Question: Let us consider the Poisson problem on the square with constant source $1$ $$ \begin{cases} - \Delta u &= 1, \qquad \text{ in } (0,1)^n \\\\ u &= 0, \qquad \text{ on } \partial (0,1)^n \end{cases} $$ By symmetry the maximum is attained in $x_0:=(1/2,\dots,1/2)$. Does there exist a formula or precise estimate for $$ \max u((0,1)^n) = u(x_0) \quad ? $$ How does this value scale in $n$? Can one recover some bound $o(1)$, for $n\to \infty$?

Background: It is easy by comparison with $1/8-(x-x_0)^2/(2n)$ and the maximum principle to deduce the bound $$ u(1/2, \dots, 1/2 ) \leq 1/8 , $$ which is also sharp for $n=1$, where $u(x) = 1/2 \; x (1-x)$ is the solution. By the same method one finds for the solution of the Poisson problem on the unit ball the bound $1/(2n)$, which is the second part of the question.

Strategy so far: By the product-ansatz one can deduce an explicit solution for $u$, given by $$ u(x) = \frac{4^n}{\pi^{n+2}}\sum_{\alpha\in (2 \mathbb{N}+1)^n} \frac{1}{\sum_{i=1}^n \alpha_i^2} \prod_{j=1}^n \frac{\sin(\pi \alpha_j x_j)}{\alpha_j} . $$ Hereby, I denote with $2\mathbb{N}+1$ the odd integers starting with $1$, hence $\alpha\in (2\mathbb{N}+1)^n$ is a multiindex with odd entries. In principle, setting $x=(1/2,\dots,1/2)$ should give the solution.

How does one calculate the resulting alternating series or find estimates recovering at least the scaling in $n$?

Numerical results: For small $n$ from series representation:

n=1: 0.125
n=2: 0.0737
n=3: 0.0562
n=4: 0.0473

Probabilistic interpretation of the solution: $u(x)$ is the mean hitting time of a Brownian particle starting in $x$, which may open estimates from the stochastic side.

One more thought wrt. volume: It is not clear, if the comparison of unit cube and unit ball is the right comparison, since the volume of the unit ball goes to $0$ for $n\to \infty$. Maybe, one should consider the unit-volume-ball, for which I find the scaling of the maximum like $O(n^{-1/2})$. I can add details, if there is some indication in this direction.

Bonus questions: Is there a more general method not directly based on explicit solutions, which would lead to estimates for general domains, for instance convex, star shaped or just bounded?

What is a reasonable quantity describing the scaling behaviour in $n$? For instance: Do all reasonable unit-volume domains scale like $O(n^{-1/2})$?


Here is a start of a method which should give you some asymptotics for the cube. You want to find

$$\int_{t=0}^\infty p(t)^n dt$$

where $p(t)$ is the probability that a Brownian motion has not left a unit interval by time $t$ after starting at the center. As $n\to \infty$, what matters is the behavior of $p(t)$ near $0$. While the first exit time density is a little messy for two barriers, it's easy for one barrier by reflection, and you get a Lévy distribution. For small $t$, it's very unlikely that a Brownian motion would hit both endpoints of the interval, so it's a good approximation to say that the probability of exiting the interval is twice the probability of leaving in one particular direction.

  • $\begingroup$ This is nice approach. In this way I obtained yet another series representation, which is not absolutely convergent. Some numerical estimates are feasible, although the scaling in $n is not obvious. $\endgroup$ – André Schlichting Apr 23 '13 at 7:46

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.