Lecture 19 L 2 -Stochastic integration

Transcription

1 Lecture 19: L 2 -Stochastic integration 1 of 12 Course: Theory of Probability II Term: Spring 215 Instructor: Gordan Zitkovic Lecture 19 L 2 -Stochastic integration The stochastic integral for processes of finite variation For T >, let F : [, T] R be càdlàg function of finite variation, and let F : [, T] [, ) be its total variation. For a < b T we define µ((a, b]) = F(b) F(a), µ({}) = F(). A version of the Caratheodory extension theorem guarantees that the functional µ can be extended in a unique way to a signed (real) measure on B([, T]); this measure is called the Stieltjes measure induced by F, and it is typically denoted by df. It is not hard to see (we leave the details to the reader) that there exist unique càdlàg functions F +, F : [, T] R of finite variation such that 1. df + and df are finite positive measures (do nottake negative values) with df = df + df, and 2. df + and df are absolutely continuous with respect to the measure d F (the Stieltjes measure induced by the total variation F of F) and df + + df = d F. The decomposition df = df + df is called the Hahn-Jordan decomposition of the Stieltjes measure F. Definition 19.1 (Lebesgue-Stieltjes integral). A function h : [, T] R is said to be df-integrable (denoted by h L 1 (df)), if h(t) d F (t) <. For h L 1 (df), its Stieltjes integral with respect to df is defined by h(u) df(u) = h(u) df + (u) h(u) df (u). Last Updated: March 29, 215

2 Lecture 19: L 2 -Stochastic integration 2 of 12 When the function h is continuous, one can approximate the Lebesgue-Stieltjes integral by Riemann-like sums. More precisely, let = t n < tn 1 < < tn d n < t n d n +1 = T be a sequence of partitions of [, T] such that lim sup t n n k+1 tn k =, k d n and suppose that h : [, T] R is a continuous function. We leave it to the reader to show that h L 1 (df) and that h(x) df(x) = lim d n n k= If, additionally, F C 1 ([, T]), we have for all h L 1 (df). h(t) df(t) = h(t n k )(F(tn k+1 ) F(tn k )). h(t)f (t) dt, We turn to processes, now. Let {X t } t [,) be an adapted càdlàg process of finite variation, and let {H t } t [,) be a progressively measurable process such that H t (ω) d X t (ω) <, for all T, for almost all ω, where d X t (ω) is the Stieltjes measure induced by the non-decreasing function t X t (ω) (the total-variation function of X(ω)). Therefore, for almost all ω, and all t, we can define the Stieltjes integral Y t (ω) = t H t (ω) dx t (ω), where dx t (ω) is the Stieltjes measure induced by the FV function t X t (ω). We set Y t (ω) =, for all t for ω in the exceptional set. It is not hard to show that the process {Y} t [,) is an adapted and càdlàg process of finite variation, and that it is continuous when X is continuous. The process {Y t } t [,) is called the stochastic integral of {H t } t [,) with respect to {X t } t [,) and is sometimes denoted by Y = (H X). The importance of being of finite variation Unfortunately, the Lebesgue-style integration stops with functions of finite variation. Here is why. A sequence = t < t 1 < < t k < t k+1 < < of real numbers with the property that lim k t k = +, is called a partition of [, ), and the set of all partitions of [, ) is denoted by P [,). Last Updated: March 29, 215

3 Lecture 19: L 2 -Stochastic integration 3 of 12 The elements t, t 1,... of a partition are referred to as its nodes. For a partition = {t k } k N P [,), and t, we define [,t] = sup k N t k+1 t t k t and note that [,t] = max( t 1 t, t 2 t 1,..., t t k(t) ), where k(t) = sup{k N : t k < t}. A sequence { n } n N in P [,) is said to converge to identity, denoted by n Id if n [,t], for each t. Let F : [, ) R be a continuous function, and let D simp denote the set of all functions h : [, ) R for which there exists a partition = {t k } k N P [,), such that h(t) = α 1 {} (t) + For h S and t > we naturally define t h(u) df(u) = α F() + α k 1 (tk,t k+1 ](t), n N, α k R, k N. (19.1) α k (F(t t k+1 ) F(t t k )). (19.2) Note that the sum is really finite (we could have stopped at the index of the first t k > t). For a continuous function h : [, ) R and a partition P [,), we define the approximation h to h in D simp by h (t) = h()1 {} (t) + h(t k )1 (tk,t k+1 ](t), n N, (19.3) and note that h h uniformly on compact sets. We ask the following question Question What properties does F need to have for the limit t lim h (u) df(u), Id to exist for each t and each continuous h? Remark The reader should note that the question we are asking above really deals with a weak continuity property for the eventual extension of the simple integral (19.2) to a larger class of integrands h. Indeed, if a functional h h df is to be called an integral, it should be linear, and somewhat continuous. The requirement that the uniform convergence of integrands h h implies convergence of the integrals is a particularly weak form of the dominated convergence theorem. Nevertheless, as we will see, it restricts the class of functions F considerably. Last Updated: March 29, 215

4 Lecture 19: L 2 -Stochastic integration 4 of 12 The main analytic tool we are going to use is the celebrated Banach- Steinhaus (aka uniform-boundedness theorem). Remember that for a linear operator T : X Y between two normed spaces (X, X ) and (Y, Y ), we define its operator norm A by T = sup{ T f Y : f X, f X 1}. Theorem 19.4 (Banach, Steinhaus). Let X be a Banach space, let Y be a normed space, and let (B α ) α A be a family of continuous linear operators B α : X Y. Then sup α B B α < if and only if f X, sup B α f Y <. α B Proof. (*) One direction is trivial. For the other, suppose that sup B α f Y < for each f X. α B Define the sequence {D n } n N of subsets of X by D n = { f X : B α f n}, α A which, as intersections of closed sets are themselves closed. By the assumption, n N D n = X. Since X is complete, Baire s theorem 1 implies that at least one of {D n } n N has nonempty interior. More explicitly, there exists n N, f D n and ε > such that f + εb 1 (X) D n, where B 1 (X) = { f X : f X 1} denotes the unit ball of X. Consequently, 1 Baire s (category) theorem states that the intersection of a countable collections of dense open sets in a complete metric space is dense itself. B α ( 1 ε f + f ) Y 1 ε n for all f B 1(X) and all α A. The triangle inequality now implies that B α f Y 1 ε n + B α 1 ε f n ε + 1 ε sup B α f, for all α A, f B 1 (X), α A and so sup α A B α n ε + 1 ε sup α A B α f <. Proposition Let F be a continuous function on [, 1], such that for each continuous function h : [, 1] R, the family 1 h n (u) df(u), n N, is bounded in R for each sequence { n } n N in P [,) such that n Id. Then F is a function of finite variation on [, 1]. Proof. (*) Let X = C[, 1] be the Banach space of all continuous functions on [, 1] normed with h = sup t [,1] h(t), for h C[, 1], and Last Updated: March 29, 215

5 Lecture 19: L 2 -Stochastic integration 5 of 12 let Y = R. We fix a sequence { n } n N in P [,) with n Id, and assume, without loss of generality that 1 n, for all n N. A sequence of operators can be defined by B n : X Y, B n h = 1 h n (u) df(u), n N. These operators are continuous (why?), and, by the assumption, the sequence {B n h} n N is bounded, for each h X. Therefore, by Theorem 19.4, there exists a constant K such that 1 h n (u) df(u) K h, h C[, 1], n N. (19.4) For each n N, one can construct a function h n C[, 1] such that h n () = sgn(f()) and h n (t n k ) = sgn(f(tn k+1 ) F(tn k )), for all k k n(1), where n = {t n k } k N. Moreover, such a function can be constructed with h n = 1. Using (19.4), we get K 1 h n n (u) df(u) k n (1) = F() + F(t n k+1 ) F(t n k ). We have proved that the sequence of variations of F along any sequence of partitions converging to identity remains bounded. It follows (why?) that F must be a function of finite variation. We can repeat the above argument on each [, t] instead on [, 1], and, since any convergent sequence is bounded, the answer to Question 19.2 must be It is necessary that F be of finite variation. On the other hand, when F is of finite variation, the dominated convergence theorem implies that the sequence in Question 19.2 indeed converges towards t h(u) df(u). k= Square-integrable martingales as integrators In spite of the impossibility result of the previous subsection, we will still be able to construct a satisfactory integration theory when the integrand is not of finite variation. In that case, it will be important that some randomness is present and that the class of integrands be restricted to those that do not depend on the future. This way, the terms that would otherwise accumulate and lead to the explosion in the approximating sequence will cancel each other (in the law-of-largenumbers manner). Before we move on to the general case, here is an example in a simplified (but still fully-featured) setting: Example Without knowing it, we have constructed a very simple version of a stochastic integral when we proved the martingale convergence theorem. Indeed, let {ξ n } n N be an iid sequence of coin-tosses Last Updated: March 29, 215

6 Lecture 19: L 2 -Stochastic integration 6 of 12 (P[ξ 1 = 1] = P[ξ 1 = 1] = 1 2 ) and let {M t} t [,1] be defined as follows: M t = n S k 1 [1 2 k+1,1 2 k ) (t), where S 1 k = k ξ k, for t < 1. The martingale convergence theorem guarantees that the limit M 1 = lim t 1 M t exists a.s. The reader can check that the variation of the path M (ω) on [, 1] is given by 1 k ξ k(ω) = 1 k = +, so that no trajectory of {M t } t [,1] is of finite variation 2. On the other 2 The reader will also note that hand, let {H {M t } t [,1] is not continuous; a similar t } t [,1) be any bounded (say by K ) and {Ft M } t [,1] - example with {M t } t [,1] continuous adapted left-continuous stochastic process. On each segment [, t], can be concocted, but it would not t < 1, {M t } t [,1] is a process of finite variation and so we have the following expression t H u dm u = k(t) h k k ξ k, where h k = H 1 2 k+1 and k(t) is such that t [1 2 k+1, 1 2 k ). Moreover, h k σ(ξ 1, ξ 2,..., ξ k 1 ), so that the process N n = n h k k ξ k, is a martingale (it is, really, a martingale transform of {S n } n N ). Moreover, it is a martingale bounded in L 2 ; indeed, by the orthogonality of martingale increments, we have E[N 2 n] = n E[ h2 k ξ 2 n k 2 k ] = 1 K 2 K 2 k 2 1 k 2 <, so that sup n E[Nn] 2 <. Consequently, the martingale convergence theorem implies that the limit N = lim n N N n exists a.s., and in L 2. Equivalently, the random variable 1 t H u dm u = lim H u dm u = lim N n = N, t 1 n is well defined, even though no path of M is of finite variation. It is important to note that the sequence {N n } n N only converges a.s. There is no way to conclude that the limit exists for each ω. Indeed, take h n = 1, for all n N, and note that for those ω for which ξ k (ω) = 1, for all k N, the sequence 1 1 k ξ k (ω) = k = +, does not converge. Moreover, the series 1 k ξ k(ω) will never converge absolutely. It is the fact that the sequence {ξ k } k N fluctuates so much and the fact that h k does not see ξ k that guarantee convergence N n N, and, equivalently, allow for the integral 1 H u dm u to be well-defined. be as transparent as the one we give here. Also, if one prefers to work with [, ) instead of [, 1], one can define M t = S k1 [k 1,k) (t). Last Updated: March 29, 215

7 Lecture 19: L 2 -Stochastic integration 7 of 12 Turning to the general case, we assume that a filtered probability space (Ω, F, {F t } t [,), P), where {F t } t [,) satisfies the usual conditions, is given and fixed throughout. When the integrand is simple the definition of the stochastic integral is easy. More precisely, a stochastic process {H t } t [,) is said to be simple predictable if there exists a partition = {t n } n N P [,) and a sequence {K n } n N of random variables such that K n F tn, and H t = K n 1 (tn,t n+1 ](t). (19.5) n= The set of all simple predictable processes is denoted by H simp. The subset Hsimp of H simp consists of all processes {H t } t [,) H simp such that H H = sup simp n N K n L <. Definition 19.7 (Stochastic integration of simple processes). Suppose that {M t } t [,) is an arbitrary stochastic process, and let {H t } t [,) be a simple predictable process with representation (19.5). The (simple) stochastic integral of H with respect to M is the stochastic process {(H M) t } t [,), given by (H M) t = K n (M t tn+1 M t tn ). (19.6) n= Remark Note that simple predictable processes are predictable (measurable in the σ-algebra generated by left-continuous adapted processes). Moreover, H simp and Hsimp are vector spaces and the functional H is a norm on H simp simp (if indistinguishable processes are identified). 2. One of the best ways to visualize the stochastic integral is to remember the notion of the martingale transform from our discussion of discrete-time martingales. The gambling interpretation used there easily transfers to continuous time, at least when the integrands are simple predictable, as above. Note, though, that one has to think of gambles in an infinitesimal manner in continuous time. 3. The value at t of the stochastic integral is usually denoted by (H M) t or t H u dm u. Sometimes the terminology is abused and the random variable (H M) t (as opposed to the whole process {(H M) t } t [,) or, simply, H M) is called the stochastic integral. The stochastic integral for square-integrable martingales We turn now to a general construction. Let M 2,c denote the family of all continuous martingales M with M = such that M M 2,c := Last Updated: March 29, 215

8 Lecture 19: L 2 -Stochastic integration 8 of 12 sup t E[Mt 2 ] <. Here is the fundamental property of the (simple) stochastic integral with respect to M M 2,c : Proposition 19.9 (Preservation of M 2,c - simple predictable integrands). For M M 2,c and {H t } t [,) Hsimp we have H M M2,c and E[(H M) 2 t ] = E[Kn(M 2 tn+1 t M tn t) 2 ]. (19.7) n= Proof. Continuity, value at time, and the martingale property of H M are evident from the definition. To establish (19.7), we assume that t = t N for some N (otherwise just add it to the partition), and use the fact that the increments of square-integrable martingales are orthogonal, i.e., E[K n (M tn+1 M tn )K m (M tm+1 M tm )] = for n < m. Finally, (H M) M 2,c because, for t, we have E[(H M) 2 t ] = E[Kn(M 2 tn+1 M tn ) 2 ] n N H 2 H simp n = H 2 Hsimp M 2 M 2,c. E[(M tn+1 M tn ) 2 ] We turn now to the notion of quadratic variation of a continuous martingale. The main theorem listing its properties is stated without proof, but the reader is instructed to draw the parallel with the case of the Brownian motion we covered in previous lectures. Before we start, we need to introduce some notation: let A c denote the set of all continuous, adapted, nondecreasing processes {A t } t [,) with A =. We also remind the reader that, for a RCLL of RLCC process X, we define the maximal process X by Xt = sup s t X s. For a sequence {X n } n N of RCLL or RLCC processes, we say that X n converges to X, ucp uniformly on compacts in probability (ucp), and write X n X, if (X n X) t in probability, for each t. Theorem 19.1 (Quadratic variation of square-integrable martingales). For M M 2,c, there exists a unique process M Ac such that such that Mt 2 M t is a uniformly-integrable martingale. Furthermore, for each sequence { n } n N in P [,) with n Id, we have M n ucp M, where, for = {t k } k N P [,), M t := k= (M t k+1 t M tk t) 2. Last Updated: March 29, 215

9 Lecture 19: L 2 -Stochastic integration 9 of 12 The process M A c is called the quadratic variation (process) of M. It admits a natural interpretation as the increasing part in the (continuous-time) Doob-Meyer decomposition of M 2 and plays a central role in stochastic analysis. The reader should probably do Problem?? at this point. Here is how quadratic variation can be used to construct the stochastic integral: Given M M 2,c, it follows from the fact that both M and M2 M are a martingales that E[(M tk+1 M tk ) 2 F tk ] = E[M 2 t k+1 M 2 t k F tk ] = E[ M tk +1 M tk F tk ]. Therefore, the relation (19.7) can be written as [ E[(H M) 2 ] ] = E Kk 2 E[ M t k+1 M tk F tk ] n= (19.8) [ = E K 2 ( ) ] k M tk+1 M tk = E[ Ht 2 d M t ]. n= Inspired by the last expression in (19.8) above, for a progressively-measurable process H, we define H L 2 (M) E[ := Ht 2 d M t]. It is not hard to check that the family L 2 (M) of all progressively-measurable processes H for which H L 2 (M) < forms a vector space, and that L 2 (M) is a norm there. We also note that H simp L2 (M), for each M M 2,c (why?). In this, new, notation, the conclusion of (19.8) looks like this: H M M 2,c = H L 2 (M), H H simp, (19.9) and is called Itô s isometry. We use it to extend the domain of the stochastic integral from Hsimp to a much larger set. To do that, we need a simple, but very powerful, result from real analysis: Proposition Let X be a metric space and Y a complete metric space. Moreover, let f : A Y, with = A X be a uniformly-continuous function. Then, there exists a unique uniformly-continuous function f : Cl A Y such that f (x) = f (x), for all x A. Proof. For x Cl A, let {x n } n N be a sequence in A such that x n x, and let y n = f (x n ). Since x n is convergent in X and f is uniformly continuous, {y n } n N is a Cauchy sequence in Y. Since Y is complete it admits a unique limit y, and we set f (x) = y. Last Updated: March 29, 215

10 Lecture 19: L 2 -Stochastic integration 1 of 12 To show that f is well-defined, we need to argue that y does not depend on the choice of the sequence {x n } n N. Indeed, let {x n} n N be another sequence in A with x n x. Then d(x n, x n), and, so, by uniform continuty, d( f (x n ), f (x n)) =, for each y. It follows that y = lim n f (x n ) = lim n f (x n). Next, we argue that that f is uniformly continuous. Given ε >, let δ > be such that d(x, ˆx) < δ implies d( f (x), f ( ˆx)) < ε, for all x, ˆx A. For any x, x Cl A with d(x, x ) < δ, we can choose two sequences {x n } n N and {x n} n N in A such that x n x and x n x and d(x n, x n) < δ, for all n. It follows that d( f (x n ), f (x n)) < ε, for all n, and, so, d( f (x), f (x )) ε, estalishing the uniform continuity of f on Cl A. Finally, given x A, we may take x n = x, for all n, to conclude that f (x) = f (x). The completeness of the space Y is crucial for the Proposition to work. To check that it is applicable in our case, we need the following result: Proposition (M 2,c, M 2,c ) is a Banach space. Proof. Simple properties of square-integrable martingales suffice to show that M 2,c is a linear space, and that M 2,c satisfies all the axioms of a norm. The more delicate matter is completeness. We start with a Cauchy sequence {M n } n N and conclude immediately that their last elements {M} n n N form a Cauchy sequence in L 2 (F). Thus, M n M in L 2, for some M L 2 (F). We define {M} t [,) as the RCLL version of the Lévy martingale M t = E[M F t ] and use Doob s inequality to get that sup Mt n M t 2 M L n M L 2, 2 t Hence, possibly through a subsequence, we have sup Mt n M t, a.s. t It follows that almost all trajectories of M are uniform limits of the corresponding trajectories of M n. By assumption, all M n are continuous martingales, and uniform limits of continuous functions are continuous, so M is a continuous martingale and M = lim n M n in M 2,c. If we combine the Itô s isometry (19.9) (which yields the uniform continuity of the map H H M on Hsimp ) with Proposition (which asserts that the target space is complete) and use Proposition we immediately conclude that the map H H M M 2,c Last Updated: March 29, 215

11 Lecture 19: L 2 -Stochastic integration 11 of 12 admits a unique linear and continuous extension of the simple predictable integral from H simp to its closure in L2 (M), with values in M 2,c. The only question still remaining is to give a nice description of the closure of Hsimp in L2 (M): Proposition For any M M 2,c, H simp is dense in L2 (M). Proof. Without loss of generality, we assume that E[ M ] = 1. Let { n } n N be a sequence of nested partitions in P [,) such that n Id; for concreteness, we take n = {k2 n : k n2 n }). For each n, let P n be the σ-algebra on the product [, ) Ω generated by all simple predictable processes H of the form H t = n2 n K k 1 (k2 n,(k+1)2 n ](t), K k F k2 n, for all k. (19.1) k= In fact, it is not hard to see that a process H is measurable with respect to P n, precisely when it is of the form (19.1). Also, since all left-continuous processes can be obtained as pointwise limits of the processes of the form (19.1), we have P = σ( n P n ), where P denotes the predictable σ-algebra. The product space Ω = [, ) Ω, together with the σ-algebra P, the filtration {P n } n N, and the probability measure P, given by P [A] = E[ 1 A (t, ω) d M t (ω)], for A P, forms a filtered probability space. Moreover, random variables on this space correspond precisely to the predictable processes on the original space, and their expectations to the expected d M -integrals, i.e., if E denotes the expectation under P, we have In particular, we have Given H L 2 (P ), define E [H] = E[ H t d M t ] where. H n H in L 2 (P ) iff H n H in L 2 (M). H n = E [ H P n ]. This is a square-integrable martingale and H P, so H n H in L 2 (P ). Therefore, by the above characterization of convergence in L 2 (M), we have H n H in L 2 (M), as required. Last Updated: March 29, 215

12 Lecture 19: L 2 -Stochastic integration 12 of 12 We can now summarize all our findings in a compact statement: Theorem Given M M 2,c, there exist a unique linear isometry L 2 (M) H H M M 2,c which extends the simple predictable integral (19.6). Last Updated: March 29, 215