrix. A correlation matrix of a vector of random variable X is defined as the matrix of pairwise correlations between the elements of X. Explicitly, 1 ρ ρ 1n ρ crr(x) = ρ 2n, (3.1.3).... ρ n1 ρ n where ρ jk = corr(x j,x k ) = σ jk /(σ j σ k ), j, k = 1, 2,...,n. Example If only successive random variables in the random vector X are correlated and have the same correlation ρ, then the correlation matrix corr(x) is given by corr(x) = 1 ρ ρ 1 ρ ρ , (3.1.4) Chapter 3 79

6 Example If every pair of random variables in the random vector X have the same correlation ρ, then the correlation matrix corr(x) is given by corr(x) = 1 ρ ρ... ρ ρ 1 ρ... ρ ρ ρ 1... ρ,..... (3.1.5) ρ ρ ρ... 1 and the random variables are said to be exchangeable. 3.2 Multivariate Normal Distribution Definition Multivariate Normal Distribution. A random vector X = (X 1,X 2,...,X n ) T is said to follow a multivariate normal distribution with mean µ and covariance matrix Σ if X can be expressed as X = AZ + µ, where Σ = AA T and Z = (Z 1,Z 2,...,Z n ) with Z i, i = 1, 2,...,n iid N(0, 1) variables. Chapter 3 80

7 BIOS 2083 Linear Models Abdus S. Wahed Bivariate normal distribution with mean (0, 0) T and covariance matrix Probability Density x x Definition Multivariate Normal Distribution. A random vector X = (X 1,X 2,...,X n ) T is said to follow a multivariate normal distribution with mean µ and a positive definite covariance matrix Σ if X has the density. f X (x) = [ 1 (2π) n/2 Σ 1/2exp 1 ] 2 (x µ)t Σ 1 (x µ) (3.2.1) Chapter 3 81

8 Properties 1. Moment generating function of a N(µ,Σ) random variable X is given by { M X (t) = exp µ T t + 1 } 2 tt Σt. (3.2.2) 2. E(X) = µ and cov(x) = Σ. 3. If X 1,X 2,...,X n are i.i.d N(0, 1) random variables, then their joint distribution can be characterized by X = (X 1,X 2,...,X n ) T N(0,I n ). 4. X N n (µ,σ) if and only if all non-zero linear combinations of the components of X are normally distributed. Chapter 3 82

9 Linear transformation 5. If X N n (µ,σ) and A m n is a constant matrix of rank m, then Y = Ax N p (Aµ,AΣA T ). Proof. Use definition or property 1 above. Orthogonal linear transformation 6. If X N n (µ,i n ) and A n n is an orthogonal matrix and Σ = I n, then Y = Ax N n (Aµ,I n ). Chapter 3 83

10 Marginal and Conditional distributions Suppose X is N n (µ,σ) and X is partitioned as follows, X = X 1, where X 1 is of dimension p 1 and X 2 is of dimension n p 1. Suppose X 2 the corresponding partitions for µ and Σ are given by respectively. Then, µ = µ 1, and Σ = µ 2 Σ 11 Σ 12 Σ 21 Σ Marginal distribution. X 1 is multivariate normal - N p (µ 1,Σ 11 ). Proof. Use the result from property 5 above. 8. Conditional distribution. The distribution of X 1 X 2 is p-variate normal - N p (µ 1 2,Σ 1 2 ), where, and provided Σ is positive definite. µ 1 2 = µ 1 + Σ 12 Σ 1 22 (X 2 µ 2 ), Σ 1 2 = Σ 11 Σ 12 Σ 1 22 Σ 21, Proof. See Result , page 156 (Ravishanker and Dey). Chapter 3 84

11 Uncorrelated implies independence for multivariate normal random variables 9. If X, µ, and Σ are partitioned as above, then X 1 and X 2 are independent if and only if Σ 12 = 0 = Σ T 21. Proof. We will use m.g.f to prove this result. Two random vectors X 1 and X 2 are independent iff M (X1,X 2 )(t 1, t 2 ) = M X1 (t 1 )M X2 (t 2 ). Chapter 3 85

12 3.3 Non-central distributions We will start with the standard chi-square distribution. Definition Chi-square distribution. If X 1, X 2,...,X n be n independent N(0, 1) variables, then the distribution of n i=1 X2 i is χ 2 n (ch-square with degrees of freedom n). χ 2 n-distribution is a special case of gamma distribution when the scale parameter is set to 1/2 and the shape parameter is set to be n/2. That is, the density of χ 2 n is given by f χ 2 n (x) = (1/2)n/2 Γ(n/2) e x/2 x n/2 1, x 0; n = 1, 2,...,. (3.3.1) Example The distribution of (n 1)S 2 /σ 2, where S 2 = n i=1 (X i X) 2 /(n 1) is the sample variance of a random sample of size n from a normal distribution with mean µ and variance σ 2, follows a χ 2 n 1. The moment generating function of a chi-square distribution with n d.f. is given by M χ 2 n (t) = (1 2t) n/2, t < 1/2. (3.3.2) The m.g.f (3.3.2 shows that the sum of two independent ch-square random variables is also a ch-square. Therefore, differences of sequantial sums of squares of independent normal random variables will be distributed independently as chi-squares. Chapter 3 86

13 Theorem If X N n (µ,σ) and Σ is positive definite, then (X µ) T Σ 1 (X µ) χ 2 n. (3.3.3) Proof. Since Σ is positive definite, there exists a non-singular A n n such that Σ = AA T (Cholesky decomposition). Then, by definition of multivariate normal distribution, X = AZ + µ, where Z is a random sample from a N(0, 1) distribution. Now, Chapter 3 87

14 λ= λ= λ= λ=6 λ=8 λ= Figure 3.1: Non-central chi-square densities with df 5 and non-centrality parameter λ. Definition Non-central chi-square distribution. Suppose X s are as in Definition (3.3.1) except that each X i has mean µ i, i = 1, 2,...,n. Equivalently, suppose, X = (X 1,...,X n ) T be a random vector distributed as N n (µ,i n ), where µ = (µ 1,...,µ n ) T. Then the distribution of n i=1 X2 i = X T X is referred to as non-central chi-square with d.f. n and non-centrality parameter λ = n i=1 µ2 i /2 = 1 2 µt µ. The density of such a non-central chisquare variable χ 2 n(λ) can be written as a infinite poisson mixture of central chi-square densities as follows: f χ 2 n (λ)(x) = e λ λ j (1/2) (n+2j)/2 j! Γ((n + 2j)/2) e x/2 x (n+2j)/2 1. (3.3.4) j=1 Chapter 3 88

15 Properties 1. The moment generating function of a non-central chi-square variable χ 2 n(λ) is given by M χ 2 n (n,λ)(t) = (1 2t) n/2 exp 2. E [ χ 2 n(λ) ] = n + 2λ. 3. V ar [ χ 2 n(λ) ] = 2(n + 4λ). 4. χ 2 n(0) χ 2 n. 5. For a given constant c, (a) P(χ 2 n(λ) > c) is an increasing function of λ. (b) P(χ 2 n(λ) > c) P(χ 2 n > c). { } 2λt, t < 1/2. (3.3.5) 1 2t Chapter 3 89

16 Theorem If X N n (µ,σ) and Σ is positive definite, then X T Σ 1 X χ 2 n(λ = µ T Σ 1 µ/2). (3.3.6) Proof. Since Σ is positive definite, there exists a non-singular matrix A n n such that Σ = AA T (Cholesky decomposition). Define, Then, and Y = A 1 X. X = AY, X T Σ 1 X = Y T A T Σ 1 AY = Y T A T {AA T } 1 AY n = Y T Y = Yi 2. But notice that Y is a linear combination of X and hence is distributed as N n (A 1 µ,a 1 Σ { A 1} T = In ). Therefore components of Y are independently distributed. Specifically, Y i N(τ i, 1), where τ = A 1 µ. Thus, by definition of non-central chi-square, X T Σ 1 X = n i=1 Y i 2 λ = n i=1 τ2 i /2 = τt τ/2 = µ T Σ 1 µ/2. i=1 χ 2 n(λ), where Chapter 3 90

17 λ=0 0.5 λ= λ=4 λ=6 λ=8 λ= Figure 3.2: Non-central F-densities with df 5 and 15 and non-centrality parameter λ. Definition Non-central F-distribution. If U 1 χ 2 n 1 (λ) and U 2 χ 2 n 2 and U 1 and U 2 are independent, then, the distribution of F = U 1/n 1 U 2 /n 2 (3.3.7) is referred to as non-central F-distribution with df n 1 and n 2, and noncentrality parameter λ. Chapter 3 91

18 λ=0 λ=2 λ=4 λ=6 λ=8 0.1 λ= Figure 3.3: Non-central t-densities with df 5 and non-centrality parameter λ. Definition Non-central t-distribution. If U 1 N(λ, 1) and U 2 χ 2 n and U 1 and U 2 are independent, then, the distribution of T = U 1 U2 /n (3.3.8) is referred to as non-central t-distribution with df n and non-centrality parameter λ. Chapter 3 92

19 3.4 Distribution of quadratic forms Caution: We assume that our matrix of quadratic form is symmetric. Lemma If A n n is symmetric and idempotent with rank r, then r of its eigenvalues are exactly equal to 1 and n r are equal to zero. Proof. Use spectral decomposition theorem. (See Result on page 51 of Ravishanker and Dey). Theorem Let X N n (0,I n ). The quadratic form X T AX χ 2 r iff A is idempotent with rank(a) = r. Proof. Let A be (symmetric) idempotent matrix of rank r. Then, by spectral decomposition theorem, there exists an orthogonal matrix P such that P T AP = Λ = I r (3.4.1) Define Y = P T X = PT 1 X P T 2 X = Y 1 Y 2, so that P T 1 P 1 = I r. Thus, X = Chapter 3 93

20 PY and Y 1 N r (0,I r ). Now, X T Ax = (PY) T APY = Y T I r 0 Y 0 0 = Y T 1 Y 1 χ 2 r. (3.4.2) Now suppose X T AX χ 2 r. This means that the moment generating function of X T AX is given by M XT AX(t) = (1 2t) r/2. (3.4.3) But, one can calculate the m.g.f. of X T AX directly using the multivariate normal density as M XT AX(t) = E [ exp { (X T AX)t }] = exp { (X T AX)t } f X (x)dx = exp { (X T AX)t } [ 1 (2π) n/2exp 1 ] 2 xt x dx [ 1 = (2π) n/2exp 1 ] 2 xt (I n 2tA)x dx = I n 2tA 1/2 n = (1 2tλ i ) 1/2. (3.4.4) i=1 Equate (3.4.3) and (3.4.4) to obtain the desired result. Chapter 3 94

21 Theorem Let X N n (µ,σ) where Σ is positive definite. The quadratic form X T AX χ 2 r(λ) where λ = µ T Aµ/2, iff AΣ is idempotent with rank(aσ) = r. Proof. Omitted. Theorem Independence of two quadratic forms. Let X N n (µ,σ) where Σ is positive definite. The two quadratic forms X T AX and X T BX are independent if and only if AΣB = 0 = BΣA. (3.4.5) Proof. Omitted. Remark Note that in the above theorem, the two quadratic forms need not have a chi-square distribution. When they are, the theorem is referred to as Craig s theorem. Theorem Independence of linear and quadratic forms. Let X N n (µ,σ) where Σ is positive definite. The quadratic form X T AX and the linear form BX are independently distributed if and only if BΣA = 0. (3.4.6) Proof. Omitted. Remark Note that in the above theorem, the quadratic form need not have a chi-square distribution. Chapter 3 95

22 Example Independence of sample mean and sample variance. Suppose X N n (0,I n ). Then X = n i=1 X i/n = 1 T X/n and S 2 X = n i=1 (X i X) 2 /(n 1) are independently distributed. Proof. Chapter 3 96

23 Theorem Let X N n (µ,σ). Then E [ X T AX ] = µ T Aµ + trace(aσ). (3.4.7) Remark Note that in the above theorem, the quadratic form need not have a chi-square distribution. Proof. Chapter 3 97

24 Theorem Fisher-Cochran theorem. Suppose X N n (µ,i n ). Let Q j = X T A j X, j = 1, 2,...,k be k quadratic forms with rank(a j ) = r j such that X T X = k j=1 Q j. Then, Q j s are independently distributed as χ 2 r j (λ j ) where λ j = µ T A j µ/2 if and only if k j=1 r j = n. Proof. Omitted. Theorem Generalization of Fisher-Cochran theorem. Suppose X N n (µ,i n ). Let A j, j = 1, 2,...,k be k n n symmetric matrices with rank(a j ) = r j such that A = k j=1 A j with rank(a) = r. Then, 1. X T A j X s are independently distributed as χ 2 r j (λ j ) where λ j = µ T A j µ/2, and 2. X T AX χ 2 r(λ) where λ = k j=1 λ j if and only if any one of the following conditions is satisfied. C1. A j Σ is idempotent for all j and A j ΣA k = 0 for all j < k. C2. A j Σ is idempotent for all j and AΣ is idempotent. C3. A j ΣA k = 0 for all j < k and AΣ is idempotent. C4. r = k j=1 r j and AΣ is idempotent. C5. the matrices AΣ, A j Σ, j = 1, 2,...,k 1 are idempotent and A k Σ is non-negative definite. Chapter 3 98