Stat/219 Math 136 - Stochastic Processes Note on Continuity of Processes Kolmogorov’s criteria for a continuous modification Kolmogorov’s criteria, as stated in class, is the following. Given a stochastic process {Xt , t ≥ 0} if there exist constants α > 0, β > 0, and C > 0 for which IE(|Xt − Xs |α ) ≤ C|t − s|1+β , for all 0 ≤ s, t < ∞ (0.1) then there exists a continuous modification of {Xt , t ≥ 0}. Theorem 3.3.3 in the Lecture Notes is just a refinement of this criteria, relating the “degree of smoothness” (e.g. Holder continuity) in terms of the constants α and β. Note that we must require β > 0. As an example of why β = 0 does not work, consider Ω = [0, 1] with its Borel σ-field and the uniform probability measure. Let U (ω) = ω; that is, U is a Uniform[0,1] random variable. Define Xt (ω) = I{U ≤t} (ω), t ≥ 0. Note that |Xt+h − Xt | is 1 if 0 ≤ t < 1 and t < U ≤ t + h, and |Xt+h − Xt | is 0 otherwise. So for 0 ≤ t < 1 and any h > 0, IE|Xt+h − Xt | = IP(t < U ≤ t + h) ≤ h, where the inequality is due to the possibility that t + h > 1. So {Xt , t ≥ 0} satisfies (0.1) with C = 1, α = 1, β = 0, but clearly it has no continuous modification (see below). Continuity in probability does not imply continuity of sample paths Consider again the example above. Using Markov’s inequality, we can show that {Xt , t ≥ 0} is continuous in probability: For ε > 0, IP(|Xt+h − Xt | > ε) ≤ 1 h IE|Xt+h − Xt | ≤ ε ε which approaches 0 as h → 0 for any ε > 0. However, consider A = {ω : t 7→ Xt (ω) is continuous }. By definition of Xt , A = {0}. For if ω = 0 then U (ω) = 0 and Xt (ω) = 1 for all t ≥ 0; so this particular sample path is continuous. (If ω > 0 then U (ω) > 0 and clearly the associated sample path is not continuous.) So IP(A) = IP({0}) = 0, and hence {Xt ≥ 0} does not have continuous sample paths almost surely. (We have actually shown the stronger statement: that almost surely {Xt } has discontinuous sample paths.) Kolmogorov’s criteria is stronger than just continuity in probability If {Xt , t ≥ 0} satisfies Kolmogorov’s criteria for some α, β, C > 0, then there is a continuous modification. The Kolmogorov criteria also implies that the process is continuous in probability. This is true since for any 1 ε>0 IP(|Xt − Xs | > ε) ≤ ≤ 1 IE(|Xt − Xs |α ) by Markov’s inequality εα C |t − s|1+β by Kolmogorov’s criteria εα Since β > 0, this last term approaches 0 as |t − s| → 0. Therefore the process is continuous in probability. What is stronger about the Kolmogorov criteria is that it puts a requirement (in terms of α and β) on the rate of convergence. 2