From local energy to global divergence equations

A divergence equation can be meaningful before its solution has even one square-integrable derivative. Its positive energy then supplies that derivative, and local elliptic regularity supplies the next one. Global existence has a different difficulty: data may grow arbitrarily towards infinity or the boundary of an open set. We solve that difficulty by assigning a sufficiently large weight to the adjoint residual, with controlled changes on each previously treated compact region.

The argument permits complex coefficients and uses only local Lipschitz bounds. On a manifold it requires a density to specify divergence. We make that convention explicit and separate the existence theorem on noncompact components from the exact obstruction on compact components. No boundary condition or global integrability of the solution is imposed.

1. The equation and the tools used below

Let $X$ be a Hausdorff, second-countable $C^2$ manifold without boundary of dimension $n\geq1$. An arbitrary open subset of $\mathbb R^n$ is included. Fix a positive $C^1$ density $d\mu$ and a complex symmetric contravariant tensor $A$, with locally Lipschitz coordinate entries. Symmetric means transpose-symmetric, $a^{jk}=a^{kj}$, and does not mean Hermitian. We require

$$\operatorname{Re}\sum_{j,k}a^{jk}(x)\xi_k\xi_j>0 \quad(0\ne\xi\in\mathbb R^n). \tag{D1}$$

Continuity and compactness make the lower bound uniform on each compact coordinate neighborhood. Since the real and imaginary parts of $A$ are real symmetric matrices, (D1) also gives

$$\operatorname{Re}\sum_{j,k}a^{jk}z_k\overline{z_j} \geq \lambda\sum_j|z_j|^2\quad(z\in\mathbb C^n) \tag{D2}$$

on that neighborhood. Indeed the imaginary symmetric part contributes a purely imaginary number, and the real part acts separately on the real and imaginary parts of $z$.

In a chart write $d\mu=\rho(x)\,dx$. Our operator and its Hilbert adjoint on compactly supported tests are

$$\begin{split} Lu&=-\rho^{-1}\partial_j(\rho a^{jk}\partial_k u),\\ L^*v&=-\rho^{-1}\partial_j(\rho\overline{a^{jk}}\partial_k v). \end{split} \tag{D3}$$

Repeated indices in this lesson are summed from $1$ to $n$. The pairing $(u,v)_\mu=\int u\overline v\,d\mu$ is linear in its first entry. For Euclidean Lebesgue density, (D3) is $D_j(a^{jk}D_k u)$, with $D_j=-i\partial_j$, as in Local inverses and distance-weighted elliptic estimates.

For $u\in L^2_{\mathrm{loc}}$, the equation $Lu=f$ means

$$(u,L^*\phi)_\mu=(f,\phi)_\mu \quad(\phi\in C_c^2(X)). \tag{D4}$$

This pairing is defined: $L^*\phi$ is a bounded, compactly supported measurable function. In each chart (D4) agrees with distributional differentiation. More explicitly, for a Lipschitz multiplier $b$, define

$$b\partial_k u=\partial_k(bu)- (\partial_k b)u. \tag{D5}$$

The right side belongs locally to $H^{-1}$. Multiplication by $\rho^{-1}$ and changes of charts in (D3) can be interpreted through (D4); no product of an arbitrary second-order distribution by a merely $C^1$ function is being postulated. The equivalent chart equation is $-\partial_j(\rho a^{jk}\partial_k u)=\rho f$.

There is no need to choose an unmentioned smooth structure. Compactly supported $C^2$ tests suffice for an operator of order two acting on $L^2_{\mathrm{loc}}$. In a coordinate chart, approximation in $H^2$ shows that an equation against smooth tests also holds against these $C^2$ tests. Finite partitions of unity make the equivalence global. The coordinate changes are $C^2$; their first derivatives and inverse first derivatives are bounded on compact subcharts, and their second derivatives are bounded there. The ordinary chain rule, first on smooth functions and then by Sobolev approximation, therefore preserves local $H^1$ and $H^2$. The transformation of $A$ uses first derivatives of the coordinate map, so its new entries remain locally Lipschitz. The density has the usual absolute-Jacobian transformation. These facts make (D3) and (D4) intrinsic, including on a nonorientable manifold.

The following dependencies specify the base of the proofs.

• Weak Lipschitz products. Section 7 of Local inverses and distance-weighted elliptic estimates gives the weak Lipschitz product (D5) and the Friedrichs commutator bound, including convergence, on $L^2$.
• Weak elliptic regularity. Section 8 of Local inverses and distance-weighted elliptic estimates proves that for a second-order elliptic principal operator with locally Lipschitz complex coefficients, $u\in H^1_{\mathrm{loc}}$ and $P_2u\in L^2_{\mathrm{loc}}$ imply $u\in H^2_{\mathrm{loc}}$, with the local estimates proved there. Its quantitative estimate comes from Section 6 of Local inverses and distance-weighted elliptic estimates.
• Support continuation. Sections 7–8 of Curved weights and the directions in which support can end prove continuation from a nonempty open subset of a connected domain for an $H^1_{\mathrm{loc}}$ solution satisfying $|P_2u|\leq C(|u|+|Du|)$, when every normal is admissible. Positive real part gives this admissibility also in dimension two: the path $\operatorname{Re}p+it\operatorname{Im}p$, $0\leq t\leq1$, remains elliptic and preserves the number of roots in each half-plane. No reality assumption at a point is needed in this positive-real-part case.
• Sobolev and compactness facts. The local $L^2$ and weak-derivative calculus, smooth approximation in integer Sobolev spaces, boundedness of convolution on $H^{-1}(\mathbb R^n)$, weak subsequential compactness of bounded Hilbert-space sequences, and compactness of bounded $H^1$ sequences in $L^2$, or bounded $H^2$ sequences in $H^1$, when their supports lie in one fixed compact interior region. The compactness statements transfer through a finite collection of charts. A weakly differentiable function with zero gradient on a connected open set is almost everywhere constant.
• Manifold geometry. Locally finite $C^2$ partitions of unity and compactly supported $C^2$ cutoffs on second-countable $C^2$ manifolds; a positive $C^1$ density and a $C^1$ auxiliary Riemannian metric exist. Only their local coordinate bounds are used. A manifold is locally path connected and locally compact; its connected components are open, and each connected component is path connected. The compact-hull construction needed below is proved in this lesson.
• Hilbert-space tools. The complex Hahn–Banach theorem for a bounded linear functional on any linear subspace of a normed space, and the Riesz representation theorem for bounded linear functionals on a Hilbert space. The extension in the global proof is generally from an infinite-dimensional subspace; a finite-dimensional extension contract would not suffice. We also use completeness, the Cauchy–Schwarz inequality and the convergent geometric series for an operator of norm less than one.

The Hilbert-space tools have complete proofs in this course. Section 5 of Banach estimates, quotient spaces and compact parameter arguments proves norm-preserving complex Hahn–Banach extension from any linear subspace, including one that is neither closed nor finite dimensional. Section 2.3 of Spectral measures with the original operator domain retained proves Hilbert representation with the linear-first convention; Section 6.5 proves the operator geometric series and both inverse products. These proofs supply exactly the functional extension, represented vector and inverse used below.

The argument uses these stated entry results and the named prior lessons. No global solvability theorem is assumed.

2. Energy recovers the first derivative

Local gain. If $u\in L^2_{\mathrm{loc}}(X)$ and $Lu\in H^{-1}_{\mathrm{loc}}(X)$, then $u\in H^1_{\mathrm{loc}}(X)$.

We first work in a Euclidean chart and multiply the equation by $\rho$. Set $b^{jk}=\rho a^{jk}$ and $Q=-\partial_j(b^{jk}\partial_k)$. The entries of $b$ are Lipschitz and their real part is uniformly positive on any fixed compact subchart. A $C^1$ multiplier acts boundedly on local $H^{-1}$: by duality its action is the product on $H^1$, where the first-order product rule gives the bound. Thus $Qu\in H^{-1}_{\mathrm{loc}}$.

Choose a smooth cutoff $\chi$ supported in the chart. Distributional product rules, using (D5), give

$$Q(\chi u)=\chi Qu -\partial_j\bigl(b^{jk}(\partial_k\chi)u\bigr) -(\partial_j\chi)b^{jk}\partial_k u. \tag{D6}$$

All terms on the right belong to $H^{-1}$ after zero extension. For the last term put $c^k=(\partial_j\chi)b^{jk}$ and write $c^k\partial_k u=\partial_k(c^ku)-(\partial_kc^k)u$. Both $c^ku$ and $(\partial_kc^k)u$ are $L^2$ with fixed compact support. The same observation justifies the product rule itself by local smooth approximation of $u$ in $L^2$.

It therefore suffices to treat $v\in L^2(\mathbb R^n)$ supported in a compact set $K_0$ inside the chart, with $Qv\in H^{-1}(\mathbb R^n)$. Extend the coefficients by a cutoff on a slightly larger chart; they become globally Lipschitz and agree with the original positive matrix on a fixed compact neighborhood $K$ of $K_0$. Positivity of the extension outside $K$ is irrelevant.

Let $J_\epsilon v=v*\eta_\epsilon$, where $\eta\in C_c^\infty$ has integral one. For small $\epsilon$, $v_\epsilon=J_\epsilon v$ is smooth with support in $K$. Define the vector commutators

$$r_{\epsilon,j}=b^{jk}\partial_kv_\epsilon -J_\epsilon(b^{jk}\partial_kv). \qquad \sup_{0<\epsilon<\epsilon_0}\|r_\epsilon\|_2\leq C\|v\|_2. \tag{D7}$$

The bound is exactly Section 7 of Local inverses and distance-weighted elliptic estimates, applied to each entry and derivative; replacing $D_k$ by $\partial_k$ only multiplies by a scalar of modulus one. Hence

$$Qv_\epsilon=J_\epsilon(Qv)-\partial_jr_{\epsilon,j}, \qquad \|Qv_\epsilon\|_{H^{-1}} \leq C\bigl(\|Qv\|_{H^{-1}}+\|v\|_2\bigr). \tag{D8}$$

In particular no derivative of a Lipschitz coefficient of order two has appeared. Choose a fixed box containing $K$. The elementary compact-support Poincaré estimate $\|w\|_2\leq C_K\|\nabla w\|_2$ follows by integrating $w$ along lines parallel to one coordinate axis from a face of that box, where its zero extension vanishes, and applying Cauchy–Schwarz. Integration by parts and (D2) now give

$$\begin{split} \|v_\epsilon\|_{H^1}^2 &\leq C\|\nabla v_\epsilon\|_2^2 \leq C\operatorname{Re}(Qv_\epsilon,v_\epsilon)_{dx}\\ &\leq C\|Qv_\epsilon\|_{H^{-1}}\|v_\epsilon\|_{H^1}. \end{split} \tag{D9}$$

If the last $H^1$ norm is zero there is nothing to divide by; otherwise cancellation gives a uniform $H^1$ bound. A sequence $\epsilon\downarrow0$ has a weak $H^1$ limit. Its strong $L^2$ limit is $v$, so uniqueness of distributional limits identifies the weak limit with $v$. This proves $v\in H^1$, and (D6) proves the local assertion. The argument also yields, for nested compact subcharts,

$$\|u\|_{H^1(K_0)}\leq C\bigl(\|Lu\|_{H^{-1}(V)} +\|u\|_{L^2(V)}\bigr), \tag{D10}$$

where $K_0\Subset V$ and the local negative norm is taken after a fixed cutoff equal to one near $K_0$. A finite chart cover supplies the same statement on $X$. $\square$

3. The second derivative and the local estimate

Regularity theorem. Under (D1)–(D3), if $u,f\in L^2_{\mathrm{loc}}(X)$ satisfy (D4), then $u\in H^2_{\mathrm{loc}}(X)$.

By Section 2, $u\in H^1_{\mathrm{loc}}$. In a chart the weak product rule gives

$$-a^{jk}\partial_j\partial_k u =f+\bigl(\partial_j a^{jk} +a^{jk}\partial_j\log\rho\bigr)\partial_k u. \tag{D11}$$

The parenthesized coefficients are locally bounded. Its right side is now $L^2_{\mathrm{loc}}$, and the left side is the weak nondivergence principal operator of Section 8 of Local inverses and distance-weighted elliptic estimates, with $m=p=2$. The real part of its principal symbol is positive in the sense (D1), so it is elliptic. That result gives $H^2_{\mathrm{loc}}$, with no differentiability of the first-order coefficients beyond boundedness. Its quantitative local estimate yields

$$\|u\|_{H^2(K)}\leq C_{K,V} \bigl(\|Lu\|_{L^2(V)}+\|u\|_{H^1(V)}\bigr), \qquad K\Subset V\Subset X, \tag{D12}$$

using finitely many nested coordinate neighborhoods when necessary. Here and below Sobolev norms on relatively compact regions may be defined using any fixed finite chart partition; the choices give equivalent norms. Applied after (D10) on a slightly larger region, (D12) also gives the familiar interior estimate with $\|u\|_{L^2}$ on that larger region. The proof has established finiteness of $H^1$ before using (D11), and finiteness of $H^2$ before applying any estimate to its highest derivatives. $\square$

There is a useful support-preserving approximation consequence. If $u\in H^2_{\mathrm{loc}}(X)$ has compact support $K$ inside an open set $V$, then there are $u_\nu\in C_c^2(V)$, supported in one fixed compact subset of $V$, with

$$u_\nu\longrightarrow u\text{ in }H^2, \qquad Lu_\nu\longrightarrow Lu\text{ in }L^2. \tag{D13}$$

To construct them, choose a finite $C^2$ partition near $K$, subordinate to coordinate charts compactly contained in $V$. Each partitioned function has compact support in its chart. Convolve its zero extension there and take the sum after returning to $X$. The margin between the supports and chart boundaries keeps all approximants in $V$. Chartwise mollification converges in $H^2$, and multiplication by a $C^2$ cutoff is bounded on $H^2$. Finally (D3) expanded as in (D11) defines a bounded map from $H^2$ to $L^2$ on that fixed compact region. This proves the second convergence. The same statement holds with $L^*$.

4. Fixed-support coercivity and continuation

For the next three sections assume that every connected component of $X$ is noncompact. No uniform ellipticity constant on all of $X$ is required.

For a fixed compact $K\subset X$, choose a relatively compact neighborhood $V$ of $K$. Every $v\in H^1(X)$ supported in $K$ satisfies

$$\|v\|_{H^1}\leq C_K\|dv\|_{L^2}. \tag{D14}$$

Here $|dv|$ is measured by an auxiliary $C^1$ metric and the norm can be computed within $V$. For a proof, failure would give supported $v_\nu$ with $\|v_\nu\|_2=1$ and $\|dv_\nu\|_2\to0$. This sequence is bounded in $H^1$. Compactness on a finite chart cover gives a strongly $L^2$ convergent subsequence with limit $v$, of norm one, supported in $K$, and with weak gradient zero. On every connected component it is constant. A component has points outside $K$, since it is noncompact; near such a point the supported function is zero. Thus every constant is zero, a contradiction. A compact set meets only finitely many components because those components form an open cover of it. This also justifies the finite chart argument if $X$ is disconnected.

The energy identity extends by $H^1$ approximation from tests to supported $H^1$ functions:

$$\operatorname{Re}(L^*v,v)_\mu =\operatorname{Re}\int\overline{a^{jk}}\partial_kv \overline{\partial_jv}\,d\mu \geq c_K\|dv\|_2^2. \tag{D15}$$

When the first entry is in $H^{-1}$, the pairing in (D15) denotes its dual action. Combining (D14) and (D15), and defining the negative norm by duality with $H_0^1(V)$, gives

$$\|v\|_{H^1}\leq C_K\|L^*v\|_{H^{-1}(V)} \leq C_K'\|L^*v\|_2. \tag{D16}$$

The last inequality is used only when $L^*v\in L^2$. For compactly supported $C^2$ tests, (D12) and (D16) imply

$$\|v\|_{H^2}\leq C_K''\|L^*v\|_2, \qquad\operatorname{supp}v\subset K. \tag{D17}$$

The constants depend on a compact neighborhood of $K$, its coefficient bounds, and the fixed norm conventions. Formula (D17) is precisely the compactness bound needed in the global construction.

The second ingredient is support propagation. If $w\in H^1_{\mathrm{loc}}(U)$, $U\subset X$ is connected and open, and $L^*w=0$, then vanishing on a nonempty open subset of $U$ implies $w=0$ throughout $U$. Indeed (D11), conjugating the principal coefficients, gives $|P_2w|\leq C|Dw|$ on every compact subchart. Its principal form has positive real part. Section 8 of Curved weights and the directions in which support can end verifies every normal using the positive-real-part path, and Section 7 of Curved weights and the directions in which support can end gives continuation. These are local statements, so coordinate neighborhoods propagate the conclusion along any path in $U$. This checks the complex coefficients and the density term in the exact use of unique continuation.

5. Filling the bounded complementary components

For a compact set $K\subset X$, define its filled hull by adjoining every connected component of $X\setminus K$ whose closure in $X$ is compact:

$$\mathcal H(K)=K\ \cup\! \bigcup_{\substack{C\text{ component of }X\setminus K\\ \overline C\text{ compact}}} C. \tag{D18}$$

We retain the assumption that $X$ has no compact connected components. Then $\mathcal H(K)$ is compact, and $X\setminus\mathcal H(K)$ has no relatively compact connected component. We prove compactness rather than impose it as a further geometric condition.

Choose a relatively compact open neighborhood $U$ of $K$, with $K\subset U$. Such a neighborhood is a finite union of precompact chart neighborhoods. Its boundary is compact and disjoint from $K$. Cover $\partial U$ by finitely many connected open neighborhoods $W_1,\ldots,W_r$ disjoint from $K$. Each $W_i$ lies in one component of $X\setminus K$, so at most $r$ components of that complement meet $\partial U$.

If a component $C$ is relatively compact, the ambient connected component $Y$ of $X$ containing it must meet $K$. Otherwise $C=Y$, contradicting the noncompactness of $Y$. If also $C\not\subset U$, choose $x\in C\setminus U$ and join $x$ to a point of $K\cap Y$ by a path in $Y$. Up to its first meeting with $K$, the path remains in $C$. It must cross $\partial U$ before meeting $K$, since $K\subset U$. Thus $C$ is among the finitely many components just found. Consequently every component added in (D18) is contained either in $U$ or in one of finitely many relatively compact components. Their closures together with $\overline U$ form a compact set containing $\mathcal H(K)$.

The hull is closed: its complement is the union of the other components of $X\setminus K$, each open by local connectedness. Therefore the hull is compact. Those remaining components are exactly the components of its complement and none is relatively compact, by their selection. Notice that $K$ need not have a smooth boundary, finitely many boundary components, or even nonempty interior.

We obtain a sequence of compact sets

$$K_j\subset\operatorname{int}K_{j+1},\qquad X=\bigcup_{j\geq1}K_j,\qquad X\setminus K_j\text{ has no relatively compact component}. \tag{D19}$$

Indeed choose a countable cover by precompact coordinate neighborhoods. Given $K_j$, enclose it and the closures of the first $j+1$ members of the cover in a precompact open set $U_{j+1}$, and set $K_{j+1}=\mathcal H(\overline U_{j+1})$. Start in the same way with the first neighborhood. The hull contains $\overline U_{j+1}$, so $K_j$ lies in its interior. The countable cover proves exhaustion. This construction uses no regular-value theorem and preserves the $C^2$ setting. $\square$

6. Extending an adjoint weight without disturbing the interior

Fix $f\in L^2_{\mathrm{loc}}(X)$ and an exhaustion (D19). Write $\mathcal T(K)$ for the $C_c^2(X)$ functions supported in $K$. Suppose $j\geq3$ and a continuous positive function $M$ on $X$ satisfies

$$|(f,\phi)_\mu|\leq\|ML^*\phi\|_2 \quad(\phi\in\mathcal T(K_j)). \tag{D20}$$

For every $\epsilon>0$ there is a continuous positive $\widetilde M$ such that

$$\begin{gathered} \widetilde M\geq(1+\epsilon)M\text{ on }X, \qquad \widetilde M=(1+\epsilon)M\text{ on }K_{j-2},\\ |(f,\phi)_\mu|\leq\|\widetilde M L^*\phi\|_2 \quad(\phi\in\mathcal T(K_{j+1})). \end{gathered} \tag{D21}$$

Proof. Choose $\chi\in C^2(X)$, $0\leq\chi\leq1$, equal to zero on a neighborhood of $K_{j-2}$ and equal to one outside $\operatorname{int}K_{j-1}$. Set

$$M_N=(1+\epsilon)M+N\chi,\qquad N\geq1. \tag{D22}$$

If no $M_N$ satisfies the last line of (D21), choose a violating $\phi_N\in\mathcal T(K_{j+1})$. Its pairing with $f$ is nonzero; multiply it by the reciprocal of the conjugate of that pairing to arrange

$$(f,\phi_N)_\mu=1,\qquad \|M_NL^*\phi_N\|_2<1. \tag{D23}$$

The positive minimum of $M$ on $K_{j+1}$ gives a uniform bound for $\|L^*\phi_N\|_2$. Formula (D17) then bounds $\phi_N$ in $H^2$, all with support in that same compact set. Pass to a subsequence weakly convergent in $H^2$ and strongly in $H^1$, with limit $\Phi$. The limit has support in $K_{j+1}$, and $(f,\Phi)_\mu=1$, since $f$ is square integrable there. The bounded coefficient formula for $L^*:H^2\to L^2$ on a compact neighborhood gives weak $L^2$ convergence of $L^*\phi_N$ to $L^*\Phi$. Since $M$ is bounded on that neighborhood, weak lower semicontinuity gives

$$\|(1+\epsilon)ML^*\Phi\|_2\leq1. \tag{D24}$$

On $X\setminus K_{j-1}$, (D22) and (D23) give

$$\|L^*\phi_N\|_{L^2(X\setminus K_{j-1})}\leq N^{-1}. \tag{D25}$$

Thus $L^*\Phi=0$ on $X\setminus K_{j-1}$. Every component $C$ of this open set is not relatively compact. It cannot be contained in the compact set $K_{j+1}$; hence $C\setminus K_{j+1}$ is a nonempty open set. There $\Phi=0$. Continuation from Section 4 gives $\Phi=0$ throughout $C$. This establishes

$$\operatorname{supp}\Phi\subset K_{j-1}. \tag{D26}$$

One must use continuation from an open zero set at this point: $\Phi|_C$ need not have compact support relative to $C$, since its support may approach $\partial C$.

Since $K_{j-1}\subset\operatorname{int}K_j$, approximation (D13) gives $\psi_\nu\in\mathcal T(K_j)$ converging to $\Phi$ in $H^2$. It also gives $ML^*\psi_\nu\to ML^*\Phi$ in $L^2$, because $M$ is bounded on the fixed approximation support. Passing to the limit in (D20) yields

$$1=|(f,\Phi)_\mu|\leq\|ML^*\Phi\|_2 \leq(1+\epsilon)^{-1}, \tag{D27}$$

which is impossible. At least one $M_N$ therefore works; take it as $\widetilde M$. $\square$

The argument explains the two inner compact sets in (D21). The weight remains controlled on $K_{j-2}$; the limiting adjoint solution is supported in $K_{j-1}$; the extra margin to $K_j$ permits approximation in the previously proved estimate.

The weight step when compact components are present. The implication (D20)–(D21) holds on any manifold satisfying (D1)–(D3) with an exhaustion (D19), even if compact connected components are present. A filled exhaustion does not itself exclude them. Each compact connected component $Y$ must lie entirely in $K_1$: otherwise a component of $Y\setminus K_1$ would be a relatively compact component of $X\setminus K_1$, contradicting (D19). There are finitely many such $Y$, because they all meet the compact set $K_1$ and the components of $X$ are open. Their indicator functions $1_Y$ belong to $C_c^2(X)$. Testing (D20) with $1_Y$ shows that $\int_Y f\,d\mu=0$.

For any test $\phi$, replace it by

$$\Pi\phi=\phi-\sum_{Y\text{ compact component}} \left(\frac{1}{\mu(Y)}\int_Y\phi\,d\mu\right)1_Y.$$

For every $r\geq1$, this operation preserves $\mathcal T(K_r)$, the residual $L^*\phi$, and the pairing $(f,\phi)_\mu$. Its image has mean zero on each compact component. The proof of (D14) applies to supported functions in this image: a limit with zero gradient is constant on each component, vanishes on each noncompact component by its support, and vanishes on each compact component by its mean. Therefore (D16)–(D17) hold for these projected functions. In the contradiction argument, project each violating test before imposing (D23). The projection leaves both quantities in (D23) unchanged, and its mean-zero conditions pass to the weak limit. The remaining steps (D24)–(D27) are unchanged; in particular, (D19) supplies the same nonempty open zero set on every complementary component. This proves the weight extension in its full filled-exhaustion setting. It does not construct an initial estimate for data of nonzero mean on a compact component.

7. Global existence on noncompact components

Solvability theorem. Assume (D1)–(D3) and that every connected component of $X$ is noncompact. For every $f\in L^2_{\mathrm{loc}}(X)$ there exist a continuous function $M>0$, a function $g\in L^2(X,d\mu)$ with $\|g\|_2\leq1$, and $u=Mg\in H^2_{\mathrm{loc}}(X)$ satisfying $Lu=f$.

Construction of the weight. By Cauchy–Schwarz and (D16),

$$|(f,\phi)_\mu|\leq\|f\|_{L^2(K_3)}\|\phi\|_2 \leq C_3\|f\|_{L^2(K_3)}\|L^*\phi\|_2 \quad(\phi\in\mathcal T(K_3)). \tag{D28}$$

Thus a positive constant function $M_3$ larger than the last coefficient establishes (D20) for $j=3$. Choose positive numbers $\epsilon_j$, $j\geq3$, with $\sum_j\epsilon_j<\infty$, for example $\epsilon_j=2^{-j}$. Inductively use Section 6 to obtain $M_{j+1}$ from $M_j$. Put

$$P_3=1,\qquad P_j=\prod_{k=3}^{j-1}(1+\epsilon_k), \qquad N_j=P_j^{-1}M_j. \tag{D29}$$

The positive numbers $P_j$ increase to a finite $P$: use $\log(1+t)\leq t$ for $t\geq0$. Moreover $N_{j+1}\geq N_j$ everywhere and $N_{j+1}=N_j$ on $K_{j-2}$. On any fixed compact neighborhood this sequence is therefore eventually stationary. Its limit $N$ is continuous and positive, since locally it is exactly one of the continuous positive functions $N_j$. Set $M=PN$. Equivalently, $M_j\to M$ locally uniformly: after stationarity, only the scalar $P_j$ is changing. In particular $M\geq M_j$, and every compactly supported test belongs to $\mathcal T(K_j)$ for some $j$. We have proved

$$|(f,\phi)_\mu|\leq\|ML^*\phi\|_2 \quad(\phi\in C_c^2(X)). \tag{D30}$$

Dual construction of the solution. Let $E$ be the linear subspace of $L^2(X,d\mu)$ consisting of the vectors $ML^*\phi$, $\phi\in C_c^2(X)$. Define a complex linear functional on it by

$$F(ML^*\phi)=(\phi,f)_\mu. \tag{D31}$$

If two tests give the same vector in $E$, their difference has zero right side in (D30), so (D31) is well-defined. Also $|F(h)|\leq\|h\|_2$. Hahn–Banach extends $F$ to $L^2$ with norm at most one. Riesz representation gives $g\in L^2$, $\|g\|_2\leq1$, such that $F(h)=(h,g)_\mu$. Conjugating (D31) now gives the correct pairing order:

$$(f,\phi)_\mu=(g,ML^*\phi)_\mu=(Mg,L^*\phi)_\mu. \tag{D32}$$

Since $M$ is locally bounded, $u=Mg$ belongs to $L^2_{\mathrm{loc}}$; (D32) is its equation (D4). Section 3 supplies $u\in H^2_{\mathrm{loc}}$. The earlier sentence that differentiation of $M$ was not justified was too weak: the actual recursion gives a twice continuously differentiable weight, as the separately identified strengthening in Section14 proves. Regularity comes from the equation, even though the weight itself was only continuous. $\square$

The theorem gives existence for each datum. It does not assert uniqueness, a globally bounded inverse on unweighted $L^2$, a solution satisfying prescribed boundary values, or a datum-independent choice of $M$. These distinctions matter when this theorem is used as a local ingredient in a boundary or Fredholm problem.

8. The exact compact-component obstruction

The local regularity theorem applies on compact manifolds too. The unrestricted global existence assertion does not: on a compact connected component $Y$, the test $\phi=1$ has compact support in $X$ and $L^*1=0$. Thus a necessary condition is

$$\int_Y f\,d\mu=0. \tag{D33}$$

For the pure divergence operator (D3) this is also sufficient. We include the proof to identify the entire obstruction rather than merely give a counterexample.

Let $H$ be the closed subspace of $H^1(Y)$ with mean zero. On $H$ the gradient norm is equivalent to the full $H^1$ norm. Indeed failure of Poincaré would give a normalized sequence converging strongly in $L^2$ to a constant of mean zero, exactly as in the proof of (D14), a contradiction. Define

$$B(v,w)=\int_Y a^{jk}\partial_kv \overline{\partial_jw}\,d\mu. \tag{D34}$$

Its boundedness follows from the bounded coefficients on $Y$. Its real part satisfies $\operatorname{Re}B(v,v)\geq c\|v\|_H^2$ for some $c>0$. Fix any Hilbert norm on $H$ equivalent to $H^1$, with pairing linear first. Riesz representation in the second entry supplies a bounded operator $T:H\to H$ with $(Tv,w)_H=B(v,w)$. Let $\|T\|\leq C$, where $C\geq c$. For $t=c/C^2$,

$$\|(I-tT)v\|_H^2 \leq(1-2tc+t^2C^2)\|v\|_H^2 =(1-c^2/C^2)\|v\|_H^2. \tag{D35}$$

Thus $I-tT$ has norm less than one (possibly zero), and the geometric series gives an inverse for $T$. The bounded conjugate-linear functional $w\mapsto(f,w)_\mu$ on $H$ is represented by a vector $z\in H$. Solve $Tv=z$; then $B(v,w)=(f,w)_\mu$ for every $w\in H$. Any $H^1$ test differs from an element of $H$ by a constant. Both sides vanish on constants, the right side by (D33). Hence $Lv=f$ on $Y$, and Section 3 gives $v\in H^2(Y)$.

If $Lw=0$ on $Y$, local regularity and compactness give $w\in H^2(Y)$ whenever $w\in L^2(Y)$. Its energy is zero, so (D2) implies $dw=0$; it is constant. The same proof applies to $L^*$. Consequently the adjoint obstruction is exactly the constants on each compact connected component, and the normalized solution there is unique.

We have proved the following complete form for a possibly disconnected manifold:

$$\begin{split} &\exists u\in H^2_{\mathrm{loc}}(X):Lu=f\\ &\hspace{1em}\Longleftrightarrow\quad \int_Y f\,d\mu=0\text{ for every compact connected component }Y. \end{split} \tag{D36}$$

For sufficiency solve separately on compact components by (D34)–(D35), and on noncompact components by Section 7. There are at most countably many components by second countability. Every compact subset meets only finitely many of them, so their componentwise solutions assemble to an $H^2_{\mathrm{loc}}$ function. Necessity was (D33).

For Euclidean open sets no component is compact, so (D36) gives arbitrary-data solvability without an additional condition. For the manifold formulation, the noncompact-component convention or the obstruction in (D36) must be stated. An unqualified assertion of arbitrary-data solvability on every boundaryless manifold would be false, even for the Laplacian. This is a mathematical qualification of that formulation, independent of any unstated convention in a reference.

9. Worked models

Example 1: a complex Lipschitz coefficient and rapidly growing data. On $\mathbb R$, take

$$a(x)=2+|x|+ix,\qquad \rho(x)=e^{\sin x},\qquad f(x)=e^{x^2}(1+i\sin x). \tag{D37}$$

The real part of $a$ is at least two, but $a$ is not differentiable at zero. Define

$$F(x)=\int_0^x\rho(s)f(s)\,ds,\qquad u(x)=-\int_0^x\frac{F(t)}{\rho(t)a(t)}\,dt. \tag{D38}$$

These integrals are finite on every compact interval. Since $F$ is continuously differentiable and $1/(\rho a)$ is locally Lipschitz, $u'=-F/(\rho a)$ is locally Lipschitz and $u\in H^2_{\mathrm{loc}}$. Its flux is $\rho a u'=-F$, so differentiating once gives $-\rho^{-1}(\rho a u')'=f$, including across zero. No pointwise value of $a'(0)$ is required. Arbitrary constants may be added both to $F$ and to $u$, producing the two local homogeneous degrees of freedom. The theorem accommodates the growth of $f$ because it asks for local, rather than global, square integrability.

Example 2: bounded in coordinates need not mean relatively compact. Let $X=\mathbb R^2\setminus\{0\}$ and $K=\{x:1\leq|x|\leq2\}$. Its inner complementary component $0<|x|<1$ is bounded in $\mathbb R^2$ but is not relatively compact in $X$: a sequence tending to the missing origin has no convergent subsequence in $X$. The outer complementary component is also not relatively compact. Thus $\mathcal H(K)=K$ in this manifold. The distinction explains why the support argument is expressed intrinsically, with compact closure in $X$, rather than by a Euclidean size test. For any later compact $K'\subset X$, the inner component has points outside $K'$, as the continuation step requires.

Example 3: mixed compact and noncompact components. Let $X=\mathbb R^2\sqcup\mathbb T^2$, use the standard densities, and choose $A=(1+i)I$ on both components; both torus coordinates have period $2\pi$. Give the plane the datum $f_1(x)=1$, and the torus the datum $f_2(\theta)=\cos\theta_1$. A solution is

$$u_1(x)=-\frac{x_1^2}{2(1+i)},\qquad u_2(\theta)=\frac{\cos\theta_1}{1+i}. \tag{D39}$$

Each is locally $H^2$; direct differentiation verifies the equation. Replacing $f_2$ by $1+\cos\theta_1$ makes its integral nonzero and destroys solvability on that component, while the plane equation remains solvable. A compact-component obstruction is a condition on that component’s datum, not on the behavior at infinity of other components.

10. Problems and full solutions

Problem 1. On the line let $a(x)=2+|x|$ and $u(x)=|x|$. Compute $-\partial_x(a\partial_xu)$ as a distribution. Verify that its local regularity is compatible with Section 2 but does not trigger Section 3.

Solution. Almost everywhere $u'=\operatorname{sgn}x$, and $au'=2\operatorname{sgn}x+x$. Since the distributional derivative of $\operatorname{sgn}x$ is $2\delta_0$,

$$-\partial_x(au')=-4\delta_0-1. \tag{D40}$$

The Dirac mass belongs to $H^{-1}_{\mathrm{loc}}(\mathbb R)$. For instance, for a smooth compactly supported $\psi$, integrate $(|\psi|^2)'$ from an endpoint beyond its support to zero; this gives $|\psi(0)|^2\leq2\|\psi\|_2\|\psi'\|_2\leq\|\psi\|_{H^1}^2$. Evaluation therefore extends continuously to $H^1$. Thus (D40) has the negative regularity needed for the first gain. In fact $|x|\in H^1_{\mathrm{loc}}$. But $u''=2\delta_0$ is not represented by an $L^2$ function, so $u\notin H^2_{\mathrm{loc}}$ near zero. The datum in (D40) is likewise not $L^2_{\mathrm{loc}}$. The two stages of the theorem have different data assumptions, and this example distinguishes them.

Problem 2. Start with $L=-\partial_x^2$, Lebesgue density, and the coordinate change $y=F(x)=x+x^3$. Determine the transformed density and tensor coefficient and check the first-order term of the transformed operator.

Solution. Write $x=G(y)=F^{-1}(y)$. Since $F'=1+3x^2>0$, this is a global smooth coordinate change. The density and coefficient are

$$\rho_y(y)=\frac1{F'(G(y))},\qquad a_y(y)=F'(G(y))^2. \tag{D41}$$

Thus $\rho_y a_y=F'$, with the right side evaluated at $G(y)$. The divergence formula gives

$$-\rho_y^{-1}\partial_y(\rho_y a_y\partial_y) =-F'^2\partial_y^2-F''\partial_y. \tag{D42}$$

Here $\partial_y F'(G(y))=F''/F'$. The same result follows from $\partial_x=F'\partial_y$, applied twice. If the density were omitted and one wrote $-\partial_y(a_y\partial_y)$, the first-order coefficient would instead be $-2F''$, giving a different operator. This is why a divergence formula with coordinate-dependent coefficients needs its density convention.

Problem 3. For a fixed real $\kappa$, consider on $\mathbb R^2$ the form $q(\xi)=(1+i\kappa)\xi_1^2+(2-i\kappa)\xi_2^2$. Check the normal-root condition for $N=e_2$, then explain why it holds for every real normal without choosing roots continuously over all normals.

Solution. For $\xi=e_1$ the line polynomial is $(1+i\kappa)+(2-i\kappa)z^2$. Its two roots are opposite, nonzero numbers. Neither is real, since a real $z$ would give real part $1+2z^2>0$. Hence one lies in each half-plane and the roots are distinct. For arbitrary independent real $\xi,N$, use $q_t=\operatorname{Re}q+it\operatorname{Im}q$, $0\leq t\leq1$. The leading coefficient $q_t(N)$ is nonzero and the line polynomial has no real root throughout the path. At $t=0$ its real coefficients yield a nonreal conjugate pair. Continuity of the unordered pair of roots preserves the number in each half-plane, so the same is true at $t=1$. When $\xi$ is parallel to $N$, the root corresponds to the zero argument and is excluded from the normal criterion. The path argument is pointwise in the chosen pair and requires no global root labeling.

Problem 4. Determine solvability of $-u''=1$ on $\mathbb R$ if the desired solution is (a) locally $H^2$, (b) globally $L^2$. Compare with the same equation on the circle. Explain what the global theorem actually supplies.

Solution. Integrating distributionally twice gives every solution on the line as $u(x)=-x^2/2+cx+d$, where $c,d\in\mathbb C$. These functions are smooth and therefore locally $H^2$. None is globally $L^2$: its quadratic leading coefficient cannot be canceled by the affine part, so for sufficiently large $|x|$, $|u(x)|\geq |x|^2/4$. On the circle, integration against the constant test gives $0=\int1$, an impossibility. Thus arbitrary-data local-Sobolev existence on a noncompact component gives neither unweighted global integrability nor arbitrary-data existence on a compact component. In the weighted representation $u=Mg$, one may have $g\in L^2$ precisely because the allowed $M$ grows; boundedness of $M$ was never asserted.

Problem 5. In the weight construction, replace $\epsilon_j=2^{-j}$ by $\epsilon_j=1/j$, $j\geq3$. Show why the argument no longer produces a finite continuous weight on the previously treated interior. Does this refute the solvability theorem?

Solution. On $K_1$ every step $j\geq3$ satisfies the exact equality $M_{j+1}=(1+1/j)M_j$, since $K_1\subset K_{j-2}$. Hence there

$$M_j=M_3\prod_{k=3}^{j-1}\frac{k+1}{k} =\frac j3M_3. \tag{D43}$$

The initial weight is strictly positive, so this tends to infinity at every point of $K_1$. Normalized weights still stabilize there, but the compensating scalar product does not have a finite limit. The argument therefore needs a bounded product of interior enlargement factors, supplied by summability of the positive $\epsilon_j$. This failure concerns that choice of iterative construction. Choosing the summable sequence used in Section 7 proves the theorem and avoids the divergence.

Problem 6. Explain concretely why one cannot replace the filled exhaustion by an arbitrary compact exhaustion in the continuation step. Use the Laplacian on $\mathbb R^2$ and a compact annulus.

Solution. Take $K=\{1\leq|x|\leq2\}$ and a smooth radial function $\Phi$ equal to one on $|x|\leq1$ and zero on $|x|\geq2$, with its transition strictly between these radii. Then $\operatorname{supp}(-\Delta\Phi)\subset K$. Thus $-\Delta\Phi=0$ on $X\setminus K$, and $\Phi$ has compact support. But $\Phi$ is nonzero on the bounded complementary component $|x|<1$. Continuation does not force zero there, because there is no nonempty open zero set within that connected component. The support conclusion $\operatorname{supp}\Phi\subset K$ would be false. Filling the inner disk replaces $K$ by $\{|x|\leq2\}$, for which the support conclusion is correct. The failure occurs in the support inference, not in local regularity or the energy estimate.

11. The manifold geometry used by the equation

We prove all the geometry used in Section 1 on its original Hausdorff, second-countable $C^2$ manifold without boundary and of dimension $n\geq1$; the notation below retains those original objects. If $X$ is empty, the empty partition has sum one at every point vacuously, the unique empty metric and density have all the stated positivity and regularity properties, every compact subset is empty and has zero cutoff, and there are no components. All the assertions therefore hold in that case. For the remaining construction assume $X$ is nonempty. The proof uses the original coordinate compactness and finite scalar calculus; it does not select a smoother atlas.

11.1. Countable compact exhaustion from the original charts

Every point has a coordinate neighborhood and a smaller coordinate ball whose closed ball is contained in the coordinate image. Its inverse image is compact, since the inverse chart is continuous and the closed Euclidean ball is compact. It is closed in $X$, since $X$ is Hausdorff. Its interior is an open neighborhood of the point. This proves local compactness and gives an open cover by relatively compact coordinate balls.

Here is the countable-subcover argument used in the construction. Fix the given countable topological base. For each base member contained in some member of the cover, choose one such cover member. This gives at most countably many chosen members. Any point of the covered space belongs to a base neighborhood contained in a cover member, so belongs to one of the chosen members. The chosen family covers $X$. This argument applies to every open cover, with no countability assumption on that cover.

Write the resulting relatively compact coordinate balls as $V_1,V_2,\ldots$, allowing a finite list when it is finite. Put $K_0=K_{-1}=\varnothing$. Choose $K_1=\overline V_1$. Given compact $K_j$, cover it by finitely many relatively compact coordinate balls and take the union of their compact closures together with the compact closures of the first $j+1$ members of the original cover. Call this finite union $K_{j+1}$. It is compact; the open balls covering $K_j$ show $K_j\subset\operatorname{int}K_{j+1}$. Including the first $j+1$ original balls makes the interiors exhaust $X$. In a finite cover, include every available member once the index exceeds its length; the same construction works, and may become constant at $X$ if $X$ is compact. Thus $$K_j\subset\operatorname{int}K_{j+1},\qquad X=\bigcup_{j\geq1}\operatorname{int}K_j . \tag{MG1}$$ There is no connectedness requirement. Each compact subset of $X$ lies in the interior of some $K_N$, by a finite subcover of these increasing interiors.

Define the original compact shells and their open neighborhoods by $$A_j=K_j\setminus\operatorname{int}K_{j-1},\qquad W_j=\operatorname{int}K_{j+1}\setminus K_{j-2}. \tag{MG2}$$ The shell is compact, and $A_j\subset W_j$: it lies in $\operatorname{int}K_{j+1}$, while $K_{j-2}\subset\operatorname{int}K_{j-1}$ makes it disjoint from $K_{j-2}$. Every point lies in some shell, by choosing the first index for which it belongs to $K_j$.

11.2. Compactly supported $C^2$ partition subordinate to any open cover

Let $\mathcal O$ be the original open cover. For every point of $A_j$, choose a coordinate ball with closed outer ball inside $W_j$, inside one assigned member $O\in\mathcal O$, and inside its original chart. Choose concentric coordinate balls of radii $r,2r,3r$, with the closed radius-$3r$ ball inside this open intersection. The radius-$r$ balls cover $A_j$; compactness selects finitely many, indexed by $r'\in\{1,\ldots,N_j\}$. Denote the smaller balls by $V_{j,r'}$ and the radius-$3r$ balls by $U_{j,r'}$. The chosen finite families together cover $X$, and $\overline U_{j,r'}\subset W_j\cap O_{j,r'}$. Empty shells need no balls.

This family is locally finite. If $x\in\operatorname{int}K_N$, that same open set is a neighborhood meeting none of the $U_{j,r'}$ with $j\geq N+2$, because their $W_j$’s exclude $K_{j-2}\supset K_N$. Only finitely many shells remain, and each has finitely many balls. This proves local finiteness on the original space, rather than assuming a paracompactness theorem.

To construct the needed coordinate functions, put $$\theta(t)= \begin{cases}e^{-1/t},&t>0,\\0,&t\leq0,\end{cases} \qquad s(t)=\frac{\theta(t)}{\theta(t)+\theta(1-t)} . \tag{MG3}$$ For $t>0$, each derivative of $\theta$ is a finite polynomial in $1/t$ times $e^{-1/t}$, by induction using the product and chain rules. Every such expression tends to zero as $t\downarrow0$: for each positive integer $k$, the exponential series gives $e^u\geq u^k/k!$, and choosing $k$ larger than the polynomial degree proves the limit. Thus extension by zero makes $\theta$ smooth with all derivatives zero at zero. The denominator in (MG3) is positive for every real $t$; $s$ is smooth, equals zero for $t\leq0$, equals one for $t\geq1$, and lies between zero and one.

In the original chart of a selected ball with center $a$ and small radius $r$, define $$\beta_{j,r'}(x) =s\!\left(\frac{(2r)^2-|x-a|^2}{(2r)^2-r^2}\right). \tag{MG4}$$ It equals one on the closed radius-$r$ ball, and is zero outside the radius-$2r$ ball. Its support is compact inside $U_{j,r'}$. Composing with the $C^2$ chart and extending by zero gives a $C^2$ function on $X$: the support has an open neighborhood inside the chart, and the function is identically zero near every point outside that compact support. This retains the manifold’s $C^2$ structure and introduces no unmentioned smoother atlas.

Let $I$ denote these countably many indices and put $$S(x)=\sum_{i\in I}\beta_i(x),\qquad \phi_i(x)=\frac{\beta_i(x)}{S(x)},\qquad \sum_{i\in I}\phi_i(x)=1 . \tag{MG5}$$ Each sum is a finite sum on a neighborhood of each point by local finiteness. The inner balls cover $X$, so $S(x)\geq1$. The quotient is therefore $C^2$. The $\phi_i$’s are nonnegative, their compact supports stay in the original assigned cover members, and their family is locally finite. Their exact derivatives in any original chart are $$\begin{split} \partial_a\phi_i &=\frac{\partial_a\beta_i}{S} -\frac{\beta_i\partial_aS}{S^2},\\ \partial_a\partial_b\phi_i &=\frac{\partial_a\partial_b\beta_i}{S} -\frac{(\partial_a\beta_i)(\partial_bS) +(\partial_b\beta_i)(\partial_aS) +\beta_i\partial_a\partial_bS}{S^2} +\frac{2\beta_i(\partial_aS)(\partial_bS)}{S^3}. \end{split} \tag{MG6}$$ All sums and all denominator factors remain present. On a compact subchart only finitely many supports meet the compact set, so every derivative through order two has a finite supremum there. This proves precisely the subordinate locally finite $C^2$ partition asserted by the original contract.

11.3. The original compact cutoff and positive local-radius minorant

If a compact $K\subset X$ lies in an open $O$, apply the partition construction to the cover $\{O,X\setminus K\}$. A locally finite family meets a compact set in only finitely many members: cover the compact set by finitely many neighborhoods each meeting finitely many supports. Take the finite collection $J$ of supports assigned to $O$ that meet $K$, and define $\chi=\sum_{i\in J}\phi_i$. Then $0\leq\chi\leq1$, $\chi\in C_c^2(O)$, and its support is contained in the finite union of the selected original compact supports.

In fact $\chi=1$ on a neighborhood of $K$, not just on $K$. Every unselected support is disjoint from $K$: a support assigned to $X\setminus K$ is contained there, and every other support disjoint from $K$ was not selected. A locally finite union of closed sets is closed, because near each point that union is a finite closed union. The union of the unselected supports is therefore closed and misses $K$. Its open complement is a neighborhood of $K$; there every unselected $\phi_i$ vanishes, so (MG5) gives $\chi=1$. If $K=\varnothing$, use $\chi=0$. This proves every required compact cutoff.

There is also a useful proved consequence for the local radii used in the continuation and kernel constructions. Let $R:X\to(0,\infty)$ have a positive lower bound on some neighborhood of every point, as does a positive lower-semicontinuous function. Choose such neighborhoods $O_i$, with constants $a_i>0$ for which $R(x)\geq a_i$ on $O_i$, and take the proved subordinate partition, assigning each support its original $a_i$. Then $$r(x)=\sum_i\phi_i(x)\frac{a_i}{2},\qquad 0<r(x)\leq\frac{R(x)}2<R(x). \tag{MG7}$$ The sum is locally finite and $C^2$, and at least one positive summand occurs at every point. Each term active at $x$ has $a_i\leq R(x)$, and summing with (MG5) proves the upper bound. No positive global lower bound or compactness of $X$ is assumed.