I have found a theorem (see below) in two papers an I try to figure how it could be proved. The result seems to be intuitive, but I'm not able to prove it in a rigorous way.

Assumptions:

Consider a continuous stochastic process $(X_t)$ together with a Brownian motion $(B_t)$. The two processes are assumed to be independent. Their natural filtrations are denoted by $(\mathcal F_t^X)$ and $(\mathcal F_t^B)$

Theorem:

Under the above assumption, the following relation holds: $$\mathbb E\left(\exp\left(\int_0^tX_sdB_s\right) \mid \mathcal F_t^X\right)=\exp\left(\frac12\int_0^tX_s^2ds\right)$$

Proof 1

I have found a proof in Piterbarg page 1 (see here for the main article). The proof is based on computing the expectation path by path. That is, they assume that the path $(X_t)$ is known and compute the integral.

However, I believe that this argument only works when the number of paths $(X_t)$ is finite, using this result.

Proof 2

In Sin this result is proved (in particular) when $(X_t)$ is given by the dynamic $dX_t=X_td\tilde B_t$, where $(\tilde B_t)$ is a Brownian motion independent of $(B_t)$. They define the processes $$S_t=\exp\left(\int_0^tX_sdB_s-\frac12\int_0^tX_s^2ds\right)\quad\text{and}\quad S_t^{(n)}=S_{t\wedge\tau_n},$$ where we define the following stopping time $$\tau_n=\inf\left\{t>0:\int_0^t X_s^2ds\geq n\right\}.$$ In the paper, they use the following step, based on the Lebesgue dominated convergence: $$\mathbb E(\lim_{n\to\infty}S_T^{(n)})=\lim_{n\to\infty}\mathbb E(S_T^{(n)})$$

However I don't directly see how the Lebesgue dominated convergence can be applied in this case.

Question:

Can someone explain me why this results holds or has a good reference book about it?

Attempt of a proof:

Following this post, one can prove the result for simple processes, that is when $$X_s =\sum_{j=1}^n 1_{[t_{j-1},t_j)} \xi_j.$$ Now, we can take a sequence of processes $(X_s^n)$ converging to $(X_s)$ (with respect to the sup norm): $$(X_s^n)\to(X_s)\quad(n\to\infty)$$ Using the Itô isometry, we obtain the following convergence in probability: $$\int_0^tX_s^ndB_s\to\int_0^tX_sdB_s\quad (n\to\infty)$$ For each subsequence, we can find a subsubsequence (wlog $n_k$) such that the following convergence holds almost surely: $$\int_0^tX_s^{n_k}dB_s\to\int_0^tX_sdB_s\quad (k\to\infty)$$ If the Lebesgue dominated convergence theorem did hold, then we could conclude that $$\mathbb E\left(\exp\left(\int_0^tX_sdB_s\right) \mid \mathcal F_t^X\right)=\exp\left(\frac12\int_0^tX_s^2ds\right).$$

However I can't find an integrable majoring function.

Remark: What I basically did above is use the following property, which is true when $X$ is a discrete-valued random variable (see here).

The conditional expectation of $Y$ given $X$ is given by $\mathbb E(Y \mid \sigma(X))=f(X)$ where $f(x)=E(Y \mid X=x)$.

This computation seems to work above, by conditioning on the process $(X_t)$, even though it is not mathematically correct. In my view, this way of reasoning would fail in the following case if we say that $$\mathbb E\left(\exp\left(\int_0^tX_sdB_s\right) \mid \mathcal F_t^B\right)=f((B_s)_{0<s<t})$$ where $$f((b_s)_{0<s<t})=\mathbb E\left(\exp\left(\int_0^tX_sdB_s\right) \mid(B_s)_{0<s<t}=(b_s)_{0<s<t}\right)=\mathbb E\left(\exp\left(\int_0^tX_sdb_s\right)\right)$$

Namely, we would (almost surely) integrate with respect to a realization of the Brownian path $(b_s)_{0<s<t}$ which isn't of locally bounded variation. Thus the function $f$ is not well defined.

  • It seems that you missed at least a square in your equation or where do you apply Ito isometry? – user190080 Jun 25 '15 at 20:25
  • I have used the Ito isometry to compute the variance of the term in the exponential, which follows a normal distribution. There are surely other ways to compute it. – Mots du Jour Jun 25 '15 at 20:27
  • 1
    It seems that what you need is a definition of conditional expectation in the non trivial case, that is, when the sigma-algebra one is conditioning on is not generated by a discrete random variable. Have you got this at your disposal? – Did Jun 25 '15 at 21:28
  • ((OP stays silent. Well...)) – Did Jun 26 '15 at 8:34
  • 2
    See this question: math.stackexchange.com/q/1287843 – saz Jun 26 '15 at 11:05

First of all, note that we have to ensure that

$$\exp \left( \int_0^t X_s \, dB_s \right) \in L^1. \tag{1}$$

If your claim is true, then

$$\mathbb{E} \exp \left( \int_0^t X_s \, dB_s \right) = \mathbb{E}\exp \left( \frac{1}{2} \int_0^t X_s^2 \, ds \right),$$

i.e. $(1)$ holds if

$$\mathbb{E}\exp \left( \frac{1}{2} \int_0^t X_s^2 \, ds \right)< \infty. \tag{2}$$

Through the remaining part of my answer, I'll assume that $(2)$ holds. The last part of the following proof is very close to the proof of Novikov's condition in René Schilling/Lothar Partzsch: Brownian Motion - An Introduction to Stochastic Processes.


We denote by $(M_t)_{t \geq 0}$ the stochastic exponential of $(X_t)_t$:

$$M_t := \mathcal{E}(X)_t := \exp \left( \int_0^t X_s \, dB_s- \frac{1}{2} \int_0^t X_s^2 \, ds \right).$$

Then the claim is equivalent to $\mathbb{E}(M_t \mid \mathcal{F}_t^X)=1$. Moreover, we set $$Y_t := \int_0^t X_s \, dB_s \qquad \text{and} \qquad \langle Y \rangle_t := \int_0^t X_s^2 \, ds.$$

Using e.g. the approximation procedure described in the OP and (conditional) Fatou's lemma, it is not difficult to see that $$\mathbb{E}(M_t \mid \mathcal{F}_t^X) \leq 1.$$ Consequently, it remains to show that $$\mathbb{E}(M_t \mid \mathcal{F}_t^X) \geq 1. \tag{3}$$

For $n \in \mathbb{N}$ we define

$$\tau_n := \inf\{t \geq 0; \max\{|B_t|,|X_t|\} \geq n\}.$$

Since, by Itô's formula,

$$M_{t \wedge \tau_n} -1 = \int_0^{t \wedge \tau_n} M_s \, dB_s$$

we find for any $F \in \mathcal{F}_t^X$

$$M_{t \wedge \tau_n}1_F-1_F = \int_0^{t \wedge \tau_n} (1_F M_s) \, dB_s.$$

(Note: Because of the independence of $X$ and $B$, we can put the indicator function $1_F$ under the integral sign.) As the right-hand side is a martingale, we get

$$\mathbb{E}(M_{t \wedge \tau_n} 1_F) = \mathbb{P}(F).$$

Since this holds for any $F \in \mathcal{F}_t^X$, this shows

$$\mathbb{E}(M_{t \wedge \tau_n} \mid \mathcal{F}_t^X) = 1. \tag{4}$$

In particular,

$$\mathbb{E}(M_{t \wedge \tau_n}) = 1. \tag{5}$$

For fixed $c \in (0,1)$, we pick $p=p(c)>1$ such that $p < \frac{1}{c(2-c)}$. By Hölder's inequality (for the exponents $1/pc$ and $1/(1-pc)$), we obtain

$$\begin{align*} \mathbb{E}[\mathcal{E}(c X_{t \wedge \tau_n})^p] &\stackrel{\text{def}}{=} \mathbb{E} \exp \left( pc Y_{t \wedge \tau_n} - \frac{1}{2} pc^2 \langle Y \rangle_{t \wedge \tau_n} \right) \\ &= \mathbb{E} \left[\exp \left(pc (Y_{t \wedge \tau_n} - \frac{1}{2} \langle Y \rangle_{t \wedge \tau_n} \right) \exp \left( \frac{1}{2} pc(1-c) \langle Y \rangle_{t \wedge \tau_n} \right) \right] \\ &\leq \underbrace{ \left[\mathbb{E}\exp \left( Y_{t \wedge \tau_n} - \frac{1}{2} \langle Y \rangle_{t \wedge \tau_n} \right) \right]^{pc}}_{[\mathbb{E}(M_{t \wedge \tau_n}]^{pc} \stackrel{(5)}{=} 1} \left[ \mathbb{E} \exp \left( \frac{1}{2} \frac{pc(1-c)}{1-pc} \langle Y \rangle_t \right) \right]^{1-pc} \\ &\leq \left[ \mathbb{E} \exp \left( \frac{1}{2} \frac{pc(1-c)}{1-pc} \langle Y \rangle_T \right) \right]^{1-pc} \end{align*}$$

for any $t \leq T$. This shows that the $p$-th moments of the family $(\mathcal{E}(c X_{t \wedge \tau_n}))_{n \in \mathbb{N}}$ is uniformly integrable. It follows from Vitali's convergence theorem that

$$\mathcal{E}(c X_{t \wedge \tau_n}) \to \mathcal{E}(cX_t) \qquad \text{in $L^1$} \tag{6}$$

as $n \to \infty$. By $(4)$ (applied for $\tilde{X} := cX$), this implies

$$\begin{align*} 1 &\stackrel{(4)}{=} \lim_{n \to \infty} \mathbb{E}\left[ \exp \left( cY_{t \wedge \tau_n} - \frac{1}{2} c^2 \langle Y \rangle_{t \wedge \tau_n} \right) \mid \mathcal{F}_t^X \right] \\ &\stackrel{(6)}{=} \mathbb{E}\left[ \exp \left( cY_{t} - \frac{1}{2} c^2 \langle Y \rangle_{t} \right) \mid \mathcal{F}_t^X \right] \end{align*}$$

Finally, we use a last time (conditional) Hölder's inequality (for $1/c$ and $1/(1-c)$) to obtain

$$\begin{align*} 1 &= \mathbb{E} \left[ \exp \left( c Y_t - \frac{1}{2} c \langle Y \rangle_t \right) \exp \left( \frac{1}{2} c(1-c) \langle Y \rangle_t \right) \mid \mathcal{F}_t^X \right] \\ &\leq \left[ \mathbb{E} \left( \exp \left( Y_t- \frac{1}{2} \langle Y \rangle_t \right) \mid \mathcal{F}_t^X \right) \right]^c \left[ \mathbb{E} \left( \exp \left( \frac{1}{2} \langle Y \rangle_t \right) \mid \mathcal{F}_t^X \right) \right]^{1-c} \end{align*}$$

Since, by $(2)$, the last term is bounded for $t \leq T$, we can let $c \uparrow 1$:

$$1 \leq \mathbb{E} \left[ \exp \left( Y_t - \frac{1}{2} \langle Y \rangle_t \right) \mid \mathcal{F}_t^X \right] \stackrel{\text{def}}{=} \mathbb{E}(M_t \mid \mathcal{F}_t^X).$$

  • Thank you for the answer, however the aim of the papers claiming the above theorem is (in particular) to prove that the equation (2) holds. – Mots du Jour Jul 1 '15 at 22:57
  • 6
    @MotsduJour ... would have been nice if you had made this more clear in your question. – saz Jul 2 '15 at 5:19

Your Answer

 

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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