Maximum Principles and Uniqueness for Elliptic and Parabolic PDEs

A second-order linear PDE is elliptic when its principal symbol is sign-definite, and parabolic when that symbol degenerates in exactly one, time-like, direction. Neither type supports free travelling waves, and that single structural fact has a sharp physical payoff: the solution everywhere is dictated by its data on the boundary of the region (elliptic) or on the parabolic boundary — the initial slice plus the walls, but not the final instant (parabolic). The maximum principle is the rigorous statement of this idea, and it is the analytic engine of the whole subject. Its charm is that it is at once elementary and powerful. Elementary, because its weakest form is nothing more than calculus applied to the Hessian at a peak: at an interior maximum the gradient vanishes and the second-derivative matrix is negative semidefinite, so a positive-definite leading operator cannot produce a positive value there. Powerful, because from that one observation flow uniqueness, continuous dependence on data (well-posedness), and comparison theorems ordering solutions by their data — all without ever solving the equation. This article works on a bounded domain Ω ⊂ ℝⁿ and the general operator $Lu = \sum a_{ij}\partial_i\partial_j u + \sum b_i\partial_i u + c\,u$ with $(a_{ij})$ uniformly positive definite. The weak principle sends a source-free extremum to the boundary; Eberhard Hopf's strong principle and boundary-point lemma sharpen it to forbid interior extrema outright and control the sign of the normal derivative. For the heat equation the same machinery, respecting the arrow of time, sends extrema to the parabolic boundary — the analytic face of irreversibility. The sign of the reaction coefficient $c$ is decisive throughout, and its failure is exactly why the Helmholtz equation escapes the principle.

Key takeaways

The maximum principle says a source-free elliptic or parabolic solution cannot have an interior extremum beyond its boundary values — peaks and troughs migrate to the boundary — and this at once delivers uniqueness for the Dirichlet problem and sup-norm stability $\lVert u_1-u_2\rVert_\infty \le \lVert g_1-g_2\rVert_\infty$ .
For Laplace's equation $\nabla^2 u = 0$ the weak form gives $\min_{\partial\Omega} u \le u(x) \le \max_{\partial\Omega} u$ throughout Ω, a companion of the mean-value property $u(x_0)=\text{average of }u$ over any enclosed sphere.
Hopf's strong principle sharpens this: a non-constant subsolution attains no interior maximum, and at a boundary maximum the outward normal derivative is strictly positive (the Hopf lemma) — the key to Neumann and Robin uniqueness.
For the heat equation the extremum sits on the parabolic boundary — initial slice plus walls, never the top — encoding causality; a cooling body never exceeds max(initial, boundary) temperature.
The reaction sign is everything: the principle needs $c\le 0$ ; for Helmholtz $\nabla^2 u+k^2 u=0$ ( $c=k^2>0$ ) it fails, and eigenfunctions oscillate with interior peaks.
Key idea: At an interior maximum the gradient vanishes and the Hessian is negative semidefinite; uniform ellipticity forces the leading term $\sum a_{ij}\partial_i\partial_j u \le 0$ , contradicting $Lu>0$ . This trivial-looking calculus fact, bootstrapped with the Hopf lemma, controls existence, uniqueness, and stability across the whole elliptic–parabolic class.

The core insight

At an interior maximum the gradient vanishes and the Hessian is negative semidefinite; uniform ellipticity then forces the leading term $\sum a_{ij}\partial_i\partial_j u \le 0$ , contradicting $Lu>0$ . This trivial-looking calculus fact, when bootstrapped with the Hopf lemma, controls existence theory, uniqueness, and stability across the whole class of elliptic and parabolic equations.

Where this lives today

The maximum principle is load-bearing in modern analysis and computation. In fully nonlinear elliptic theory it is the backbone of viscosity solutions — Crandall, Ishii, and Lions define "solution" for Hamilton–Jacobi and Monge–Ampère equations by comparison against smooth test functions at extrema, the maximum principle promoted to an axiom. In scientific computing, engineers demand discrete maximum principles: a scheme that respects it produces no spurious overshoots or negative densities, is provably $L^\infty$ -stable, and yields monotone answers — the five-point stencil and $M$ -matrix conditions of the first figure are exactly this requirement. Comparison principles power the monotone (Perron) iteration and barrier methods behind free-boundary and obstacle problems, from American-option pricing to the melting front of the Stefan problem. And the humblest use is the most pervasive: every uniqueness step behind "by symmetry the potential must be…" or "the method of images gives the unique field" is the maximum principle standing silently behind the argument.

The physical picture

Picture a harmonic function as a taut, unloaded soap film over a bent wire frame. With no pressure on it the film satisfies $\nabla^2 u=0$ , and its highest and lowest points are inevitably on the wire — it cannot bulge into a dome or sag into a pit inside, because at such a peak the surface is concave in every direction, giving $\nabla^2 u<0$ , not zero. The same fact wears a second face, the mean-value property: a harmonic function at any interior point equals its average over any sphere centred there. An average never exceeds the largest value averaged, so an interior point can only match its surroundings, never top them. The first figure makes this quantitative through the discrete stencil.

Figure

The mean-value property in discrete form: the interior value equals the average of its four neighbours, u₀ = (100 + 0 + 60 + 20)/4 = 45, which cannot exceed the largest boundary value 100 — no interior peak is possible

Diffusion adds the arrow of time. Temperature obeys $u_t = D\nabla^2 u$ , so a hot interior spot — concave, $\nabla^2 u<0$ — has $u_t<0$ and must cool. A peak cannot spontaneously grow; the hottest the body ever gets is fixed by its initial state or by the walls. In space-time the maximum therefore lives on the parabolic boundary — the base at $t=0$ plus the lateral walls, not the top face at $t=T$ , where the smoothed data merely come to rest. The second figure draws that cylinder and pins the bound to a concrete temperature. Hopf's boundary-point lemma supplies the last ingredient: at a smooth boundary maximum the function is genuinely climbing toward it, $\partial u/\partial n>0$ — the gradient cannot quietly vanish there, which is exactly what makes uniqueness work under Neumann and Robin data.

Figure

The parabolic maximum principle: on the space-time cylinder Q_T the maximum is attained on the parabolic boundary (base t = 0 plus the walls, drawn heavy), never on the top t = T. An initial 100 °C peak only decays — after 60 s a steel bar has fallen to 49 °C

The core equations

The whole theory is built around one operator and one sign condition. We write it in non-divergence form, so the principal part acts directly on the Hessian — the object we control at an extremum.

Lu \;=\; \sum_{i,j=1}^{n} a_{ij}(x)\,\partial_i\partial_j u \;+\; \sum_{i=1}^{n} b_i(x)\,\partial_i u \;+\; c(x)\,u \qquad (1)

where: $a_{ij}$ — symmetric coefficient matrix of the principal (second-order) part; $b_i$ — first-order drift/convection coefficients; $c$ — zeroth-order reaction coefficient, whose sign decides whether the principle holds; Ω — bounded open domain in ℝⁿ. Everything below is a statement about $L$ under a positivity hypothesis on $(a_{ij})$ and a sign hypothesis on $c$ . Ellipticity is the condition that the principal symbol never degenerates: the quadratic form built from $(a_{ij})$ is bounded below by a fixed positive multiple of $|\xi|^2$ in every direction.

\exists\,\lambda>0:\quad \sum_{i,j} a_{ij}(x)\,\xi_i\xi_j \;\ge\; \lambda\,|\xi|^2 \quad \forall\,\xi\in\mathbb{R}^n,\ x\in\Omega \qquad (2)

where: λ — ellipticity constant, the uniform floor on the smallest eigenvalue of $(a_{ij})$ ; ξ — arbitrary real cotangent vector. A uniform positivity gap is what turns "the Hessian is $\preceq 0$ at a peak" into "the leading term is $\le 0$ at a peak" — the seed of every proof. Harmonic functions satisfy an exact averaging identity — value equals mean over any enclosed sphere or ball — both a consequence and a restatement of the maximum principle.

\nabla^2 u = 0 \ \text{in}\ \Omega \ \Longrightarrow\ u(x_0)=\frac{1}{|\partial B_r|}\!\int_{\partial B_r(x_0)}\! u\, \mathrm{d}S = \frac{1}{|B_r|}\!\int_{B_r(x_0)}\! u\, \mathrm{d}V \qquad (3)

where: $B_r(x_0)$ — ball of radius $r$ about the interior point $x_0$ with $\overline{B_r}\subset\Omega$ ; $|\partial B_r|$ — surface area of the bounding sphere; $|B_r|$ — its volume. An interior value is a genuine average of surrounding values, so it cannot strictly exceed them. The weak maximum principle packages the averaging fact into a two-sided bound by the boundary data.

\nabla^2 u = 0 \ \text{in}\ \Omega,\ \ u\in C(\bar\Omega) \ \Longrightarrow\ \min_{\partial\Omega} u \;\le\; u(x) \;\le\; \max_{\partial\Omega} u \quad \forall x\in\bar\Omega \qquad (4)

where: $C(\bar\Omega)$ — continuous up to and including the boundary; $\partial\Omega$ — topological boundary, the only place extrema may be attained. The interior is trapped between the least and greatest boundary values. Hopf's strong principle upgrades " $\le$ " to a strict interior inequality for any non-constant subsolution — an interior maximum is not merely bounded, it is impossible.

Lu \ge 0,\ c\equiv 0,\ u\ \text{non-constant} \ \Longrightarrow\ u(x) < \max_{\bar\Omega} u \quad \forall x\in\Omega \qquad (5)

where: $Lu\ge 0$ — $u$ is a subsolution; $\max_{\bar\Omega} u$ — the global maximum over the closed domain, attained only on $\partial\Omega$ for non-constant $u$ . The Hopf boundary-point lemma controls the geometry at the boundary maximum itself: there the function strictly increases along the outward normal.

Lu\ge 0\ \text{in}\ \Omega,\ \ u(x_0)=\max_{\bar\Omega}u,\ x_0\in\partial\Omega,\ u<u(x_0)\ \text{inside} \ \Longrightarrow\ \frac{\partial u}{\partial n}(x_0) > 0 \qquad (6)

where: $x_0$ — a boundary point satisfying an interior-ball condition; $\partial/\partial n$ — the outward normal derivative at $x_0$ . The normal derivative cannot vanish at a boundary extremum — what forces uniqueness for Neumann and Robin data. For the heat operator the same logic runs in space-time, but the extremum set shrinks to the parabolic boundary — base plus walls, not the final slice.

u_t - D\,\nabla^2 u = 0 \ \text{in}\ Q_T=\Omega\times(0,T] \ \Longrightarrow\ \max_{\overline{Q_T}} u = \max_{\partial_p Q_T} u \qquad (7)

where: $Q_T$ — the space-time cylinder up to time $T$ ; $\partial_p Q_T=(\bar\Omega\times\{0\})\cup(\partial\Omega\times[0,T])$ — the parabolic boundary, base plus lateral walls but not the top; $D>0$ — diffusivity. A temperature extremum can always be traced back in time or out to a wall. The immediate payoff is sup-norm stability of the Dirichlet problem: a contraction from boundary data to interior values.

\|u_1 - u_2\|_{L^\infty(\Omega)} \;\le\; \|g_1 - g_2\|_{L^\infty(\partial\Omega)} \qquad (8)

where: $u_1,u_2$ — harmonic solutions with boundary data $g_1,g_2$ ; $\lVert\cdot\rVert_{L^\infty}$ — supremum norm. A small change in the data produces an at-most-equal change in the solution — continuous dependence, hence well-posedness.

Derivation AWeak maximum principle by an epsilon-perturbation

The goal is to prove that a subsolution of a uniformly elliptic operator with no reaction term attains its maximum on the boundary, and to do it honestly — handling the borderline case where the Hessian at an interior peak is only negative semidefinite, so the naive contradiction $\operatorname{tr}(AH)\le 0$ versus $Lu>0$ does not immediately close. The trick is to dispose first of the strict case $Lu>0$ , then convert a mere subsolution into a strict one with a small convex exponential barrier, and finally let the perturbation vanish. Take $Lu\ge 0$ in Ω with ellipticity constant λ, $c\equiv 0$ , $u\in C^2(\Omega)\cap C(\bar\Omega)$ .

Step 1. Suppose first the strict inequality $Lu>0$ . If $u$ had an interior maximum at $x_0$ then $\nabla u(x_0)=0$ and the Hessian $H$ is negative semidefinite, so $Lu(x_0)=\sum a_{ij}H_{ij}=\operatorname{tr}(AH)$ . Writing $A=B^2$ gives $\operatorname{tr}(AH)=\operatorname{tr}(BHB)\le 0$ , contradicting $Lu(x_0)>0$ . Why: a strict subsolution cannot peak inside, because the leading operator reads off the concavity a maximum forces.

Step 2. Remove the strictness. For a genuine subsolution set $u_\varepsilon=u+\varepsilon\,e^{\gamma x_1}$ with $\gamma>0$ ; then $Lu_\varepsilon = Lu + \varepsilon e^{\gamma x_1}(a_{11}\gamma^2+b_1\gamma)\ge \varepsilon e^{\gamma x_1}(\lambda\gamma^2-\beta\gamma)$ , using $a_{11}\ge\lambda$ and $|b_1|\le\beta$ . Why: the exponential is convex in $x_1$ , so an elliptic operator gives it a strictly positive value we can dial up.

Step 3. Choose $\gamma>\beta/\lambda$ so the bracket $\lambda\gamma^2-\beta\gamma>0$ ; then $Lu_\varepsilon>0$ and Step 1 applies to $u_\varepsilon$ , giving $\max_{\bar\Omega}u_\varepsilon=\max_{\partial\Omega}u_\varepsilon\le \max_{\partial\Omega}u+\varepsilon e^{\gamma M}$ with $M=\max_{\bar\Omega}x_1$ finite (Ω is bounded). Why: the perturbed function is a strict subsolution, so its maximum is honestly on the boundary, where we can bound the added term.

Step 4. Since $u\le u_\varepsilon$ , $\max_{\bar\Omega}u\le \max_{\partial\Omega}u+\varepsilon e^{\gamma M}$ ; letting $\varepsilon\to 0$ yields $\max_{\bar\Omega}u=\max_{\partial\Omega}u$ , and the same argument on $-u$ gives the minimum principle. Why: the barrier was the only thing between "no strict interior peak" and "the maximum is literally on the boundary", and it costs nothing in the limit. Verified against Gilbarg–Trudinger Theorem 3.1 and Evans §6.4.1.

Go deeper · Derivation B · 1 minUniqueness and sup-norm stability for the Dirichlet problem

Derivation BUniqueness and sup-norm stability for the Dirichlet problem

The goal is to extract, from the weak maximum principle alone and in a few lines, the two pillars of well-posedness for Poisson's Dirichlet problem: that a solution is unique, and that it depends continuously on the boundary data in the supremum norm. The mechanism is disarmingly simple — the difference of two solutions with the same source is harmonic, so the principle traps it between its own boundary values, which are exactly the data mismatch. Let $u_1,u_2$ solve $\nabla^2 u=f$ in Ω with $u=g_1,g_2$ on $\partial\Omega$ .

Step 1. Form $w=u_1-u_2$ . Then $\nabla^2 w=f-f=0$ , so $w$ is harmonic with $w=g_1-g_2$ on $\partial\Omega$ . Why: the common source cancels, converting two inhomogeneous solutions into one source-free function — precisely where the maximum principle bites.

Step 2. Apply the weak maximum and minimum principles: $\min_{\partial\Omega}(g_1-g_2)\le w(x)\le \max_{\partial\Omega}(g_1-g_2)$ . Why: a harmonic function is pinned between its extreme boundary values, here the data mismatch itself.

Step 3. Combine into $|w(x)|\le \max_{\partial\Omega}|g_1-g_2|=\lVert g_1-g_2\rVert_{L^\infty(\partial\Omega)}$ , estimate (8). Why: continuous dependence — a data perturbation of size δ moves the solution by at most δ everywhere, so the problem is stable.

Step 4. Set $g_1=g_2$ : then $\lVert w\rVert_{L^\infty}=0$ , so $w\equiv 0$ and $u_1\equiv u_2$ — uniqueness. Why: uniqueness is the $\delta=0$ corner of stability, not a separate theorem; the parabolic principle gives the same for the heat IBVP. Verified against Riley–Hobson–Bence Ch. 21 and Evans §2.2.3.

Go deeper · Derivation C · 1 minThe energy method: an independent uniqueness proof for the heat equation

Derivation CThe energy method: an independent uniqueness proof for the heat equation

The goal is to prove uniqueness for the heat equation by a route that never mentions the maximum principle — an $L^2$ energy argument — both as a cross-check and because this method survives into settings (systems, higher-order operators) where no maximum principle exists. The idea is to watch the squared $L^2$ norm of the difference of two solutions: it starts at zero, can only decrease, and is non-negative, so it is trapped at zero forever. Let $u_1,u_2$ solve $u_t=D\nabla^2 u$ in $Q_T$ with identical initial and Dirichlet data, and set $w=u_1-u_2$ .

Step 1. The difference solves the homogeneous problem $w_t=D\nabla^2 w$ in $Q_T$ , with $w(\cdot,0)=0$ and $w=0$ on $\partial\Omega$ . Define $E(t)=\tfrac12\int_\Omega w^2\,\mathrm{d}V\ge 0$ . Why: subtracting kills source and data, leaving a clean problem whose only solution we suspect is zero.

Step 2. Differentiate: $\dfrac{\mathrm{d}E}{\mathrm{d}t}=\int_\Omega w\,w_t\,\mathrm{d}V = D\int_\Omega w\,\nabla^2 w\,\mathrm{d}V$ . Why: the energy's time derivative is governed by the spatial operator, which we integrate by parts.

Step 3. Green's first identity gives $\int_\Omega w\,\nabla^2 w\,\mathrm{d}V=\int_{\partial\Omega} w\,\partial_n w\,\mathrm{d}S-\int_\Omega|\nabla w|^2\,\mathrm{d}V$ ; the surface term vanishes since $w=0$ on $\partial\Omega$ , leaving $\dfrac{\mathrm{d}E}{\mathrm{d}t}=-D\int_\Omega|\nabla w|^2\,\mathrm{d}V\le 0$ . Why: diffusion dissipates the $L^2$ energy at a rate set by the gradient — a Lyapunov functional for the heat flow.

Step 4. Now $E(0)=0$ , $E\ge 0$ , and $E'\le 0$ force $E(t)\equiv 0$ , hence $w\equiv 0$ and $u_1\equiv u_2$ . Why: a non-negative quantity starting at zero that never increases is pinned at zero — uniqueness, by a method wholly independent of the maximum principle. Verified against Stone & Goldbart Ch. 7 and Evans §2.3.4.

Limiting cases

Whether a maximum principle holds turns sharply on the sign structure of the operator and the geometry of the domain — a map of where it lives and where it dies.

Case	Result	Meaning
Pure Laplacian, $c=0$	$\nabla^2 u=0\implies \min_{\partial\Omega}u\le u\le \max_{\partial\Omega}u$	Weak and strong principles both hold; a non-constant harmonic function takes no interior extremum.
Non-positive reaction, $c\le 0$	$Lu=\sum a_{ij}\partial_i\partial_j u+\sum b_i\partial_i u+c\,u$	The principle survives for the positive part: a non-negative interior maximum is bounded by the boundary; enough for uniqueness.
Positive reaction (Helmholtz), $c=k^2>0$	$\nabla^2 u+k^2 u=0$	Principle FAILS. Eigenfunctions like $\sin kx$ oscillate with interior extrema; uniqueness also fails at eigenvalues.
Parabolic (heat)	$u_t-D\nabla^2 u=0$	Extremum on the parabolic boundary $\partial_p Q_T$ only — the final-time top is excluded, encoding the arrow of time.
Unbounded domain	$\Omega=\mathbb{R}^n,\ u\to 0$ at ∞	Needs a growth restriction (Phragmén–Lindelöf / Liouville): a bounded harmonic function on all of ℝⁿ is constant; drop the bound and the principle fails.

Common trapThe maximum principle holds for any nice second-order equation

It does not — the reaction coefficient must satisfy $c\le 0$ , ideally $c\equiv 0$ for the cleanest statement. The moment $c>0$ , as in Helmholtz $\nabla^2 u+k^2 u=0$ , the principle collapses: at an interior maximum the Hessian term is still $\le 0$ , but the equation now reads $\sum a_{ij}\partial_i\partial_j u=-k^2 u$ , and the positive $+k^2 u$ no longer forbids the peak. Concretely $\sin(kx)$ and the drumhead modes $\sin(kx)\sin(ky)$ attain interior maxima with zero boundary data, and uniqueness fails at eigenvalues. Two further traps: the principle controls extrema of $u$ , not its derivatives (a harmonic gradient can peak inside), and it needs a bounded domain unless a growth condition is imposed. Sign, derivatives, boundedness — miss any one and the theorem evaporates.

Worked examples

1 · A hard temperature bound from the parabolic principle (fully worked)

A steel bar of length $L=0.10$ m has diffusivity $D=1.2\times10^{-5}\ \mathrm{m^2/s}$ . Its ends are held at $0\ ^\circ\mathrm{C}$ , and at $t=0$ the profile is $u(x,0)=100\sin(\pi x/L)\ ^\circ\mathrm{C}$ , a single hump peaking at $100\ ^\circ\mathrm{C}$ . Bound the temperature for all later times, and evaluate the interior peak at $t=60$ s.

Step 1. Read off the parabolic-boundary data. On $\partial_p Q_T$ the data are the two ends (at $0\ ^\circ\mathrm{C}$ for all $t$ ) and the initial slice (from $0$ up to $100\ ^\circ\mathrm{C}$ ), so $\max_{\partial_p Q_T}u=100\ ^\circ\mathrm{C}$ . By equation (7), $\max_{\overline{Q_T}}u=100\ ^\circ\mathrm{C}$ — the bar never exceeds 100 °C, at any interior point or any later time. No solving required; the bound is pure maximum principle.

Step 2. Confirm with the exact solution. The sine mode is a Laplacian eigenfunction, so $u(x,t)=100\sin(\pi x/L)\,e^{-D(\pi/L)^2 t}$ , with decay rate $D(\pi/L)^2=1.2\times10^{-5}\times 986.96=1.184\times10^{-2}\ \mathrm{s^{-1}}$ and time constant $\tau=84.4$ s.

Step 3. Evaluate at $t=60$ s. The exponent is $1.184\times10^{-2}\times60=0.711$ , so $e^{-0.711}=0.491$ and the interior peak is $100\times0.491=\mathbf{49.1\ ^\circ C}$ — the largest value anywhere at that instant, comfortably below the bound. The peak only ever decays, saturating the principle solely at $t=0$ , exactly where the ceiling is touched.

2 · A harmonic function versus a Helmholtz mode

On the unit disk, verify that $u(x,y)=x^2-y^2$ obeys the maximum principle and the mean-value property, then show that the Helmholtz mode $v(x,y)=\sin(kx)\sin(ky)$ violates the principle on a square.

Step 1. Check harmonicity: $\nabla^2 u=2-2=0$ . In polar form $u=r^2\cos 2\theta$ , so on the boundary $r=1$ it ranges over $[-1,1]$ , with $\max=1$ at $(1,0)$ and $\min=-1$ at $(0,1)$ — both extremes sit on the boundary. For any interior point $|u|=r^2|\cos 2\theta|\le r^2<1$ , strictly below the boundary maximum, as the strong principle demands.

Step 2. Check the mean-value property at the origin: $\frac{1}{2\pi}\int_0^{2\pi} r^2\cos 2\theta\,\mathrm{d}\theta=0=u(0,0)$ — the centre value equals the circular average, confirming equation (3).

Step 3. Now the counterexample. Take $v=\sin(kx)\sin(ky)$ on $[0,\pi/k]^2$ . Then $\nabla^2 v=-2k^2 v$ , i.e. $\nabla^2 v+2k^2 v=0$ — Helmholtz with $c=2k^2>0$ . On the whole boundary $v=0$ , yet at the centre $(\pi/2k,\pi/2k)$ , $v=1>0=\max_{\partial\Omega}v$ . The interior value beats every boundary value — the principle fails because $c>0$ , and this same $v$ is a zero-data eigenfunction, so uniqueness fails at that eigenvalue too.

Physics of the portrait

The maximum principle is old as an observation — Gauss knew potentials have no interior extrema, and Earnshaw built a no-go theorem on it — but it became a general theorem in the hands of one man who saw that a single inequality on the Hessian governs a whole class of equations.

Origins · 1 minEberhard Hopf and how this was discovered

Eberhard Hopf

1902–1983 · elliptic and parabolic PDE theory

Eberhard Hopf, born in Salzburg and trained in Berlin, is one of those figures whose name attaches to results in three fields at once — the Wiener–Hopf technique in integral equations, the Hopf bifurcation in dynamical systems, and, central here, the Hopf maximum principle for PDEs. In 1927, at twenty-five, he published a short paper with an unassuming title ("Elementary remarks on the solutions of partial differential equations of second order of elliptic type") proving the strong maximum principle for general uniformly elliptic operators: a non-constant subsolution attains no interior maximum, full stop. The proof needs no explicit solution, no special coordinates, nothing but the sign of the leading operator acting on the Hessian — which is exactly what made it universal. A quarter-century later, in a 1952 note, he supplied the missing boundary ingredient, the boundary-point lemma, showing the outward normal derivative at a boundary maximum is strictly positive; Olga Oleinik found the same result independently that year, and together they closed the theory for Neumann and Robin problems. Hopf's career spanned continents and a difficult era — MIT in the early 1930s, a return to Germany in 1936 and the war years at Leipzig and Munich, then emigration in 1947 to Indiana University, where he remained until his death. What endures is the lesson at the heart of the 1927 paper: the qualitative behaviour of a vast family of equations is dictated not by solving them but by reading the sign of a trace at a single point.

History

1839

— Gauss establishes the mean-value property of harmonic (potential) functions

1842

— Earnshaw's theorem forbids stable electrostatic equilibrium, a physical corollary of the maximum principle

1927

— Eberhard Hopf proves the strong maximum principle for general second-order elliptic operators

1952

— Hopf, and independently Oleinik, prove the boundary-point lemma controlling the normal derivative

1967

— Protter and Weinberger's monograph systematizes maximum principles across elliptic and parabolic types

Going further

Texts. Stone & Goldbart, Mathematics for Physics (2009), Ch. 7 (PDEs) · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 9 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., Ch. 20–21 · Evans, Partial Differential Equations, 2nd ed., §2.2, §2.3, §6.4 · Gilbarg & Trudinger, Elliptic Partial Differential Equations of Second Order, Ch. 3 · Protter & Weinberger, Maximum Principles in Differential Equations (1967). Key checkpoints in the derivations: (1) the trace inequality $\operatorname{tr}(AH)\le 0$ for $A\succ 0,\ H\preceq 0$ follows by writing $A=B^2$ ; (2) the $\varepsilon e^{\gamma x_1}$ barrier with $\gamma>\beta/\lambda$ gives $Lu_\varepsilon>0$ ; (3) the energy identity $E'(t)=-D\int|\nabla w|^2$ follows from Green's first identity; (4) the steel-bar decay rate is $D(\pi/L)^2=1.18\times10^{-2}\ \mathrm{s^{-1}}$ and the $t=60$ s interior peak is $49.1\ ^\circ\mathrm{C}$ ; (5) $u=x^2-y^2$ is harmonic with boundary extrema $\pm1$ on the unit circle, and $v=\sin kx\sin ky$ solves $\nabla^2 v+2k^2 v=0$ with interior value 1 strictly above the zero boundary data.

Builds on. Classification of Second-Order Linear PDEs · Laplace's and Poisson's Equations · The Heat (Diffusion) Equation · Boundary Conditions and Well-Posedness. Leads to. The Helmholtz Equation · Sturm–Liouville Theory and Eigenfunction Expansions · Separation of Variables in Spherical Coordinates.

Ask the tutor: Ask me to run the steel-bar temperature bound or the $x^2-y^2$ versus Helmholtz comparison step by step, or to show where the barrier makes the Hopf boundary-point lemma work.

What is the difference between the weak and strong maximum principles?

The weak principle says a source-free elliptic solution is bounded by its boundary values, $u\le \max_{\partial\Omega}u$ everywhere, allowing that an interior point might equal the maximum. The strong principle removes that loophole for non-constant solutions — an interior point can never even equal the global maximum, so the max is attained only on the boundary. The weak form suffices for uniqueness and stability, which need only the two-sided bound; the strong form is what you reach for to rule out interior plateaus, keep a positive solution strictly positive, or run the Hopf-lemma argument for Neumann data. Weak is a bound; strong is a strict dichotomy between constant and boundary-only extrema.

Why does the sign of the coefficient c matter so much?

Because the whole proof hinges on what the equation says at an interior maximum. There the gradient vanishes and the Hessian is negative semidefinite, so the leading term $\sum a_{ij}\partial_i\partial_j u$ is $\le 0$ and the drift term is zero, and the equation reduces to $\sum a_{ij}\partial_i\partial_j u=-c\,u$ . If $c\le 0$ and $u>0$ at that max, the right side is $\ge 0$ while the left is $\le 0$ , squeezing consistently and forbidding a strict interior peak. But if $c>0$ , the term $+c\,u$ supplies the positivity the Hessian denies, and an interior maximum becomes possible — the Helmholtz story. The sign of $c$ decides whether the reaction reinforces or fights the concavity a peak requires.

Why is the final-time slice excluded for the heat equation?

Because diffusion has a temporal arrow. Suppose the space-time maximum sat at an interior point at the final time $t=T$ . Approaching that instant the temperature there was not decreasing (you are climbing into a maximum), so $u_t\ge 0$ ; but at a spatial maximum the profile is concave, $\nabla^2 u\le 0$ , and $u_t=D\nabla^2 u$ forces $u_t\le 0$ . The two are compatible only if $u_t=0=\nabla^2 u$ , and a closer analysis shows the value was already attained earlier or on a wall. So the maximum always traces back to the parabolic boundary. The top face is where the smoothed-out data come to rest, not where extrema are born; excluding it is the analytic signature of irreversibility.

How does the maximum principle prove uniqueness in one line?

Subtract. If $u_1$ and $u_2$ both solve the same Dirichlet problem — same source $f$ , same boundary data $g$ — then $w=u_1-u_2$ satisfies $\nabla^2 w=0$ inside (sources cancel) and $w=0$ on the boundary (data cancel). The maximum and minimum principles trap $w$ : $0=\min_{\partial\Omega}w\le w\le \max_{\partial\Omega}w=0$ , so $w\equiv 0$ and $u_1\equiv u_2$ . This is why a physicist who finds any potential matching the boundary data — by images, a lucky guess, or separation of variables — has found the potential: there is only one, and the maximum principle is the guarantee.

Is the maximum principle the same as the energy method?

No — they are two independent routes to overlapping conclusions. The maximum principle is an $L^\infty$ (pointwise) statement: it controls the largest value directly and yields sup-norm stability and comparison theorems. The energy method is an $L^2$ (integrated) statement: it watches a quadratic functional like $\tfrac12\int w^2$ and shows it dissipates. For a single scalar Laplace or heat equation both give uniqueness. But they part ways at the edges: the maximum principle handles drift terms and gives comparison and positivity results the energy method cannot, while the energy method extends to systems and higher-order operators where no maximum principle exists. A well-equipped analyst carries both.

Does the principle work on unbounded domains?

Not without an extra hypothesis. On a bounded domain the argument uses that continuous functions attain their extrema on the closure — which fails when the domain runs to infinity. The fix is a growth condition, the Phragmén–Lindelöf principle: if a subharmonic function grows slower than a comparison harmonic function, the principle is restored. The cleanest special case is Liouville's theorem: a harmonic function bounded on all of ℝⁿ must be constant. Drop the boundedness and counterexamples appear at once — $u=x_1$ is harmonic, unbounded, with no maximum at all. So exterior and whole-space problems must always pair the equation with a decay or growth condition.