Context: Consider the advection-diffusion equation with periodic boundary conditions (PBC) over a flat square domain $L \times L$. The scalar density $\rho $ is transported by a prescribed field $\mathbf{v}=-\nabla U$, where $U(\mathbf{x})$ is a scalar potential that has the periodicity imposed by the PBC. The density $\rho$ evolves as

$$ \partial_t \rho(\mathbf{x},t) = -\nabla \cdot [ \mathbf{v}(\mathbf{x}) \rho(\mathbf{x},t) - \nabla \rho(\mathbf{x},t) ] = 0 $$

The steady-state solution is found by imposing $\partial_t \rho(\mathbf{x},t) =0$ and has the usual Gibbs form:

$$ \rho(\mathbf{x}) \, \propto \, e^{-U(\mathbf{x}) } $$

The problem: I am wondering how to find the steady-state in a slightly more general case, where

$$\mathbf{v} = -\nabla U + \mathbf{q}$$

The potential $U$ has the periodicity imposed by the PBC and $\mathbf{q} =(q_x,q_y)$ is a constant vector field (the constant \mathbf{q} defines the so-called "tilt" of the potential $U$). Hence, the equation we have to solve is

$$ \nabla \cdot [ \, \rho(x,y) \, \mathbf{q} - \rho(x,y) \nabla U(x,y) - \nabla \rho(x,y) \, ] = 0 $$

with the periodic conditions $\rho(0,y) = \rho(L,y)$, $\rho(x,0) = \rho(x,L)$, $U(0,y) = U(L,y)$, $U(x,0) = U(x,L)$. For simplicity, I tried to consider the case $\mathbf{q}=(q,0)$, but the problem still seems non-trivial.

Question: Any idea or reference about the diffusion-advection equation in periodic boundary conditions (in particular about the steady-state)? Which is the "Gibbs-like solution" in this case?

Further considerations: I have the feeling that finding a solution is not easy because the "tilt" potential that generates the constant field $\mathbf{q}$ is $-\mathbf{x}\cdot \mathbf{q}$. This "tilt" contribution males the total potential $U-\mathbf{x}\cdot \mathbf{q}$ not periodic (i.e. it does not satisfy the PBC).

Moreover, define the total current in the steady-state as

$$ \mathbf{J}(x,y) = \rho(x,y) \, [\mathbf{q} - \nabla U(x,y)] - \nabla \rho(x,y) \, , $$

so that we have to find the $\mathbf{J}$ such that

$$ \nabla \cdot \mathbf{J} = 0 \quad \Rightarrow \quad \mathbf{J} = R \nabla g $$

where $R$ is a 90-degrees rotation and $g$ is an unknown scalar potential. Note that $g$ does not have to respect the PBC, but $\mathbf{J}$ does: (probably) the most general form of $g$ is

$$ g(x,y) = G(x,y) + a x + b y $$

where $G$ respects the PBC and $a$ and $b$ are constants. Even though this problem is more likely to be studied by physicists, I have the feeling that the problem is intimately related to the topology of the 2D torus, so I posted also a similar question on math SE.

  • 1
    $\begingroup$ What happens when you try a frame transformation, i.e. rewriting the problem in terms of $x' = x - qt$? $\endgroup$
    – Daniel
    Jul 24, 2020 at 17:27
  • $\begingroup$ Hi @Daniel, sorry for late reply. Yes I tried but unfortunately it does not seem useful. $\endgroup$
    – Quillo
    Jul 26, 2020 at 15:10
  • 1
    $\begingroup$ An observation: If $U$ is constant, no solution exists (unless $q = 0$ also). And I think I can show that solutions "blow up" (have energy pushed to arbitrarily high frequencies) as $\nabla U \rightarrow \mathbf{0}$, if they exist. Have you tried studying the 1D version of this problem? $\endgroup$
    – Daniel
    Jul 29, 2020 at 3:34
  • $\begingroup$ Yes, the 1D version is basically the same. You end up with a formal solution $\rho(x) \propto e^{q x -U(x)}$ that cannot be periodic. Maybe I found something interesting here: sciencedirect.com/science/article/pii/… ..it seems the "correct" thing to do is to consider the adjoint version of the equation. $\endgroup$
    – Quillo
    Jul 29, 2020 at 10:26
  • 1
    $\begingroup$ This doesn’t quite answer the question, but in case it is of interest: I wrote a paper studying systems of this type in some detail (hence my interest in Quillo and @Daniel ‘s discussion). arxiv.org/abs/2108.06431. Theorem 3 gives (small diffusion) asymptotic information about $\rho$ (cf. Remark 4). However the paper is more focused on approximating the flux of $\mathbf{J}$ through suitable hypersurfaces (Theorems 1, 2, 4; Prop. 3). $\endgroup$ Aug 23, 2021 at 2:36

1 Answer 1


Any $\rho$ which solves the equation on the whole torus must also be a solution locally on every subset. In particular, it must be solution on the (non-toroidal) open $L \times L $ square. Since solutions on the torus are a subset of the solutions on the square, the question becomes: Do there exist solutions on the square which happen to match at the boundaries?

On this square, we can define $V = U - \mathbb{x} \cdot \mathbb{q}$, and we have an ordinary advection-diffusion equation. We know there exist solutions of the form $\alpha e^{-V(\mathbb{x})}$. We also know that $U$ is periodic, so $V$ can only be periodic if $\mathbb{q} = \mathbb{0}$. However $e^{-V}$ could still be periodic if $\mathbb{q}$ is imaginary. Specifically, we have periodic solutions for $\mathbb{q} = \frac{2\pi i}{L}\mathbb{n}, \mathbb{n} \in \mathbb{Z}^2$.

For other $\mathbb{q}$, solutions proportional to $ e^{-V(\mathbb{x})} $ cannot extend to solutions on the whole torus. The remaining question: Are such solutions the whole solution space?

Now, Matthew Kvalheim points to Zeeman, 1988. Theorem 3 reads

Let $U$ be a vector field on a compact manifold $X$ without boundary, and let $\epsilon$ > 0. Then the Fokker-Planck equation for $U$ with $\epsilon$-diffusion has a unique steady state, and all solutions tend to that steady state.

The torus is a compact manifold without boundary, Zeeman's $U$ is our $-\nabla V$, and we have $\epsilon = 1$, so the theorem tells us a solution $\rho$ must exist and is unique (up to an overall scalar). Unfortunately, this proof is not constructive.

In one dimension, variation of parameters gives the solution $$\rho = C_1 e^{-V}\left(C_2 + \int_0^x e^V\right)$$ and the requirement $\rho(0) = \rho(L)$ fixes $C_2$. We can try to extend this to two dimensions as follows: Assume $\rho$ is of the form $\alpha(x)e^{-V}$. Then the equation becomes $$ \nabla \cdot [-\nabla V \alpha(x)e^{-V} - \nabla (\alpha(x) e^{-V})] = 0 $$ which simplifies to $$ \nabla \cdot (\nabla\alpha(x) e^{-V}) = 0 $$ Solutions are $$ \nabla\alpha(x) e^{-V} = \nabla \times \mathbf{\psi} $$ for $\mathbf{\psi} = \mathbf{e}_z $ and $g$ some scalar function. Then $$ \nabla\alpha(x) = e^V(\nabla \times \mathbf{\psi}) $$ If $$ \nabla \times (e^V(\nabla \times \mathbf{\psi})) = 0 $$ then this has solution $$ \alpha(x,y) = C + \left(\int_0^x -e^V g_y dx\right) + \left(\int_0^y e^V g_x dy\right) $$ The requirement of periodic boundary conditions picks out some unique $g$, $C$ up to an overall constant. We need $$ \alpha(x,0) = \alpha(x,L)e^{-Lq_y} $$ or $$ C + \left(\int_0^x -e^V g_y dx\right) = Ce^{-Lq_y} + \left(\int_0^x -e^V g_y dx\right)e^{-Lq_y} + \left(\int_0^L e^V g_x dy\right)e^{-Lq_y} $$ At $x = 0$ this simplifies to $$ C = \frac{1}{e^{Lq_y} - 1}\int_0^L e^V g_x(0,y) dy $$ It remains to find $g$.

I'm not sure that there is a nice expression for the solution in general. Some miscellaneous thoughts:

  • When $U = 0$, $\rho = C$ is a solution, which corresponds to $\alpha = e^V, g = xq_y-yq_x$. This shows that $g$ may be defined only on the square, not on the torus.
  • When $\nabla U \gg q$ or $q \gg \nabla U $, we can start with the nearby known solution and series expand.
  • 1
    $\begingroup$ Why would it be true that $\nabla \cdot \psi = 0$? $\endgroup$ Sep 19, 2020 at 6:31
  • 1
    $\begingroup$ $\psi = \psi_z \mathbb{e}_z$, and because this is a 2D problem $\frac{\partial}{\partial z} = 0$. $\endgroup$
    – Daniel
    Sep 19, 2020 at 14:58
  • 2
    $\begingroup$ I think your conclusion $\alpha =$ const. would imply that there are no solutions on the torus since $V$ is not periodic. But this contradicts the fact that this PDE always has solutions on the torus (see e.g. Theorem 3 of "Stability of Dynamical Systems", Zeeman, 1988). So far I think I spot one error in your reasoning: on $\mathbb{R}^2$ I don't think it is true that $\nabla \cdot \mathbf{J} = 0$ $\implies$ that $\mathbf{J}$ is the curl of something. Rather, as Quillo wrote, it implies $\mathbf{J} = R \nabla g$ for some $g$. Hence I think you should have $\nabla \alpha e^{-V} = R \nabla g$. $\endgroup$ Sep 19, 2020 at 17:18
  • 1
    $\begingroup$ Wait, aren't those equivalent? $$R\nabla g = (-g_y, g_x)$$ while $$\nabla \times (g \mathbb{e}_z) = (-g_y, g_x, 0) $$ (where $g$ is a scalar function and subscripts denote partial derivatives) $\endgroup$
    – Daniel
    Sep 19, 2020 at 19:34
  • 1
    $\begingroup$ Regarding Zeeman, however, I agree that this seems to contradict. Right now I'm thinking about the $U = 0$ case - shouldn't $\rho = C$ then be a solution? $\endgroup$
    – Daniel
    Sep 19, 2020 at 19:46

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.