11

Good day.

There is a question I have already encountered twice, in very different contexts, that is relatively simple looking, but both solutions I know involve some pretty advanced theorems from the respective fields. I would like to ask if someone could think of any different proofs.

So, here is the question: Let $f: \mathbb Z^2 \rightarrow \mathbb R$ satisfy $f(m,n) = \frac14(f(m+1,n) + f(m-1,n) + f(m,n+1) + f(m,n-1))$. i.e. $f$ is defined on a grid, at a value at a point is the average of its four neighbors. Assume that $f$ is bounded (in some versions, bounded from below only). Prove that $f$ is constant.

One proof of the question uses martingales, and relies, I think on the martingale convergence theorem. In the proof, you define a two-dimensional random walk $S_n$ and look at the process $f(S_n)$.

The other proof comes from functional analysis, and uses the Krein-Milman theorem on the set $K_1 = \{f \in L_{\infty}(\mathbb Z^2) | f(m,n) = \frac14(...), ||f||\leq1 \} $, after finding its extremal set (which is the constant functions $\{+1,-1\}$).

So, any other proofs of this seemingly simple question? I should suspect there is something related to complex analysis, since for analytic functions this would follow from the Liouville theorem. Also, perhaps a combinatorial solution?

Aahz
  • 1,023
  • 7
  • 15
  • 3
    Near duplicate of http://math.stackexchange.com/questions/51926/a-stronger-version-of-discrete-liouvilles-theorem . No elementary proofs there either, though. – David E Speyer Jun 27 '14 at 15:30
  • Can you think of a suitable title which includes some non-LaTeX text? Pure LaTeX titles cause some problems. – Peter Taylor Jun 27 '14 at 16:38
  • I left an elementary proof at the other question. – David E Speyer Jun 27 '14 at 16:57
  • Well, I'm not sure 'elementary' is the correct word to use here :). But it indeed does not seem to rely on any deep theorems, so in that sense it is elementary. – Aahz Jun 27 '14 at 17:02
  • Chebyshev's inequality, Newton's inequalities, generating functions and the inequality $\binom{2r}{r} \leq 2^{2r}/\sqrt{2r+1}$, which can be proved by induction on $r$. All of these are permissible in an IMO solution, so I figure they count as elementary. (I could make it sound more elementary by never using the word random walk, but that's pointless obscurity.) – David E Speyer Jun 27 '14 at 17:07
  • This question was asked on [this forum] (https://www.ocf.berkeley.edu/~wwu/cgi-bin/yabb/YaBB.cgi?action=display;board=riddles_hard;num=1131915398;start=0), where a user claims to have shown an elementary solution via considering the difference of each square and the square left of it. Sadly I don't understand how "after reading hint 3 the solution is trivial"; if anyone understands this poster's solution this could provide an elementary solution to the question. – Aditya Gupta Nov 26 '22 at 20:44

1 Answers1

6

This is the discrete equivalent of Liouville's principle for harmonic function: every bounded function with a null laplacian is constant.

Without loss of generality you can assume that $f(0,0)=0$ (if not so, replace $f$ with $f-f(0,0)$). Now consider the square $B_N=\{(m,n):|m|+|n|\leq N\}$. Over $B_N$, $f$ attains its maximum and minimum on $\partial B_N$, and $$\Gamma_N = \max_{(m,n)\in B_N}f(m,n)=\max_{(m,n)\in \partial B_N}f(m,n)$$ is non decreasing as a function of $N$. $\Gamma_N$ is also bounded by $\Gamma$, and satisfies the almost-midpoint-convexity condition: $$\Gamma_N\leq\frac{3\Gamma_{N+1}+\Gamma_{N-1}}{4}$$ since the values of $f$ over $\partial B_N$ are completely determined by the values of $f$ over $\partial B_{N+1}$ and $\partial B_{N-1}$. Since any non-decreasing, bounded, midpoint-convex function is constant, we would expect that the same applies here. We have: $$\Gamma_{N+1}-\Gamma_{N}\geq\frac{1}{3}(\Gamma_{N}-\Gamma_{N-1})\geq\frac{\Gamma_1}{3^N}$$ so for any $\epsilon>0$ we can find a point $(a_1,b_1)$ such that $f(a_1,b_1)\geq\frac{3}{2}\Gamma_1-\varepsilon>\frac{10}{7}\Gamma_1.$ This gives that the difference of the values of $f$ in two neighbour points cannot be bounded, hence the function itself cannot be bounded.

We have just used a Borel-Caratheodory-type argument to provide lower bounds for $\Gamma_N$.

Jack D'Aurizio
  • 361,689
  • I do not quite follow. Can you explain why "over $B_N$, $f$ attains its maximum of $\partial B_N$", and why is $\Gamma_N$ non-decreasing in $N$? I do not see how those follow, not even intuitively. It seems to me that such a function, if not constant, may decrease in some directions and increase in other, as long as those average out. – Aahz Jun 27 '14 at 19:16
  • It is the same principle for which the maximum modulus of an olomorphic function is attained on the boundary of a ball: since any value of $f$ is the average of the values of $f$ in a neighbourhood, $\operatorname{argmax}{z\in B}f(z)$ and $\operatorname{argmin}{z\in B}$ cannot lie strictly inside $B$. Hence, for any compact ball $B$, $\operatorname{argmax}f$ lies on the boundary, and the function $\max_{z\in B_N}f(z)$ is non decreasing since $B_N\subset B_{N+1}$. – Jack D'Aurizio Jun 27 '14 at 19:57
  • 1
    I don't understand the end of your proof. I don't see why the difference cannot be bounded. In fact, a sequence starting with 0, 1 and with the same recurrence relation tends to 1.5. – Labo Aug 03 '17 at 18:11
  • I'm struggling with the same problem as Labo. Could you clarify the final paragraph? – Hecatonchires Mar 18 '24 at 18:28