Pointwise Dual Relaxation (PDR) Overview

• Pointwise Dual Relaxation is a method that constructs convex relaxations by dualizing constraints and enforcing them pointwise, providing reliable lower or upper bounds for complex non-convex problems.
• It leverages sum-of-squares hierarchies and dual Lagrangian formulations to convert intractable variational and control problems into tractable semidefinite programs with empirically small duality gaps.
• Extensions to stochastic control and learning applications demonstrate PDR’s scalability and effectiveness in reducing constraint violations in large-scale, high-dimensional optimization tasks.

Pointwise Dual Relaxation (PDR) refers to a family of convex relaxations that yield rigorous lower or upper bounds for intractable optimization, variational, and control problems by dualizing constraints and enforcing relaxations pointwise in finite-dimensional spaces or over individual sample paths. Originally developed for global optimization of non-convex variational problems, PDR frameworks have also been extended to optimal control with information relaxations and to large-scale learning-based constrained optimization, notably in power systems. Key features include rigorous duality properties, computable sum-of-squares (SOS) hierarchies for polynomial problems, and empirically small duality gaps. PDR has sharpness guarantees for broad classes of minimization problems, notably quadratic forms and principal eigenvalue computations, and shows practical effectiveness in both deterministic and stochastic regimes (Chernyavsky et al., 2021, Owerko et al., 23 Oct 2025, Ye et al., 2013).

1. Theoretical Foundation and Problem Setting

PDR arises in the global minimization of integral variational problems of the form

$$F^* = \inf_{u \in W^{1,p},\ \text{subject to constraints}} \int_\Omega f(x, u(x), \nabla u(x))\, dx,$$

where $u: \Omega \to \mathbb{R}^m$ is a function in a Sobolev space, $$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$ is continuous, and the feasible set is characterized by arbitrary combinations of

• integral inequality constraints $\int_\Omega a(x,u,\nabla u)\, dx \geq 0$,
• pointwise constraints $c(x,u,\nabla u)=0$ or $c(x,u,\nabla u)\geq 0$ for all $x \in \Omega$,
• boundary conditions $d_i(x,u)=0$ or $d_i(x,u)\geq 0$ for $x \in \partial \Omega_i$ (Chernyavsky et al., 2021).

The original problems are typically non-convex and intractable for direct global optimization. The key innovation of PDR is to dualize and relax the problem such that the dual feasible set is defined over pointwise or sample-path constraints in finite dimensions.

2. Construction of the PDR: Dualization and Pointwise Relaxation

The essential derivation of PDR involves:

• Formulating a max–min Lagrangian using scalar and (vector) function multipliers for constraints, notably $u: \Omega \to \mathbb{R}^m$0 for integral constraints and $u: \Omega \to \mathbb{R}^m$1 to enforce the weak gradient constraint via the divergence theorem.
• Taking the supremum over multipliers and exchanging the order of infimum/supremum, which leads to a relaxation where the constraints are imposed pointwise rather than over function spaces.
• Introducing functions $u: \Omega \to \mathbb{R}^m$2 and $u: \Omega \to \mathbb{R}^m$3 such that

$u: \Omega \to \mathbb{R}^m$4

over the pointwise constraint sets $u: \Omega \to \mathbb{R}^m$5 (Chernyavsky et al., 2021).

The resulting PDR takes the form of a convex maximization: $u: \Omega \to \mathbb{R}^m$6 subject to pointwise nonnegativity conditions on $u: \Omega \to \mathbb{R}^m$7 and boundary coordinates. Each relaxation step (dualization, infimum/supremum exchange, pointwise infima) is documented to weaken the original problem, so the PDR supremum is always a lower bound to $u: \Omega \to \mathbb{R}^m$8 (Chernyavsky et al., 2021).

3. Sum-of-Squares Hierarchy and Computational Realization

For polynomial data (i.e., $u: \Omega \to \mathbb{R}^m$9, and $$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$0 polynomial/semialgebraic), the PDR can be further relaxed and numerically implemented via SOS programming:

• All pointwise inequalities are replaced by SOS constraints using quadratic modules, yielding a tractable semidefinite program (SDP).
• The search space is parameterized by polynomials of bounded degree $$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$1, and nonnegativity is relaxed to membership in quadratic modules $$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$2.
• The resulting SDP aims to minimize $$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$3 subject to the polynomial constraints (Chernyavsky et al., 2021).

Convergence theorems guarantee that the optimal values of SOS relaxations approach the true PDR value as polynomial degree increases on compact feasible sets or under Dirichlet boundary conditions.

Step	Description	Reference
Dualization	Lagrangian + divergence multipliers	(Chernyavsky et al., 2021)
Exchange of inf/sup	Infimum becomes pointwise via relaxation	(Chernyavsky et al., 2021)
SOS relaxation	Replace polynomial nonnegativity with SOS constraints (SDP)	(Chernyavsky et al., 2021)

4. Sharpness Theorems and Typical Problem Classes

PDR admits exactness ("sharpness") for several important classes:

• Quadratic integrands (generalized Poisson): For energies of the form $$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$4 with $$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$5, the PDR yields the exact minimum value $$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$6; for polynomial data, SOS relaxations converge to the correct bound as degree increases.
• Principal eigenvalue problems: The smallest eigenvalue of $$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$7 with Dirichlet conditions is obtained exactly via the PDR; again, SOS relaxations converge.
• Poincaré constants/$$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$8-Laplacian: In 1D, the sharp constant of the $$f: \Omega \times \mathbb{R}^m \times \mathbb{R}^{m\times n} \to \mathbb{R}$$9–Poincaré inequality is attained and the sharpness holds for rational ansatzes in the SOS hierarchy (Chernyavsky et al., 2021).

Numerical experiments confirm the closing of the gap between lower (PDR) and true infima up to $\int_\Omega a(x,u,\nabla u)\, dx \geq 0$0–$\int_\Omega a(x,u,\nabla u)\, dx \geq 0$1 across 1D and 2D test problems.

5. Extensions to Controlled Markov Processes and Information Relaxation

PDR extends to stochastic control via the information-relaxed dual approach:

• For discrete-time Markov decision processes (MDPs) and continuous-time controlled diffusions, information-relaxed duality (PDR) allows feasible controls to depend on the entire noise realization, penalized pathwise by a dual function to ensure no net bias for non-anticipative (admissible) policies.
• The dual value is

$\int_\Omega a(x,u,\nabla u)\, dx \geq 0$2

subject to FEASIBLE martingale penalty $\int_\Omega a(x,u,\nabla u)\, dx \geq 0$3.

• Strong duality guarantees the dual upper bound matches the primal value for suitable (often explicit) choices of penalty, constructed from the value function via Ito calculus (martingale structure) (Ye et al., 2013).

Empirical results, e.g., in portfolio optimization with CRRA utility and factor models, exhibit small duality gaps (a few percent) for practical penalty constructions.

6. PDR in Learning and Large-Scale Constrained Optimization

Recent work applies PDR to supervised learning of neural solvers for constrained optimization, notably in optimal power flow (OPF):

• Unlike conventional methods, PDR imposes pointwise constraints over all sampled realizations (demand vectors), ensuring no cancellation of violations in expectation.
• Optimization is performed in the dual domain via augmented Lagrangian and block-coordinate primal–dual stochastic updates, with sample-specific multipliers $\int_\Omega a(x,u,\nabla u)\, dx \geq 0$4, $\int_\Omega a(x,u,\nabla u)\, dx \geq 0$5.
• Empirically, PDR achieves significantly lower mean and maximum constraint violations (e.g., max violation 5.68% versus 18.97% for baseline on large power grids), with pronounced improvements in "corner cases" (worst-case realizations) (Owerko et al., 23 Oct 2025).

This approach generalizes to large networks and high-dimensional problems, provided memory and computational resources suffice to store and update pointwise multipliers.

7. Computational Aspects and Limitations

PDR with SOS relaxations is computationally tractable for small $\int_\Omega a(x,u,\nabla u)\, dx \geq 0$6 (dimension) and constraint number $\int_\Omega a(x,u,\nabla u)\, dx \geq 0$7, with SDP scaling exponentially in degree and input dimension. In practice, $\int_\Omega a(x,u,\nabla u)\, dx \geq 0$8–$\int_\Omega a(x,u,\nabla u)\, dx \geq 0$9 and degree $c(x,u,\nabla u)=0$0–$c(x,u,\nabla u)=0$1 are feasible. PDR is strictly more general and systematic than previous low-dimensional dual relaxations (e.g., Valmorbida et al., 2016). In learning applications, the need to store $c(x,u,\nabla u)=0$2 dual variables is manageable for modern hardware up to tens of thousands of instances (Chernyavsky et al., 2021, Owerko et al., 23 Oct 2025).

8. Summary and Research Directions

Pointwise Dual Relaxation provides a unifying methodology for convex relaxation of variational, stochastic, and learning-based constrained problems. Its asymptotic sharpness for wide classes, rigorous duality theory, and practical scalability in key domains support its adoption for global optimization and upper/lower bounding purposes. A plausible implication is that future work may extend PDR to non-polynomial, high-dimensional functionals, or exploit domain-specific structure for improved scalability and expressivity (Chernyavsky et al., 2021, Owerko et al., 23 Oct 2025, Ye et al., 2013).