While reading a preprint Eldan, Lehec, and Shenfeld - Stability of the logarithmic Sobolev inequality via the Föllmer Process I came across the following SDE in Section 3: $$d X_t=d B_t+\nabla \log P_{1-t} f\left(X_t\right) d t$$ where $B_t$ is the Brownian motion and $P_t$ denotes the heat semigroup.

Given a probability measure $\mu(dx)=f(x)\gamma(dx)$ where $\gamma$ is the gaussian density, one can consider $X_t$ to be the process that is close to a Brownian motion while having the law $\mu$ at time $t=1$.

I would like to understand why the following change of measure formula holds for every test function $u$,

$$\mathbb{E}\left[u\left(X_1\right) \mid \mathcal{F}_t\right]=\frac{P_{1-t}(u f)\left(X_t\right)}{P_{1-t} f\left(X_t\right)}$$

where $\mathcal{F}_t$ denotes the natural filtration associated with the process $X_t$. I looked into Girsanov's theorem in this regard but I don't see any exponential martingales in the above expression, so I am not sure what is going on here.


1 Answer 1


One can show it using the Feynman-Kac formula. In particular, from Feynman-Kac, we know that $h(x, t) := \mathbb{E}[u(X_1) \mid \mathcal{F}_t ]$ solves the following PDE:

$ \frac{\partial}{\partial t} h(x, t) + \nabla \log P_{1-t} f(X_t)^T \nabla_x h(x, t) + \frac{1}{2} \Delta_x h(x, t) = 0.$

with the terminal condition $h(x, 1) = u(x)$.

Now, note that your proposed function for $h(x, t)$, namely $\frac{P_{1-t} (uf)}{P_{1-t}(f) }$ satisfies the terminal condition, so you just need to plug your proposed function to the above equation to show that it satisfies the PDE above.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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