Math 280B, Winter 2012 Feller’s method for the CLT 1. Let U1 , . . . , Un , V1 , . . . , Vn be independent random variables. We are going to compare the distribution of U := U1 + · · · + Un to that of V := V1 + · · · + Vn . Bearing in mind the Portmanteau Theorem, we examine E[g(U ) − g(V )] for suitable bounded continuous functions g. First, we have the telescoping sum (1.1) E[g(U ) − g(V )] = n X k=1 E[g(V1 + · · · +Vk−1 + Uk + · · · Un ) − g(V1 + · · · + Vk + Uk+1 + · · · Un )], in which the k = 1 term is understood to be E[g(U1 + · · · Un ) − g(V1 + U2 + · · · Un )] and the k = n term is understood to be E[g(V1 + · · · + Vn−1 + Un ) − g(V1 + · · · + Vn )]. The k th term in the above sum is equal to (1.2) E |g(Uk + Wk ) − g(Vk + Wk )| , where Wk := V1 + · · · + Vk−1 + Uk+1 + · · · + Un , with the obvious interpretations in the extreme cases k = 1 and k = n. If we introduce the notation Ψk g(x) := E[g(Uk + x)], Υk g(x) := E[g(Vk + x)], the the quantity in (1.2) is equal to (1.3) E [Ψk g(Wk ) − Υk g(Wk )] , because of the mutual independence of Uk , Vk , and Wk . Since g is bounded and continuous, both Ψk g and Υk g are bounded and continuous functions. Let us now write, for a bounded function u : R → R, kuk := sup{|u(x)| : x ∈ R}. 1 Then the expression in (1.3) is at most kΨk g − Υk gk. Feeding this information into (1.1) we obtain the important estimate (1.4) |E[g(U ) − g(V )]| ≤ n X k=1 kΨk g − Υk gk. 2. We now specialize the situation of section 1 to the case of iid random variables; this yields a direct route to the central Limit Theorem. So let X1 , X2 , . . . be iid random variables, each with mean 0 and variance 1. Also, let Z, Z1 , Z2 , . . . be iid standard normal random variables, independent of the Xk s. Fix n and define Xk Uk := √ , n Zk Vk := √ , n k = 1, 2, . . . , n. Evidently V = V1 + · · · Vn has the same distribution as Z; that is, the standard normal distribution. Moreover, U is just the normalized sum X + · · · + Xn S √n = 1 √ . n n In the present case the n terms in the sum (1.4) are all the same, so we have √ E[g(Sn / n) − g(Z)] ≤ nkΨ1 g − Υ1 gk. (2.1) The right side of (2.1) is dominated by nkΨ1 g − g − g ′′ /2k + nkΥ1 g − g − g ′′ /2k, and both of these terms converges to 0 as n → ∞. This is a consequence eof the following lemma. 3. Lemma. Let X be a random variable with E[X] = 0 and E[X 2 ] = 1. Let g : R → R be thrice continuously differentiable with compact support. Then (3.1) √ n E[g(x + X/ n)] − g(x) − g ′′ (x)/2 → 0, n → ∞, and the convergence is uniform in x ∈ R. Proof. The expression in (3.1) can be written as √ √ g(x + X/ n) − g(x) − (X/ n)g ′ (x) g ′′ (x) 2 (3.2) E X . − X 2 /n 2 2 For a fixed ω the integrand √ √ g(x + X/ n) − g(x) − (X/ n)g ′ (x) g ′′ (x) − X 2 /n 2 converges to 0 as n → ∞, uniformly in x. Also, the integrand is bounded (uniformly in x and ω) and the measure x2 P[X ∈ dx] has total mass 1, so by bounded convergence the expectation in (3.2) converges to 0 and the convergence is uniform in x. 4. It follows from the lemma and (2.1) that √ lim E[g(Sn / n) − g(Z)] = 0, n→∞ √ provided g is smooth and of compact support. This in turn implies that Sn / n converges in distribution to Z, which is the basic Central Limit Theorem. 3