Kolmogorov backward/forward equations are pdes, derived for the semigroups constructed from the Markov transition kernels.

Feynman-Kac formula is also a pde corresponding to a stochastic process defined by a SDE. But I was wondering if the stochastic process is also Markovian? I.e., does Feynman-Kac formula apply only to Markov process?

Does the semigroups from the Markov transition kernels also lead to Feynman-kac pde?

  • If not, what leads to Feynman-Kac pde?
  • If yes, Is Feynman-Kac pde also some kind of Kolmogorov backward/forward equation? if not, how is Feynman-Kac pde related to Kolmogorov backward/forward equations?

Thanks and regards!


1 Answer 1


First of all, it would be nice if you link the definitions of F-K formula and KBE you have in mind. Anyway:

  1. KBE is an equation used to study $P_t f(x):=\mathsf E_x[f(X_t)]$ in terms of the generator $$ \mathcal A :=\lim_{t\downarrow 0}\frac{P_t - P_0}{t}. $$ This is indeed applies to general Markov processes, as its definition depends only such notions as semigroups and generators.

  2. F-K formula (at least the one given in Oksendal, Theorem 8.2.1) is devoted to the same problem and appears to be only a generalization to the case of killed diffusions. The term killed refers to the process which is not conservative, e.g. for conservative Markov processes $P_t1 = 1$ for all times $t$, whereas for non-conservative $P_t1\leq 1$ and $P_t1\neq 1$. This refers to the case, when the process "leaves" the state space at some random time $\zeta$ (killing time) and jumps to some auxiliary cemetery state.

    You wrote that F-K formula is a PDE linked to some SDE. I would expect, though, that such SDE has to be Markovian, since otherwise you wouldn't have a nice dependence on states exclusively (as PDE requires). I neither seen F-K formula (as a PDE) for non-Markovian SDEs, so it will be interesting if you provide a link.

Although F-K formula is stated originally for SDE, nothing prevents you from formulating it over general Markov processes. However, I guess in such case it falls down to be a special case of KBE for non-conservative Markov process. Unfortunately, I don't have a monograph of Dynkin "Markov processes" in hand, but I'm pretty sure you can find it there.


W.r.t. your question on the classification. There are Markovian and non-Markovian SDEs. A class of solutions of Markovian SDEs of a special form is called Ito diffusions. In general, SDEs (in a sense of a martingale problem) can be considered as a superset of Markov processes - but such SDEs do not have to be driven by Brownian motion only.

Killing/non-conservatism is a structure added on top of the structures above. This refers to the case, when $\Bbb P(X_t \in E)<1$ where $E$ is a state space and $t>0$. I'm not an expert on that, though, and I suggest you start reading serious book on stochastic processes, rather than lurking wikipedia.

  • $\begingroup$ Thanks! Are diffusions, killed diffusions, and"non-conservative" all Markov processes? $\endgroup$
    – Tim
    Mar 5, 2013 at 6:19
  • $\begingroup$ @Tim: in such shape the question is not precise. As much as you didn't provide a source for the F-K formula you have in mind. $\endgroup$
    – Ilya
    Mar 5, 2013 at 8:13
  • $\begingroup$ I didn't know there are different Feynman-Kac formulas? I was reading en.wikipedia.org/wiki/Feynman%E2%80%93Kac_formula $\endgroup$
    – Tim
    Mar 5, 2013 at 8:19
  • $\begingroup$ @Tim: this formulation is backwards in time, whereas the one in Oksendal is forwards in time (as an example of the difference.) $\endgroup$
    – Ilya
    Mar 5, 2013 at 8:19
  • $\begingroup$ @Tim: added an answer to your second question $\endgroup$
    – Ilya
    Mar 5, 2013 at 8:28

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.