Chapter 3 Doléans measures 1. Introduction The definition of the isometric stochastic integral (in Chapter 2) with respect to L 2 -martingales depends on the existence of the Doléans measure, the sigma-finite measure µ M on the predictable sigma-field P characterized by the properties (i) P F M02 = µ M {s = 0, ω ∈ F} for each F ∈ F0 2 (ii) µ M [[0, τ ∧ t]] = PM(τ ∧ t)2 = PM02 + P Mτ ∧t − M0 for each t ∈ R+ and stopping time τ . Notice that the measure µ M depends on M only through M02 and the 2 submartingale Mt − M0 . In fact, analogous measures can be defined for any submartingale S. To simplify the basic construction, I will initially consider only cadlag submartingales indexed by [0, 1] with S0 ≡ 0. For the general case, the measure will be built from countably many finite pieces. <1> Definition. Say that the cadlag submartingale S has a Doléans measure µ, defined on the predictable sigma-field of [0, 1] × , if µ[[0, τ ]] = PSτ for every τ in the collection T of all stopping times taking values in [0, 1]. The assumption S0 ≡ 0 forces µ[[0]] = PS0 = 0, which means that µ concentrates on ((0, 1]], with finite total mass, µ((0, 1]] = PS1 < ∞. The countable additivity of µ imposes (Problem [2]) a subtle extra requirement on S, known somewhat cryptically as property [D]: [D] {Sτ : τ ∈ T} is uniformly integrable. The property will let us calculate expectations at stopping times by passing to the limit from calculations with simpler stopping times. The square integrable submartingales involved in the construction of the isometric stochastic integral all have property [D]: if Mt := P(M1 | Ft ), for 0 ≤ t ≤ 1, then Mτ = P(M1 | Fτ ) and 2 2 for each τ in T. 0 ≤ M τ − M 0 ≤ P M 1 − M 0 | Fτ (Remember that {P(X | G) : G ⊆ F} is uniformly integrable for each integrable random variable X .) 2. Existence of the Doléans measure under property [D] Let {St : 0 ≤ t ≤ 1} be a cadlag submartingale with property [D] and S0 ≡ 0. In this Section, I will construct the Doléans measure µ as an increasing linear Statistics 603a: 28 October 2001 3-1 c David Pollard Section 3.2 Existence of the Doléans measure under property [D] functional on the set H+ of all adapted processes H for which there exist finite constants C H and C H with 0 ≤ H (t, ω) ≤ C H |H (t, ω) − H (s, ω)| ≤ C H |s for all t ∈ [0, 1] and ω ∈ − t| for all s, t ∈ [0, 1] and ω ∈ . It is easy to show that H+ generates the predictable sigma-field P on [0, 1] × . The Daniell existence theorem (UGMTP, §A.6) asserts that finite measures on P correspond precisely to those increasing linear functionals µ on H+ with the “sigma-smoothness” property: µHk → 0 whenever 1 ≥ Hk ↓ 0 pointwise. <2> Theorem. For a cadlag submartingale {St : 0 ≤ t ≤ 1} with property [D], the Doléans measure (as in Definition <1>) exists. Proof. Construct µ as a limit of simpler increasing linear functionals on H+ , with property [D] ensuring its sigma-smoothness. For each n in N define i,n := S((i + 1)/2n ) − S(i/2n ) and write Pi,n (· · ·) for expectations conditional on F(i/2n ). Note that Pi,n i,n ≥ 0 almost surely, by the submartingale property. For each H in H+ , define linear functionals µn H := P H (i/2n )i,n = P H (i/2n )Pi,n i,n . 0≤i<2n 0≤i<2n The second form ensures that µn is an increasing functional. To prove that µH := limn→∞ µn H exists for each H ∈ H+ , I will show that the sequence {µn H : n ∈ N} is Cauchy. For n ≤ m, write t j for the largest multiple of 2−n smaller than j/2m . Then |µm H − µn H | = | P H ( j/2m )i,n − P H (t j ) j,m | 0≤ j<2m ≤ P |H ( j/2m ) − H (t j )| P j,m j,m j ≤ C H 2−n P j,m = C H 2−n PS1 , j which tends to zero as n tends to infinity. The functional µ inherits linearity and the increasing property from the {µn }. For each fixed > 0, property [D] will give a convenient upper bound for µH in terms of the stopping time <3> τ (H, ) := inf{t : H (t, ω) ≥ } ∧ 1. Temporarily write τn for the discretized stopping time obtained by rounding τ (H, ) up to the next multiple of 2−n . Then P {i/2n < τn }Pi,n i,n + C H {i/2n ≥ τn }Pi,n i,n µn H ≤ 0≤i<2n ≤ Pi,n + C H P {i/2n ≥ τn }i,n 0≤i<2n 0≤i<2n ≤ PS1 + C H P S1 − Sτn . <4> Let n tend to infinity. Uniform integrability of the sequence {Sτn } together with right-continuity of the sample paths of S lets us deduce in the limit that µH ≤ PS1 + C H P S1 − Sτ (H,) . Now suppose {Hk : k ∈ N} is a sequence from H+ for which 1 ≥ Hk ↓ 0 pointwise. For a fixed > 0, temporarily write σk for τ (Hk , ). By compactness of [0, 1], the pointwise convergence of the continuous functions, Hk (·, ω) ↓ 0, is actually uniform. For each ω, the sequence {σk (ω)} not only increases to 1, it actually achieves the value 1 at some finite k (depending on ω). Uniform integrability of {Sσk : k ∈ N} and the analog of <4> for each Hk then give as k → ∞. µHk ≤ PS1 + P S1 − Sσk → PS1 Statistics 603a: 28 October 2001 3-2 c David Pollard Chapter 3 Doléans measures The sigma-smoothness of µ follows. The functional corresponds to the integral with respect to a finite measure on P, with total mass µ[[0, 1]] = limn µn 1 = PS1 . It remains to prove that µ[[0, τ ]] = PSτ for τ ∈ T. For given > 0, approximate [[0, τ ]] by the continuous process H (t, ω) = min 1, (τ (ω) + − t)+ / for 0 ≤ t ≤ 1. It is adapted because {H (t, ω) ≤ c} = {τ ≤ t − (1 − c)} ∈ Ft , for each fixed t and constant 0 ≤ c < 1. It belongs to H+ , with C H = 1 and C H = 1/, and [[0, τ ]] ≤ H ≤ [[0, τ + ]]. Hε 1–Hε ε τ 0 τ+ε 1 −n When 2 < , direct calculation shows that µn H ≤ PSτ +2 , which in the limit implies µH ≤ PSτ +2 and, by Dominated Convergence, µ[[0, τ ]] = lim→0 µH ≤ PSτ . Inequality <4> applied with τ := τ (1 − H , ) gives µ((τ + , 1]] ≤ µ 1 − H ≤ PS1 + P S1 − Sτ , which, in the limit as tends to zero, implies µ((τ, 1]] ≤ P S1 − Sτ . From the fact that PS1 = µ[[0, 1]] = µ[[0, τ ]] + µ((τ, 1]] ≤ PSτ + P S1 − Sτ , <5> 3. conclude that µ[[0, τ ]] = PSτ . Example. Let {Bt : 0 ≤ t ≤ 1} be a Brownian motion with continuous sample paths. The submartingale St := Bt2 has a very simple Doléans measure, characterized for F ∈ Fα by µ S (α, β] ⊗ F = PF Bβ2 − Bα2 = PF(β − α). That is, µ S = m ⊗ P, with m equal to Lebesgue measure on B[0, 1]. Of course µ S has a further extension to the product sigma-field B[0, 1] ⊗ F. A Poisson process {Nt : 0 ≤ t ≤ 1} with intensity 1 shares with Brownian motion the independent increment property, but the increment Nt − Ns has a Poisson(t − s) distribution. The sample paths are constant, except for jumps of size 1 corresponding to points of the process. The process {Nt : 0 ≤ t ≤ 1} is a submartingale with respect to its natural filtration, with Doléans measure m ⊗ P, the same as the square of Brownian motion. Clearly the Doléans measure does not uniquely determine the submartingale: both squared Brownian motion and the Poisson process have Doléans measure m ⊗ P. But the only square integrable martingale M with continuous sample paths and Doléans measure µ M = m ⊗ P is Brownian motion: if F ∈ Fs and s < t then PF(Mt2 − Ms2 ) = µ M (s, t] ⊗ F = (t − s)PF, from which it follows that Mt2 − t is a martingale with respect to {Ft }. It follows from Lévy’s characterization that M is a Brownian motion. Doléans measure for submartingales Once more consider processes indexed by R+ . Redefine T to be the set of all stopping times taking values in [0, ∞]. Say that a submartingale {St : t ∈ R+ } Statistics 603a: 28 October 2001 3-3 c David Pollard Section 3.3 Doléans measure for submartingales has property [D] if {Sτ {τ < ∞} : τ ∈ T} is uniformly integrable. For such a submartingale, the variables St converge almost surely and in L 1 to a limit S∞ , and {St : 0 ≤ t ≤ ∞} is a submartingale (Problem [3]). Clearly {S∞ {τ = ∞} : τ ∈ T} is uniformly integrable. Thus we may reinterpret the property [D] for S as an assertion that {Sτ : τ ∈ T} is uniformly integrable. Via any one-to-one increasing map ψ from [0, ∞] onto [0, 1], we may identify S with a submartingale indexed by [0, 1] having the property [D] as it was understood in Section 2. The corresponding Doléans measure µ0 puts zero mass on [[1]], because µ0 ((ψ(n), 1]] = P(S∞ − Sn ) → 0 as n → ∞. The inverse map ψ −1 transfers µ0 back to a finite measure on the predictable sigma-field of R+ × for which µ[[0, τ ]] = PSτ for each τ ∈ T. Remark. The construction from Section 2 could have been carried out directly for R+ × , at the cost of some extra precautions to keep mass from straying off to infinity. Alternatively, we could apply an analog of the construction from Section 2 to build Doléans measures for each [[0, n]], with n ∈ N, which could then be pieced together in much the same way as described below. The construction of Doléans measures also extends to processes that are only locally submartingales with property [D], that is, processes X for which there exist an sequence of stopping times τk ↑ ∞ for which each X ∧τk is a submartingale with property [D]. The sequence {τk } is said to reduce X to a submartingale with property [D]. The extension is important partly because of the following result. <6> <7> Extend to local submartingales? Lemma. Every cadlag submartingale has property [D] locally. Proof. For a cadlag submartingale {St : t ∈ R+ }, define stopping times τk = k ∧ inf{t : |St | ≥ k}. It suffices to show that Sτk is integrable, for each fixed k, for then uniform integrability will follow from the fact that each |Sτ ∧τk | is bounded by the same integrable random variable k + |Sτk |. Let {Mt : 0 ≤ t ≤ k} be the cadlag version of the uniformly integrable martingale P(Sk | Ft ). The random variable Mτk = P Mk | F(τk ) is integrable. The process X t := Mt − St , for 0 ≤ t ≤ k, is a positive supermartingale. By the Stopping Time Lemma, 0 ≤ PX τk ≤ PX 0 < ∞. The random variable |Sτk | is bounded by the sum X τk + |Mτk |, which is integrable. Theorem. Let {St : t ∈ R+ } be a cadlag submartingale, reduced to a process with property [D] by stopping times {τk }. Suppose S0 ≡ 0. Then there exists a sigma-finite measure µ on the predictable sigma-field of R+ ⊗ characterized by the property µ[[0, τ ]] = limk→∞ PSτ ∧τk for each stopping time τ . Proof. Write µk for the Doléans measure generated on P by S∧τk . First note that µk [[0, ∞]] = PS∞∧τk = µn [[0, τk ]], which implies that the measure µk concentrates on the stochastic interval [[0, τk ]]. Then note that, for each stopping time τ , µk [[0, τ ]] = PSτ ∧τk = PS(τ ∧ τk ∧ τk+1 ) = µk+1 [[0, τ ]] ∩ [[0, τk ]] , which implies that the restriction of µk+1 to [[0, τk ]] equals µk . For each nonnegative, predictable H the sequence {µk H } is increasing. Its limit, µH , defines a linear functional with the monotone convergence property; the limit is the Doléans measure of S. By definition, µ[[0, τ ]] = limk µk [[0, τ ]] = lim PSτ ∧τk for every stopping time τ . Statistics 603a: 28 October 2001 3-4 c David Pollard Chapter 3 Doléans measures To see that the limit does not depend on the particular choice for the reducing sequence, let {σk } be another reducing sequence of stopping times, with associated Doléans measures {νk } and limit ν. For each j and k, and each stopping time τ , ν j [[0, τ ∧ τk ]] = PSτ ∧τk ∧σj = µk [[0, τ ∧ σ j ]]. 4. Let j tend to infinity. The interval [[0, τ ∧ σ j ]] increases pointwise to [[0, τ ]], even if τ is allowed to take infinite values (a benefit of excluding {∞} ⊗ from the stochastic interval). From the definition of ν and the countable additivity of µk , deduce that ν[[0, τ ∧ τk ]] = µk [[0, τ ]]. Then let k tend to infinity to deduce that ν and µ give the same value to each stochastic interval [[0, τ ]], for τ ∈ T. Also ν[[0]] = 0 = µ[[0]]. The equality for all sets in P follows by a generating class argument. Problems [1] Let {St : 0 ≤ t ≤ 1} be a cadlag martingale with S0 ≡ 0. For a fixed λ ∈ R+ , define stopping times σ := inf{t : St ≤ −λ} ∧ 1 and τ := inf{t : St ≥ λ} ∧ 1. (i) Show that λP{supt St > λ} ≤ λP{Sτ ≥ λ} ≤ PSτ {Sτ ≥ λ} ≤ PS1 {Sτ ≥ λ} ≤ P|S1 |. (ii) Show that λP{inft St < −λ} ≤ P(−Sσ ){Sσ ≤ λ} ≤ −PSσ + PS1 {Sσ ≥ −λ} ≤ P|S1 |. (iii) Deduce that λP{supt |St | > λ} ≤ 2P|S1 |. [2] Suppose a cadlag martingale {St : 0 ≤ t ≤ 1}, with S0 ≡ 1, has a Doléans measure µ in the sense of Definition <1>, that is, µ[[0, τ ]] = PSτ for every τ ∈ T. Show that S has property [D] by following these steps. (i) Invoke the continuous version of the Stopping Time Lemma (from Chapter 1) to show that 0 ≤ PSτ and PSτ+ ≤ PS1+ for each τ ∈ T. Deduce that supτ ∈T P|Sτ | ≤ κ := 2PS1+ < ∞. (ii) For each C ∈ R+ , show that PSτ {Sτ > C} ≤ PS1 {Sτ > C}. (Hint: STL again.) Deduce, via the integrability of S1 and the bound P{|Sτ | > C} ≤ κ/C that supτ ∈T PSτ {Sτ > C} → 0 as C → ∞. (iii) Show that every cadlag function on [0, 1] is bounded in absolute value. Deduce that the stopping time σC := inf{t : St < −C} ∧ 1 has σC (ω) = 1 for all C large enough. Deduce that µ((σC , 1]] → 0 as C → ∞. (iv) For a given τ ∈ T and C ∈ R+ , define Fτ := {Sτ < −C}. Show that τ := τ Fτc + Fτ is a stopping time for which P S1 − Sτ Fτ = P Sτ − Sτ = µ((τ, τ ]] ≤ µ((σC , 1]], then deduce that supτ ∈T P − Sτ {Sτ < −C} → 0 as C → ∞. Hint: Show that if ω ∈ Fτ then σC (ω) ≤ τ (ω) and if ω ∈ Fτc then τ (ω) = τ (ω). [3] Let {St : t ∈ R+ } be a submartingale of class [D]. Show that there exists an integrable random variable S∞ for which P S∞ | Ft ≥ St → S∞ almost surely and in L 1 by following these steps. (i) Show that the uniformly integrable submartingale {Sn : n ∈ N} converges almost surely and in L 1 to an S∞ for which P S∞ | Fn ≥ Sn . (ii) For t ≤ n, show that St ≤ P P(S∞ | Fn ) | Ft = P S∞ | Ft . Statistics 603a: 28 October 2001 3-5 c David Pollard Section 3.4 Problems (iii) For t ≥ n, show that P(St − Sn )− ≤ P(St − Sn )+ ≤ P(S∞ − Sn )+ → 0 as n → ∞. (iv) For each k ∈ N, choose n(k) for which P|S∞ − Sn(k) | ≤ 4−k . Invoke Problem [1] to show that k P{supt≥n(k) |St − Sn(k) | > 2−k } < ∞. Deduce that St → S∞ almost surely 5. Notes My exposition in this Chapter is based on ideas drawn from a study of Métivier (1982, §13), Dellacherie & Meyer (1982, Chapter VII), and Chung & Williams (1990, Chapter 2). The construction in Section 2 appears new, although it is clearly closely related to existing methods. References Chung, K. L. & Williams, R. J. (1990), Introduction to Stochastic Integration, Birkhäuser, Boston. Dellacherie, C. & Meyer, P. A. (1982), Probabilities and Potential B: Theory of Martingales, North-Holland, Amsterdam. Métivier, M. (1982), Semimartingales: A Course on Stochastic Processes, De Gruyter, Berlin. Pollard, D. (2001), A User’s Guide to Measure Theoretic Probability, Cambridge University Press. Statistics 603a: 28 October 2001 3-6 c David Pollard