Boundary Conditions and Well-Posedness

Well-posedness is the entry gate to computational science: the Lax equivalence theorem makes it the precondition for a numerical scheme to converge, and the Courant-Friedrichs-Lewy step-size limit echoes finite propagation speed. The theory bites hardest on the ill-posed side, because so many questions we want to answer are inverse problems wearing Hadamard's warning label. Computed tomography, electrical impedance and diffuse optical imaging, seismic inversion, and image deblurring are all, at bottom, backward-diffusion or elliptic-Cauchy problems that inherit Hadamard's exponential amplification of high-frequency noise. The modern response is not to pretend the instability away but to regularize — Tikhonov penalties, truncated spectral expansions, and Bayesian priors that trade a little bias for a large drop in variance — turning an ill-posed reconstruction into a well-posed approximation, the idea behind weather-forecast data assimilation. Hadamard's century-old distinction between the well-posed and the merely soluble is the daily boundary between the calculations we trust and those we handle with care.

Think of the PDE as a local rule and the boundary data as the global anchoring that locks the rule onto a single state. An equilibrium (elliptic) membrane is pinned everywhere along its rim: give the height at every edge point and the equation fills in the interior uniquely, one scalar per boundary point. A propagating (hyperbolic) string is launched: give its initial shape and velocity along a line of constant time and it marches forward along its characteristics. A diffusing (parabolic) temperature field is both launched and contained: give an initial profile and keep telling it what the ends do forever after. The first figure lays these three pairings side by side so the counting — how much data each type needs, and on what surface — becomes one picture.

The Hadamard requirements are a triage of three failure modes: non-existence (the data over-constrain the equation), non-uniqueness (they under-constrain it), and instability (the data-to-solution map is a valid but discontinuous function that stretches bounded inputs into unbounded outputs). Only when all three are healthy does a measured input yield a trustworthy prediction. The two universal weapons for the good cases are the energy method (multiply by the solution, integrate, and let the boundary conditions kill the surface terms) and the maximum principle (extremes occur only on the boundary, so identical data force identical interiors). When the pairing is wrong no such argument exists, and the failure turns quantitative, as the second figure shows.

Hadamard's three demands are the definition against which every problem is measured — not a formula but a checklist that turns a vague wish for a "sensible" model into a testable property.

where: $(i)$ — at least one function satisfies the PDE and all data; $(ii)$ — no two distinct functions satisfy the same data; $(iii)$ — the data-to-solution map is continuous, so small data changes give small solution changes. Reading: existence and uniqueness are about counting conditions correctly; continuous dependence is about the geometry of the solution map — the criterion Hadamard's example violates while passing the other two. The first and commonest type prescribes the value of the field on the boundary — a clamped displacement, a thermostatted wall, a fixed electrode voltage.

where: $u$ — the unknown field on $\Omega$; $\partial\Omega$ — the closed boundary; $g$ — the prescribed values. Reading: the Dirichlet condition — for an elliptic equation it supplies one scalar per boundary point, just what a unique, stable solution needs. The second prescribes the normal derivative — a heat flux, a free-edge force, a surface charge density.

where: $\hat{\mathbf n}$ — outward unit normal; $\partial u/\partial n$ — the normal derivative (flux or free-edge force); $h$ — its prescribed values. Reading: the Neumann condition. The divergence theorem forces a compatibility condition $\oint_{\partial\Omega}h\,\mathrm{d}S=\int_\Omega f\,\mathrm{d}V$ (total flux equals total source), and a constant shift in $u$ leaves $\partial u/\partial n$ unchanged, so uniqueness holds only up to that constant. The third interpolates between the two, weighting value against flux — a wall losing heat by convection or radiation (Newton cooling).

where: $\alpha,\beta$ — real weighting constants (or functions) with $\beta\neq0$; $q$ — the prescribed combination. Reading: the Robin (mixed) condition. $\beta=0$ recovers Dirichlet, $\alpha=0$ recovers Neumann; $\alpha/\beta>0$ is the Newton-cooling coefficient, and because the value term pins the level it restores full uniqueness to the elliptic problem. The fourth prescribes both value and normal derivative on one surface — twice as much data per point as the elliptic types.

where: $S$ — a surface, typically the initial slice $t=0$; $\phi,\psi$ — the prescribed value and normal derivative. Reading: Cauchy data — right for a hyperbolic equation on a non-characteristic $S$ (the initial shape and velocity of a launched string), wrong for an elliptic one, where it over-determines the problem and destroys continuous dependence (equation 8). For the hyperbolic case, the tool certifying both uniqueness and stability is a conserved energy — for the wave equation, kinetic plus potential.

where: $E(t)$ — total energy of the string; $u_t,u_x$ — time and space derivatives; $c$ — wave speed; $L$ — length. Reading: differentiating and integrating by parts gives $\dot E=\int_0^L u_t(u_{tt}-c^2u_{xx})\,\mathrm{d}x+c^2[u_tu_x]_0^L$, which vanishes by the wave equation and by fixed ($u_t=0$) or free ($u_x=0$) ends. The difference of two solutions with identical data then has $E(0)=0$, so $E\equiv0$ — uniqueness, with stability since $E$ bounds the solution. For elliptic and parabolic cases the tool is a maximum principle: a harmonic function hides no interior extremum.

where: $\overline\Omega$ — the closed domain (interior plus boundary); $u$ — a function satisfying $\nabla^2u=0$. Reading: a harmonic function attains its extremes only on $\partial\Omega$, and the same holds for the heat equation with $\partial\Omega$ replaced by the parabolic boundary (the initial slice plus the two ends). Derivation 2 turns this into uniqueness — a mechanism entirely different from energy. Finally, the mode showing why Cauchy data must never touch an elliptic equation: harmonic, with vanishing data yet unbounded growth.

where: $N$ — a large integer setting the data's oscillation frequency; $\sin(Nx)$ — the oscillatory pattern; $\sinh(Ny)$ — exponential growth away from the data line. Reading: the Cauchy data on $y=0$ are $u_N(x,0)=0$ and $\partial_yu_N(x,0)=\tfrac1N\sin(Nx)$, of amplitude $1/N\to0$; yet at any fixed $y>0$ the solution reaches $\sinh(Ny)/N^2\to\infty$. One line proves the elliptic Cauchy problem ill-posed.

The well-posed data set changes discontinuously with the equation's type. Each row pairs a canonical equation with the data that make it well-posed; the last two rows are cautionary.

1 · Quantifying Hadamard's instability (fully worked)

For $u_N=\frac{1}{N^2}\sin(Nx)\sinh(Ny)$ at height $y=1$, tabulate the Cauchy-data amplitude at $y=0$ against the solution amplitude at $y=1$ for $N=10,20,40$. How badly does the data-to-solution map fail to be continuous?

Step 1. The data live on $y=0$: $u_N(x,0)=0$ and $\partial_yu_N(x,0)=\frac1N\sin(Nx)$, so the data amplitude $1/N$ is $0.100,\ 0.050,\ 0.025$ at $N=10,20,40$ — shrinking toward zero.

Step 2. The solution amplitude at $y=1$ is $\sinh(N)/N^2$: $\sinh(10)/100=\mathbf{110}$, $\sinh(20)/400=\mathbf{6.06\times10^5}$, $\sinh(40)/1600=\mathbf{7.36\times10^{13}}$.

Step 3. The amplification factor is $A(N)=\dfrac{\sinh(N)/N^2}{1/N}=\dfrac{\sinh(N)}{N}$, giving $A(10)\approx1.1\times10^3$, $A(20)\approx1.2\times10^7$, $A(40)\approx2.9\times10^{15}$: the gain grows without bound as the data get finer.

Step 4. Halving the data (from $N=20$ to $N=40$) multiplies the solution by more than $10^8$. There is no constant $C$ with $\|\text{solution}\|\leq C\|\text{data}\|$, so continuous dependence fails and the elliptic Cauchy problem is ill-posed.

2 · Does a pure Neumann problem have a solution? (compatibility check)

On the unit disk (area $\pi$, circumference $2\pi$), solve $\nabla^2u=f$ with uniform source $f=4$ and uniform outward flux $\partial u/\partial n=h$ on the rim. (a) For what $h$ does a solution exist? (b) What if $h=3$? (c) Is the solution then unique?

Step 1. A pure Neumann problem is solvable only if the data satisfy the divergence-theorem compatibility condition $\oint_{\partial\Omega}h\,\mathrm{d}S=\int_\Omega f\,\mathrm{d}V$: net flux through the boundary must equal net source enclosed.

Step 2. For uniform data the left side is $h\times(\text{circumference})=2\pi h$ and the right side is $f\times(\text{area})=4\pi$. Equating, $2\pi h=4\pi$, so $h=2$ — the only flux consistent with a source of strength $4$ over the disk.

Step 3. Prescribe $h=3$ and the left side is $6\pi$ against $4\pi$: the boundary must carry out more flux than the interior produces. No solution exists — a steady state would need charge or heat from nowhere. An existence failure, not an instability.

Step 4. Even with $h=2$ the solution is not unique: if $u$ works so does $u+\text{const}$, since a constant changes neither $\nabla^2u$ nor $\partial u/\partial n$. The pure Neumann problem is well-posed only up to an additive constant — the free reference level (ground of a potential, zero of a temperature). Pinning $u$ at one point removes it.

Well-posedness seems obvious once stated, yet Jacques Hadamard was the first to state it — and the first to show, with one counterexample, how badly a natural-looking problem can behave.

Texts. Stone & Goldbart, Mathematics for Physics, ch. 6-7 (PDEs and boundary-value problems) · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., ch. 9 · Hassani, Mathematical Physics, 2nd ed., Part VIII (PDEs) · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., ch. 20 · Courant & Hilbert, Methods of Mathematical Physics, vol. II · John, Partial Differential Equations, 4th ed. The key checkpoints across this article are: (1) Green's first identity gives $\int_\Omega|\nabla w|^2\,\mathrm{d}V=0$, establishing uniqueness for the Dirichlet problem; (2) $\nabla^2[\tfrac{1}{N^2}\sin(Nx)\sinh(Ny)]=0$ and $\sinh(N)/N^2$ equals $110,\ 6.06\times10^5,\ 7.36\times10^{13}$ at $N=10,20,40$, quantifying Hadamard's blow-up; (3) $\mathrm{d}E/\mathrm{d}t=0$ for the wave energy under fixed or free ends; (4) Neumann compatibility $\oint h\,\mathrm{d}S=\int f\,\mathrm{d}V$, giving $h=2$ on the unit disk with $f=4$.

Prerequisites. Classification of Second-Order Linear PDEs · Laplace's and Poisson's Equations · The Wave Equation · The Heat (Diffusion) Equation. Leads to. Separation of Variables in Cartesian Coordinates · Sturm-Liouville Theory and Eigenfunction Expansions · Fundamental Solutions and Green's Functions for PDEs · Method of Characteristics.

Ask the tutor: Tell me your PDE and the surface you want to prescribe data on, and I will tell you whether the problem is well-posed — and if not, which of Hadamard's three conditions it breaks.