Continuous Space Control of Partial Differential Equations

23 concepts22 papers1 proposal

When we design a controller for a standard physical system—like a cruise control loop for a car or a balancing mechanism for a drone—we model the system using Ordinary Differential Equations (ODEs). The state of the system can be summarized by a finite vector of numbers: position, velocity, and orientation.

But when we try to control spatiotemporal phenomena—such as fluid turbulence, heat propagation, or the torsional vibrations of a miles-long drilling pipe—the state is no longer a finite vector. It is a continuous profile across space. To know the state of a fluid, you need to know its velocity at every single point in space. These systems have infinite degrees of freedom and are governed by Partial Differential Equations (PDEs).

Historically, engineers dealt with this by using a "reduce-then-design" approach: they discretized the spatial domain into a grid, turning the PDE into a massive system of coupled ODEs, and then applied standard control theory. However, this approach strips away the underlying continuous geometry of the system, often leading to high-dimensional computational bottlenecks or "spillover instabilities," where the controller inadvertently excites unmodeled high-frequency dynamics between the grid points.

A powerful paradigm shift has emerged to solve this: direct infinite-dimensional control. Instead of forcing the system into a finite-dimensional box, these methods operate directly on the continuous function spaces, optimizing controllers before any discretization occurs.


The Core Intuition: From Vectors to Functions

To understand why direct control works, we must change how we view states and actions. Instead of treating a state as a vector $x \in \mathbb{R}^n$, we treat it as a function $u(x, t)$ residing in an infinite-dimensional Hilbert space. A controller's job is no longer to find a control matrix $K$ that multiplies a state vector, but rather to find an operator $\mathcal{K}$ that maps a state function to a control function.

Traditional Control (Finite-Dimensional):
[State Vector] -------> ( Matrix K ) -------> [Control Vector]

Direct PDE Control (Infinite-Dimensional):
[State Function] -----> ( Operator K ) -----> [Control Function]

By keeping the problem in the continuous domain, we preserve the structural properties of the physics—such as dissipation, conservation laws, and boundary conditions—that guide stable control design.


Optimize-Then-Discretize via Convex Optimization

One of the most robust ways to handle continuous systems is to formulate the control problem as an optimization over functions.

For parabolic PDEs (which govern diffusion and heat transfer), Matthew M. Peet et al. demonstrated that Sum-of-Squares (SOS) polynomials can be used to search for Lyapunov functions directly within the infinite-dimensional space. This allows for the synthesis of state and output feedback controllers without ever approximating the underlying PDE, guaranteeing the stability of the true continuous system.

For systems with structural constraints—such as communication delays or localized actuation—Lauren Conger et al. extended the framework of System Level Synthesis (SLS) to Hilbert spaces. This "optimize-then-discretize" approach allows engineers to compute highly constrained controllers while maintaining mathematical rigor, bypassing the suboptimality that plagues early discretization.

To bypass explicit PDE modeling altogether while retaining systems-theoretic guarantees, Girish Chowdhary et al. introduced kernel controllers. By layering linear dynamical systems over the weights of reproducing kernel models, they created a domain-agnostic framework capable of estimating and controlling continuous spatiotemporal processes using only sparse sensor networks. Similarly, Pierre-Cyril Aubin-Frankowski extended the classical Linear Quadratic Regulator (LQR) framework to infinite dimensions by utilizing operator-valued kernels, providing an elegant mathematical bridge that encodes continuous control problems into kernel spaces.


Guaranteeing Stability at the Boundaries

In many real-world applications, we cannot inject control inputs everywhere inside the system; we can only act on the boundaries (e.g., adjusting a valve at the end of a pipeline or a motor at the top of a drill string).

Boundary control introduces severe mathematical challenges because the control input acts as a boundary condition rather than a smooth internal force. To prove that these systems will not blow up under external disturbances, researchers rely on specialized Lyapunov functionals that act as continuous energy metrics.

Boundary Control Intuition:
Disturbance / Actuation ──> [ Boundary ] ──(Energy Dissipation)──> [ Internal Domain ]
  • Hyperbolic Systems & Pipelines: Hyperbolic PDEs govern convective phenomena like gas flowing through pipelines or water in canals. Ferdinand Thein provided a unifying theoretical bridge connecting different boundary feedback frameworks in multi-dimensional spaces, resolving mathematical friction in pipeline network control. To handle nonlinearities, Wancong Yao et al. designed boundary feedback laws for quasilinear hyperbolic systems by exploiting their "partially dissipative structure," proving exponential stability around equilibrium points.
  • Mechanical Vibrations: Long physical structures behave like continuous springs. Matthieu Barreau et al. analyzed the practical stability of a drilling pipe subject to nonlinear friction, proving that finite-dimensional approximations are fundamentally inadequate for robust control design in these environments. To make these insights practical, Matthieu Barreau et al. utilized Bessel's inequality and Linear Matrix Inequalities (LMIs) to synthesize finite-dimensional controllers for coupled PDE-ODE systems, avoiding the unimplementable infinite-dimensional control laws generated by traditional analytical methods like backstepping.
  • Robustness to Noise: Real boundaries are noisy. The authors of Shorten et al. developed a spectral decomposition framework to analyze Input-to-State Stability (ISS) for boundary control systems under rough, non-smooth disturbances, ensuring that boundary noise does not destabilize the internal state.

Scaling Up with Neural Operators and Reinforcement Learning

While optimization and Lyapunov methods provide ironclad stability guarantees, they struggle to scale to highly nonlinear, chaotic regimes like turbulent fluid dynamics. This is where data-driven methods step in—not by reverting to grid-based vectors, but by designing machine learning architectures that natively understand functions.

Function-Valued Action Spaces

Standard Reinforcement Learning (RL) algorithms output a discrete vector of actions. If you want to control a continuous actuator profile, a standard vector output fails to scale. To overcome this, Yangchen Pan et al. proposed an RL framework with function-valued action spaces. By utilizing spatial "action descriptors," the agent learns to output a continuous control function, enabling stable learning at scales (such as 256-dimensional continuous actions) where conventional RL completely diverges.

Cardinality-Invariant Neural Operators

If you train a standard neural network to control a fluid using 10 sensors, that network becomes useless if you add an 11th sensor. Pietro Zanotta et al. solved this by framing control as an operator learning problem. Their architecture, CINOC (Cardinality-Invariant Neural Operator Control), maps the continuous sensor function space directly to the actuator function space. Because the network learns the underlying operator rather than a grid mapping, it achieves zero-shot scalability: a policy trained on a small number of agents can immediately control a population 60 times larger without retraining.

Grid-Based Policy (Rigid):
[10 Sensors] ───> ( Fixed Neural Net ) ───> [10 Actuators]
(Fails if sensor count changes)

Neural Operator Policy (Fluid):
[Continuous Sensor Field] ───> ( Neural Operator ) ───> [Continuous Actuator Field]
(Works zero-shot at any resolution or sensor count)

Taming Chaos and Partial Observations

When PDEs become chaotic, keeping track of the system state becomes incredibly difficult due to limited sensor placement. Max Weissenbacher et al. introduced CHAROT, an attention-based memory architecture that allows RL agents to robustly control chaotic PDEs under highly partial observations.

To make these data-driven policies safer and faster to train, researchers often combine them with physics-based insights. For instance, Xiangyuan Zhang et al. introduced a "reduce-then-design-then-adapt" framework that uses a simplified model-based controller to provide a stable "warm start" for model-free policy optimization, drastically improving training stability. Similarly, Mouhacine Benosman combined data-driven Dual Ensemble Kalman Filters (EnKF) with Lyapunov-based designs to create a highly robust, computationally efficient framework for stabilizing nonlinear PDEs under unknown external disturbances.

Summary

By treating physical systems as inherently continuous, modern infinite-dimensional control avoids the pitfalls of early discretization. Whether through the rigorous lens of operator-valued kernels and Sum-of-Squares optimization, or via the scalable power of neural operators and function-valued reinforcement learning, preserving the spatial integrity of PDEs allows us to design controllers that are more stable, more scalable, and fundamentally aligned with the physics of the real world.

Key Foundational Work

Sliding Mode Control Design for Linear Evolution Equations with Uncertain Measurements and Exogenous Perturbations (2017) — contributed 2 concept elements to this direction.
A Convex Sum-of-Squares Approach to Analysis, State Feedback and Output Feedback Control of Parabolic PDEs (2017) — contributed 1 concept elements to this direction.
Optimization With Affine Homogeneous Quadratic Integral Inequality Constraints (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_infinite_dimensional_system_control_e68c83fc,
  title = {infinite-dimensional system control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/infinite-dimensional-system-control/rdr_e68c83fc6aab428fa632968883fb5d37},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_infinite_dimensional_system_control_e68c83fc_v1,
  title = {infinite-dimensional system control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/infinite-dimensional-system-control/rdr_e68c83fc6aab428fa632968883fb5d37/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake