Homework Assignment #5 1: A family {X(t) : t ≥ 0} of square integrable, R-valued random variables on a probability space (Ω, F, P) is called a centered Gaussian process if its span in L2 (P; R) Gaussian family. is a centered, Given such a process, define its covariance function C(s, t) ≡ EP X(s)X(t) . (i) Given 0 ≤ t0 < · · · tn , let µt0 ,...,tn be the distribution of X(t0 ), . . . , X(tn ) and express the Fourier transform of µt0 ,...,tn in terms of C. In particular, conclude that C uniquely determines the distribution of {X(t) : t ≥ 0} as a stochastic process. (ii) Assuming that for each T > 0 there exist a K(T ) < ∞ and αT > 0 such that C(t, t) − 2C(s, t) + C(s, s) ≤ K(T )(t − s)αT for s, t ∈ [0, T ], show that there exists a version that is continuous in the sense that there exists a family {X̃(t) : t ≥ 0} such that t X̃(t) is continuous and X̃(t) = X(t) (a.s.,P) for each t ≥ 0. C(s,t) X(s) is independent of X(s) and (iii) Assume that C(s, s) > 0, and show that X(t) − C(s,s) therefore that, for any Borel measurable ϕ : R −→ [0, ∞), s ! Z C(s, t) C(s, t)2 ϕ X(s) + (C(t, t) − y γ0,1 (dy) C(s, s) C(s, s) R is the conditional expectation of ϕ X(t) given σ({X(s)}). (iv) Show that {X(t) : t ≥} is Markov in the sense that, for all 0 ≤ s < t and non-negative Borel measurable ϕ’s EP ϕ X(t) σ {X(τ ) : τ ∈ [0, s]} = EP ϕ X(t) σ {X(s)} , if and only if C(τ, t)C(s, s) = C(s, t)C(τ, s) for all 0 ≤ τ ≤ s < t. (v): Take Ω = C [0, ∞); R , F = BΩ , and P = W. Let h : [0, ∞)2 −→ R be a Borel measurable function for which Z ∞ |h(t, τ )|2 dτ < ∞ for all t ≥ 0, 0 and set Z X(t) = t h(t, τ ) dw(τ ). 0 Show that {X(t) : t ≥} is a centered Gaussian process with covariance Z ∞ C(s, t) = h(s, τ )h(t, τ ) dτ. 0 Show that this process will have a continuous version if for each T > 0 there exist a K(T ) < ∞ and αT > 0 such that Z ∞ |h(t, τ ) − h(s, τ )|2 dτ ≤ K(T )(t − s)αT for 0 ≤ s < t ≤ T. 0 2: Aside from Brownian motion, the most famous Gaussian process is the Ornstein-Uhlenbeck process. Namely, continue in the setting of (v) above, and, for x ∈ R and w ∈ C [0, ∞); R , show that there is precisely one solution X( · , x)(w) to the integral equations Z X(t, x) = x + w(t) − (1)1 t X(τ, x)(w) dτ. 0 In fact, use Riemann-Stieltjes integration to show that −t X(t, x) = e x + e −t Z t eτ dw(τ ). 0 In particular, conclude that {X(t, 0) : t ≥ 0} is a centered Gaussian process with covariance C(s, t) = e−|s−t| − e−(s+t . 2 Although one can use (iv) above to check that {X(t, 0) : t ≥ 0} is Markov, there is a better way. Namely, show that, for 0 ≤ s < t, X(t, x)(w) = e −(t−s) −s−t Z X(s, x)(w) + e = X t − s, X(s, x)(w) (δs w), t eτ dw(τ ) s where δs w(t) = w(s + t) − w(s), and conclude that W E ϕ X(t, x) Bs = Z ϕ(y)p t − s, X(s, x), y) dy where p(τ, x, y) = π(1 − e −2τ ) − 21 y − e−τ x exp − 1 − e−2τ Finally, show that {X(t, 0) : t ≥ 0} has the same distribution as −t e w e2t − 1 2 : t≥0 . .