Follow-up questions on Ch4 Q7 An asymmetric random walk is defined by S0 = x and Sn = x + for n > 1, where ( +1, with probability p Xn = −1, with probability q Pn k=1 Xk with p 6= q. Fix θ = q/p, we have shown that Mn := θSn is a martingale x −θ b and P(Ha < Hb ) = θθa −θ b , where Ha = inf(n > 0 : Sn = a) and Hb = inf(n > 0 : Sn = b). Some follow-up questions: 1. Does limn→ Mn exist? What can be said about this limit? 2. Let T = Ha ∧ Hb . Find E(ST ). 3. Show that Hn = Sn − n(p − q) is a martingale. Hence find E(T ). Part 1 Note that Mn = (q/p)Sn is always non-negative. Hence by Martingale Convergence Theorem (MCT), limn→ Mn exists. Let µ be the common mean of Xn . By definition of Xn we have µ = p − q. Since Xn ’s are i.i.d random variables with finite variance, strong law of large numbers gives Sn /n → µ = p − q almost surely as n → ∞. Now consider two cases: Case 1: Suppose p > q. We have Sn ≈ n(p − q) → +∞ almost surely as n → ∞. Moreover, under p > q we have θ < 1. Hence Mn = θSn → 0. Case 2: Suppose q > p. We have Sn ≈ n(p − q) → −∞ almost surely as n → ∞. Moreover under q > p we have θ > 1. Hence Mn = θSn → 0. In summary, Mn converges to 0 almost surely. Remark: the argument via Sn ≈ n(p − q) for large n is quite loose. But it can be formally proved that Sn → ±∞ when p ≷ q by tracing the formal definitions from analysis. See also Ch3 Q6 where a similar idea is used. Part 2 We know that ST is a binary random variable, which takes value of a or b with probability P(Ha < Hb ) and P(Hb < Ha ) respectively. Then E(ST ) = aP(Ha < Hb ) + bP(Hb < Ha ) = a(θx − θb ) + b(θa − θx ) . θa − θb Part 3 Verify that Hn = Sn − n(p − q) satisfies the three required properties of a martingale: 1. Hn is clearly adapted. 2. E|Hn | 6 E|Sn | + n|p − q| 6 x + n + n|p − q| < ∞. 3. Check that E(Hn+1 |Fn ) = E(Sn+1 |Fn ) − (n + 1)(p − q) = E(Sn + Xn+1 |Fn ) − (n + 1)(p − q) = Sn + E(Xn+1 ) − (n + 1)(p − q) = Sn + (p − q) − (n + 1)(p − q) = Sn − n(p − q) = Hn . We have a martingale H. Now it’s tempting to directly use optimal stopping theorem (OST) on HT to conclude E(HT ) = H0 . But it can’t be done in this way since Hn is not bounded for n 6 T so the condition for OST fails. Part 3 (cont‘) Remember our trick on Ch4 Q6. If we don’t have a bounded martingale which allows direct use of OST, the second best attempt is to consider a capped (bounded) stopping time, and see if we can take limit to get the results we want. Let N > 0 be some fixed constant and set TN = min(T , N). Then TN is a bounded stopping time. OST can now be used which gives E(HTN ) = H0 = S0 = x. Then x = E(STN − TN (p − q)) and in turn E(TN ) = 1 (E(STN ) − x). p−q Part 3 (cont‘) Take limit N → ∞ on both side. I The left-hand-side becomes lim E(TN ) = E lim TN N→∞ = E(T ). N→∞ The swap of E and lim is valid due to monotone convergence theorem (MON) as TN is positive and increasing to T as N → ∞. I The right-hand-side becomes lim E(STN ) = E N→∞ lim STN N→∞ = E(ST ). The swap of E and lim is valid due to bounded convergence theorem (BDD) as STN is bounded between a and b for all N. We thus obtain: E(T ) = = 1 1 E(ST − x) = p−q p−q a(θx − θb ) + b(θa − θx ) −x θa − θb (a − x)((q/p)x − (q/p)b ) + (b − x)((q/p)a − (q/p)x ) . (p − q)((q/p)a − (q/p)b )