Elliptic Regularity

Elliptic regularity is the theorem that weak solutions are smooth. Stated backwards from practice: a PDE is solved in a weak sense because that is what the variational method gives; a function in a Sobolev space satisfying the equation after integration by parts; and the theorem says this weak object is automatically differentiable as many times as the equation’s coefficients and data allow. The Hodge theorem used it as a black box; this page opens the box.

The theorem is also why the word elliptic is a classification. Ellipticity is a condition on the equation’s principal symbol; the coefficient matrix of the highest-order terms being definite; and it is the condition that makes the regularity true. The heat equation smooths in time but not backwards; the wave equation propagates singularities; the Laplace equation smooths everywhere. That ordering is not folklore; each is a theorem, and only the elliptic case is proved here.

A weak solution of $\Delta u = f$ is a function satisfying $\int u\,\Delta\varphi = \int f\varphi$ for every test function; no derivatives of $u$ required. The theorem’s content is that this integration-by-parts identity, holding against all test functions, forces $u$ to have classical derivatives.

The mechanism is mollification. Convolve $u$ with a smooth kernel: the result $\rho_\delta * u$ is smooth, and because $\Delta$ commutes with convolution,

where the middle equality is the weak equation read as an equality of distributions. So the mollified function solves a smooth equation, and the whole theorem is the bookkeeping that the mollified solutions converge back to $u$ while their derivatives stay bounded; bounds inherited from the equation rather than from $u$.

Theorem (Weyl’s lemma). Let $u \in L^1_{\mathrm{loc}}(\Omega)$ satisfy $\int_\Omega u\,\Delta\varphi = 0$ for all $\varphi \in C_c^\infty(\Omega)$. Then $u$ agrees almost everywhere with a smooth harmonic function.

Proof. Mollify: $u_\delta = \rho_\delta * u$ is smooth on $\Omega_\delta = \{x \in \Omega : \operatorname{dist}(x, \partial\Omega) > \delta\}$. The weak equation makes $\Delta u = 0$ as a distribution, and convolution with the smooth kernel $\rho_\delta$ gives

so each $u_\delta$ is a smooth harmonic function on $\Omega_\delta$.

Harmonic functions are smooth, and satisfy the mean-value property. For a smooth harmonic $v$, Green’s identity applied to $v$ and the constant $1$ on a ball $B(x_0, r)$ gives

so the normal derivative integrates to zero on every sphere; differentiating the identity $\int_{\partial B(x_0,r)} v\,dS = v(x_0)\,|\partial B|$-style with respect to $r$; or applying Green’s representation formula with the fundamental solution $\Gamma$; yields the mean-value property:

The mean-value property implies smoothness, with bounds. A function satisfying the mean-value property is smooth: for multi-indices $\alpha$,

with $\psi_\alpha$ bounded and $c_\alpha$ explicit from differentiating the spherical integral; each derivative of $v$ is a bounded linear functional of $v$ on a slightly larger ball, so $\sup_{B(x_0,r)} |\partial^\alpha v| \leq C_{\alpha}\, r^{-|\alpha|} \sup_{B(x_0, 2r)}|v|$.

Applying those bounds to the family $u_\delta$: on any compact $K \subseteq \Omega$, choosing $r$ with $K + B(0,2r) \subseteq \Omega$ gives

uniformly in $\delta$, the last bound by the Young inequality for the convolution. So $(u_\delta)$ is a uniformly bounded, equicontinuous family with uniformly bounded derivatives of every order on $K$; Arzelà–Ascoli extracts a subsequence converging uniformly along with all derivatives, and the limit is smooth on $K$. Covering $\Omega$ by an exhaustion of such $K$ and diagonalising gives a smooth $v$ on all of $\Omega$ with $u_\delta \to v$ in $L^1_{\mathrm{loc}}$; and since also $u_\delta \to u$ in $L^1_{\mathrm{loc}}$ (mollifiers approximate), $u = v$ almost everywhere. $\blacksquare$

The proof is long because it is really three theorems: mollifiers commute with $\Delta$, harmonic smooth functions satisfy the mean-value property, and the mean-value property gives uniform derivative bounds. Each is elementary; their composition is the theorem. Note where locality enters; every bound is on a compact subset with room to spare, which is why the theorem is “interior” regularity: nothing is claimed at the boundary.

Theorem. Let $u \in L^1_{\mathrm{loc}}(\Omega)$ weakly solve $\Delta u = f$ with $f \in C^\infty(\Omega)$. Then $u \in C^\infty(\Omega)$.

Proof. The mollified functions solve $\Delta u_\delta = \rho_\delta * f$ smoothly on $\Omega_\delta$. Split $u_\delta = v_\delta + w_\delta$ where $v_\delta$ solves $\Delta v_\delta = \rho_\delta * f$ with smooth data; Newtonian potential; and $w_\delta$ is harmonic. The Newtonian potential of smooth compactly-supported data is smooth by differentiating under the integral (the singularity of $\Gamma$ is integrable and its derivatives stay integrable away from the diagonal, which the mollifier keeps off). The harmonic parts $w_\delta$ have the uniform derivative bounds of Weyl’s lemma. So both halves have uniformly convergent subsequences on compacts, the limits sum to $u$, and the sum is smooth. $\blacksquare$

The difference-quotient method gives the Sobolev version, which is what the variational pipeline consumes:

Theorem (interior $H^2$ regularity). If $u \in H_0^1(\Omega)$ weakly solves $\Delta u = f$ with $f \in L^2(\Omega)$, then $u \in H^2_{\mathrm{loc}}(\Omega)$ and $\|u\|_{H^2(K)} \leq C(\|f\|_{L^2} + \|u\|_{H^1})$ for every compact $K \subseteq \Omega$.

Proof sketch. Apply the difference quotient $D_k^h u(x) = \frac{u(x + he_k) - u(x)}{h}$ to the weak equation. The test function $\varphi = -D_k^{-h}(D_k^h u\,\psi)$ for a cutoff $\psi$ supported in $\Omega$ is admissible, and after the discrete integration by parts $\int D_k^{-h}a \cdot b = -\int a\, D_k^{h}b$:

using that difference quotients commute with $\Delta$ and that $|D_k^h v(x)| \leq \|\nabla v\|_{L^\infty}$ bounds the cutoff’s derivative terms. Absorbing the common factor (small-$h$ control from the equation’s coercivity) gives uniform $H^1$ bounds on the difference quotients, and the compactness argument of Rellich–Kondrachov extracts a limit in $H^1$; which is the weak second derivative. Interior because the cutoff $\psi$ vanishes at the boundary. $\blacksquare$

The two theorems compose into the pipeline the Sobolev page assembled:

Each iteration gains two derivatives: $H^m \Rightarrow H^{m+2}$ locally, by the difference-quotient argument applied to $\partial^\alpha u$ for $|\alpha| = m$. The Sobolev embedding page 3 of the functional-analysis track then converts the $H^m$ control into classical differentiability at the threshold $m > n/2 + j$; the threshold that is the whole bookkeeping of the subject.

For the Hodge theorem, the same argument applies to the form-Laplacian $dd^* + d^*d$ because it is elliptic: the principal symbol is $\pm|\xi|^2$, definite and nonzero off $\xi = 0$. The harmonic forms are smooth because the operator is elliptic, and the theorem’s “analytic input” is exactly this page. The same applies to the eigenfunctions of the Dirichlet Laplacian on the spectral theory page: kernel elements of an elliptic operator are smooth, which is why “eigenfunction” is meaningful as a classical object.

The principal symbol determines everything. For a second-order operator $\sum a_{ij}\partial_i\partial_j$, ellipticity means the coefficient matrix $\{a_{ij}\}$ is definite; the symbol never vanishes for $\xi \neq 0$, and the difference-quotient estimates close.

• The heat equation $\partial_t u = \Delta u$ is parabolic: smoothing forward in time, none backward. A weak solution with rough initial data becomes smooth for $t > 0$; the forward smoothing is the same mollifier-and-bound technology; and a backward-in-time problem is ill-posed.
• The wave equation $\partial_t^2 u = \Delta u$ is hyperbolic: singularities propagate along characteristics and never smooth. A weak solution stays as rough as its initial data; the wire’s pluck travels.
• Elliptic is the no-time case: the smoothing is unconditional and global, which is why boundary-value problems (the Dirichlet problem) and geometry live here.

The Ricci flow on the geometrisation page is weakly parabolic, which is exactly why its maximum principle works and why Hamilton’s theorem could prove anything; a degenerate-elliptic equation with the smoothing surviving the degeneracy.

• It is interior. Boundary regularity needs the boundary to be regular ($C^{2}$ at least) and the boundary data compatible; a corner singularity stays.
• It needs smooth coefficients and data. With rough $f$, the solution is regular only as $f$ allows: $f \in L^2$ gives $H^2_{\mathrm{loc}}$ and no further. The machine “solve weakly, then regularise” cannot exceed the data.
• It does not make all weak solutions unique. Uniqueness is coercivity - Lax–Milgram; not regularity. Regularity is about the solution’s quality after existence and uniqueness are established.

• Quoting regularity at the boundary. The interior theorems say nothing where $\operatorname{dist}(x, \partial\Omega) \to 0$; boundary regularity is a separate theorem with separate hypotheses.
• Assuming all weak solutions are smooth. Only elliptic equations have the theorem. The wave equation’s weak solutions are exactly as rough as their data, and quoting elliptic regularity there is false.
• Differentiating under the integral without the uniform bound. The proof’s Arzelà–Ascoli step is what licenses the limit’s smoothness; without the uniform derivative bounds the mollified limit is only $L^1$.
• Expecting harmonic functions to be analytic without using the equation. Real analyticity of harmonic functions is a sharper theorem than smoothness and needs elliptic regularity with analytic coefficients; smoothness alone does not self-improve.

• Sobolev Spaces; weak derivatives, the embedding threshold, and Rellich–Kondrachov’s compactness used in §1.
• Hodge Theory; the black box this page opens: smooth harmonic representatives.
• De Rham Cohomology; the cohomology whose classes get the smooth representatives.
• The Fundamental Theorems; Lax–Milgram, which produces the weak solution regularity then promotes.
• Ricci Flow and Geometrisation; the maximum principle, the parabolic sibling of the elliptic theory.
• Spectral Theory of Self-Adjoint Operators; eigenfunctions of the Laplacian, smooth by this page.

• Applications
• Curvature
• De Rham Cohomology
• Differential Forms