Imagine you're not moving a single object from point A to point B, but reshaping an entire cloud of something — heat, chemical concentration, a population of particles — from one spatial distribution into another, while an adversary is actively trying to push it off course and while it must stay within physical bounds the entire time. That is the problem a new paper by Siddhartha Ganguly, Vaibhav Upadhyay, Kenji Kashima, and Debasish Chatterjee takes on, for systems governed by noisy linear parabolic PDEs — the classic example being the heat equation with disturbances.
The practical stakes are concrete. Systems like temperature fields in a furnace, drug concentration in tissue, or particle density in a reactor are often described by these diffusion-type PDEs, and engineers frequently need to drive them from a known starting distribution to a target one within a fixed time window without violating safety limits — a maximum temperature, a minimum concentration, a bounded actuation rate. Doing this under worst-case disturbances, rather than assuming a clean nominal model, is what turns an already hard PDE control problem into a genuinely difficult minmax one.
The authors frame this as a minmax problem: minimize a cost over the choice of control actions while maximizing over the choice of disturbance, subject to convex constraints that must hold for every point in space and time. In plain terms, you want a control strategy that performs well even against the worst disturbance the system could throw at it, not just the average or expected one. That worst-case framing is what gives the resulting solution its robustness guarantee, but it's also what makes the problem numerically hard — you're not just solving one optimization, you're solving one nested inside another.
To make this tractable, the paper takes the PDE and applies finite-difference discretization to the spatial derivatives, turning an infinite-dimensional PDE into a large but finite system of ordinary differential equations evolving in time — a standard and well-understood way to bring PDE control problems into a computational setting. The control and disturbance trajectories are then represented using a finite set of parameters (a finite-dimensional basis or parametrization), rather than treated as arbitrary functions, which keeps the search space workable. The result, under stated mild assumptions, is a convex semi-infinite program (SIP): a convex optimization problem with finitely many decision variables but infinitely many constraints, one for each possible disturbance realization and each point in the constraint set.
This is the crux of the technical contribution: convex SIPs are notoriously awkward because you cannot literally check infinitely many constraints, yet the paper shows the problem retains enough convex structure that recently developed SIP tools apply. Formally, a convex SIP takes the shape $$\min_{x \in X} f(x) \quad \text{s.t. } g(x, w) \le 0 \ \ \forall w \in W,$$ where $x$ collects the finite control/disturbance parameters, $f$ is a convex cost, and $W$ is an (often infinite or continuous) index set — here, the space of admissible disturbance realizations and constraint locations. Solving this exactly for every $w \in W$ is generally impossible, so the paper leans on a method — drawn from prior work on near-optimal solutions to convex SIPs via targeted sampling — that identifies a finite, provably sufficient subset of the most binding constraints through global maximization, rather than exhaustively enumerating the constraint family.
That distinction matters for anyone who has tried to apply robust or minmax optimal control in practice: many prior approaches to constrained density transport either ignored input/actuation constraints, ignored density-path constraints, or handled disturbances only heuristically. A related 2024 paper on optimal mass transport under input and density constraints made a similar point — that earlier constrained optimal-transport work hadn't offered numerical algorithms handling both constraint types together. By building on convex SIP theory to get exact-solution guarantees with a finite, computable constraint set, this paper's authors argue their algorithm handles the combination — convex path constraints, disturbance robustness, and PDE dynamics — within one coherent computational pipeline, and they back this with numerical examples in the paper.
It's worth being precise about what "guarantee" means here: the authors establish that their finite-parameter solution satisfies the original infinite family of constraints (across all disturbance realizations) under the stated assumptions, and they provide an optimization-based algorithm that is computationally efficient relative to naive constraint enumeration — not that the method scales trivially to arbitrarily fine discretizations or high-dimensional PDEs, which remains an open engineering question for any finite-difference PDE control approach. The paper is presented as a numerical, algorithmic companion to the authors' broader research program on density control for PDEs and SDEs (including Schrödinger-bridge-based approaches), and a closely related companion paper — described as a more "direct" optimal control treatment of the same minmax density transportation problem — is reportedly under review at IEEE Transactions on Automatic Control, suggesting this line of work is being developed on multiple fronts simultaneously.
The bigger picture: as robust and adversarial framings become more common across control and optimization — from robust MPC to distributionally robust supply chain models — the bottleneck is rarely the modeling insight and almost always the computational tractability of the resulting semi-infinite or minmax formulation. Papers like this one, that import newer SIP-specific numerical machinery rather than falling back on generic robust-optimization approximations, are a signal of where PDE-constrained control is heading: toward solvers that give real guarantees on infinite constraint families instead of just plausible-looking finite approximations.
Sources: Ganguly, Upadhyay, Kashima, Chatterjee, 'Constrained minmax density transportation for linear parabolic PDEs: a numerical optimal control perspective', arXiv:2608.19170 · arXiv recent submissions listing, Optimization and Control (math.OC), https://arxiv.org/list/cs/pastweek?show=500&skip=518 · Das, Aravind, Cherukuri, Chatterjee, 'Near-optimal solutions of convex semi-infinite programs via targeted sampling', Annals of Operations Research (2022), https://research.rug.nl/en/publications/near-optimal-solutions-of-convex-semi-infinite-programs-via-targe/ · 'Optimal Mass Transport of Nonlinear Systems under Input and Density Constraints', arXiv:2403.16683 · Siddhartha Ganguly, author homepage and publications page, https://sites.google.com/view/siddhartha-ganguly/home
