In what follows, let the probability space $(\Omega, \mathcal{F}, \mathbb{P})$ as well as the chosen filtration $(\mathcal{F}_t)_{t \ge 0}$ be known, and let $f$ denote an arbitrary locally bounded progressively measurable process (i.e. bounded on compact intervals and for all $t$ measurable with respect to $\mathcal{B}([0,t])\otimes \mathcal{F_t}$).

Consider the process $(Y_t)_{t \ge 0}$ defined by the stochastic integral of $f$ with respect to $(X_t)_{t\ge 0}$: $$Y_t = \int_0^t f(s) \mathrm{d}X_s$$ Which of the following are true?

1. $X_t$ is a martingale $\implies$ $Y_t$ is a local martingale but not necessarily a true martingale.

2. $X_t$ is a locally $L^2-$bounded true martingale $\implies Y_t$ is a locally $L^2$-bounded true martingale.

3. $X_t$ is a local martingale $\implies$ $Y_t$ is a local martingale.

4. $X_t$ is a semimartingale $\implies$ $Y_t$ is a semimartingale.

I just want to know which results are true, and then I will supply the proofs on my own later.

It seems like 2. only holds when we have the additional condition on $f$ that it is locally $L^2$ bounded. In fact, it seems like I may have had it mixed-up: $X_t$ and $Y_t$ need to be globally $L^2$ bounded i.e. square integrable, but $f$ only needs to be locally $L^2$ bounded i.e. square integrable on compact intervals -- do I have this right? I'm still not sure. https://fabricebaudoin.wordpress.com/2012/09/14/lecture-19-stochastic-integrals-with-respect-to-square-integrable-martingales/

EDIT: OK, fixing the assumptions in the manner mentioned above, this is almost certainly true, see Theorem 19 on p.34 of this document: http://math.bu.edu/people/prakashb/Math/stochint.pdf; this also follows from Theorem 6.3 and Lemma 6.4 on p.35 here: https://staff.fnwi.uva.nl/p.j.c.spreij/onderwijs/master/si.pdf; also pp. 137-140 in Revuz and Yor, where being locally bounded also supposedly implies being locally $L^2-$bounded (see the last paragraph of p.140 and consider that $\langle B, B \rangle=t$ and that Brownian motion is a continuous local martingale. So seemingly the condition of being locally bounded is unnecessarily strong in many places where it is used. I am not sure if it can be replaced by locally $L^2-$bounded in all places.

Actually a closer reading of Definition 2.1 on p. 137 of Revuz and Yor implies that 2. holds only for square-integrable integrands $f$, not just locally $L^2-$bounded. If we relax the condition to locally $L^2-$bounded, then only Proposition 2.7 on p.140 is applicable (since martingales are local martingales after all), and we can only conclude that the stochastic integral is a local martingale (although the example given between the end of p.139 and the beginning of p.140 about Brownian motion suggests otherwise -- perhaps if both the integrator and the integrand are locally $L^2$-bounded martingales then we get a martingale, but not square-integrable, as the result?

Also it is worth noting that a corollary of the Martingale representation theorem and the associativity of the stochastic integral (when applicable) allows us to phrase most of the conditions given in terms of integration w.r.t. $\langle M, M \rangle_s$ as conditions in terms of $s$, i.e. $L^2$-boundedness, because the quadratic variation of Brownian motion is $t$.


Well, in the form stated above, none of the statements are true, because you're only assuming $f$ to be progressive and not predictable, and you're not assuming that the integrator $X$ has continuous sample paths.

I'd say that point (4) is neither true nor false but undefined, as the stochastic integral is not necessarily well-defined for integrands which are only progressive and not predictable.

As regards the other three points, as a counterexample, take e.g. $N$ to be a standard Poisson process and let $f(t) = N_t$, $X_t = N_t - t$ and let $(\mathcal{F}_t)$ be the filtration induced by $N$. Then $f$ is locally bounded (by e.g. the sequence of stopping times corresponding to the jump times of $N$), bounded on compacts (because it has cadlag sample paths) and is progressive (because it is cadlag and adapted). Furthermore, the integral is well-defined since $X$ has sample paths of finite variation, so the integral can be defined as a pathwise Lebesgue integral. It holds that $$ Y_t = \int_0^t f(s) dX_s = \int_0^t (N_{s-} + \Delta N_s) dX_s \\ = \int_0^t N_{s-}dX_s + \sum_{0<s\le t}(\Delta N_s)^2 = \int_0^t N_{s-}dX_s + N_t. $$

This functions as a counterexample for points (1-3) because even though $X$ is a locally $L^2$-bounded martingale, $Y$ is not even a local martingale. The problem is that $f$ is not predictable. See also this question for more on this.

If $f$ was assumed predictable, the answers would be:

(1): True. Intuitively, this is because the integral of a predictable process with respect to a local martingale is a martingale, and if $f$ is sufficiently rough, the integral process will not yield integrability, and so a true martingale cannot be expected.

(2): True. Intuitively, this is because the integral process is a local martingale, and by localising so that $f$ is bounded and $X$ is $L^2$-bounded, one obtains $L^2$ boundedness of the integral process.

(3): True. This is almost a defining property of the stochastic integral (depending on the method of construction), but certainly true in any case.

(4): True, also almost by construction, depending on the method of construction.

  • 1
    $\begingroup$ This is a really insightful (perspicacious) and comprehensive answer, with a very accessible counterexample. I appreciate you taking the time to create it. $\endgroup$ – Chill2Macht Nov 28 '17 at 18:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.