Chapter 6 Independence 6.11.P By the Kolmogorov maximal inequality, 2 P{maxj6n |Sj | > } 6 ES2n 6 1 2 i=1 EXi = A (Fubini–Tonelli). Therefore, P sup n6j6n+m ✏ |Sj - Sn | > 6 Pn+m i=n E[X2i ] 8 2 > 0. Continuity properties of P imply that P sup |Sj - Sn | > j>n ✏ We can find nk " 1 so fast that then, P 6 P1 i=n P i>nk 8 > 0. E[X2i ] 6 e-k . By Borel-Cantelli sup |Sj - Snk | > k-2 j>nk E[X2i ] 2 ✏ 6 k4 e-k , which is summable in k. Therefore, the Borel–Cantelli lemma ensures that a.s. forP all but a finite number of k’s, supj>nk |Sj - Snk | 6 k-2 . In particular, k supj>nk |Sj - Snk | < 1 a.s. This implies that {Sn }1 n=1 is a.s. a Cauchy sequence (why?). 6.13. Fix some lim inf n (Sn /n) > lim inf n (S⌫ n /n), where Pn ⌫ ⌫> 0. Clearly, ⌫ ⌫ Sn := X , and X := X 1 . By the law of large numbers, i {X 6⌫} i=1 i i i ⌫ lim inf n (S⌫ /n) = lim (S /n) = E[X ; X 6 ⌫]. Let ⌫ " 1 to finish. n 1 1 n n 6.29. Solution 1. We will need Lemma: As n ! 1, Proof Rn 1 (dx/x) 6 Pn i=1 (1/i) 61+ 13 Rn 1 Pn (dx/x). i=1 (1/i) ⇠ ln(n). ⇤ 14 CHAPTER 6. INDEPENDENCE Now let us first ] = 0. Define Xi0 = Xi 1{|Xi |6i} . Also 1P Pn assume that E[X 0 0 define Sn = i=1 (Xi /i) and Sn = n i=1 (Xi /i). By the Kolmogorov maximal inequality, P max |Sk0 - E[Sk0 ]| > " ln n 16k6n 6 VarSn0 2 " (ln n)2 ⇥ ⇤ n X E X21 ; |X1 | 6 i kSn0 k22 1 6 2 2 . = 2 2 i2 " ln (n) " ln (n) i=1 Therefore, 1 X n=1 P max n |Sk0 - E[Sk0 ]| > " ln 22 ⇥ ⇤ 1 2n X 1 X E X21 ; |X1 | 6 i 4n i2 n=1 i=1 ⇥ ⇤ 1 X X E X21 ; |X1 | 6 i 1 6A i2 4n i=1 n>log2 (i) ⇥ 2 ⇤ 1 X E X1 ; |X1 | 6 i 6B i2 i=1 2 3 X 1 5 6 CEX1 . = BE 4X21 i2 n 6A 16k622 i>|X1 | By the Borel–Cantelli lemma, a.s., for all n large, max n |Sk0 - E[Sk0 ]| 6 " ln 22 n 16k622 n Any m is between some 22 and 22 0 0 |Sm - E[Sm ]| 6 max 16k622n+1 n+1 . . Therefore, |Sk0 - E[Sk0 ]| ⇣ n+1 ⌘ 6 " ln 22 = 2" ln(2) · 2n 6 2" ln(2) ln m. P 0 0 0 This P proves that a.s., |Sm - ESm |/ ln(m) ! 0. But k P{Sk 6= Sk0 } = k P{|X1 | > k} < 1 because kX1 k1 < 1. So almost surely, Sk = Sk for 0 all k large. It suffices to prove that |ESm |/ ln m ! 0. But 0 |ESm |= m X E[X1 ; |X1 | 6 i] i=1 i = m X E[X1 ; |X1 | > i] i i=1 since EX1 = 0. Thus, 0 |ESm |6 m X E[|X1 |; |X1 | > i] i=1 i . , 15 For all ⌘ > 0, there exists i0 such that for al i > i0 , E{|X1 |; |X1 | > i} 6 ⌘. Pi 0 P Pm Therefore, m i=i0 (· · · ) 6 ⌘ i=1 (1/i) ⇠ ⌘ ln m. But i=1 (· · · ) 6 i0 kX1 k1 . 0 Therefore, lim supm |ESm |/ ln m 6 ⌘ for all ⌘, whence the result. If µ = EX1 6=P 0, then consider instead Xi := Xi - µ. The preceding proves Pn n -1 -1 that ln (n) (X /i) ! 0 a.s. Equivalently, ln (n){ (X i /i) i=1 i=1 i P Pn µ n (1/i)} ! 0. Because (1/i) ⇠ ln(n), we obtain the result in i=1 i=1 general. Solution 2. P We may start with P Lemma: If limP n!1 an = µ, bn > 0 for all n n n > 1, and 1 b = 1, then a b ⇠ µ i i i i=1 i=1 i=1 bi . Proof 8" > 0, 9n0 so that |ai - µ| 6 µ + " for all i > n0 . Then, n X ai bi ⇠ i=1 n X i=n0 ai bi = (µ ± ") n X i=n0 bi ⇠ (µ ± ") n X bi , i=1 ⇤ notation being clear. Now let S0 = 0, and Sn = n X Xi i=1 n X Pn j=1 Sj (n > 1), so that n n 1 X 1 X 1 = (Si - Si-1 ) = Si Si-1 i i i i i=1 i=1 i=1 ◆ n n-1 n-1 X X ✓1 1 X 1 1 Sn = Si Si = S1 + Si . i i+1 i i+1 n+1 i=1 i=2 i=2 By the strong law, Sn /(n + 1) ! µ a.s. as n ! 1. Therefore, n n-1 n-1 1 X Xi 1 X Si 1 X µ ⇠ ⇠ !µ ln n i ln n i(i + 1) ln n i+1 i=1 i=2 a.s. i=2 6.32. If Sn := the number of Xi ’s (1 6 i 6 n) that are equal to one, then Sn = Bin(n, 1/2). By the De Moivre–Laplace central limit theorem, for all > 0, ⌦ p ↵ n lim P Sn > + n = P{N(0 , 1) > } > 0. n!1 2 Thus, by the Kolmogorov 0–1 law, lim sup n!1 Sn - (n/2) p =1 n a.s. P P p Because n follows that lim supn!1 n i=1 Xi = 2Sn -n, it P i=1 Xi / n = 1 p a.s. By symmetry, lim inf n!1 n i=1 Xi / n = -1 a.s. Thus, lim sup n!1 n X i=1 Xi = - lim inf n!1 n X i=1 Xi = 1 a.s. 16 CHAPTER 6. INDEPENDENCE 6.33. 1. First of all we note that if n > 1 and t > 0, then for any > 0, ⇥ ⇤ P{Sn > t} = P e Sn > e t 6 e- t E e Sn , owing to Chebyshev’s inequality. But the expectation of a product of independent random variables is the product of the expectations; Q therefore, E exp( Sn ) = n E exp( Xj )}. Problem 4.30 of the text j=1 implies that E exp( Xj ) 6 exp( 2 c2j /2). Therefore, This proves that ⇥ E e Sn ⇤ 6e P{Sn > t} 6 e- 2 2 sn /2 t+ . 2 2 sn /2 . The left-hand side is independent of . So we can minimize the right-hand side over all > 0; the minimum occurs at = t/s2n , and yields P{Sn > t} 6 exp(-t/s2n ). The result follows from replacing Sn by -Sn . 2. Let I1 , . . . , In be i.i.d. Bin(1 , p) random variables. Then B has the same distribution as I1 + · · · + In . Let Xi := Ii - p. Then the Xi ’s are i.i.d., have mean zero, and are bounded between -1 and 1. Therefore, we can apply the first part to Sn := X1 + · · · + Xn to find that s2n := n and P{|Sn | > t} 6 2e-t/(2n) for all t > 0. p Replace t by t n and observe that Sn has the same distribution as B - np to finish. c c 3. Note that, in (6.48), |T2 | 6 KP(A p ), and A = {|Sn -np| > n }. Apply the previous part with t := n to finish. 4. Similar to part 3.