Stochastic Processes in Vector Lattices Lectures 5: Doob-Meyer decomposition JJ Grobler1 and CCA Labuschagne2 18-26 September, 2015 Spring School on Stochastic processes in Functional Analysis, Potchefstroom 1 North-West University, Potchefstroom Campus, Potchefstroom, South Africa. 2 University of Johannesburg, South Africa. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 Outline Optional times Step elements of Stopping times Doob-Meyer docomposition Kolmogorov-C̆entsov inequality JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 Optional times Proposition If S is an optional time and T is a stopping time with S T then PS+ ⊂ PT . Proposition If (Sn ) is a sequence of optional times and S = inf Sn , then ∞ \ PS+ = PSn +. n=1 Moreover, if each Sn is a stopping time, and S Sn , then PS+ = ∞ \ PSn . n=1 JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 Proposition If (Sn ) is a sequence of optional times, then, for S = sup Sn (which is optional), we have [ PSn + ⊂ PS+ . JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 Step elements of Stopping times Proposition Let (Ft ) be a filtration in the Riesz space E and let {Pk : k = 1, . . . , n} in P be a partition of I. Let S ∈ Orth(E) be a step element of the form S := n X tk Pk , k=1 with the tk arranged in strict increasing order, i.e., 0 < tk < tk+1 . Then, S is an optional time if and only if Pk ∈ Pt + for k = 1, . . . , n k and a stopping time if and only if Pk ∈ Ptk for k = 1, . . . , n. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 What is XS ? Classical case: For a stopping time t = S(ω) and a stochastic process Xt (ω) = X (t, ω) one defines effortless XS (ω) := X (S(ω), ω). In the classical case we have Pn no pointset Ω so this is not so simple. However, if S := k=1 tk Pk ∈ Orth(E) the element XS is defined as n X XS = Pk Xtk k=1 The process (XS∧t )t∈T is called the stopped process with reference to the stopping time (or optional time) S. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 The Doob-Meyer decomposition In order to prove this famous theorem in vector lattices, we need to define a number of notions. The Doob-Meyer decomposition plays an important part in the construction of Itô integrals. Definition Let (Ft , Ft )t∈T be a filtration on E and let A = (At ) be an adapted stochastic process. Let Π be the set of all partitions of the interval [0, t]. We call (At ) tractable on [0, t] whenever IA (φt − ) := lim π∈Π n X hφti−1 , Ati − Ati−1 i for i=1 π = {0 = t0 < t1 < · · · < tn = t}, exists for every adapted bounded dual martingale φ = (φt ). The process (At ) is called tractable if it is tractable on every finite interval [0, t]. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 Every martingale M = (Mt ) is tractable since each of the sums in the definition equals zero; hence we have IM (φt − ) = 0 for all φ and for all t. The Poisson process N = (Nt , Ft ) with intensity λ is tractable and if (Bt ) is a Brownian motion, then the submartingale (Bt2 ) is tractable; in fact, in the classical case every increasing process (in the sense of Karatzas and Shreve) is tractable. A simple calculation shows Theorem If X = (Xt ) and A = (At ) are adapted processes such that M = (Xt − At ) is a martingale, then (Xt ) is tractable if and only if (At ) is tractable and for each t, IX (φt − ) = IA (φt − ). JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 Definition The adapted increasing process A = (At ) is called predictable on the interval [0, t] if it is tractable and IA (φt − ) = hφt , At i. It is called predictable if it is predictable on every interval [0, t]. If F is a conditional expectation on E and F = R(F) we let I be ∼ the ideal generated by F∼ in E∼ 00 and note that F maps I into itself. By Nakano’s theorem we have that the canonical image of E ∼ in I∼ 00 is an order dense ideal in I00 . With the notation introduced here, we have JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 Theorem ∼ If E is perfect, i.e., E = (E∼ 00 )00 , and F-universally complete, then, in the duality hE, Ii, we have E = I∼ 00 . If hE, τ i is a locally convex-solid Riesz space, a subset A ⊂ E0 is called order-equicontinuous on E if (Un ) is a ρA -Cauchy sequence whenever 0 ≤ Un ↑≤ U holds in E with ρA (U) := sup{|φ(U)| : φ ∈ A}, U ∈ E. Denote by Sa the set of all bounded step elements S of stopping times with respect to the filtration (Ft , Ft ). ∼ Taking the dual pair hI∼ 00 , Ii = hE, Ii, with the σ(E00 , E) topology, the right continuous stochastic process (Xt , Ft ) is said to be of class DL if the family {XS }S∈Sa is order-equicontinuous on I. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 We have now defined all the necessary concepts in order to state the celebrated Doob-Meyer result in Riesz spaces. Theorem (Doob-Meyer-decomposition) Let (Ft , Ft ) be a filtration on the perfect Riesz space E and assume that E is F0 -universally complete. Let X = (Xt ) be a tractable class DL submartingale, then Xt = Mt + At , 0 ≤ t < ∞, with M := {Mt , Ft , t ≥ 0} a right continuous martingale and A := {At , Ft , t ≥ 0} an increasing predictable process. The decomposition is unique. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 We have defined the element XS for a stochastic process and a stopping time S which is a step element. It is possible to define the element XS for a submartingale that has a Doob-Meyer decomposition. It may be mentioned at this point that it is still a problem to define this in a more general setting. However in the case one has such a decomposition we have a proof of Doob’s optional sampling theorem: Theorem Let (Xt , Ft ) be a right continuous submartingale with the Doob-Meyer decomposition property and let S ≤ T be two optional times of the filtration (Ft , Ft ). Then, if either 1 T is bounded or 2 (Xt ) has a last element X∞ , we have FS+ XT ≥ XS . JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 Stochastic Processes in Vector Lattices Lectures 6: Kolmogorov-C̆entsov inequality JJ Grobler1 and CCA Labuschagne2 18-26 September, 2015 Spring School on Stochastic processes in Functional Analysis, Potchefstroom 1 North-West University, Potchefstroom Campus, Potchefstroom, South Africa. 2 University of Johannesburg, South Africa. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 Outline Optional Skipping Theorem Doob’s Lp -inequality Kolmogorov-C̆entsov inequality JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 Theorem (Optional Skipping Theorem) Let (Xn , Fn ) be a countable submartingale and let Sn ∈ Orth(Fn ) be such that 0 ≤ Sn ≤ I for all n. Set Y1 := X1 , Yn := X1 + n−1 X Sk (Xk+1 − Xk ) k=1 Then 1 (Yn , Fn ) is a submartingale and F1 (Yn ) ≤ F1 (Xn ) for all n. 2 If (Xn , Fn ) is a martingale, so is (Yn , Fn ) and F1 (Yn ) = F1 (Xn ) for all n. 3 If Zn = Yn − X1 then (Zn , Fn ) is a submartingale and F1 (Zn ) ≥ F1 (Zn−1 ) ≥ · · · ≥ F1 (Z1 ) = 0. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 For any two elements X and Y in E we define B(X < Y ) to be the band generated by (Y − X )+ and B(Y ≤ X ) = B(X < Y )d . For the intersection of several of these bands, we write for instance B(X < Y , Z ≤ Y , X > Z ), etc. The corresponding band projection for the last band will be denoted by P(X < Y , Z ≤ Y , X > Z ). Consider the finite submartingale (Xk , Fk )nk=1 and two elements F , G ∈ F1 , F < G . We will construct disjoint upcrossing bands in E as follows: E = B(X1 > F ) + B(X1 ≤ F ); E = B(X1 > F , X2 > F ) + B(X1 > F , X2 ≤ F )+ + B(X1 ≤ F , X2 < G ) + B(X1 ≤ F , X2 ≥ G ); .. . JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 We continue the process by splitting each band in the following way in two disjoint bands in each consecutive step: B(. . . , Xk−1 > F ) = B(. . . , Xk−1 > F , Xk > F )+ B(. . . , Xk−1 > F , Xk ≤ F ), B(. . . , Xk−1 ≤ F ) = B(. . . , Xk−1 ≤ F , Xk < G )+ B(. . . , Xk−1 ≤ F , Xk ≥ G ), B(. . . , Xk−1 < G ) = B(. . . , Xk−1 < G , Xk < G )+ B(. . . , Xk−1 < G , Xk ≥ G ), B(. . . , Xk−1 ≥ G ) = B(. . . , Xk−1 ≥ G , Xk > F )+ B(. . . , Xk−1 ≥ G , Xk ≤ F ), JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 In this manner we obtain 2n disjoint bands, which we shall call the upcrossing bands. We shall say that the process is in an upward phase at k on the bands B(. . . , Xk ≤ F ) and B(. . . , Xk < G ) and in a downward phase at k on the bands B(. . . , Xk > F ) and B(. . . , Xk ≥ G ). The number of upcrossings u(B) on the upcrossing band B is the number of times that the process changes from an upward phase to a downward phase on B. We define the upcrossing orthomorphism X UFG = u(B)PB . B JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 With these preparations the proof of the inequality is easy Theorem (Upcrossing inequality) Let (Xk , Fk )nk=1 be a finite submartingale and let F , G be two elements in F1 = F1 (E) such that F < G . Let UFG be the upcrossing orthomorphism of the process. Then F1 (UFG (G − F )) ≤ F1 [(Xn − F )+ ]. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 We now consider a continuous time submartingale (Xt , Ft )t∈T , an arbitrary closed interval [σ, τ ] and elements F , G ∈ F0 with F < G . Let π = {σ = t1 < t2 < . . . < tn = τ } be a partition of [σ, τ ]. Then (Xtk )tk ∈π is a finite submartingale with upcrossing orthomorphism UFG ,π . We note that if π2 is finer than π1 , then the number of upcrossings increase, i.e., we have UFG ,π1 ≤ UFG ,π2 ; it follows that (UFG ,π )π is an upwards directed system of orthomorphisms. For any Z ≥ 0 we then define the upcrossing orthomorphism of the submartingale (Xt , Ft )t∈T over [F , G ] on the interval [σ, τ ] ⊂ T as UFG ,[σ,τ ] Z = sup UFG ,π Z ∈ Es . π JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 We then have Theorem (Upcrossing inequality: the continuous case) Let (Xt , Ft )t∈T be a submartingale in the Dedekind complete Riesz space E. Let F , G be two elements in F1 = F1 (E) such that F < G and let UFG ,[σ,τ ] be the upcrossing orthomorphism of the process over [F , G ] on the interval [σ, τ ] ⊂ R. Then F1 (UFG ,[σ,τ ] (G − F )) ∈ E and F1 (UFG ,[σ,τ ] (G − F )) ≤ F1 [(Xτ − F )+ ]. Moreover, if E is F1 universally complete, then UFG ,[σ,τ ] (G − F ) ∈ E. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 Similarly, we construct downcrossing bands for a finite submartingale (Xk , Fk )nk=1 and two elements F , G ∈ F1 = F1 (E). Counting the number of downcrossings d(B) on the downcrossing band B we define the downcrossing orthomorphism X DFG = d(B)PB . B We then obtain downcrossing inequalities; the first inequality is obtained by replacing UFG by DFG on the left and F by G on the right; the second inequality is obtained by replacing UFG ,[σ,τ ] by DFG ,[σ,τ ] (defined in the obvious way) on the left and again F by G on the right. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 Theorem (i) Let (Xk , Fk )N k=1 be a finite submartingale. Then, for every λ > 0 we have λF1 [P(sup Xk ≥ λE )(E )] ≤ F1 [P(sup Xk ≥ λE )(XN )] k k ≤ F1 [P(sup Xk ≥ λE )(|XN |)]. k (ii) Let (Xk , Fk )N k=1 be a martingale or a positive submartingale. Then, for every λ > 0 and p ≥ 1 we have λp F1 [P(sup |Xk | ≥ λE )(E )] ≤ F1 [|XN |p ]. k (iii) Let (Xk , Fk )N k=1 be a finite submartingale. Then, for every λ > 0 and p > 1 we have p p p p F1 (|XN | ) ≤ F1 [(sup |Xk |) ] ≤ F1 (|XN |p ). p − 1 k JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 Theorem (i) Let X = (Xt )t∈T be a martingale or a positive submartingale, [σ, τ ] ⊂ T be a closed interval and let λ > 0 be a real number. Then, if X ∗ = sup |Xt | ∈ Es we have for p ≥ 1 t∈[σ,τ ] that λp F1 (P(X ∗ ≥ λE )E) ≤ F1 [|Xτ |p ]. For p > 1, F1 [(X ∗ )p ] ≤ p p−1 p F1 (|Xτ |p ). (ii) (Doob’s Lp -inequality) Let X = (Xt )t∈T be a martingale or a positive submartingale with T an interval in R and let λ > 0 be a real number. Then, if X ∗ = sup |Xt | ∈ Es we have for t∈T ∗ ≥ λE )E ) ≤ sup F [|X |p ]. If p > 1 we p ≥ 1 that λp F1 (P(X t t 1 p p ∗ p p have F1 [(X ) ] ≤ p−1 supt F1 (|Xt | ). JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 Kolmogorov-C̆entsov inequality. Definition Let E be a Dedekind complete Riesz space with weak order unit E . We say that the stochastic process (Xt ) is locally Hölder-continuous with exponent γ if there exist a number δ > 0 and a strictly positive orthomorphism S ∈ Orth E such that, for all s, t ∈ T satisfying |t − s|I ≤ S on a band C, we have |Xt − Xs | ≤ δ|t − s|γ E on the band C. JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6 Theorem (Kolmogorov-C̆entsov) Let E be a Dedekind complete Riesz space with weak order unit E and let F be a conditional expectation defined on E with FE = E . Suppose that (Xt ) is a stochastic process in E such that for some positive constants α, β and C , we have F(|Xt − Xs |α ) ≤ C |t − s|1+β E , 0 ≤ s, t ∈ T . Then the process is locally Hölder-continuous with exponent γ for every γ ∈ (0, β/α); in fact, there exists a maximal disjoint sequence of bands Ck in E such that for n ≥ k and for s, t ∈ [a, b] with |s − t| < 2−n we have PCk (|Xt − Xs |) ≤ δ|t − s|γ E for some constant δ and for every γ ∈ (0, β/α) and ∞ X S := 2−k PCk which is strictly positive. k=1 JJ Grobler and CCA Labuschagne Stochastic Processes, Lecture 5 and 6