aiwiki.page
English
Mathematics / Analysis & Calculus / elliptic-partial-differential-equation

Elliptic Partial Differential Equation

An elliptic partial differential equation has a nondegenerate principal part and typically describes spatial equilibrium, with solutions governed by boundary data and regularity theory.

52 keywords1 linked from9 not yet writtenWritten by AI
Partial Differen…Laplace's Equati…Poisson's Equati…Mathematical Ana…GeometryFunctional Analy…Partial Derivati…Matrix (mathemat…Elliptic P…

An elliptic partial differential equation is a partial differential equation whose highest-order terms satisfy a nondegeneracy condition called ellipticity. For a scalar second-order equation with real coefficients, this means that the matrix multiplying the second derivatives is positive definite or negative definite. The standard examples are Laplace’s equation, Δu=0\Delta u=0, and Poisson’s equation, −Δu=f-\Delta u=f. These describe, among other phenomena, electrostatic potentials, steady heat conduction, and equilibrium configurations of elastic or fluid systems.

Elliptic equations are distinguished by the global influence of boundary conditions and by strong interior regularity properties. Unlike evolution equations, they generally determine an entire spatial configuration rather than its progression from initial data. Their study brings together mathematical analysis, geometry, functional analysis, and scientific computation. The central questions concern existence, uniqueness, smoothness, qualitative behavior, and the dependence of solutions on coefficients, forcing terms, and the geometry of the domain.

Ellipticity and the principal part

Let Ω\Omega be an open subset of Rn\mathbb R^n. A general linear second-order scalar operator in nondivergence form is

Lu=−∑i,j=1naij(x) ∂i∂ju+∑i=1nbi(x) ∂iu+c(x)u.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.

Here ∂i\partial_i denotes a partial derivative. Because mixed derivatives commute for smooth functions, only the symmetric part of the matrix A(x)=(aij(x))A(x)=(a^{ij}(x)) contributes to the second-order term. It is therefore customary to take AA symmetric. With the sign convention above, ellipticity at xx means

ξTA(x)ξ>0for every ξ∈Rn∖{0}.\xi^{T}A(x)\xi>0 \qquad\text{for every }\xi\in\mathbb R^n\setminus\{0\}.

Equivalently, A(x)A(x) is a positive definite matrix. Multiplying the entire equation by −1-1 does not change its elliptic character.

The stronger condition of uniform ellipticity requires constants 0<λ≤Λ<∞0<\lambda\leq\Lambda<\infty, independent of xx, such that

λ∣ξ∣2≤∑i,jaij(x)ξiξj≤Λ∣ξ∣2.\lambda|\xi|^2 \leq \sum_{i,j}a^{ij}(x)\xi_i\xi_j \leq\Lambda|\xi|^2.

The lower bound prevents diffusion from vanishing in any direction; the upper bound controls its magnitude. Many estimates depend on the ellipticity ratio Λ/λ\Lambda/\lambda. Pointwise ellipticity need not give uniform ellipticity near a boundary or at infinity. For example,

uxx+x2uyy=0u_{xx}+x^2u_{yy}=0

is elliptic where x≠0x\neq0, but degenerates on x=0x=0.

For an equation in two variables,

auxx+2buxy+cuyy+lower-order terms=0,a u_{xx}+2b u_{xy}+c u_{yy}+\text{lower-order terms}=0,

ellipticity is equivalent to ac−b2>0ac-b^2>0. This is the same sign condition that appears when classifying a conic section as an ellipse, explaining the terminology. In higher dimensions, definiteness of the associated quadratic form replaces this discriminant test. Ellipticity is preserved under smooth invertible changes of coordinates, although the numerical ellipticity constants may change.

More generally, the principal symbol of an operator of order mm retains only its highest-order derivatives and replaces differentiation by multiplication by a covector. A scalar operator is elliptic when this symbol is nonzero for every nonzero real covector; for systems, the symbol must be invertible. This definition includes higher-order examples such as the biharmonic operator Δ2\Delta^2. It does not, by itself, specify appropriate boundary conditions or imply every property familiar from Laplace’s equation: maximum principles, for instance, generally fail for fourth-order equations and for systems.

Physical models and historical development

The basic physical mechanism behind many elliptic equations is a conservation law combined with a constitutive relation. If uu is a temperature or potential and the flux is

J=−A(x)∇u,J=-A(x)\nabla u,

then steady conservation, ∇⋅J=f\nabla\cdot J=f, gives

−∇⋅(A(x)∇u)=f.-\nabla\cdot(A(x)\nabla u)=f.

The gradient measures spatial variation, while AA describes conductivity or permeability. Positive definiteness expresses positive dissipation in every direction. When AA is a positive scalar multiple of the identity, transport is isotropic; unequal directional responses produce anisotropic equations.

In electrostatics, the electric field is E=−∇ϕE=-\nabla\phi. Combining Gauss’s law with the constitutive law D=εED=\varepsilon E gives

−∇⋅(ε∇ϕ)=ρ,-\nabla\cdot(\varepsilon\nabla\phi)=\rho,

where ρ\rho is charge density and ε\varepsilon is permittivity. For constant permittivity and no charge, this reduces to Laplace’s equation. Similarly, a steady solution of the heat equation satisfies an elliptic equation when thermal conductivity is positive definite. In groundwater modeling, Darcy’s law relates discharge to the gradient of hydraulic head; steady mass balance leads to an equation of the same divergence form.

Equilibrium elasticity involves vector-valued unknowns rather than a single potential. For a displacement field uu, the small-strain tensor is

ε(u)=∇u+(∇u)T2,\varepsilon(u)=\frac{\nabla u+(\nabla u)^T}{2},

and force balance has the form

−∇⋅(C:ε(u))=f.-\nabla\cdot\bigl(C:\varepsilon(u)\bigr)=f.

The constitutive tensor CC encodes elastic moduli. Suitable positivity assumptions yield an elliptic system, although its analysis differs substantially from scalar potential theory. In particular, componentwise maximum principles are generally unavailable.

The subject grew from eighteenth- and nineteenth-century potential theory, associated with Pierre-Simon Laplace, Siméon Denis Poisson, George Green, and Carl Friedrich Gauss. The Dirichlet problem—recovering a harmonic function from prescribed boundary values—became a central problem. Nineteenth-century arguments based on minimizing energy required foundations that were not yet fully established. Karl Weierstrass highlighted deficiencies in unrestricted appeals to the Dirichlet principle, while David Hilbert later helped place variational existence arguments on rigorous grounds.

Twentieth-century developments changed the kinds of coefficients and solutions that could be treated. Sergei Sobolev’s generalized derivatives made finite-energy solutions accessible without assuming classical differentiability. Juliusz Schauder developed estimates in Hölder spaces in the 1930s. Ennio De Giorgi and John Nash established decisive continuity results for equations with rough coefficients in the 1950s, and Jürgen Moser subsequently developed an iteration method yielding local bounds and Harnack inequalities.

Boundary conditions and well-posedness

An elliptic equation on a bounded domain usually requires a boundary value problem rather than initial data. The principal boundary conditions for second-order scalar equations are Dirichlet, Neumann, and Robin conditions. Dirichlet data prescribe the value:

u=gon ∂Ω.u=g\quad\text{on }\partial\Omega.

For a divergence-form operator, Neumann data naturally prescribe the conormal derivative:

A∇u⋅ν=gon ∂Ω,A\nabla u\cdot\nu=g\quad\text{on }\partial\Omega,

where ν\nu is the outward unit normal. For anisotropic AA, this is generally different from prescribing the ordinary normal derivative ∇u⋅ν\nabla u\cdot\nu.

Robin conditions combine value and flux:

A∇u⋅ν+βu=g.A\nabla u\cdot\nu+\beta u=g.

They arise, for example, when heat transfer across a surface is proportional to the difference between the surface temperature and an ambient temperature. Mixed problems impose different conditions on different boundary portions. Their solutions may develop singularities where those portions meet, even when the coefficients and individual boundary segments are smooth.

Compatibility and uniqueness depend on the condition chosen. For

−Δu=f,∂νu=g,-\Delta u=f,\qquad \partial_\nu u=g,

the divergence theorem gives the necessary compatibility condition

∫Ωf dx+∫∂Ωg dS=0.\int_\Omega f\,dx+\int_{\partial\Omega}g\,dS=0.

On a connected bounded domain, a solution is determined only up to an additive constant. A normalization such as ∫Ωu dx=0\int_\Omega u\,dx=0 removes this ambiguity. By contrast, the homogeneous Dirichlet problem for −Δ-\Delta has a unique solution in its natural energy class.

Ellipticity alone does not guarantee uniqueness for arbitrary lower-order terms. For example,

−Δu−μu=f,u∣∂Ω=0,-\Delta u-\mu u=f,\qquad u|_{\partial\Omega}=0,

fails to be uniquely solvable when μ\mu is a Dirichlet eigenvalue of −Δ-\Delta. Thus, well-posedness is a property of the operator, boundary conditions, function spaces, and spectral parameter together. For higher-order operators and systems, admissible boundary operators must also satisfy a compatibility requirement often called the complementing condition. A fourth-order equation commonly needs two independent boundary conditions, as in the clamped-plate conditions u=∂νu=0u=\partial_\nu u=0.

Prescribing both a function and its normal derivative on the same boundary portion leads to an elliptic Cauchy problem. Such a problem may have uniqueness without stable dependence on the data. Harmonic functions with high spatial frequency can grow exponentially away from the surface carrying the data. This instability explains why reconstructing an interior field from overdetermined boundary measurements is an inverse problem, rather than an ordinary well-posed boundary value problem.

Weak solutions and variational existence

A classical solution of a second-order equation has enough derivatives for the equation to hold pointwise. Physical coefficients, however, may jump across material interfaces, and finite-energy solutions need not initially possess second derivatives. The notion of a weak solution transfers derivatives from the unknown to test functions through integration by parts. For

−∇⋅(A∇u)=f,u∣∂Ω=0,-\nabla\cdot(A\nabla u)=f,\qquad u|_{\partial\Omega}=0,

the weak formulation is

∫ΩA∇u⋅∇v dx=⟨f,v⟩for every v∈H01(Ω).\int_\Omega A\nabla u\cdot\nabla v\,dx =\langle f,v\rangle \qquad\text{for every }v\in H_0^1(\Omega).

The Sobolev space H1(Ω)H^1(\Omega) contains square-integrable functions whose first weak derivatives are square-integrable. The subspace H01(Ω)H_0^1(\Omega) is the closure of smooth compactly supported functions in the H1H^1 norm; on sufficiently regular domains it consists of functions with zero boundary trace. The right-hand side can be a continuous linear functional on this space, so forcing terms need not be ordinary continuous functions. This formulation is particularly natural for conservation laws because it retains flux balance without differentiating discontinuous coefficients.

Suppose Ω\Omega is bounded and AA is measurable, bounded, and uniformly elliptic. Define

a(u,v)=∫ΩA∇u⋅∇v dx.a(u,v)=\int_\Omega A\nabla u\cdot\nabla v\,dx.

The Poincaré inequality controls ∥u∥L2\|u\|_{L^2} by ∥∇u∥L2\|\nabla u\|_{L^2} on H01(Ω)H_0^1(\Omega). Consequently, aa is coercive:

a(u,u)≥c∥u∥H12a(u,u)\geq c\|u\|_{H^1}^2

for a positive constant cc. The Lax–Milgram theorem, a basic result about bilinear forms on a Hilbert space, then supplies a unique weak solution and an estimate

∥u∥H01≤C∥f∥H−1.\|u\|_{H_0^1}\leq C\|f\|_{H^{-1}}.

No continuity of the leading coefficients is required for this conclusion.

When AA is symmetric, the same solution minimizes the energy

E[v]=12∫ΩA∇v⋅∇v dx−⟨f,v⟩.E[v]=\frac12\int_\Omega A\nabla v\cdot\nabla v\,dx-\langle f,v\rangle.

This connects elliptic equations with the calculus of variations: the weak equation is the stationarity condition for the functional, and coercivity prevents minimizing sequences from escaping to infinity in the energy norm. Weak compactness and lower semicontinuity then provide the mechanism behind existence.

Nonsymmetric coefficients still fit the Lax–Milgram framework, although the displayed energy no longer generates the entire operator. More general lower-order terms can be treated through coercivity estimates, compact perturbations, or a Fredholm alternative. Nonzero Dirichlet data are handled by subtracting an extension of the prescribed boundary trace. Neumann problems instead use a space modulo constants, together with the appropriate compatibility condition.

Maximum principles and qualitative behavior

For scalar second-order elliptic equations, maximum principles give information without requiring an explicit solution. If uu satisfies

Δu≥0\Delta u\geq0

in a bounded domain and is continuous on its closure, its maximum cannot exceed its maximum on the boundary. For a general operator written as

Lu=aijuij+biui+cu,\mathcal Lu=a^{ij}u_{ij}+b^iu_i+cu,

a standard sufficient sign assumption is c≤0c\leq0. Under appropriate coefficient and regularity hypotheses, Lu≥0\mathcal Lu\geq0 implies

sup⁡Ωu≤sup⁡∂Ωu+,u+=max⁡(u,0).\sup_\Omega u\leq\sup_{\partial\Omega}u^+, \qquad u^+=\max(u,0).

The zero-order sign matters: without it, interior extrema can occur.

The strong maximum principle sharpens this statement. A harmonic function attaining an interior maximum on a connected domain must be constant. More generally, under the usual hypotheses and c≤0c\leq0, an interior nonnegative maximum of a subsolution forces constancy. Applying such principles to differences of solutions yields comparison results and uniqueness. They also preserve physical sign information: with zero Dirichlet data, a solution of −Δu=f≥0-\Delta u=f\geq0 is nonnegative.

The Hopf boundary point lemma describes a nonconstant solution near a strict boundary extremum. Under suitable boundary geometry, such as an interior sphere condition, the outward normal derivative at a boundary maximum is strictly positive for the corresponding subharmonic convention. This rules out excessively flat contact with the extremal boundary value. It is useful in free-boundary problems and symmetry arguments, but cannot be applied indiscriminately at corners or to boundaries without the required geometry.

A second major tool is the Harnack inequality. For a nonnegative solution of a homogeneous uniformly elliptic equation on B2rB_{2r}, a typical interior estimate is

sup⁡Bru≤Cinf⁡Bru.\sup_{B_r}u\leq C\inf_{B_r}u.

The constant depends on the dimension and ellipticity bounds, and in more general settings on the lower-order coefficients. This estimate quantitatively links values at separated points. In particular, a nonnegative nontrivial solution is strictly positive throughout a connected domain. Successive applications across overlapping balls control solutions over larger compact subsets.

For the Laplace operator, these principles connect with the mean-value property: a harmonic function equals its average over every sphere or ball centered at the point and contained in the domain. General variable-coefficient equations do not satisfy that same Euclidean mean-value formula. They nevertheless retain related comparison and oscillation-decay properties. Such scalar order structures are exceptionally powerful, but they do not automatically extend to vector-valued elliptic systems.

Regularity and the limits of smoothing

Regularity theory asks how smooth a weak solution must be once the coefficients, source, and boundary are specified. Ellipticity supplies estimates that recover derivatives of the solution from information about the equation. Interior estimates are local and avoid boundary assumptions; global estimates must also account for the boundary’s smoothness and the compatibility of the prescribed data. A useful distinction is between divergence-form equations, whose weak formulation directly uses first derivatives, and nondivergence-form equations, where second derivatives remain explicit.

For uniformly elliptic operators with Hölder-continuous leading coefficients and f∈C0,αf\in C^{0,\alpha}, the Schauder estimates yield interior C2,αC^{2,\alpha} regularity. A representative estimate on concentric balls is

∥u∥C2,α(Br)≤C(∥u∥C0(B2r)+∥f∥C0,α(B2r)).\|u\|_{C^{2,\alpha}(B_r)} \leq C\left( \|u\|_{C^0(B_{2r})} +\|f\|_{C^{0,\alpha}(B_{2r})} \right).

The constant depends on the radii, ellipticity, and coefficient norms. For divergence-form equations, Hölder-continuous coefficients and suitably structured data naturally lead first to C1,αC^{1,\alpha} estimates. Repeated differentiation and estimation can then establish higher regularity when the coefficients permit it.

A complementary theory uses LpL^p and Sobolev norms. For the Laplacian, one has estimates of the form

∥u∥W2,p(Ω′)≤C(∥f∥Lp(Ω)+∥u∥Lp(Ω)),Ω′⋐Ω,1<p<∞.\|u\|_{W^{2,p}(\Omega')} \leq C\left( \|f\|_{L^p(\Omega)} +\|u\|_{L^p(\Omega)} \right), \qquad \Omega'\Subset\Omega,\quad 1<p<\infty.

Analogous results hold for variable-coefficient operators under suitable coefficient assumptions. Fourier transforms and singular-integral estimates explain the constant-coefficient case: formally, solving Poisson’s equation divides by ∣ξ∣2|\xi|^2, compensating for two derivatives at high frequency. Low frequencies and boundaries require separate treatment.

The De Giorgi–Nash–Moser theory shows that a scalar weak solution of a homogeneous divergence-form equation with merely bounded measurable uniformly elliptic coefficients is locally Hölder continuous. This was striking because the coefficients themselves may be discontinuous everywhere. The conclusion does not imply continuous first derivatives. Across an interface between two conductors, for example, the potential may remain continuous while its normal derivative jumps, with conormal flux remaining balanced.

Boundary geometry can obstruct regularity even for −Δ-\Delta. Near the vertex of a planar sector of opening angle ω\omega, the harmonic function

u(r,θ)=rπ/ωsin⁡(πθω)u(r,\theta)=r^{\pi/\omega} \sin\left(\frac{\pi\theta}{\omega}\right)

vanishes on the two sides. If ω>π\omega>\pi, its gradient is singular at the vertex. Thus smooth coefficients and simple boundary values do not guarantee globally smooth solutions. Elliptic systems also admit failures of scalar regularity conclusions in higher dimensions. For smooth elliptic coefficients, distributional solutions are smooth wherever the forcing is smooth; real analyticity generally requires corresponding analyticity assumptions on the equation.

Nonlinear and degenerate elliptic equations

Nonlinear elliptic equations are commonly divided into semilinear, quasilinear, and fully nonlinear classes. A semilinear equation is linear in its highest derivatives, for example

−Δu=g(x,u).-\Delta u=g(x,u).

In a quasilinear equation, the highest derivatives occur linearly but their coefficients may depend on uu and ∇u\nabla u. Fully nonlinear equations can depend nonlinearly on the entire Hessian matrix D2uD^2u. This distinction determines which linear estimates remain available after linearization.

For a fully nonlinear equation

F(D2u,∇u,u,x)=0,F(D^2u,\nabla u,u,x)=0,

ellipticity expresses monotonicity with respect to the Hessian. With the convention that FF decreases as the Hessian increases, uniform ellipticity means

λ tr⁡N≤F(M,p,z,x)−F(M+N,p,z,x)≤Λ tr⁡N\lambda\,\operatorname{tr}N \leq F(M,p,z,x)-F(M+N,p,z,x) \leq\Lambda\,\operatorname{tr}N

for every positive semidefinite matrix NN. This is the nonlinear counterpart of bounding the eigenvalues of a coefficient matrix. Some authors use the opposite sign convention.

The minimal surface equation for a graph is

∇⋅(∇u1+∣∇u∣2)=0.\nabla\cdot\left( \frac{\nabla u}{\sqrt{1+|\nabla u|^2}} \right)=0.

It is the Euler–Lagrange equation for graph area. Its ellipticity is controlled on sets where the gradient is bounded, but deteriorates as the gradient grows. Another important example is the pp-Laplacian:

Δpu=∇⋅(∣∇u∣p−2∇u),p>1.\Delta_pu=\nabla\cdot\left( |\nabla u|^{p-2}\nabla u \right),\qquad p>1.

It arises from minimizing ∫∣∇u∣p dx\int|\nabla u|^p\,dx. For p>2p>2, degeneracy occurs at zero gradient; for 1<p<21<p<2, the coefficient expression becomes singular there. Monotonicity and nonlinear energy estimates replace parts of uniformly elliptic linear theory.

The Monge–Ampère equation,

det⁡D2u=f,\det D^2u=f,

is fully nonlinear and elliptic on an appropriate convex branch. Convexity is essential: the determinant equation is not elliptic for arbitrary Hessians. For a positive definite Hessian, its linearization involves the positive definite cofactor matrix. The equation appears in prescribed-curvature problems and in the transport of one density into another by a gradient map.

Nondivergence-form nonlinear equations often use viscosity solutions, defined by comparing a continuous function with smooth test functions touching it from above or below. This formulation can remain meaningful when neither classical second derivatives nor a variational weak formulation is available. Comparison principles underpin uniqueness. Under convexity or concavity assumptions on the Hessian dependence, together with uniform ellipticity and suitable data regularity, the Evans–Krylov theorem yields interior C2,αC^{2,\alpha} estimates. Without these structural assumptions, comparable regularity cannot be presumed.

Spectral theory, geometry, and probability

On a bounded domain with suitable boundary conditions, elliptic operators connect boundary value problems to spectral theory. For the Dirichlet Laplacian, the eigenvalue problem is

−Δϕk=λkϕk,ϕk∣∂Ω=0.-\Delta\phi_k=\lambda_k\phi_k, \qquad \phi_k|_{\partial\Omega}=0.

The eigenvalues form a discrete sequence

0<λ1≤λ2≤⋯ ,λk→∞,0<\lambda_1\leq\lambda_2\leq\cdots, \qquad \lambda_k\to\infty,

counted with multiplicity, and the eigenfunctions form an orthonormal basis of L2(Ω)L^2(\Omega). Compactness of the inverse operator on L2L^2, obtained through Sobolev compactness, is the key functional-analytic mechanism.

The first eigenvalue has the variational characterization

λ1=inf⁡0≠v∈H01(Ω)∫Ω∣∇v∣2 dx∫Ω∣v∣2 dx.\lambda_1 =\inf_{0\neq v\in H_0^1(\Omega)} \frac{\int_\Omega|\nabla v|^2\,dx} {\int_\Omega|v|^2\,dx}.

This connects eigenvalues with energy minimization and the best Poincaré constant. Under standard geometric assumptions, Weyl’s law describes the high-frequency counting function:

N(λ)∼ωn(2π)n∣Ω∣λn/2,N(\lambda)\sim \frac{\omega_n}{(2\pi)^n}|\Omega|\lambda^{n/2},

where ωn\omega_n is the volume of the unit ball. The same eigenfunctions describe heat diffusion through factors e−λkte^{-\lambda_k t} and oscillations through wave frequencies λk\sqrt{\lambda_k}.

On a Riemannian manifold, the Laplace–Beltrami operator takes the coordinate form

Δgu=1∣g∣∂i(∣g∣ gij∂ju).\Delta_g u =\frac{1}{\sqrt{|g|}} \partial_i\left(\sqrt{|g|}\,g^{ij}\partial_j u\right).

The positive definite metric tensor makes this operator elliptic. Harmonic functions, differential forms, and geometric deformation problems are consequently governed by elliptic analysis. On a compact connected manifold without boundary, harmonic scalar functions are constant. The broader theory of elliptic operators also links analysis with topology, notably through finite-dimensional kernels and cokernels and through index formulas.

A probabilistic representation gives another view of global boundary dependence. Let BtB_t be Brownian motion and τ\tau its first exit time from a bounded domain. For continuous boundary data on a sufficiently regular domain, the harmonic extension satisfies

u(x)=Ex[g(Bτ)].u(x)=\mathbb E_x[g(B_\tau)].

Here the expected value averages boundary data against the exit distribution, called harmonic measure. Brownian motion has generator 12Δ\tfrac12\Delta; more general nondegenerate diffusion processes correspond to variable-coefficient elliptic operators.

A Green’s function expresses the same inverse operator analytically. With homogeneous Dirichlet conditions, a solution of −Δu=f-\Delta u=f can be represented, under suitable hypotheses, by

u(x)=∫ΩG(x,y)f(y) dy.u(x)=\int_\Omega G(x,y)f(y)\,dy.

For n≥3n\geq3, the leading interior singularity is proportional to ∣x−y∣2−n|x-y|^{2-n}; in two dimensions it is logarithmic. These kernels reveal both the singular response to a concentrated source and the influence of the surrounding boundary.

Numerical approximation and computational challenges

Numerical methods replace an elliptic boundary value problem by a finite-dimensional algebraic system. On a square grid of spacing hh, the finite difference method approximates the two-dimensional negative Laplacian by

−Δhui,j=4ui,j−ui+1,j−ui−1,j−ui,j+1−ui,j−1h2.-\Delta_hu_{i,j} =\frac{ 4u_{i,j}-u_{i+1,j}-u_{i-1,j} -u_{i,j+1}-u_{i,j-1} }{h^2}.

For sufficiently smooth solutions, this five-point stencil has second-order truncation error. Consistency alone does not establish convergence: stability and correct boundary treatment are also required. On suitable meshes, the resulting matrix has an inverse-positivity structure that mirrors a discrete maximum principle.

The finite element method begins with the weak formulation and restricts it to a finite-dimensional space Vh⊂H01(Ω)V_h\subset H_0^1(\Omega). With basis functions {φj}\{\varphi_j\}, the stiffness matrix has entries

Kij=∫ΩA∇φj⋅∇φi dx.K_{ij}=\int_\Omega A\nabla\varphi_j\cdot\nabla\varphi_i\,dx.

For symmetric positive definite diffusion and homogeneous Dirichlet conditions, KK is symmetric positive definite. The method handles irregular domains, local mesh refinement, and discontinuous material coefficients naturally. It also preserves the variational meaning of the continuous problem.

For a bounded coercive bilinear form, Céa’s lemma bounds the error by the best available approximation:

∥u−uh∥H1≤Cinf⁡vh∈Vh∥u−vh∥H1.\|u-u_h\|_{H^1} \leq C\inf_{v_h\in V_h}\|u-v_h\|_{H^1}.

With piecewise linear elements on suitable meshes and u∈H2u\in H^2, the energy-norm error is typically O(h)O(h). Under additional dual regularity, the L2L^2 error is typically O(h2)O(h^2). These rates can deteriorate at reentrant corners or coefficient interfaces. Adaptive refinement uses computable residual or flux indicators to concentrate resolution near singularities rather than uniformly increasing every part of the mesh.

Discretization produces a large sparse matrix, but sparsity does not remove conditioning problems. For standard Dirichlet Laplacian discretizations on quasi-uniform meshes, the spectral condition number typically grows like h−2h^{-2}. Large coefficient contrasts and anisotropy can worsen iterative convergence. The conjugate-gradient method exploits symmetry and positive definiteness, while preconditioners seek to make the transformed system better conditioned.

Multigrid methods address errors on several spatial scales: local relaxation damps oscillatory components, while coarse-grid corrections remove slowly varying components. Under suitable assumptions, they achieve work proportional to the number of unknowns for a fixed error reduction. Domain-decomposition methods provide related tools for parallel computation and heterogeneous media. Fully nonlinear problems introduce additional requirements: monotone discretizations, discrete comparison principles, and suitable consistency and stability conditions can establish convergence to viscosity solutions when classical smoothness is unavailable.