Proces Wienera
Transkrypt
Proces Wienera
. Proces Wienera Karol Dziedziul Gdansk Karol Dziedziul Proces Wienera . A schedule of the lecture Hitting times Reflection principle Hitting time of a two sided barrier for BM Karol Dziedziul Proces Wienera . Wiener process Defintion Let Wt be a stochastic process which satisfies the following conditions (a) It starts at zero W (0) = 0 (b) It has independent increments: for any partition 0 ≤ t0 < t1 < · · · < tm ≤ 1 and any m the random variables Wt1 − Wt0 , . . . , Wtm − Wtm−1 are independent. (c) For every 0 ≤ s < t ≤ 1, Wt − Ws has a Gaussian distribution with mean zero and variance σ 2 = t − s. ♣ Use Program 20 to create the estimation of a sample path of the √ Wiener process. In algorithm we use a random walk with a δ = 1/ n step. We start from 0, Sn (0) = 0. At the discrete moment t = kδ we are in position Sn (t) and we decide where to go in a moment t + δ = (k + 1)δ. Hence, we use the rule ξ Sn (t + δ) = Sn−1 (t) + √ = Sn−1 (t) + δξ n where the next step ξ ∈ N(0, 1) is independent of the past by the time t . Next, we join two points with an interval. Karol Dziedziul Proces Wienera . Random walk Namely, we create a piecewise linear process given by Sn (t) = ξ1 + · · · ξ[nt] ξ[nt]+1 √ + (nt − [nt]) √ , n n where ξj , j = 1, 2, ... are independent with the standard normal distribution N(0, 1). Note that Sn : Ω → C ([0, ∞)), where C ([0, T ]) denotes the space of continuous function defined on interval [0, 1]. To formulate the next construction of a Wiener process Wt , t ∈ [0, 1] we need to define a Haar system. Karol Dziedziul Proces Wienera . Haar system A wavelet system on R starts with one 1 for −1 for h1 (t) = 0 function. Let 0 ≤ t < 2−1 2−1 ≤ t ≤ 1 t ∈ R \ [0, 1] We define a wavelet (Haar) system on R: for all k, j ∈ Z hk,j (x) = 2k/2 h1 (2k x − j), x ∈ R. This system is an orthonormal basis in L2 (R). This concept will be explained later. Note that j j +1 supphk,j := {x ∈ R : hk,j (x) 6= 0} = k , k . 2 2 Note that supphk,j ⊂ [0, 1] if and only if 0 ≤ j < 2k and k ≥ 0. Karol Dziedziul Proces Wienera . We define a Haar system on interval [0, 1] by choosing the functions hk,j , where 0 ≤ j < 2k and k ≥ 0 and restricting these functions to the interval [0, 1]. We have to add also h0 , a constant function h0 (t) = 1, t ∈ [0, 1]. The equivalent definition for Haar system on [0, 1] is the following: for any natural number n ≥ 1 there is a unique representation of n by n = 2k + j, 0 ≤ j < 2k , and k, j ∈ N. For example 13 = 23 + 5, 20 = 24 + 4. A Haar function with an index n ≥ 1 is defined for t ∈ [0, 1] by k 2j 2j+1 2 2 for 2k+1 ≤ t < 2k+1 k hn (t) = hk,j (t)|[0,1] = −2 2 for 2j+1 ≤ t ≤ 2j+2 2k+1 2k+1 0 otherwise and h0 is defined above. Karol Dziedziul Proces Wienera . We define an inner product for square integrable functions Z 1 < f , g >= f (x)g (x)dx. 0 By a square integrable function f we understand a function such that Z 1 |f (x)|2 dx < ∞. 0 The space of such functions is denoted by L2 [0, 1]. Note that for all n 6= m < hn , hm >= 0 and for all n < hn , hn >= 1. The Haar system is an orthonormal basis in L2 [0, 1], i.e. for any f ∈ L2 [0, 1] we have following representation called a reconstruction formula ∞ X f = < f , hn > hn . n=0 The convergence is in the norm of L2 [0, 1]. Karol Dziedziul Proces Wienera . 2 Orthonormal basis in L [0, 1]2 There is a similar Haar system Hn is in L2 [0, 1]2 . When transmitting data (a picture f ) by internet, we send N coefficient < f , Hn > and then we reconstruct f by fN f ≈ fN = N X < f , Hn > Hn . n=0 If we want to enlarge a picture (in google earth, for instance), it is sufficient to transmit only the next coefficients: f ≈ fN+M = N+M X < f , Hn > Hn = fN + n=0 N+M X n=N+1 Karol Dziedziul Proces Wienera < f , Hn > Hn . . Ciesielski’s construction of Wiener process A sample path of the Wiener process is piecewiese linear. Ciesielski’s construction of Wiener process allows us to overcome the difficulty of estimation. Let {ξj } be a sequence of independent variables with the standard normal distribution. Then Z t ∞ ∞ X X W (t) = ξn hn (s)ds = ξn ϕn (t) n=0 0 n=0 is the Wiener process. The series converges a.s. in C [0, 1]. The functions Z t hn (s)ds ϕn (t) = 0 are known as Schauder’s functions. They are continuous and piecewise linear. Moreover, if n = 2k + j, 0 ≤ j < 2k , then for all 0 ≤ m ≤ 2k , m 6= 2j + 1 m 2j + 1 = 1. ϕn k+1 = 0, ϕn 2 2k+1 Karol Dziedziul Proces Wienera . Let us assume that we have a process ψK (t) := K 2X −1 ξn ϕn (t). n=0 To construct a piecewise linear trajectory of the process ψK (t) it is sufficient to know the value of the trajectory in points t= i , 2K 0 ≤ i ≤ 2K . Note that ψK +1 (t) = +1 2KX −1 ξn ϕn (t) = ψK (t) + +1 2KX −1 ξn ϕn (t). n=2K n=0 To draw a piecewise linear trajectory of the process ψK +1 we ONLY have to know the value of a trajectory in points t= 2i + 1 , 2K +1 Karol Dziedziul 0 ≤ i < 2K . Proces Wienera . Why? The reason is simple ψK (t) = ψK +1 (t), for t= i , 2K 0 ≤ i ≤ 2K . From the definition of ϕn we obtain that for t= 2i + 1 , 2K +1 0 ≤ i < 2K , we have ψK +1 2i + 1 2K +1 = ψK i 2K + ψK 2 i+1 2K + 1 √ 2 K +1 ξ, where a random variable ξ has the normal standard distribution N(0, 1) and ξ is independent of previous generated random variables. This is a key formula in Program 21. ♣ Use Program 21 to create a sample path of Wiener process. Karol Dziedziul Proces Wienera . Multivariate Gaussian distribution We consider two dimensional vector T (X1 , X2 ) = X1 X2 Let Σ denote the covariance matrix of (X1 , X2 ) a11 a12 Σ= a21 a22 , where ai,j = Cov (Xi , Xj ) = EXi Xj − EXi EXj . Recall that det Σ = a11 a22 − a12 a21 . When Σ is nonsingular (det Σ 6= 0), then the density of Gaussian distribution is defined for (x, y ) ∈ R 2 by f (x, y ) = T 1 1 √ e − 2 (x−µ1 ,y −µ2 ))Q(x−µ1 ,y −µ2 ) , 2π detΣ where µj = EXj . Karol Dziedziul Proces Wienera . The matrix Q = Σ−1 is invertible to matrix Σ, i..e. 1 0 QΣ = ΣQ = . 0 1. From algebra we know that there exits a nonnegative symmetric square root of Σ, i.e. a matrix A such that AAT = Σ If (Z1 , Z2 ) are independent and have the standard normal distribution, then a random vector A(Z1 , Z2 )T + µ has the same distribution as the vector (X1 , X2 )T . This is clue formula how to generate a sample. ♣ Use Program 22 to generate a sample of (X1 , X2 )T , where µ = (1, 2) 1 −1 A= . 0 2 Observe a histogram and a smooth estimator of the density. Karol Dziedziul Proces Wienera . Karol Dziedziul Proces Wienera . Brownian motion Let (Zt1 , Zt2 ) be two independent copies of the Wiener process. A sample path of t → (Zt1 , Zt2 ) ∈ R 2 is a visualization of a Brownian motion. ♣ Use Program 27 to generate sample paths of two-dimensional Brownian motion. Karol Dziedziul Proces Wienera . Two-dimensional correlated Wiener process We create two independent copies of the Wiener process Zt1 and Zt2 . Next, we define a process Wt = ρ1 Zt1 + ρ2 Zt2 , where ρ1 , ρ2 > 0 and ρ21 + ρ22 = 1. It turns out that Wt is the Wiener process. For each i = 1, 2 processes (Zti , Wt ) define two dimensional correlated Wiener process. Therefore, (Zti , Wt ) for i = 1, 2 are two dimensional Gaussian vectors and such that Cov (Zti , Wt ) = ρi t, i = 1, 2. ♣ Use Program 23 to generate sample paths of these three Wiener processes. Change the parameter ρi . Karol Dziedziul Proces Wienera . Laws of the iterated logarithm Theorem (Laws of the iterated logarithm) Let Wt be the Wiener process. Then Wt lim sup p =1 t→∞ 2t log log t ♣ Use Program 24 to check the thesis of Theorem Theorem (Maximum process) Let Wt be the Wiener process. Let Mt = sup Ws . 0≤s≤t Then for all a ≥ 0 P(Mt > a) = 2P(Wt > a) = P(|Wt | > a). Karol Dziedziul Proces Wienera . Karol Dziedziul Proces Wienera . ♣ Use Program 25 to check that random variables M1 and |W1 | have the same distribution. ♣ Use Program 3 to create the arcsine law. This distribution is an example of the beta distribution with parameter a = b = 0.5. Theorem ( Levy Theorem) Let Wt be the Wiener process, t ∈ [0, 1]. Then the optional times τ1 = inf{t ∈ [0, 1] : Wt = M1 }, τ2 = sup{t ∈ [0, 1] : Wt = 0} have the arcsine distribution. ♣ Use Program 26 to check the statement of the Theorem for the optional time τ2 . Karol Dziedziul Proces Wienera . Hitting times Let Wt be a Brownian motion. For Wt we denote by Ta the first hitting time of a level (a barrier) a > 0 by Wt , i.e. Ta = inf{t > 0 : Wt = a}. Program 39. Ta is a random variable such that for every scenario ω ∈ Ω, Ta (ω) = t, where Wt (ω) = a is the first moment of reaching the level a. Question: what is a distribution of RV Ta for given a? Let us create an empirical distribution of Ta for a = 0.5. We observed that there is a chance that a trajectory does not reach the level a. If we observe trajectories in time interval [0, 10] and if a trajectory does not reach the level a = 0.5 then in this case we will put t = 10.5. Program 40. Karol Dziedziul Proces Wienera . Reflection principle To calculate the distribution of Ta we need to use a reflection principle. Let us define the maximum process of the Brownian motion by Mt = sup Wt . s≤t Note that for y ≥ 0 {Mt ≥ y } = {Ty ≤ t}. Obviously if Ty (ω) ≤ t it means that the Brownian motion reached the level y to the time t, hence the maximum Mt ≥ y . Theorem For y ≥ 0, x ≤ y we have P(Wt ≤ x, Mt ≥ y ) = P(Wt ≥ 2y − x). Consequently if we take in the above formula x = y then for y ≥ 0 P(Wt ≤ y , Mt ≥ y ) = P(Wt ≥ y ). Karol Dziedziul Proces Wienera . The distribution function of hitting time On the other hand for y ≥ 0 P(Wt > y , Mt ≥ y ) = P(Wt > y ) = P(Wt ≥ t). The sets {Wt > y , Mt ≥ y } and {Wt ≤ y , Mt ≥ y } are disjoint. Thus for y ≥ 0 P(Mt ≥ y ) = P(Wt > y , Mt ≥ y ) + P(Wt ≤ y , Mt ≥ y ) = 2P(Wt > y ). √ But Wt ∈ N(0, t). Hence for y ≥ 0 √ √ √ P(Wt > y ) = P(Wt / t > y / t) = 1 − Φ(y / t), where Φ is a distribution function of the standard normal distribution. Hence for y ≥ 0 √ Fy (t) = P(Ty ≤ t) = P(Mt ≥ y ) = 2(1 − Φ(y / t)) Karol Dziedziul Proces Wienera . The density function of hitting time The density function for t ≥ 0 √ f (t) = Fy0 (t) = φ(y / t)yt −3/2 , where φ is the density function of the standard normal distribution. Recall 2 1 φ(x) = √ e −x /2 . 2π We can compare the empirical distribution with the theoretical one. Program 41 y = 0.5. For given y > 0 the expected time of reaching the barrier y is infinity i.e. Z ∞ Z ∞ t −y 2 /(2t) y 1 √ dt = ∞, ETy = e dt ∼ 3/2 2π t t 0 1 There is not a singularity at t = 0, since √ e −1/t lim+ √ = lim e −u u = 0. u→∞ t→0 t Karol Dziedziul Proces Wienera . Hitting time of a two sided barrier for BM Let a < 0 < b and Ta , Tb the two hitting times. Recall Ta = inf{t > 0 : Wt = a}. Use Program 42 and 43 to see the following phenomena Theorem Let T ∗ = min{Ta , Tb } then P(Ta < Tb ) = b . a−b Moreover ET ∗ = −ab. Compare two cases |a| > |b| and |a| < |b|. The density function is a rather complicated. Karol Dziedziul Proces Wienera