28.4. PROKHOROV AND LEVY THEOREMS 773

Note how it was only necessary to assume ψ (0) = 1 and ψ is continuous at 0 in orderto conclude that ψ is a characteristic function. This helps to see why Prokhorov’s andLevy’s theorems are so amazing. Limits of characteristic functions tend to be characteristicfunctions. What about random variables?

If you have a probability measure λ on the Borel sets of Rp, is there a random variableX such that λ = λX? Yes. You could let Ω =Rp and X (x) = x and P(E)≡ λ (E) for allE Borel. Then λX

(X−1 (E)

)≡ P(E) ≡ λ (E) so this is indeed a random variable such

that λ = λX . Thus for a probability measure λ , you can generally get a random variablewhich has λ as its distribution measure. Later, this is considered more. You might havemore than one random variable having λ as its distribution measure.

In this next corollary, it suffices to have the random variables have values in a Banachspace. However, I will write |X| rather than ∥X∥.

Corollary 28.4.11 In the context of Theorem 28.4.10, suppose µn is the distributionmeasure of the random variable Xn and that supn E (|Xn|q) =Mq <∞ for all q≥ 1 and thatµn converges weakly to the probability measure µ . Then if µ is the distribution measurefor a random variableX, then E (|X|q)< ∞ for all q≥ 1.

Proof:E (|X|q) =

∫∞

0P(|X|q > α)dα =

∫∞

0µ ([|x|q > α])dα

≤∫

0µ ([|x|q > α])dα ≤

∫∞

0

∫Rp

(1−ψα)dµdα

where ψα = 1 on B(0, 1

2 α1/q)

is nonnegative, and is in Cc(B(0,α1/q

)). Thus if |x|q > α,

then 1−ψα (x) = 1 which shows the above inequality holds. Also, if (1−ψα (x)) > 0,then |x| > 1

2 α1/q and so |x|q > 12q α . Since weak convergence holds and 1−ψα is a

bounded continuous function,∫Rp

(1−ψα)dµ = limn→∞

∫Rp

(1−ψα)dµn (28.10)

Therefore, from the above and Fatou’s lemma,∫Ω

|X|q dP≤∫

0limn→∞

∫Rp

(1−ψα)dµndα

≤ lim infn→∞

∫∞

0µn

([|x|q > 1

2q α

])dα

Changing the variable,

= lim infn→∞

2q∫

0µn (|x|

q > δ )dδ = lim infn→∞

2qE (|Xn|q)< ∞ ■

Now recall the multivariate normal distribution.

Definition 28.4.12 A random vectorX, with values in Rp has a multivariate nor-mal distribution written as

X ∼ Np (m,Σ)

if for all Borel E ⊆ Rp, the distribution measure is given by

λX (E) =∫Rp

XE (x)1

(2π)p/2 det(Σ)1/2 e−12 (x−m)∗Σ−1(x−m)dx

28.4. PROKHOROV AND LEVY THEOREMS 773Note how it was only necessary to assume y (0) = 1 and y is continuous at O in orderto conclude that y is a characteristic function. This helps to see why Prokhorov’s andLevy’s theorems are so amazing. Limits of characteristic functions tend to be characteristicfunctions. What about random variables?If you have a probability measure J on the Borel sets of R”, is there a random variableX such that A = 1.x? Yes. You could let Q = R? and X (a) = x and P(E) =A (E) for allE Borel. Then 2.x (X~'(E)) = P(E) =A(E) so this is indeed a random variable suchthat A = Ax. Thus for a probability measure 2, you can generally get a random variablewhich has A as its distribution measure. Later, this is considered more. You might havemore than one random variable having A as its distribution measure.In this next corollary, it suffices to have the random variables have values in a Banachspace. However, I will write |X| rather than ||X'|].Corollary 28.4.11 In the context of Theorem 28.4.10, suppose |, is the distributionmeasure of the random variable X, and that sup, E (|X n|4) = Mg < © for all g > 1 and thatLL, converges weakly to the probability measure Ub. Then if w is the distribution measurefor a random variable X , then E(|X|*) < for allq > 1.Proof: so soE(X\)= [ P(X|">a)da= |" w(ja\" > al) aa< [ ulllel>a)da< [| —wa)duaawhere WY, =lonB (0,5a!/2) is nonnegative, and is in C, (B (0, a'/7)) . Thus if |x|? > a,then 1 — y, (a) = 1 which shows the above inequality holds. Also, if (1— W, (a)) > 0,then |x| > 5a!/4 and so |x|? > 37a. Since weak convergence holds and | — Wy is abounded continuous function,[Wada = tim [= Wa) dey (28.10)n—o JIRpTherefore, from the above and Fatou’s lemma,qd oer _as ap< [ im [ Wy) du, da~ 1. qi<lim int Ln (c > sa ) da= lim inf 2# | u, (||? > 6) d5 =lim inf 27E (|X,)|2) < co Mln—oo 0 n—ooChanging the variable,Now recall the multivariate normal distribution.Definition 28.4.12 4 random vector X 5 with values in R? has a multivariate nor-mal distribution written asX ~N,(m,z)if for all Borel E CR’, the distribution measure is given by1 =1 —Ax (E =| Qe (ae) ——_b et e-my'E (em) gyx(E) RP e(@) (2m)? det (2) 72" *