Let's say I have a Langevin equation that reads $$\frac{dx(t)}{dt} = A(t) + B(t) + \eta(t)$$

Where $A(t)$ and $B(t)$ are some non-random functions that we know the functional form of, and $\eta(t)$ is a random term. For the sake of this question, lets say it has Gaussian properties such that it has zero mean, a finite standard deviation and $\langle\eta(t_1)\eta(t_2)\rangle = D\delta(t_1-t_2)$.

If $A(t) = B(t) = 0$, the equation just describes a random walk, and we can calculate

$$\langle x(t_1)x(t_2)\rangle = \int_{0}^{t_1}\int_{0}^{t_2}\langle\eta(t'_1)\eta(t'_2)\rangle dt'_1dt'_2$$ $$\langle x(t_1)x(t_2)\rangle = \int_{0}^{t_1}\int_{0}^{t_2}D\delta(t'_1-t'_2) dt'_1dt'_2$$ $$ \langle x(t_1)x(t_2)\rangle= Dt_1$$

But now lets say we know that $A(t)\neq 0$ and $B(t)\neq 0$. If we now calculate $\langle x(t_1)x(t_2)\rangle$ we have

$$\langle x(t_1)x(t_2)\rangle = \int_{0}^{t_1}\int_{0}^{t_2}\Big\langle(A(t'_1) + B(t'_1) + \eta(t'_1))(A(t'_2) + B(t'_2) + \eta(t'_2))\Big\rangle dt'_1dt'_2$$

And we find we have ensemble average terms where we mix random and non-random terms for example $$\langle A(t'_1)\eta(t'_2)\rangle$$

My question is: how do you deal with these terms? In the case of only random terms, only one term $\langle\eta(t_1)\eta(t_2)\rangle = D\delta(t_1-t_2)$ appears under the integral which we've defined before. But how do we deal with the ensemble average terms with both random and non-random terms in it?

  • $\begingroup$ Why differentiate between $A$ and $B$? You can reformulate the question in terms of $C(t)=A(t)+B(t)$ without loss of generality. The functions do not depend on an stochastic process (e.g. $A(\eta)$ or $A(x)$)? $\endgroup$
    – Javi
    Commented Sep 13, 2022 at 12:34

2 Answers 2


A non-random function behaves as a constant in respect to statistical averaging, that is $$ \langle A(t_1)\eta(t_2)\rangle=A(t_1)\langle \eta(t_2)\rangle=0 $$ (the second equality is because the OP assumes that the random process has the zero average.)


  • A word of caution - why are there explicitly two different non-random functions in the equation? It might be that one of them is a function of $x(t)$, in which case it is no more a non-random term, and the problem is more complex.
  • As linear differential equations are exactly solvable, the corresponding Langevin equations can be formally solved and the solutions used for calculating all the necessary averages.
  • $\begingroup$ Thank you for the answer! I added two non-random functions because I also was wondering about the case for $\langle A(t_1)B(t_2)\rangle$, though this case should be answered by your response, and it just comes out to be $\langle A(t_1)B(t_2)\rangle =A(t_1)B(t_2) $? If there was another term, $C(x(t))$, how would this be dealt with in calculating $\langle C(x(t_1))\eta(t_2)\rangle$ and also $\langle C(x(t_1))A(t_2)\rangle$? $\endgroup$ Commented Sep 13, 2022 at 18:40
  • $\begingroup$ @physics_fan_123 if there were a term dependent on $x(t)$, the problem would be a lot more complex, since $x(t)$ depends on $\eta (t)$. However, if the equation is linear, it would be still tractable - see my remarks. Otherwise, one would have to use more special techniques... Or work with the Fokker Planck equation. $\endgroup$
    – Roger V.
    Commented Sep 13, 2022 at 18:58

Given a deterministic function, $f(t)$, then it follows from Ito isometry & independent increments that $$ \int_0^tf(u)\,\mathrm{d}W_u\sim\mathcal{N}\left(0,\,\int_0^tf^2(u)\,\mathrm{d}u\right).\tag{1}$$ where $W_t$ is the Wiener process (such that one could at least loosely say $\mathrm{d}W_t=\eta(t)\,\mathrm{d}t$) and $\mathcal{N}(\mu,\,\sigma)$ is the normal distribution. This is to say, the integral in Eq (1) is a Gaussian process with zero mean--see, for instance, these lecture notes. So one can safely ignore these terms when we average them out, $$\left\langle\int f(u)\,\mathrm{d}W_u\right\rangle\equiv0.$$ In the cases we don't take an average, we could replace the integral with an arbitrary term, e.g. $$Z(t)\sim\int f(u)\,\mathrm{d}W,$$ and not really miss anything when solving stochastic differential equation (though a final solution might need to be solved numerically or through an approximation such as being small compared to another term).


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.