Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Indeterminism wikipedia , lookup
Inductive probability wikipedia , lookup
Birthday problem wikipedia , lookup
Ars Conjectandi wikipedia , lookup
Random variable wikipedia , lookup
Probability interpretations wikipedia , lookup
Infinite monkey theorem wikipedia , lookup
Random walk wikipedia , lookup
Central limit theorem wikipedia , lookup
Optional Stopping Theorem REU. 07/27/2011 Suppose every minute you toss a symmetric coin. If it comes heads (with probability 1/2), you win 1$. If it comes tails (also with probability 1/2), you lose 1$. Let Xk be your win (or loss) at the moment k. So Xk takes values ±1 with equal probability. All Xk are independent. Your total capital at the moment n is n X Sn := Xk , S0 = 0. k=1 This random process S = (Sn )n≥0 is called a random walk. You can stop this game at any moment. This moment needs not be deterministic; it can be random. Denote this moment by τ . This is a nonnegative integer-valued random variable. And your total gain in this game will be Sτ . This τ is your strategy: you cannot influence the values of Xk , but you can choose when to withdraw yourself from this game. It is also called a stopping time. You can decide whether to exit the game at the moment n only basing on the past: using the values of X1 , . . . , Xn , which you already know by this moment. Speaking formally, the event {τ = n} depends only on X1 , . . . , Xn . So we consider the random walk S = (Sn )0≤n≤N . There are two cases: when the time horizon N is finite and when it is infinite. Can we find a strategy τ such that we will gain something? But we will gain some random sum, so let us use the expectation. Question: can we find a strategy τ such that EXτ > 0? Theorem 1. [Optional Stopping Theorem] For finite time horizon, this is not possible: for every strategy τ , we have ESτ = 0. Lemma. For every n = 0, . . . , N we have: E Sn I{τ =n} = E SN I{τ =n} . (1) Suppose we proved this lemma. Let us complete the proof. The variable τ attains values 0, . . . , N . Therefore, N N X X ESτ = E Sn I{τ =n} = E Sn I{τ =n} . n=0 n=0 Recall that for any event A, the indicator of A is the random variable which is 1 when A happens and 0 when A does not happens. It follows from (1) that ! N N N X X X E Sn I{τ =n} = E SN I{τ =n} = E SN I{τ =n} = ESN = 0. n=0 n=0 n=0 The proof is complete. Now we only need to prove the lemma. Proof of the lemma. It suffices to prove that E (SN − Sn )I{τ =n} = 0. Note that SN − Sn = N X k=n+1 1 Xk , so it depends only on Xn+1 , . . . , XN . And {τ = n} depends only on X1 , . . . , Xn . So SN − Sn and I{τ =n} are independent. Recall: for independent variables X, Y , we have EXY = EXEY . And E(SN − Sn ) = E N X Xk = k=n+1 N X EXk = 0. k=n+1 Thus, E (SN − Sn )I{τ =n} = E ((SN − Sn )) EI{τ =n} = 0. This completes the proof of the lemma and of the theorem. In the case of infinite time horizon, the situation is different. The random walk will eventually come to the value 1. You should wait until it comes there and then exit the game. Your gain will be 1. Formally speaking, τ = min{n | Sn = 1}, and Sτ = 1. Certainly, the main question is whether τ always exists. Does this random walk eventually come to 1? Theorem 2. With probability 1, there exists n such that Sn = 1. Proof. Let us split the proof into five steps: √ Step 1. Consider the event A = {Sn ≥ n for infinitely many n}. Then P(A) > 0. Step 2. Consider the event Sn B = lim √ ≥ 1 . n→∞ n Then P(B) > 0. Step 3. Fix N . Then B does not depend on X1 , . . . , XN . Step 4. P(B) = 0 or 1. Step 5. Finish the proof. √ Proof of step 1. Let Am = {Sm ≥ m}. Then A= ∞ [ ∞ \ Am = ∞ \ Bn , Bn = Am . m=n n=1 n=1 m=n ∞ [ The sequence of events B1 , B2 , . . . is decreasing: B1 ⊇ B2 ⊇ B3 ⊇ . . .. Therefore, P(A) = lim P(Bn ) (this is a property of probability). Note that, by Central Limit Theorem, √ Sn P(Bn ) ≥ P(An ) = P{Sn ≥ n} = P √ ≥ 1 → P{ξ ≥ 1}, n as n → ∞. In the first inequality, we used the fact that Bn ⊇ An . It suffices to note that 1 P{ξ ≥ 1} = √ 2π Z∞ e−x 2 /2 dx ≈ 0.16 > 0. 1 Thus, P(A) = lim P(Bn ) ≥ P{ξ ≥ 1} > 0. Proof of step 2. Since A ⊆ B, we have: 0 < P(A) ≤ P(B). Proof of step 3. Sn lim √ = lim n→∞ n n→∞ N 1 X Sn − SN √ Xk + √ n k=1 n 2 ! Sn − SN √ , n→∞ n = lim and Sn − SN = n X Xk k=N +1 Sn is independent of X1 , . . . , XN . Therefore, lim √ does not depend on X1 , . . . , XN , and the event n→∞ n B is also independent of these variables. Such events are called tail events. Proof of step 4. It follows from step 3 that B is independent of any event generated by X1 , . . . , XN , for any N . Let N → ∞. Then B is independent of any event generated by X1 , X2 , . . .. But B is itself generated by these variables; so B is independent of itself! Recall that any two events C1 , C2 are independent if P(C1 ∩ C2 ) = P(C1 )P(C2 ). Let C1 = C2 = B. Then P(B) = P2 (B), and P(B) = 0 or 1. This is called Kolmogorov’s 0 − 1 law. Proof of step 5. Combining the results of steps 2 and 4, we get: P(B) = 1. Thus, with probability 1, we have: Sn lim √ ≥ 1. n→∞ n Therefore, with probability 1, Sn > 0 for some n. But this random walk starts from 0. Thus, with probability 1, there exists n such that Sn = 1. Remark. You can actually see that S attains every integer value infinitely many times. This can be easily seen from the fact that (by symmetry of the random walk) Sn lim √ ≥ 1, n→∞ n Sn lim √ ≤ −1. n n→∞ 3