Solution Set # 3 (2.1.11) Suppose that µ ∈ P, and let {Xn : n ≥ 1} be a sequence of independent random variables with distribution µ. Then T2n µ is the distribution of S̆2n , and so, by the Central Limit Theorem, T2n µ =⇒ γ0,1 . In particular, if µ = T2 µ and therefore µ = T2n µ for all n ≥ 1, then µ = γ0,1 . (2.1.13) (i) Let λSn−1 denote surface measure on Sn−1 . Using cyllindrical coordinates, one can show (cf. Exercise 5.2.21 in my Essential of Integration Theory for Analysis) that, for n ≥ 2, λSn−1 (dθ) = 1 − θ12 n−3 2 λSn−2 (dθ0 )λ(−1,1) (dθn ), where λ(−1,1) is Lebesgue measure on (−1, 1) and θ0 = (θ2 , . . . , θn ). Hence, because Z Sn−1 (r) ϕdλSn−1 (r) = r n−1 Z Sn−1 ϕ(rθ) λSn−1 (dθ), (1) we obtain the asserted expression for λn , and, given that expression, the rest is easy. (ii)) Simply follow the outline. (iii) First reduce to the case when hϕ, γ0,1 i = 0. In that case, note that, for any 1 ≤ k 6= ` ≤ n, the (2) distribution of (xk , x` ) under λn is λn , and therefore Z n 1X ϕ(xk ) n k=1 !2 1 λn (dx) = n Z n−1 ϕ(x1 ) λn (dx) + n 2 Z ϕ(x1 )ϕ(x2 ) λn (dx) −→ hϕ, γ0,1 i2 = 0. (2.3.21) For parts (i) and (ii), just follow the outline. To do part (iii), for each e ∈ RN , let µe be the 2 σe distribution of x (e, x)RN under µ, apply (ii) to µe , and conclude that µ ceR(ξ) = e− 2 , where σe2 = R 2 (e, x)RN µ(dx). Thus, µ ∈ N (0, C), where C is determined by (ξ, Cη)RN = (ξ, x)RN (η, x)RN µ(dx). (2.3.23) If µ = Tα µ for some µ on R, then µN = Tα µN , and so there exists a non-trivial solution on RN if there exists one on R. Conversely, if µ on RN is a non-trivial solution, then there exists a e ∈ SN −1 such that the distribution µe of x ∈ RN 7−→ (e, x)RN ∈ R under µ is non-trivial, and µe is a solution for all e ∈ SN −1 . Thus, the question of existence reduces to the case when N = 1, in which case the outline shows how to solve it. (2.3.25) If σ = 0 there is nothing to do, and X σ ∈ N (0, 1) when σ > 0. Hence, we may assume that σ = 1. Next, when σ = 1, r Z ∞ p x2 2 P xp e− 2 dx E |X| = π 0 and Z (*) 2 p − x2 x e 0 Hence, (2.3.26) follows. ∞ dx = 2 p−1 2 Z ∞ t 0 p−1 2 e −t dt = 2 p−1 2 Γ p+1 2 . (i) First note that X β is 1-sub-Gaussian, and so it is enough to handle the case when β = 1. In this case, the first estimate in Lemma 2.3.18 says that P(X ≥ R) ≤ 2e− P p E |X| Z ∞ t =p p−1 P |X| > t dt ≤ 2p Z ∞ R2 2 , and so, by (*), t2 p tp−1 e− 2 dt = p2 2 Γ p 0 0 2 . σ (ii) Again, by considering X β , one can reduce to the case when β = 1 after replacing σ by β . Thus, assume that β = 1. By Hölder’s inequality, EP [|X|p ] ≥ σ p for p ≥ 2. Now assume that p < 2. If q ∈ (1, ∞), then 2 = pq + q10 2q−p q−1 . Thus, by Hölder’s inequality, 1 1 2q−p p 2q−p 1 σ 2 = EP |X| q |X| q0 q−1 ≤ EP |X|p q EP |X| q−1 q0 . Now choose q = 4−p 2 . Then 2q−p q−1 = 4, and the result follows by simple arithmetic. (iii) Because S is B-sub-Gaussian and has variance Σ2 , there is nothing to do. (iv) Follow the outline. (2.2.28) and (2.3.29) Follow the outline.