774 CHAPTER 28. THE NORMAL DISTRIBUTION

form a given vector and Σ a given positive definite symmetric matrix. Recall also that thecharacteristic function of this random variable is

E(eit·X)= eit·me−

12 t∗Σt (28.11)

So what if det(Σ) = 0? Is there a probability measure having characteristic functioneit·me−

12 t∗Σt? Let Σn→ Σ in the Frobenius norm, det(Σn)> 0. That is the i jth components

converge and all the eigenvalues are positive. Then from the definition of the characteristicfunction,

φ λXn(t) = eit·me−

12 t∗Σnt→ ψ (t)≡ eit·me−

12 t∗Σt

Now clearly ψ (0) = 1 and ψ is continuous so by Levy’s theorem, Theorem 28.4.10, thereis a probability measure µ such that ψ (t) = φ µ (t) . As noted above, there is also a randomvariableX with λX = µ . Consider the moments forXn.

Lemma 28.4.13 LetX ∼ N (0,Σ) where Σ is positive definite. Then the moments ofXall exist and are dominated by an expression which is continuously dependent on det(Σ).

Proof: Let q ≥ 1. E (|X|q) =∫Rp

|x|q

(2π)p/2 det(Σ)1/2 e−12 x∗Σ−1xdx. Let R be an orthogonal

matrix with Σ = R∗DR where D is a diagonal matrix having the positive eigenvalues σ j onthe diagonal. Thus x∗Σ−1x= x∗R∗D−1Rx so let Rx≡ y. Changing the variable in theintegral and assuming q = 2m for m a positive integer,

E(|X|2m

)=

∫Rp

(∑

pk=1 y2

k

)m

(2π)p/2∏

pj=1 σ

1/2j

e−12 y∗D−1ydy

= 2p 1

(2π)p/2

∫∞

0· · ·∫

0

(∑

pk=1 y2

k

)m

∏pj=1 σ

1/2j

e−12 y∗D−1ydy

From convexity of x→ xm

= 2p 1

(2π)p/2

∫∞

0· · ·∫

0

(p∑

pk=1

1p y2

k

)m

∏pj=1 σ

1/2j

e−12 y∗D−1ydy

≤ 2p pm−1

(2π)p/2

∫∞

0· · ·∫

0

∑pk=1 y2m

k

∏pj=1 σ

1/2j

e−12 ∑

pk=1 y2

kσ−1dy

= 2p pm−1

(2π)p/2

∫∞

0· · ·∫

0

∑k ̸=l y2mk

∏pj=1 σ

1/2j

e−12 ∑k ̸=l y2

kσ−1dx1 · · · d̂xl · · ·dxp

·∫

0

1

σ1/2l

y2ml e−

12 y2

l σ−1l dyl

Now letting u = ylσ−1/2l ,dyl = σ

1/2l du and so that last integral is of the form∫

ml u2me−

12 u2

du = Ĉmσml

774 CHAPTER 28. THE NORMAL DISTRIBUTIONfor m a given vector and X a given positive definite symmetric matrix. Recall also that thecharacteristic function of this random variable isE (eX) — pitm,— ste (28.11)So what if det(£) = 0? Is there a probability measure having characteristic functioni; 1 px : : ‘ selt ™e~ 2X9 | et ©, > L in the Frobenius norm, det (Z,) > 0. That is the ij” componentsconverge and all the eigenvalues are positive. Then from the definition of the characteristicfunction,Pix (t) — eitm,- At Ynt + wit j= elt mo— st “xtNow clearly y(0) = | and y is continuous so by Levy’s theorem, Theorem 28.4.10, thereis a probability measure f such that y(t) = @,, (€). As noted above, there is also a randomvariable X with Ax = UW. Consider the moments for X ;.Lemma 28.4.13 Let X ~ N (0,2) where ¥ is positive definite. Then the moments of Xall exist and are dominated by an expression which is continuously dependent on det (2).Proof: Let g >1.E (|X|?) = Jrp Saran 2 Pde. Let R be an orthogonale|matrix with © = R*DR where D is a diagonal matrix having the positive eigenvalues 0; onthe diagonal. Thus 2*£~!a = 2*R*D~'Rzx so let Rx = y. Changing the variable in theintegral and assuming g = 2m for m a positive integer,P 2 mE (|x?") _ (ree 1» Yi) - ser? UdyRP ont? v_ k= We) syD yg= wef ri! > LI? ym)P! TT 10; /From convexity of x > x”)"_ Yee 1 pt) MEETPTEY gat 'ydy"Om rar ri am—1 00 200 Y yon - > 1 pe— oP [ _ ee 2 bane dx, ++-dxj---dXp0 0 Il; oO.j=l j1/2Now letting u = yo, dy, = 0,’ du and so that last integral is of the form[ one “2 du=Cyo7"