Let \(\Omega\subset\mathbb R^n\) be a domain with outer unit normal \(\nu\), let \(f\) be a \(C^1\) function, and let \(\kappa>0\). Serrin’s overdetermined problem asks for a function \(u\) with
\[\begin{cases} -\Delta u=f(u),\quad u>0 & \text{in }\Omega,\\ u=0,\quad \partial_\nu u=-\kappa & \text{on }\partial\Omega. \end{cases} \qquad\text{(S)}\]Serrin proved in 1971 the following surprising result: if \(\Omega\) is a bounded domain whose boundary is of class $C^2$ and (S) has a solution, then \(\Omega\) is a ball and \(u\) is radial.
It is thus natural to ask what happens when $\Omega$ is unbounded. From now on, let us assume that \(\Omega\) is the epigraph of a function \(g:\mathbb R^{n-1}\to\mathbb R\),
\[\Omega=\{(x',x_n):\ x_n>g(x')\}.\]In this case, the conjecture is that $\Omega$ is a half-space if (S) has a solution. Here is some partial progress on this problem from the literature.
- Ros, Ruiz and Sicbaldi proved that a planar $C^{1,\alpha}$ domain with connected, unbounded boundary carrying a bounded solution of (S) is a half-plane.
- In dimension three, Farina and Valdinoci proved it when $g$ is both $C^3$ and uniformly Lipschitz, assuming also that $u$ is bounded.
- For \(n\le8\), Wang and Wei proved that the answer is yes for $g \in C^1$ and Allen–Cahn-type reactions such as \(u-u^p\), provided \(u\) is monotone in the vertical direction, that is, \(\partial_nu>0\) in \(\Omega\).
- Beuvin and Farina proved that for \(n\ge9\), if \(g\in C^1\) is bounded on one side by a linear function, that is, \(g(x')\ge-C(1+\lvert x'\rvert)\) for all \(x'\) or \(g(x')\le C(1+\lvert x'\rvert)\) for all \(x'\), for some \(C>0\), and $u$ is monotone in the vertical direction, then $\Omega$ is a half-space.
- On the other hand, the conjecture fails in general for \(n\ge9\): del Pino, Pacard and Wei constructed, for every \(n\ge9\) and reactions such as \(u-u^3\) and \(u-u^2\), a smooth epigraph that is not a half-space on which (S) has a bounded solution.
A natural question is whether we can remove the assumption on monotonicity. Monotonicity of positive solutions in epigraphs goes back to Berestycki, Caffarelli and Nirenberg, for globally Lipschitz graphs. Recent works relax the Lipschitz condition: Beuvin, Farina and Sciunzi treat uniformly continuous graphs that are bounded from below, and Berestycki and Graham treat asymptotically uniformly Lipschitz graphs. Beuvin and Farina list, as an open research direction, the search for conditions on \(g\) weaker than global Lipschitz continuity that guarantee monotonicity. The theorem below answers this question when \(f(s)/s\) is nonincreasing and \(g\) is simply assumed to be $C^1$.
Theorem 1 (automatic monotonicity). Let \(n\ge2\), \(g\in C^1(\mathbb R^{n-1})\), and let \(\Omega\) be the epigraph of \(g\). Let \(f\) be locally Lipschitz on \((0,\infty)\) and such that $s\mapsto f(s)/s$ is nonincreasing on $(0,\infty)$. Let \(u\in C^2(\Omega)\cap C^1(\overline\Omega)\) solve (S), and assume that the square mass of \(u\) grows subexponentially:
\[\liminf_{R\to\infty}\frac1R\log\left(1+\int_{\Omega\cap B_R}u^2\,dx\right)=0 . \qquad\text{(G)}\]Then \(\partial_nu>0\) in \(\Omega\).
Combined with the theorem of Beuvin and Farina, Theorem 1 gives rigidity without any monotonicity assumption.
Corollary 2 (rigidity). Let \(g\in C^1(\mathbb R^{n-1})\), and let \(u\) be a solution of (S) in the epigraph of \(g\) with \(f(u)=u-u^p\) for some $p>1$. Assume that either \(2\le n\le8\), or \(n\ge9\) and \(g\) is bounded on one side by a linear function in the sense above. Then \(g\) is affine and \(u\) is one-dimensional.
Proof. Condition (G) holds since, for this class of reactions, Beuvin and Farina (Corollary 5.4) showed that every positive solution is automatically bounded. \(\square\)
Two principles
This post is part of a series of blog posts concerned with the automatic monotonicity of PDEs, and two principles used in the proof of Theorem 1 will be used again in later posts. Therefore, we summarize them here before we prove Theorem 1.
- Let $\tau:= \partial_n(\log u)=\partial_nu/u$. We measure the failure of monotonicity by $z:=\tau^- = \max(-\tau,0)$. The problem of proving $\partial_nu\ge0$ is then turned into the equivalent problem of proving \(z\equiv 0\).
- We’ll prove an estimate for a high moment of \(z\) on the ball \(B_R\) whose right-hand side involves a factor of $1/R$. The growth condition (G) enters to make sure the right-hand side decays fast enough so that we can conclude $z\equiv 0$ locally.
Proof of Theorem 1
Step 1: the equation for $\tau$. Differentiating \(-\Delta u=f(u)\) in \(x_n\) gives \(-\Delta\partial_nu=f'(u)\,\partial_nu\) in the sense of distributions. Set $h(s):=f(s)-sf’(s)$. Since \(u^2\nabla\tau=u\nabla\partial_nu-\partial_nu\,\nabla u\),
\[\begin{aligned} \operatorname{div}(u^2\nabla\tau) &=u\,\Delta\partial_nu-\partial_nu\,\Delta u\\ &=-u\,f'(u)\,\partial_nu+\partial_nu\,f(u)\\ &=u\,h(u)\,\tau . \end{aligned}\]So \(\tau\) solves the linear equation \(-\operatorname{div}(u^2\nabla\tau)+u\,h(u)\,\tau=0\).
Step 2: the inequality for $z$. The negative part of a solution of such an equation is a subsolution in the weak sense:
\[-\operatorname{div}(u^2\nabla z)+u\,h(u)\,z\le0. \qquad\text{(1)}\]To see this, let \(\Phi_\varepsilon(s)=\frac12\bigl(\sqrt{s^2+\varepsilon^2}-s\bigr)\), a smooth approximation of \(s^-\), which satisfies \(0\le\Phi_\varepsilon(s)-s\Phi_\varepsilon'(s)\le\varepsilon/2\). Using this estimate, the equation for $\tau$, and the positivity of both \(u\,h(u)\) (the nonnegativity of $h$ follows from the monotonicity of \(f(s)/s\)) and $\Phi_\varepsilon’’$ (as $\Phi_\varepsilon$ is convex), we compute
\[\begin{align*} -\operatorname{div}\bigl(u^2\nabla\Phi_\varepsilon(\tau)\bigr)+u\,h(u)\,\Phi_\varepsilon(\tau) &=-\operatorname{div}\bigl(\Phi_\varepsilon'(\tau)\,u^2\nabla\tau\bigr)+u\,h(u)\,\Phi_\varepsilon(\tau)\\ &=-\Phi_\varepsilon'(\tau)\operatorname{div}(u^2\nabla\tau)-\Phi_\varepsilon''(\tau)\,u^2|\nabla\tau|^2+u\,h(u)\,\Phi_\varepsilon(\tau)\\ &=u\,h(u)\bigl[\Phi_\varepsilon(\tau)-\tau\Phi_\varepsilon'(\tau)\bigr]-u^2\,\Phi_\varepsilon''(\tau)\,|\nabla\tau|^2\\ &\le uh(u)\frac{\varepsilon}2. \end{align*}\]Letting \(\varepsilon\to0\) gives (1).
Step 3: a Caccioppoli inequality. Fix \(q\ge2\) and a nonnegative Lipschitz function \(\eta\) with compact support. We want to test (1) with \(z^{q-1}\eta^2\). In order to integrate by parts without taking care of boundary terms, we first show that \(z\) vanishes near \(\partial\Omega\). On \(\partial\Omega\) we have \(u=0\), so \(\nabla u\) is normal to the boundary, and the Neumann condition gives \(\nabla u=-\kappa\nu\). The outer normal of an epigraph points downward,
\[\nu=\frac{(\nabla g,-1)}{\sqrt{1+|\nabla g|^2}}.\]Hence,
\[\partial_nu=\frac{\kappa}{\sqrt{1+|\nabla g|^2}}>0\quad\text{on }\partial\Omega .\]Since \(\nabla u\) is continuous up to the boundary, for every compact set \(K\subset\mathbb R^n\) there is a neighborhood of the compact set \(\partial\Omega\cap K\) on which \(\partial_nu>0\). Then \(z\) vanishes near \(\partial\Omega\cap K\). Consequently, \(z^{q-1}\eta^2\) is supported in a compact subset of the open set \(\Omega\), so they produce no boundary terms.
Now test (1) with \(z^{q-1}\eta^2\). Since
\[\nabla\bigl(z^{q-1}\eta^2\bigr)=(q-1)\,z^{q-2}\eta^2\,\nabla z+2\,z^{q-1}\eta\,\nabla\eta,\]we obtain
\[(q-1)\int u^2z^{q-2}|\nabla z|^2\eta^2+2\int u^2z^{q-1}\eta\,\nabla z\cdot\nabla\eta+\int u\,h(u)\,z^q\eta^2\le0 .\]Multiplying by \(q/2\), this bounds the cross term:
\[q\int u^2z^{q-1}\eta\,\nabla z\cdot\nabla\eta \le-\frac{q(q-1)}2\int u^2z^{q-2}|\nabla z|^2\eta^2-\frac q2\int u\,h(u)\,z^q\eta^2 .\]On the other hand, since \(\nabla(z^{q/2}\eta)=\frac q2\,z^{q/2-1}\eta\,\nabla z+z^{q/2}\nabla\eta\), expanding the square gives
\[\begin{aligned} \int u^2\bigl|\nabla(z^{q/2}\eta)\bigr|^2 &=\frac{q^2}4\int u^2z^{q-2}|\nabla z|^2\eta^2+q\int u^2z^{q-1}\eta\,\nabla z\cdot\nabla\eta+\int u^2z^q|\nabla\eta|^2 . \end{aligned}\]Substituting the bound for the cross term and using \(\frac{q^2}4-\frac{q(q-1)}2=-\frac{q(q-2)}4\), we get
\[\int u^2\bigl|\nabla(z^{q/2}\eta)\bigr|^2+\frac{q(q-2)}4\int u^2z^{q-2}|\nabla z|^2\eta^2+\frac q2\int u\,h(u)\,z^q\eta^2 \le\int u^2z^q|\nabla\eta|^2 .\]Since \(q\ge2\) and \(u\,h(u)\geq0\), the second and third terms on the left are nonnegative, and we may drop them:
\[\int u^2\bigl|\nabla(z^{q/2}\eta)\bigr|^2\le\int u^2z^q|\nabla\eta|^2 . \qquad\text{(2)}\]Step 4: one more power of \(z\). Where \(\tau<0\) we have \(\partial_nu=-zu\), so \(\partial_n(u^2)\,z^q=-2u^2z^{q+1}\) everywhere. Integrating by parts in \(x_n\) (no boundary terms, by Step 3),
\[\begin{aligned} \int u^2z^{q+1}\eta^2 &=-\frac12\int\partial_n(u^2)\,z^q\eta^2\\ &=\frac12\int u^2\,\partial_n\bigl(z^q\eta^2\bigr)\\ &=\int u^2\,\bigl(z^{q/2}\eta\bigr)\,\partial_n\bigl(z^{q/2}\eta\bigr). \end{aligned}\]By the Cauchy–Schwarz inequality and (2),
\[\int u^2z^{q+1}\eta^2 \le\left(\int u^2z^q\eta^2\right)^{1/2}\left(\int u^2z^q|\nabla\eta|^2\right)^{1/2}. \qquad\text{(3)}\]Step 5: the key estimate. Let \(\chi_R\) be a Lipschitz cutoff with \(0\le\chi_R\le1\), \(\chi_R=1\) on \(B_R\), \(\chi_R=0\) outside \(B_{2R}\), and \(\lvert\nabla\chi_R\rvert\le1/R\). Fix an integer \(m\ge3\) and apply (3) with
\[q=m-1,\qquad \eta=\chi_R^m,\qquad |\nabla\eta|\le\frac mR\,\chi_R^{m-1}.\]Using (3), \(\chi_R^{2m}\le\chi_R^{2m-2}\), and Hölder’s inequality with exponents \(\frac m{m-1}\) and \(m\) applied to the factorization
\[u^2z^{m-1}\chi_R^{2m-2}=\bigl(u^2z^m\chi_R^{2m}\bigr)^{\frac{m-1}m}\bigl(u^2\mathbf 1_{B_{2R}}\bigr)^{\frac1m},\]we get
\[\begin{aligned} \int u^2z^m\chi_R^{2m} &\le\left(\int u^2z^{m-1}\chi_R^{2m}\right)^{1/2}\left(\int u^2z^{m-1}\Bigl(\frac mR\Bigr)^2\chi_R^{2m-2}\right)^{1/2}\\ &\le\frac mR\int u^2z^{m-1}\chi_R^{2m-2}\\ &\le\frac mR\left(\int u^2z^m\chi_R^{2m}\right)^{\frac{m-1}m}\left(\int_{\Omega\cap B_{2R}}u^2\right)^{\frac1m}. \end{aligned}\]The integral \(\int u^2z^m\chi_R^{2m}\) is finite by Step 3; if it vanishes there is nothing to prove, and otherwise we divide by its power \(\frac{m-1}m\). Writing
\[V(r)=\int_{\Omega\cap B_r}u^2\,dx,\]we obtain
\[\int_{\Omega\cap B_R}u^2z^m\,dx \le\int u^2z^m\chi_R^{2m} \le\left(\frac mR\right)^mV(2R). \qquad\text{(4)}\]Step 6: a strict decrease forces exponential growth. Suppose that \(z(x_0)>0\) for some \(x_0\in\Omega\). By continuity there are a ball \(A\subset\Omega\) and a number \(a>0\) with \(z\geq a\) on \(A\); let \(\mu_A=\int_Au^2>0\). For large \(R\) we have \(A\subset B_R\), and (4) gives
\[\mu_A\,a^m\le\left(\frac mR\right)^mV(2R),\]that is,
\[V(2R)\geq\mu_A\left(\frac{aR}m\right)^m .\]Choose \(m=\lfloor aR/2\rfloor\). Then \(V(2R)\geq\mu_A\,2^{\lfloor aR/2\rfloor},\) and we can deduce that,
\[\liminf_{R\to\infty}\frac1R\log\bigl(1+V(R)\bigr)\geq\frac{a\log2}4>0,\]which contradicts (G). Hence \(z\equiv0\), that is, \(\partial_nu\geq0\) in \(\Omega\). (If \(u\) grows at most polynomially, say \(u(x)\le C(1+\lvert x\rvert)^k\), there is a simpler argument that avoids letting \(m\) depend on \(R\): then \(V(2R)\le C'R^{n+2k}\) for \(R\ge1\), and (4) with a fixed integer \(m>n+2k\) gives \(\int_{\Omega\cap B_R}u^2z^m\le C''R^{n+2k-m}\to0\) as \(R\to\infty\).)
Step 7: strict monotonicity. The derivative \(\partial_nu\geq0\) solves the linear equation \(-\Delta\partial_nu=f'(u)\,\partial_nu\) with a locally bounded coefficient. By the strong maximum principle, either \(\partial_nu\equiv0\) or \(\partial_nu>0\) in the connected open set \(\Omega\). The first alternative is ruled out by \(\partial_nu>0\) on \(\partial\Omega\) (Step 3). \(\square\)
References
- James Serrin, A symmetry problem in potential theory, Archive for Rational Mechanics and Analysis 43 (1971), 304–318.
- Henri Berestycki, Luis Caffarelli and Louis Nirenberg, Monotonicity for elliptic equations in unbounded Lipschitz domains, Communications on Pure and Applied Mathematics 50 (1997), 1089–1111. The conjecture, the monotonicity theorem for Lipschitz epigraphs, and the first rigidity results for epigraphs.
- Antonio Ros, David Ruiz and Pieralberto Sicbaldi, A rigidity result for overdetermined elliptic problems in the plane, Communications on Pure and Applied Mathematics 70 (2017), 1223–1252.
- Alberto Farina and Enrico Valdinoci, Flattening results for elliptic PDEs in unbounded domains with applications to overdetermined problems, Archive for Rational Mechanics and Analysis 195 (2010), 1025–1058.
- Kelei Wang and Juncheng Wei, On Serrin’s overdetermined problem and a conjecture of Berestycki, Caffarelli and Nirenberg, Communications in Partial Differential Equations 44 (2019), 837–858. arXiv:1502.04680
- Nicolas Beuvin and Alberto Farina, Serrin’s overdetermined problems on epigraphs, Mathematische Annalen 396 (2026), article 28. arXiv:2502.04812
- Nicolas Beuvin, Alberto Farina and Berardino Sciunzi, Monotonicity for solutions to semilinear problems in epigraphs, Journal de Mathématiques Pures et Appliquées 210 (2026), 103868.
- Henri Berestycki and Cole Graham, A stable-compact method for qualitative properties of semilinear elliptic equations, Archive for Rational Mechanics and Analysis 250 (2026), article 21.
- Manuel del Pino, Frank Pacard and Juncheng Wei, Serrin’s overdetermined problem and constant mean curvature surfaces, Duke Mathematical Journal 164 (2015), 2643–2722. arXiv:1310.4528
Comments
Questions, corrections, and further ideas are welcome. Sign in with GitHub to join the discussion.