Weak Solutions, Sobolev Spaces, and Well-Posedness

A weak solution satisfies an equation after derivatives have been transferred to test functions, in a declared function space that makes every pairing continuous. Sobolev spaces record which distributional derivatives are represented by $L^p$ functions; trace spaces say what boundary data mean. Existence, uniqueness, and stability are then separate statements about a specific operator, domain, data class, and topology. A problem is well-posed only when all three hold.

This page gives a bounded advanced introduction through two controlled linear models: a coercive elliptic boundary-value problem and an energy estimate for the wave equation. It does not attempt a general theory of nonlinear weak solutions, gauge systems, rough domains, or global hyperbolic evolution.

Required background. Test-Function Spaces, Distributions, Support, and Convergence supplies distributional derivatives and test-function pairings; Lp Spaces, Inequalities, and Weak Convergence supplies Hölder-type estimates, completeness, and weak convergence.

Weak derivatives are distributional derivatives with added regularity

Let $\Omega\subset\mathbb R^d$ be open, let $u\in L^1_{\mathrm{loc}}(\Omega)$, and let $\alpha$ be a multi-index. A function $v\in L^1_{\mathrm{loc}}(\Omega)$ is the weak derivative $\partial^\alpha u$ when

$$\int_\Omega u(x)\partial^\alpha\varphi(x)\,\mathrm d^d x = (-1)^{|\alpha|} \int_\Omega v(x)\varphi(x)\,\mathrm d^d x$$

for every $\varphi\in C_c^\infty(\Omega)$. Compact support removes the boundary term. The definition says exactly that the distributional derivative of the regular distribution defined by $u$ is represented by the locally integrable function $v$. Such a representative is unique almost everywhere.

Every distribution has distributional derivatives. It need not have weak derivatives represented by functions in a chosen $L^p$ space. For example, the Heaviside function has distributional derivative $\delta_0$, which is not an $L^p_{\mathrm{loc}}$ function. By contrast, $u(x)=|x|$ on $(-1,1)$ has weak derivative $\operatorname{sgn}x\in L^\infty$, although its second distributional derivative is $2\delta_0$.

This distinction and the integration-by-parts definition are developed in Chen 2013, §1.2, printed pp. 2–5, PDF. Terminology varies: some authors call any distributional solution “weak,” while others reserve “weak solution” for a variational identity in specified Sobolev spaces. This page uses weak derivative only when the distributional derivative has a locally integrable function representative, and always states the spaces in a weak-solution claim. The distributional distinction is developed further in Delta Distributions, Weak Derivatives, Pullbacks, and Pushforwards.

The Sobolev scale

For an integer $k\geq0$ and $1\leq p\leq\infty$,

$$W^{k,p}(\Omega) = \left\{ u\in L^p(\Omega): \partial^\alpha u\in L^p(\Omega) \text{ for }|\alpha|\leq k \right\},$$

where the derivatives are weak derivatives. For $1\leq p<\infty$, one standard norm is

$$\|u\|_{W^{k,p}}^p = \sum_{|\alpha|\leq k} \|\partial^\alpha u\|_{L^p}^p.$$

For $p=\infty$, replace the sum of powers by the maximum of the $L^\infty$ norms. The Hilbert-space case is

$$H^k(\Omega)=W^{k,2}(\Omega).$$

The estimates below repeatedly use Hölder’s inequality with $1/p+1/q=1$ and its $p=q=2$ case, Cauchy–Schwarz. For a continuous functional $F$ on a normed space $V$,

$$\|F\|_{V'} = \sup_{0\neq v\in V} \frac{|F(v)|}{\|v\|_V}, \qquad |F(v)| \leq \|F\|_{V'}\|v\|_V.$$

These are spaces of equivalence classes modulo equality almost everywhere. Membership controls derivatives in an integral norm; it does not generally supply a preferred pointwise representative.

On all of $\mathbb R^d$, real-order Sobolev spaces have a convenient Fourier description. With the site’s positive forward phase,

$$\widetilde u(p) = \int_{\mathbb R^d} \mathrm d^d x\, e^{+ip\cdot x}u(x),$$

define

$$\|u\|_{H^s(\mathbb R^d)}^2 = \int_{\mathbb R^d} \frac{\mathrm d^d p}{(2\pi)^d} \left( 1+|p|^2 \right)^s \left| \widetilde u(p) \right|^2.$$

Then $H^s(\mathbb R^d)$ consists of tempered distributions for which this quantity is finite. For nonnegative integer $s=k$, this norm is equivalent to the weak-derivative $H^k$ norm. The Fourier phase changes the sign of the derivative multiplier but not this norm. Dyatlov, Dyatlov 2022, Definition 12.3 and Eqs. (12.4)–(12.7), PDF, gives the real-order definition and the derivative mapping $\partial_j:H^{s+1}\to H^s$.

Fractional spaces on a domain can be defined by restriction, extension, interpolation, or difference quotients. Those constructions agree only under appropriate domain and parameter hypotheses. The whole-space Fourier formula should not be copied to an arbitrary $\Omega$ without specifying an extension or an intrinsic definition.

Negative-order spaces describe rough sources. On a bounded domain, set

$$H_0^1(\Omega) = \overline{ C_c^\infty(\Omega) }^{\,H^1}, \qquad H^{-1}(\Omega) = \left( H_0^1(\Omega) \right)'.$$

Thus an element of $H^{-1}$ is a continuous linear functional on the energy space $H_0^1$. It need not be an ordinary function.

Boundary values require a trace

An element of $H^1(\Omega)$ is defined only almost everywhere in the open set $\Omega$ and has no a priori boundary values. A phrase such as “$u=0$ on $\partial\Omega$” therefore requires a trace theorem.

For a bounded Lipschitz domain, classical restriction extends to a bounded trace map

$$\operatorname{Tr}: H^1(\Omega) \longrightarrow H^{1/2}(\partial\Omega).$$

It is surjective and has a continuous right inverse; on such a domain,

$$H_0^1(\Omega) = \ker\operatorname{Tr}.$$

Chen, Chen 2013, Theorem 1.19, printed p. 8, PDF, states the trace theorem and the range $H^{s-1/2}(\partial\Omega)$ for $1/2<s<3/2$. McLean 2000, Chapter 3, pp. 100–106 gives a specialist treatment.

This distinction organizes boundary conditions.

• An essential Dirichlet condition is built into the trial space, such as $u\in H_0^1(\Omega)$.
• A natural Neumann or flux condition appears in the boundary functional after integration by parts.
• For nonzero Dirichlet data $g\in H^{1/2}(\partial\Omega)$, choose a bounded trace lift $G\in H^1(\Omega)$ and solve for $u-G\in H_0^1(\Omega)$.

A generic $H^1$ function has a trace, but not a classical normal derivative. Normal-flux traces require additional information about the field and its divergence.

A weak elliptic boundary-value problem

Let $\Omega$ be bounded and Lipschitz. For clarity, work with real-valued functions and consider

$$-\nabla\cdot \left( A(x)\nabla u \right) +c(x)u =f \quad\text{in }\Omega, \qquad \operatorname{Tr}u=0.$$

Assume $A\in L^\infty(\Omega;\mathbb R^{d\times d})$ is symmetric and uniformly elliptic:

$$\lambda|\xi|^2 \leq \xi^{\mathsf T}A(x)\xi \leq \Lambda|\xi|^2$$

for almost every $x$, every $\xi\in\mathbb R^d$, and fixed $0<\lambda\leq\Lambda$. Let $c\in L^\infty(\Omega)$ with $c\geq0$, and let $f\in H^{-1}(\Omega)$.

For smooth $u$ and a compactly supported test function $v$, multiplication and integration by parts give

$$\int_\Omega \left[ \nabla u\cdot A\nabla v +cuv \right] \mathrm d^d x = \langle f,v\rangle.$$

The right side denotes the $H^{-1}$–$H_0^1$ duality pairing. The formula continues to make sense for $u,v\in H_0^1(\Omega)$, so it becomes the definition:

Weak Dirichlet problem. Find $u\in V:=H_0^1(\Omega)$ such that

$$a(u,v)=F(v) \qquad \text{for every }v\in V,$$

where

$$\begin{aligned} a(u,v) &= \int_\Omega \left[ \nabla u\cdot A\nabla v+cuv \right]\mathrm d^d x,\\ F(v) &= \langle f,v\rangle. \end{aligned}$$

The distributional equation supplies the bulk relation; the choice $V=H_0^1$ supplies the boundary condition. Chen derives the Poisson case in Chen 2013, §2.1, Eqs. (11)–(15), printed pp. 9–10, PDF.

Lax–Milgram: existence, uniqueness, and stability together

The variational problem fits one reusable theorem.

Lax–Milgram theorem, real form. Let $V$ be a real Hilbert space, $F\in V'$, and let $a:V\times V\to\mathbb R$ be bilinear. If

$$|a(u,v)| \leq M\|u\|_V\|v\|_V$$

and

$$a(v,v) \geq \alpha\|v\|_V^2 \qquad (\alpha>0),$$

then there is a unique $u\in V$ satisfying $a(u,v)=F(v)$ for every $v\in V$, and

$$\|u\|_V \leq \frac{1}{\alpha} \|F\|_{V'}.$$

For complex spaces, in the site’s convention one may take $a$ conjugate-linear in its first argument and linear in its second, match $F$ to that linear argument, and require $\operatorname{Re}a(v,v)\geq\alpha\|v\|_V^2$. The exact real theorem and estimate are given in Chen, Chen 2024, Inf-Sup Conditions for Operator Equations, Lemma 2.6, printed p. 11, PDF.

For the elliptic model, equip $V=H_0^1(\Omega)$ with

$$\|v\|_V=\|\nabla v\|_{L^2}.$$

Poincaré’s inequality makes this equivalent to the full $H^1$ norm. The $L^\infty$ bounds on $A,c$, Cauchy–Schwarz, and Poincaré give boundedness of $a$. Uniform ellipticity gives

$$a(v,v) \geq \lambda \|\nabla v\|_{L^2}^2 = \lambda\|v\|_V^2.$$

Lax–Milgram therefore produces one weak solution and the estimate

$$\|u\|_V \leq \lambda^{-1} \|F\|_{V'}.$$

Applying the estimate to the difference of two solutions gives

$$\|u_1-u_2\|_V \leq \lambda^{-1} \|F_1-F_2\|_{V'}.$$

This is a quantitative stability statement, not merely uniqueness.

When $a$ is symmetric, the same solution is the unique minimizer of

$$\mathcal I[v] = \frac12a(v,v)-F(v).$$

Coercivity prevents minimizing sequences from escaping to infinite norm; weak compactness and lower semicontinuity supply a minimizer; strict convexity gives uniqueness. The Euler–Lagrange identity of $\mathcal I$ is the weak equation.

What well-posedness means

Hadamard well-posedness is relative to a declared solution map

$$\mathcal S:\mathcal D\longrightarrow\mathcal X, \qquad \text{data}\longmapsto\text{solution}.$$

The problem is well-posed in the stated data and solution topologies when

1. a solution exists for every admissible datum;
2. it is unique in $\mathcal X$; and
3. $\mathcal S$ depends continuously on the data.

The norms are part of the statement. A problem can be stable from $H^{-1}$ to $H_0^1$ and unstable from a much weaker data topology to a pointwise solution topology. The definition and its norm dependence are stated explicitly in MIT OpenCourseWare 2009 18.336, Lecture 2, printed p. 1, PDF.

The elliptic problem above is well-posed because Lax–Milgram proves all three items and gives a bounded solution map

$$\mathcal S: H^{-1}(\Omega) \longrightarrow H_0^1(\Omega).$$

Other PDE types need other estimates. Ellipticity alone does not prove coercivity on the selected domain, and coercive elliptic theory does not apply directly to hyperbolic evolution.

Hyperbolic stability from an energy estimate

Let $\Omega$ be bounded and Lipschitz, and consider a real Klein–Gordon field with homogeneous Dirichlet data. In the global $(+---)$ convention,

$$\begin{aligned} (\Box+m^2)u &= F &&\text{in }(0,T)\times\Omega,\\ \operatorname{Tr}u &= 0 &&\text{on }(0,T)\times\partial\Omega,\\ u(0) &= u_0, \qquad \partial_tu(0)=u_1, \end{aligned}$$

Here $\Box=\partial_t^2-\Delta$ and $m\geq0$.

Take

$$(u_0,u_1) \in H_0^1(\Omega)\times L^2(\Omega), \qquad F\in L^2\!\left(0,T;L^2(\Omega)\right).$$

For a sufficiently regular solution, define

$$E(t) = \frac12 \int_\Omega \left[ |\partial_tu|^2 +|\nabla u|^2 +m^2|u|^2 \right] \mathrm d^d x.$$

Multiplying the equation by $\partial_tu$, integrating over space, and integrating the Laplacian by parts gives no boundary contribution because of the zero trace:

$$\frac{\mathrm dE}{\mathrm dt} = \int_\Omega F\,\partial_tu\, \mathrm d^d x.$$

If

$$\mathcal E(t):=\sqrt{2E(t)},$$

then Cauchy–Schwarz yields

$$\mathcal E(t) \leq \mathcal E(0) + \int_0^t \|F(s)\|_{L^2}\,\mathrm ds.$$

For $F=0$, the energy is conserved. Apply the same inequality to the difference of two solutions: equal data imply uniqueness, and nearby initial data and forcing give nearby solutions in $H_0^1(\Omega)\times L^2(\Omega)$. Poincaré’s inequality supplies the low-frequency control even when $m=0$. For complex fields the right side of the energy identity becomes $\operatorname{Re}\int_\Omega F\,\overline{\partial_tu}\,\mathrm d^d x$.

For weak solutions, approximation extends the estimate to the energy class. More importantly, an existence theorem is available: Hunter, Hunter 2014, Definition 7.2 and Theorem 7.3, printed pp. 213–214, PDF, specialized to $L=-\Delta+m^2$, gives a unique weak solution with

$$\begin{aligned} u &\in C\!\left([0,T];H_0^1(\Omega)\right),\\ \partial_tu &\in C\!\left([0,T];L^2(\Omega)\right),\\ \partial_t^2u &\in L^2\!\left(0,T;H^{-1}(\Omega)\right), \end{aligned}$$

and bounds these norms by the stated data norms. Its Galerkin construction is the separate existence step; the energy estimate supplies the a priori control and, when applied to a difference, continuous dependence. This separation matters for variable coefficients, other boundary conditions, constraints, and nonlinear equations.

QFT-facing example: a sourced static scalar field

Let $\Omega$ be a bounded Lipschitz spatial region, $J\in H^{-1}(\Omega)$, and $m^2\geq0$. On the real energy space $H_0^1(\Omega)$, define

$$I[\phi] = \frac12 \int_\Omega \left[ |\nabla\phi|^2+m^2\phi^2 \right] \mathrm d^d x - \langle J,\phi\rangle.$$

For every variation $v\in H_0^1(\Omega)$,

$$\left. \frac{\mathrm d}{\mathrm d\varepsilon} I[\phi+\varepsilon v] \right|_{\varepsilon=0} = \int_\Omega \left[ \nabla\phi\cdot\nabla v +m^2\phi v \right] \mathrm d^d x - \langle J,v\rangle.$$

Stationarity is therefore the weak equation

$$\int_\Omega \left[ \nabla\phi\cdot\nabla v +m^2\phi v \right] \mathrm d^d x = \langle J,v\rangle \qquad \text{for every }v\in H_0^1(\Omega),$$

equivalently

$$\left( -\Delta+m^2 \right)\phi =J$$

as a distribution, with zero Dirichlet trace. Poincaré’s inequality makes the quadratic form coercive even at $m=0$, so Lax–Milgram gives a unique stable weak solution and the variational functional has that solution as its unique minimizer.

Tong derives the classical scalar action, its boundary term, and the Klein–Gordon equation in Tong 2006, Quantum Field Theory, §1.1–§1.1.1, Eqs. (1.3)–(1.14), PDF. The bounded-domain coercive formulation here is a Mathematical Methods adaptation. It does not construct a quantum path-integral measure. The physical treatment of boundary terms and admissible field variations continues in Boundaries, Variations, and Well-Posed Actions. For a kernel representation of a selected inverse, continue to Fundamental Solutions and Green Operators. For self-adjoint elliptic boundary realizations, spectra, and heat evolution, continue to Elliptic Boundary Problems and Heat Kernels.

The contrast with three nearby problems is instructive.

• In a massless Neumann formulation the trial space is $H^1(\Omega)$, not $H_0^1(\Omega)$, and constants lie in the kernel. On a connected domain, a source functional $J_N\in(H^1(\Omega))'$ must satisfy $J_N(1)=0$, and the solution is unique only modulo a constant. On a disconnected domain, impose the condition separately on every component.
• If the zeroth-order coefficient is negative, say $c=-\mu^2$, the Dirichlet form remains coercive only below the first spectral threshold $\mu^2<\lambda_1^D$, where $\lambda_1^D$ is the lowest Dirichlet eigenvalue of $-\Delta$. Indeed, $\|\nabla v\|_2^2\geq\lambda_1^D\|v\|_2^2$ makes the lower bound vanish at equality, where a zero mode appears. Formal ellipticity of the principal part does not remove this obstruction.
• A Lorentzian action is indefinite in spacetime derivatives, so its stationary equation is not a direct application of coercive Lax–Milgram theory. Hyperbolic Cauchy estimates, such as the energy estimate above, supply the relevant control.

What a weak solution does not guarantee

Classical regularity. An $H_0^1$ weak solution need not lie in $H^2$ or have pointwise second derivatives. Extra regularity depends on the coefficients, data, and boundary geometry. Chen, Chen 2013, Example 2.4, printed p. 12, PDF, gives a re-entrant domain where $L^2$ data do not produce an $H^2$ solution.

Nonlinear limits. Weak convergence passes continuous linear functionals but generally not products. On $(0,2\pi)$, for example, $u_n(x)=\sin(nx)$ converges weakly to zero in $L^2$, whereas $u_n^2\rightharpoonup 1/2$, not zero. Nonlinear equations therefore require additional, problem-specific analysis beyond the linear theorems proved here.

Common pitfalls

Leaving the spaces implicit. The phrase “solve $Pu=f$” does not specify whether equality is classical, almost everywhere, distributional, or variational. Name the trial space, test space, data space, and topology.

Writing boundary values pointwise. Sobolev functions are equivalence classes. Use a trace map and state the regularity of the domain and boundary data.

Calling an energy estimate an existence proof. An a priori estimate controls any solution that exists. A construction, compactness argument, semigroup theorem, or other existence step is still needed.

Using Lax–Milgram without coercivity. Formal ellipticity or a Gårding inequality is not the displayed coercivity bound. Zero modes, negative potentials, and gauge redundancy must be handled explicitly.

Replacing weak convergence by convergence of products. From $u_n\rightharpoonup u$ and $v_n\rightharpoonup v$, it does not follow in general that $u_nv_n$ converges to $uv$; the sine sequence above is already a counterexample with $v_n=u_n$.

Exercises

1. Show that $u(x)=|x|$ belongs to $H^1(-1,1)$, identify its weak first derivative, and explain why $u\notin H^2(-1,1)$.

   Check

   Both $|x|$ and $\operatorname{sgn}x$ lie in $L^2(-1,1)$. Splitting the integral at zero shows

   $$-\int_{-1}^{1} |x|\varphi'(x)\,\mathrm dx = \int_{-1}^{1} \operatorname{sgn}(x)\varphi(x)\,\mathrm dx,$$

   so $u'=\operatorname{sgn}x$ weakly. Its next distributional derivative is $2\delta_0$, which is not represented by an $L^2$ function; hence $u\notin H^2$.

2. Starting from $-\nabla\cdot(A\nabla u)+cu=f$ with zero Dirichlet trace, derive the weak identity and identify where the boundary condition enters.

   Check

   Multiply by $v\in C_c^\infty(\Omega)$ and integrate:

   $$\int_\Omega \left[ \nabla u\cdot A\nabla v+cuv \right]\mathrm d^d x = \langle f,v\rangle.$$

   Continuity extends this identity to all $v\in H_0^1(\Omega)$. The distributional bulk equation does not encode the boundary condition; requiring the trial solution $u\in H_0^1(\Omega)$ gives it zero trace.

3. Let $u_i$ solve $a(u_i,v)=F_i(v)$ for the same coercive form. Derive the stability estimate.

   Check

   Subtract the two equations and take $v=u_1-u_2$:

   $$\begin{aligned} \alpha\|u_1-u_2\|_V^2 &\leq a(u_1-u_2,u_1-u_2)\\ &= (F_1-F_2)(u_1-u_2)\\ &\leq \|F_1-F_2\|_{V'} \|u_1-u_2\|_V. \end{aligned}$$

   Cancel the final norm unless it is zero to obtain $\|u_1-u_2\|_V\leq \alpha^{-1}\|F_1-F_2\|_{V'}$.

4. Derive the Klein–Gordon energy identity and explain what it proves before an existence theorem is supplied.

   Check

   Multiply $(\partial_t^2-\Delta+m^2)u=F$ by $\partial_tu$ and integrate over space. The three terms on the left are the time derivatives of $\|\partial_tu\|_2^2/2$, $\|\nabla u\|_2^2/2$, and $m^2\|u\|_2^2/2$. Thus

   $$E'(t) = \int F\,\partial_tu\,\mathrm d^d x.$$

   Cauchy–Schwarz gives the displayed energy inequality. Applied to a difference, it proves uniqueness and continuous dependence in the energy topology. It does not construct a solution.

References

• Long Chen (2013), Sobolev Spaces and Elliptic Equations, PDF, §§1.2, 1.5–1.6, and 2.1–2.2. This is the teaching source for weak derivatives, Sobolev and trace spaces, Poincaré’s inequality, variational Poisson problems, and the regularity counterexample.
• Long Chen (2024), Inf-Sup Conditions for Operator Equations, PDF, §2.2, especially Lemma 2.6. This is the theorem-level source for Lax–Milgram and its stability estimate.
• Semyon Dyatlov (2022), Lecture Notes for 18.155: Distributions, Elliptic Regularity, and Applications to PDEs, PDF, §12.1. This is the structural source for the Fourier definition and mapping properties of real-order Sobolev spaces.
• John K. Hunter (2014), Notes on Partial Differential Equations, PDF, §§7.1–7.5, especially Definition 7.2 and Theorem 7.3. This is the structural source for bounded-domain weak hyperbolic solutions, their energy bounds, existence, and uniqueness.
• William McLean, Strongly Elliptic Systems and Boundary Integral Equations, Chapter 3, Cambridge University Press, 2000. This is the specialist reference for Lipschitz domains, boundary Sobolev spaces, and trace operators.
• MIT OpenCourseWare (2009), 18.336 Lecture 2: Well-Posedness, PDF, printed p. 1, for the Hadamard criteria and their dependence on the chosen norm.
• David Tong (2006), Quantum Field Theory, PDF, §§1.1–1.1.1. This supplies the QFT source for the real scalar action, boundary term, spatial energy, and Klein–Gordon equation.