Paper Reading - Optimization problems for elliptic PDEs (2601.01591)
Table of Contents
This paper is a panoramic tour of three families of optimal control problems for elliptic PDEs: where the control is the coefficient, the potential, or the source term, unifying and sharpening results from the authors’ previous works.
Three ways to control an elliptic PDE #
The authors always consider a Dirichlet problem on a bounded domain $\Omega \subset \mathbb{R}^d$, with the solution $u$ as the state and a function (or measure) as the control. They study three settings:
Optimal coefficients $a(x)$: $$ -\mathrm{div}(a(x)\nabla u) = f \text{ in } \Omega, \quad u=0 \text{ on } \partial\Omega, $$ cost function $J(u,a) = \int_\Omega j(u,a),dx$, with a constraint $\int_\Omega \psi(a),dx \le 1$.
Optimal potentials $V(x)$: $$ -\Delta u + V(x)u = f \text{ in } \Omega, \quad u\in H_0^1(\Omega), $$ cost function $J(u,V) = \int_\Omega (j(x,u) + \psi(V)),dx$.
Optimal sources $f$: $$ -\Delta u = f \text{ in } \Omega, \quad u\in H_0^1(\Omega), $$ cost function $J(f) = \int_\Omega j(x,u_f,f),dx$ with $\int_\Omega \psi(f),dx \le m$.
In all cases, $\psi$ is convex and lower semi-continuous (l.s.c), encoding constraints and penalizations on the control. The paper focuses on existence of optimal controls (sometimes as measures), characterization via auxiliary variational problems and adjoint states, bang–bang behavior, and regularity of optimal controls and their induced interfaces.
Optimal Coefficients: Where to Put the Good Material? #
Minimal Compliance and Measure-Valued Coefficients #
The model problem is compliance minimization for $-\mathrm{div}(a(x)\nabla u) = f$, $u=0$, with non-negative $a$.
Compliance is defined as: $$ C(a) = \int_\Omega f u_a,dx, $$ and it relates to the energy $$ E(a) = \inf_{u\in H_0^1} \int_\Omega \left(\tfrac{1}{2} a|\nabla u|^2 - f u\right)dx $$ via $C(a) = -2E(a)$.
The optimization problem is written as: $$ \min_{a \geq 0} \left\{ C(a) + \int_\Omega \psi(a)dx \right\}, $$ or equivalently as a max–min problem in $(a,u)$.
Two growth regimes of $\psi$ are crucial:
- Superlinear: $\psi(s)/s \to +\infty$. Then admissible coefficients are in $L^1(\Omega)$, and there exists an optimal $a_{\mathrm{opt}}\in L^1(\Omega)$.
- Linear growth: $\psi(s)/s \to k>0$. Then it is natural to extend the problem to measures $\mu\ge 0$, allowing “thin” structures on lower-dimensional sets. The cost $\int \psi(\mu)$ is interpreted through the Lebesgue–singular decomposition and the recession function $\psi_\infty$. An optimal measure $\mu_{\mathrm{opt}}\in \mathcal{M}^+(\Omega)$ still exists.
Because the functional is convex in $u$ and concave in $a$, the authors exchange inf and sup and reduce to an auxiliary minimization problem in $u$ alone: $$ \inf_{u} \int_\Omega \psi^{*}(|\nabla u|^2)dx - 2\int_\Omega u df, $$ where $\psi^{*}$ is the Legendre–Fenchel conjugate. Under mild assumptions this problem has a unique minimizer $\bar u$, and the optimal coefficient is recovered point-wise from the optimality condition: $$ a_{\mathrm{opt}}|\nabla\bar u|^2 = \psi(a_{\mathrm{opt}}) + \psi^*(|\nabla\bar u|^2). $$
Examples:
- Power penalization $\psi(s) = s^p/p$, $p>1$: The auxiliary problem involves a nonlinear PDE $$-\Delta_{2p/(p-1)} u = \tfrac{2p}{p-1} f,$$ and the optimal coefficient is $a_{\mathrm{opt}}(x) = |\nabla \bar u(x)|^{2/(p-1)}$. For $\Omega$ a ball and $f=1$ or $f=\delta_0$, the authors give explicit radial formulas and plots for $\bar u$ and $a_{\mathrm{opt}}$.
- Two-phase box constraint $\psi(s) = s$ on $[\alpha,\beta]$, $+\infty$ otherwise: The auxiliary problem yields an optimal coefficient $a_{\mathrm{opt}}\in L^\infty(\Omega)$ taking values in $[\alpha,\beta]$, and under regularity of $\Omega$ and $f$ one gets extra smoothness (e.g. $\nabla a_{\mathrm{opt}}\cdot \nabla \bar u \in L^2(\Omega)$).
General Coefficients and G-Closure #
For a general cost: $$\min_{a\ge 0}\min_{u} \int_\Omega (j(x,u)+\psi(a)),dx \quad \text{s.t. } u \text{ solves } -\mathrm{div}(a\nabla u)=f,$$ existence of an optimal $a$ may fail.
The relaxed problem is naturally expressed via G-convergence: sequences of scalar coefficients $a_n\in[\alpha,\beta]$ can generate limit operators with matrix-valued coefficients $A(x)$, described by the celebrated Murat–Tartar G-closure.
The G-closure set $\mathcal{A}$ consists of symmetric matrices $A(x)$ whose eigenvalues $\lambda_1\le\cdots\le\lambda_d$ lie in $[\alpha,\beta]$ and satisfy a family of inequalities depending on a mixing parameter $t\in[0,1]$, involving the arithmetic and harmonic means $\mu_t, \nu_t$ of $\alpha,\beta$. For $d=2$, this gives an explicit admissible region in the $(\lambda_1,\lambda_2)$-plane.
Relaxed functionals of the form $\int \psi(x,a),dx$ over G-limits have been studied in special cases, e.g. $\psi(x,a)=g(x)a$, where one can express the relaxation in terms of the largest eigenvalue $\lambda_{\max}(A(x))$. The authors show a numerical example where the relaxed optimal matrix $A_{\mathrm{opt}}$ has eigenvalues $\lambda_1\neq \lambda_2$ on a set of positive measure, revealing genuine microstructure.
Optimal Potentials: Shaping the “Landscape” $V(x)$ #
Here the control is a nonnegative potential $V$ in $$-\Delta u + V u = f, \quad u\in H_0^1(\Omega).$$ The cost is: $$\min \int_\Omega (j(x,u) + \psi(V)),dx,$$ with $V\ge 0$ and $\psi$ convex, l.s.c., super-linear (so any finite-cost $V$ lies in $L^1(\Omega)$).
Compliance Case: Eliminating the Control #
For the compliance choice $j(x,u) = f(x)u$, the problem can again be reduced to a variational problem in $u$ only.
Define: $$ E(V) = \min_{u\in H_0^1(\Omega)} \int_\Omega \left(\tfrac{1}{2} |\nabla u|^2 + \tfrac{1}{2} V u^2 - f u\right)dx, \quad \Psi(V)=\int_\Omega \psi(V),dx. $$
Minimizing $-2E(V)+\Psi(V)$ over $V\ge 0$ is equivalent to: $$ \min_{u\in H_0^1(\Omega)} \int_\Omega \left(|\nabla u|^2 + \psi^*(u^2) - 2 f u\right)dx, $$ a semi-linear elliptic problem in $u$ with nonlinearity $g(s)=s(\psi^*)’(s^2)$. The optimal state $\bar u$ solves: $$ -\Delta u + g(u) = f, \quad u\in H_0^1(\Omega), $$ and the optimal potential is: $$ V_{\mathrm{opt}} = (\psi^*)’(\bar u^2). $$ So in this special case the control can be explicitly reconstructed from the state.
General Costs, Adjoint Equation, and Regularity #
For a general $j(x,u)$, the authors prove an existence theorem of an optimal $V_{\mathrm{opt}}\in L^1(\Omega)$ under natural growth and coercivity assumptions on $j$ and super-linearity of $\psi$.
Optimality conditions involve:
- The state $\bar u$ solving $-\Delta u + V_{\mathrm{opt}}u = f$.
- An adjoint state $v$ solving $-\Delta v + V_{\mathrm{opt}} v = \partial_s j(x,\bar u)$.
- A sub-differential relation $\bar u v \in \partial\psi(V_{\mathrm{opt}})$, rewritten as a point-wise inequality $h^{-}(\bar u v) \le V_{\mathrm{opt}} \le h(\bar u v)$, where $h$ is built from the sub-differential of $\psi$.
From here, regularity of $V_{\mathrm{opt}}$ is linked to properties of $h$ and to elliptic regularity for $\bar u$ and $v$. Under strengthened assumptions on $j$, $f$, and $\Omega$, the authors show that $\bar u, v \in W^{2,q}(\Omega)$ for some $q>d/2$ (hence continuous), and the product $\bar u v V_{\mathrm{opt}}$ is in $BV(\Omega)$, so $V_{\mathrm{opt}}\in BV_{\mathrm{loc}}(\Omega\setminus K)$ where $K = {\bar u v =0}$. This identifies the “degeneracy set” $K$ as the core where singularities of the optimal potential may concentrate.
Bang–Bang Potentials: If $\psi$ is flat on an interval $[\alpha,\beta]$ (e.g. $\psi(s) = s$ on $[\alpha,\beta]$, $+\infty$ otherwise), the function $h$ becomes multi-valued and the optimal potential is bang–bang: $$ V_{\mathrm{opt}} = \alpha + (\beta-\alpha)\mathbf{1}_E $$ for some set $E$ of finite perimeter. The paper includes numerical simulations showing the geometry of such sets for specific loads $f$.
Optimal Sources: Choosing the Right-Hand Side #
Finally, the control is the source $f$ in $-\Delta u = f$, $u\in H_0^1(\Omega)$, with cost $J(f) = \int_\Omega j(x,u_f,f),dx$ and constraint $\int_\Omega \psi(f),dx\le m$.
Existence with Superlinear and Linear $\psi$: If $\psi$ is super-linear and $j$ satisfies suitable lower bounds and convexity in $f$, then an optimal $f_{\mathrm{opt}}\in L^1(\Omega)$ exists.
If $\psi$ has linear growth, the natural admissible class is signed measures $f$ with finite total variation, and $\int \psi(f)$ is defined via the Lebesgue–singular decomposition and recession coefficients $c_-(\psi), c_+(\psi)$. Under a decomposition $j(x,s,z)=A(x,s)+B(x,z)$ with specific structure and lower bounds, the functional is lower semi-continuous under weak-* convergence of measures, and there exists an optimal measure-valued source $f_{\mathrm{opt}}$.
Optimality Conditions and Bang–Bang Description: Introduce the self-adjoint resolvent operator $R$ mapping a source $f$ to the solution $u_f$. Under differentiability and growth conditions on $j$, the authors derive necessary (and, under convexity, sufficient) conditions for optimality. For super-linear $\psi$, define: $$ w := R\big(\partial_s j(x, R(f_{\mathrm{opt}}), f_{\mathrm{opt}})\big) + \partial_z j(x, R(f_{\mathrm{opt}}), f_{\mathrm{opt}}). $$ Then there is $\lambda \ge 0$ such that either:
- $\lambda=0$: $w$ has a fixed sign and $f_{\mathrm{opt}}$ saturates the endpoints of $\mathrm{dom}(\psi)$ on the regions where $w$ is strictly positive/negative — a pure bang–bang behavior.
- $\lambda>0$: the constraint is saturated, $\int \psi(f_{\mathrm{opt}})=m$, and $f_{\mathrm{opt}}$ satisfies a point-wise equality involving $\psi$, its conjugate $\psi^*$, and $w$.
For linear-growth $\psi$, a similar structure holds, but the singular part of $f_{\mathrm{opt}}$ is supported on level sets where $w$ hits thresholds determined by the slopes $c_-(\psi), c_+(\psi)$.
Spectral Example: Maximizing Energy Under an $L^2$ Constraint
For: $$ j(u) = -\tfrac{1}{2} u^2, \quad \psi(s)=\tfrac{1}{2} s^2, $$ the problem becomes: $$ \max \left\{\frac{1}{2}\int_\Omega u_f^2 f,dx : \int_\Omega f^2,dx \right\}. $$
The optimality system shows that the optimal source $f$ satisfies a fourth-order eigenvalue problem $\Delta^2 f = f/\lambda$, equivalent to an eigenvalue problem for the Laplacian. The maximizer is a multiple of the first Dirichlet eigenfunction $\varphi$ of $-\Delta$: $$ f = \pm \sqrt{2m},\varphi, \quad \lambda = 1/\mu_1^2, $$ where $\mu_1$ is the first eigenvalue. The paper includes a numerical plot for such an optimal source in an ellipse.
Compliance with Box Constraints on the Source: For compliance with box constraints: $$ \min \left\{\int_\Omega f,R(f),dx : \int_\Omega f,dx \ge m,\ f\in[\alpha,\beta]\right\}, \quad 0\le \alpha<\beta, $$ the optimal source is bang–bang: $$ f _{\mathrm{opt}} = \alpha,\mathbf{1} _E + \beta,\mathbf{1} _{\Omega\setminus E}, $$ with $E = {R(f _{\mathrm{opt}}) < s}$ and $s$ chosen to fit the mass constraint. The corresponding state solves: $$ -\Delta u = \beta,\mathbf{1} _{\{u<s\}} + \alpha,\mathbf{1} _{\{u>s\}}. $$
Using results from their previous work on optimal potentials, the authors prove that $f _{\mathrm{opt}} \in BV(\Omega)$: the interface between the regions where $f=\alpha$ and $f=\beta$ has finite perimeter.
If $\Omega$ is convex, they go further: in the special case $\alpha = 0$, $f _{\mathrm{opt}} = \mathbf{1} _E$ with $E = {w < s}$, where $w$ solves $-\Delta w = \mathbf{1} _{\{w<s\}}$. They show that the optimal set $E$ is convex and its boundary is of class $C^1$. So in convex domains, the region where you “turn on” the source to maximize stiffness is itself a smooth convex set.
References #
[1] Buttazzo, G., Casado-Díaz, J., & Maestre, F. (2025). Optimal sources for elliptic PDEs. arXiv preprint arXiv:2509.01521.
| |
[2] Buttazzo, G., Casado-Díaz, J., & Maestre, F. (2025). Optimal coefficients for elliptic PDEs. arXiv preprint arXiv:2512.08431.
| |
[3] Buttazzo, G., Casado-Díaz, J., & Maestre, F. (2026). Optimization problems for elliptic PDEs. arXiv preprint arXiv:2601.01591.
| |