Show that $(S_t - M_t) \phi(S_t) = \Phi(S_t) - \int_{0}^{t} \phi(S_s) dM_s$

(This is from Le Gall's book, Brownian Motion, Martingales, and Stochastic Calculus.)

Here, $M$ is a continuous local martingale, $S_t = \sup_{0 \leq s \leq t} M_s$, $\phi$ is a twice continuously differentiable function, and $\Phi(x) = \int_0^x \phi(t)dt$.

In previous parts of this question, I showed that, for a continuous function $m$ and $s(t) = \sup_{0 \leq s \leq t} m(s)$, and for every bounded Borel function $h$, the Riemann-Stieltjes integral

$$\int_0^t (s(r) - m(r)) h(r) ds(r) = 0$$

I also showed (I think)

$$\phi(S_t) = \phi(S_0) + \int_0^t \phi'(S_s)dS_s$$

Unfortunately I can't identify how either of these facts would help with this questions so I've pursued a different course.

My idea is to argue that $(\phi(S) \cdot M)_t$ (where $(H \cdot M)_t = \int_0^t H(s) dM_s$ is a stochastic integral) satisfies

$$\langle \phi(S) \cdot M, N\rangle_t = (\phi(S))_t \cdot \langle M, N \rangle_t$$ $\forall N$ ($N$ is a cont. local m'gale) and that $\left(\Phi(S) - (S - M) \phi(S)\right)_t$ also satisfies this, so by uniqueless, $(\phi(S) \cdot M)_t = \left(\Phi(S) - (S - M) \phi(S)\right)_t$. ($\langle M, N\rangle_t$ is the quadratic variation between the cont. local martingales $M$ and $N$ and $H \cdot \langle M, N \rangle = \int_0^t H(s) d\langle M, N \rangle_s$ is a Riemann-Stieltjes integral.)

The work is in showing that $\left(\Phi(S) - (S - M) \phi(S)\right)_t$ satisfies this relation. Since for a sequence of partitions with mesh tending to zero we have

$$\lim_{n \to \infty} \sum_{i = 0}^{p_n - 1} (M_{t_{i + 1}^n} - M_{t_i^n})(N_{t_{i + 1}^n} - N_{t_i^n}) = \langle M, N \rangle_t$$

I thought I would try to compute this directly. Applying a mean-value theorem we have that:

$$\Phi(S_{t_{i + 1}^n}) - \Phi(S_{t_i^n}) = \phi(S_{t_i^n} + c_i^n S_{t_{i + 1}^n})(S_{t_{i + 1}^n} - S_{t_i^n})$$

($c_i^n \in [0,1]$.)

I write

$$\Phi(S_{t_{i + 1}^n}) - \Phi(S_{t_i^n}) - (\phi(S_{t_{i + 1}^n})S_{t_{i + 1}^n} - \phi(S_{t_{i}^n})S_{t_{i}^n}) + \phi(S_{t_{i + 1}^n}) M_{t_{i + 1}^n} - \phi(S_{t_{i}^n})M_{t_i^n}$$

This will be multiplied with $N_{t_{i + 1}^n} - N_{t_i^n}$ in the sum whose limit is the quadratic variation between the two processes. I notice that

$$\Phi(S_{t_{i + 1}^n}) - \Phi(S_{t_i^n}) - (\phi(S_{t_{i + 1}^n})S_{t_{i + 1}^n} - \phi(S_{t_{i}^n})S_{t_{i}^n}) = \phi(S_{t_i^n} + c_i^n S_{t_{i + 1}^n})(S_{t_{i + 1}^n} - S_{t_i^n}) - (\phi(S_{t_{i + 1}^n})S_{t_{i + 1}^n} - \phi(S_{t_{i}^n})S_{t_{i}^n}) \approx 0$$

This should hold for all $i$ for large $n$ since the mesh of the partition tends to zero. Thus a rewrite simplifies an earlier expression to

$$\phi(S_{t_{i + 1}^n}) M_{t_{i + 1}^n} - \phi(S_{t_{i}^n})M_{t_i^n} \approx \phi(S_{t_i^n}) (M_{t_{i + 1}^n} - M_{t_i^n})$$

So now I consider

$$\lim_{n \to \infty} \sum_{i = 0}^{p_n - 1} \phi(S_{t_i^n}) (M_{t_{i + 1}^n} - M_{t_i^n}) (N_{t_{i + 1}^n} - N_{t_i^n})$$

If I can believe that $(M_{t_{i + 1}^n} - M_{t_i^n}) (N_{t_{i + 1}^n} - N_{t_i^n}) \approx \langle M, N \rangle_{t_{i + 1}^n} - \langle M, N \rangle_{t_i^n}$, then I would have the above some being approximately

$$\lim_{n \to \infty} \sum_{i = 0}^{p_n - 1} \phi(S_{t_i^n}) (\langle M, N \rangle_{t_{i + 1}^n} - \langle M, N \rangle_{t_i^n}) = \int_0^t \phi(S_s) d\langle M, N \rangle_s = (\phi(S))_t \cdot \langle M, N \rangle_t$$

Then I would have established what I want.

For one thing, I use $\approx$ a lot and that's not how one writes a rigorous proof (at least not without a proper definition), and I don't even know if $(M_{t_{i + 1}^n} - M_{t_i^n}) (N_{t_{i + 1}^n} - N_{t_i^n}) \approx \langle M, N \rangle_{t_{i + 1}^n} - \langle M, N \rangle_{t_i^n}$. This also seems way hard, harder than it should be. So, am I doing the right thing? How can I make my argument rigorous, or is there an easier way?


In this answer I present a proof which uses the facts which you mentioned in the first part of your question.

By Itô's formula, we have

$$M_t \phi(S_t) = \int_0^t \phi(S_s) \,d M_s + \int_0^t M_s \phi'(S_s) \, dS_s \tag{1}$$


$$S_t \phi(S_t) = \int_0^t (\phi(S_s)+S_s \phi'(S_s)) \, dS_s. \tag{2}$$

(Note that Itô's formula is applicable because $t \mapsto S_t$ is increasing, and hence of bounded variation. There is no quadratic covariation term $[S,M]$ since $M$ has continuous sample paths and $S$ is of bounded variation.) Combining $(1)$ and $(2)$, we find

$$(S_t-M_t) \phi(S_t) = - \int_0^t \phi(S_s) \, dM_s + \int_0^t \big[ \phi(S_s) + (S_s-M_s) \phi'(S_s) \big] dS_s. $$

Using the identity (which you already showed)

$$\int_0^t (s(r)-m(r)) h(r) \, ds(r)=0$$

for $m(r) := M_r(\omega)$, $s(r) := \sup_{u \leq r} M_u(\omega)$ and $h(r) = \phi'(S_r(\omega))$, we get

$$(S_t-M_t) \phi(S_t) = - \int_0^t \phi(S_s) \, dM_s + \int_0^t \phi(S_s) \, dS_s.$$

Finally, note that

$$\Phi(S_t) = \Phi(S_0) + \int_0^t \Phi'(S_s) \, dS_s = \int_0^t \phi(S_s) \, dS_s$$

(this follows from the second identity you mentioned, with $\phi$ replaced by $\Phi$, or you can use Itô's formula). Consequently, we conclude that

$$(S_t-M_t) \phi(S_t) = - \int_0^t \phi(S_s) \, dM_s + \Phi(S_t).$$

  • $\begingroup$ It looks like you are using integration by parts for Itô's formula in (1) but the other process ($S_t$) is not a local martingale. In general are we able to mix martingales and bounded variation processes like this? $\endgroup$ – cgmil Feb 23 '18 at 16:23
  • $\begingroup$ @cgmil Itô's formula applies for semimartingales; $(S_t)_{t \geq 0}$ is clearly a semimartingale since it is of bounded variation. $\endgroup$ – saz Feb 23 '18 at 16:32
  • $\begingroup$ good to know, thanks! $\endgroup$ – cgmil Feb 23 '18 at 21:05

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.