Pith. sign in

REVIEW 2 major objections 2 minor 5 references

Neural operator policies for PDE control trained on small agent groups transfer directly to much larger groups without retraining.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-29 20:41 UTC pith:NFMF2DSN

load-bearing objection CINOC reframes PDE control as operator learning to get cardinality-invariant policies that transfer across agent counts, with experiments on several regimes, but the mean-field theorem's regularity assumptions look unverified for the architectures used. the 2 major comments →

arxiv 2605.25867 v1 pith:NFMF2DSN submitted 2026-05-25 eess.SY cs.SY

CINOC: Cardinality-Invariant Neural Operator Policies for Scalable PDE Control

Pietro Zanotta , Dibakar Roy Sarkar , Honghui Zheng , Somdatta Goswami , J\'an Drgo\v{n}a This is my paper
classification eess.SY cs.SY
keywords neural operatorsPDE controlcardinality invariancemulti-agent systemsmean-field theorydifferentiable solversswarm control
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The work reformulates multi-agent PDE control as learning an operator that maps entire state fields to continuous control fields. Policies are trained end-to-end with differentiable PDE solvers so that the same network parameters work for any number of sensors or actuators. When agents share one policy and interact only through the physical field, an emergent self-normalization appears that keeps behavior stable as population size changes. A mean-field theorem is supplied to show that the policy gradients obtained from any finite number of agents converge to the gradients of the corresponding continuous control problem. The resulting policies are demonstrated on tracking, stabilization, and density transport tasks across linear, nonlinear, chaotic, and turbulent PDEs.

Core claim

Reformulating PDE control as an operator learning problem that maps state fields to continuous control functions produces policies that are invariant to the number of agents. Because every agent uses the identical policy and agents coordinate solely through the shared physical environment, the learned behavior exhibits self-normalization: a policy trained on a small swarm applies unchanged to a swarm of arbitrary size and tolerates the loss of individual agents. This invariance is supported by a mean-field theorem establishing that policy gradients computed on finite-agent systems converge to those of the continuous control limit.

What carries the argument

The neural operator policy mapping state fields to continuous control functions, trained end-to-end through differentiable PDE solvers, whose cardinality invariance follows from convergence of finite-agent policy gradients to the mean-field limit.

Load-bearing premise

The mean-field convergence result holds for the specific neural operator architectures used and the differentiable PDE solvers faithfully reproduce the dynamics required for the observed invariance.

What would settle it

An explicit numerical check that policy gradients obtained from a finite swarm of agents diverge from the gradients computed in the corresponding continuous mean-field formulation would falsify the claimed explanation for the observed transfer.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A single policy trained on a small population can be deployed on populations of any larger size without retraining or architectural change.
  • The same policy remains functional after the loss of some agents, providing robustness to partial failure.
  • The method applies uniformly to linear, nonlinear, chaotic, and turbulent PDEs for tasks including tracking, stabilization, and density transport.
  • Training through differentiable solvers removes the need for separate discretization or hand-designed controllers when the sensor-actuator count changes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach could remove the need to retrain controllers whenever the number of physical devices changes in deployed systems.
  • Similar self-normalization may appear in other shared-policy multi-agent problems where agents interact only through a common medium.
  • Training cost could be further reduced by simulating only the smallest viable swarm size rather than the target population size.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Cardinality-Invariant Neural Operator Control (CINOC), reformulating multi-agent PDE control as an operator-learning task in which a shared neural operator maps the PDE state field to a continuous control function. Policies are trained end-to-end via differentiable PDE solvers; the resulting policies exhibit cardinality invariance, permitting zero-shot transfer from small training swarms to substantially larger populations and robustness to partial agent loss. A mean-field theorem is supplied to show that policy gradients computed on finite-N systems converge to those of the corresponding continuous control limit. Empirical results are reported for tracking, stabilization, and density-transport tasks on linear, nonlinear, chaotic, and turbulent PDEs.

Significance. If the mean-field convergence holds for the specific operator architectures and the empirical invariance is reproducible, the work would provide a practical route to scalable distributed control without retraining when actuator or sensor counts change. The combination of neural-operator representations with a supporting convergence theorem is a notable strength relative to fixed-dimensional policy approaches.

major comments (2)
  1. [Theorem section] Theorem (mean-field section): the stated convergence of finite-N policy gradients to the continuous limit is asserted without verification of the uniform Lipschitz or compactness conditions on the neural operator that are required for standard mean-field arguments to apply when the operator is composed with the PDE dynamics. The manuscript does not establish that these conditions hold independently of N for the DeepONet- or FNO-style layers employed in CINOC.
  2. [§5] §5 (Experiments): the reported zero-shot transfer and robustness results rely on the differentiable solver faithfully reproducing the continuous limit; no ablation or error analysis quantifies how discretization artifacts in the solver affect gradient convergence or the observed cardinality invariance as N increases.
minor comments (2)
  1. [Notation] Notation for the state field, control operator, and mean-field limit is introduced piecemeal; a consolidated symbol table would improve readability.
  2. [Introduction] Related-work discussion omits several recent papers on mean-field control of PDEs and on neural operators for distributed actuation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and indicate the revisions that will be incorporated.

read point-by-point responses

  1. Referee: [Theorem section] Theorem (mean-field section): the stated convergence of finite-N policy gradients to the continuous limit is asserted without verification of the uniform Lipschitz or compactness conditions on the neural operator that are required for standard mean-field arguments to apply when the operator is composed with the PDE dynamics. The manuscript does not establish that these conditions hold independently of N for the DeepONet- or FNO-style layers employed in CINOC.

    Authors: We acknowledge that the theorem relies on standard mean-field regularity assumptions without an explicit verification that the uniform Lipschitz and compactness conditions hold independently of N for the specific DeepONet and FNO architectures. In the revised manuscript we will add a remark immediately following the theorem statement that lists these conditions and supplies a brief argument that they are satisfied: the operators are finite-depth with globally Lipschitz activations and parameter norms bounded independently of N, which implies the required uniform bounds when composed with the PDE flow under the problem's standing assumptions on the dynamics. If the referee finds this insufficient we are prepared to move a full proof sketch to the appendix. revision: yes

  2. Referee: [§5] §5 (Experiments): the reported zero-shot transfer and robustness results rely on the differentiable solver faithfully reproducing the continuous limit; no ablation or error analysis quantifies how discretization artifacts in the solver affect gradient convergence or the observed cardinality invariance as N increases.

    Authors: We agree that an explicit quantification of discretization effects would strengthen the empirical support for the observed cardinality invariance. In the revised Section 5 we will include a new ablation that reports cardinality-invariance metrics and policy-gradient norms for increasing N under three different spatial/temporal discretization resolutions of the differentiable solver. This will directly illustrate the sensitivity (or lack thereof) of the reported zero-shot transfer results to solver artifacts. revision: yes

Circularity Check

0 steps flagged

No significant circularity; mean-field theorem presented as independent explanatory result

full rationale

The paper's central claim is that CINOC policies, formulated as neural operator mappings from state fields to continuous controls and trained via differentiable PDE solvers, exhibit cardinality invariance due to shared policies and environmental coordination. The mean-field theorem is introduced to explain gradient convergence from finite to continuous limits. No quoted equations or derivations reduce the invariance to a definitional identity, a fitted parameter renamed as prediction, or a self-citation chain. The architecture choice and theorem are independent of the target scalability result, with empirical validation on multiple PDEs providing external content. This matches the default case of a self-contained derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no concrete information on free parameters, axioms, or invented entities; all arrays left empty.

pith-pipeline@v0.9.1-grok · 5761 in / 1140 out tokens · 45188 ms · 2026-06-29T20:41:26.037005+00:00 · methodology

0 comments
read the original abstract

Controlling partial differential equations (PDEs) with learning-based policies remains fundamentally limited by fixed-dimensional representations: policies trained for a specific sensor, actuator, or agent configuration typically fail when the configuration changes. This limitation is particularly severe in multi-agent PDE control, where policies do not scale across population sizes without retraining. We address this challenge by introducing Cardinality Invariant Neural Operator Control (CINOC), reformulating PDE control as an operator learning problem that maps state fields to continuous control functions and trains them end-to-end through differentiable PDE solvers, yielding policies that naturally adapt to varying sensor and actuator configurations. Remarkably, CINOC policies trained on small swarms exhibit cardinality invariance, allowing for zero-shot transfer to significantly larger populations as well as robustness to partial agent failure. This scalability arises from agents sharing a common policy and coordinating through their physical environment, which produces an emergent self-normalization effect. To explain this phenomenon, we provide a theorem grounded in mean-field theory demonstrating that policy gradients computed from finite-agent systems converge to those of a continuous control limit. Empirically, we validate CINOC on tracking, stabilization, and density transport across linear, nonlinear, chaotic, and turbulent PDEs.

Figures

Figures reproduced from arXiv: 2605.25867 by Dibakar Roy Sarkar, Honghui Zheng, J\'an Drgo\v{n}a, Pietro Zanotta, Somdatta Goswami.

Figure 1
Figure 1. Figure 1: Cardinality-Invariant Neural Operator Policy. A shared neural operator maps PDE field observations to control actions for multiple agents, enabling zero-shot scalability across agent populations. Framework handles diverse tasks: (top right) chaos stabilization, (middle right) reference tracking, and (bottom right) density transport. Pope, 2001) to heat transfer and chemical reactions (Crank, 1979; Turing, … view at source ↗
Figure 2
Figure 2. Figure 2: Schematic of the multi-agent operator learning for PDE control. The PDE solver acts as a differentiable layer. The operator policy Gθ (shared among the agents) takes error observations and outputs control actions. Gradients propagate through the solver to update θ. the vector-valued function c enforces constraints such as maximum control amplitudes, kinematic limits, and spatial obstacle avoidance. Through… view at source ↗
Figure 3
Figure 3. Figure 3: Stabilization of 2D Kuramoto-Sivashinsky chaos. Top: Natural (uncontrolled) evolution of u(x, t) showing persistent chaotic dynamics. Bottom: Controlled evolution of u(x, t) with policy applied at t = 0. The system, initially in a developed chaotic state, rapidly converges to the zero equilibrium within approximately 5 seconds. Furthermore, as detailed in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Density Transport via Mobile Advection. (Top) Un￾controlled evolution of density. (Middle) Controlled evolution where agents (dots) coordinate to “blow” the density blob into the target zone (green circle). (Bottom) Performance metrics: (left) Wasserstein distance showing controlled vs. uncontrolled conver￾gence; (center) average velocity ˙ξi of the actuators; (right) average control action |vi|. policy le… view at source ↗
Figure 6
Figure 6. Figure 6: Cardinality invariance across diverse PDE bench￾marks. The plots show the Relative MSE as a function of the swarm size relative to the original training scale. Left: Perfor￾mance regimes for decreasing swarm sizes (M ≤ Mtrain). Right: Performance regimes for increasing swarm sizes (M ≥ Mtrain). The 100% dashed line indicates the baseline performance at the training population. Policies deployed in 2D envir… view at source ↗
Figure 7
Figure 7. Figure 7: Fisher-KPP 1D Tracking. (Top) Uncontrolled evolution of the reaction-diffusion wavefront. (Bottom) Controlled trajectory wherein mobile actuators modulate the density field. The performance metrics demonstrate an efficient convergence: an initial high￾velocity approach rapidly minimizes the tracking error, followed by a transition to a low-energy station-keeping configuration. 0 100 200 300 Time Step Natur… view at source ↗
Figure 8
Figure 8. Figure 8: Heat 1D Tracking. (Top) Natural evolution demonstrating characteristic thermal dissipation. (Bottom) Controlled evolution maintaining a multi-modal target distribution. The policy exhibits rapid convergence (MSE attenuation) followed by a decay in control effort as the system reaches a steady-state maintenance phase. C.4. Kuramoto-Sivashinsky 1D: Multi-Scale Stabilization of Chaotic Dynamics Finally, we in… view at source ↗
Figure 9
Figure 9. Figure 9: Unconstrained Heat 2D Tracking. Decentralized DPC performance on 2D heat equation control. (Left) Tracking MSE distribution shows significant error reduction for controlled scenarios regardless of obstacle presence (N=100). (Middle) Obstacle-free agents exhibit more aggressive motion toward high-error regions. (Right) Obstacle-aware agents execute prompt collision avoidance near safety boundaries (dashed c… view at source ↗
Figure 10
Figure 10. Figure 10: Scale-Invariant Stabilization (KS 1D). (Left) Natural evolution heatmaps illustrating the increase in chaotic complexity with domain size L. (Right) Energy evolution for natural (dashed) vs. controlled (solid) trajectories. The policy consistently achieves stabilization regardless of domain scale, demonstrating effective control of the underlying chaotic manifold. D. Ablation studies D.1. Empirical Verifi… view at source ↗
Figure 11
Figure 11. Figure 11: Scaling of Steady-State Total Effort (Fisher-KPP). The plot shows the mean Pu 2 i computed over the final 70% of trajectories against the number of agents M on a log-log scale. For sufficient effort penalties (λu ≥ 0.05), the total effort decreases with a slope indicating O(1/M) scaling, confirming cooperative self-normalization. Insufficient penalties (λu ≤ 0.005) on control inputs lead to a breakdown of… view at source ↗
Figure 12
Figure 12. Figure 12: (Left) illustrates the performance consistency across solver fidelities. Visually, the spatiotemporal evolution of the field in the top row (64 × 64) and bottom row (512 × 512) exhibits a high degree of coherence, indicating that the policy does not rely on grid-scale artifacts of the coarse solver. Quantitatively, this robustness is confirmed by the system energy trajectories shown in [PITH_FULL_IMAGE:f… view at source ↗
Figure 13
Figure 13. Figure 13: Robustness Analysis. Performance comparison between models trained with different noise levels. Inference is performed in a noise-less environment. Policies trained with marginal sensor noise exhibit better cardinality invariance. the self-normalization property for larger swarm sizes compared to the baseline policy, which fails to maintain the necessary inverse effort scaling (O(1/M)) as the population g… view at source ↗
Figure 14
Figure 14. Figure 14: Sensitivity to Field of View (FOV). Comparison of control error (MSE) against swarm size for different observation radii on the FKPP task. While Global FOV (100%) performs well for small swarms, it creates oscillatory instability at large M due to spurious coupling. Local FOV (10% - 20%) acts as a necessary regularizer, maintaining stability as population scales. D.4. What is the Role of Field of View Dim… view at source ↗
Figure 15
Figure 15. Figure 15: Robustness Analysis. Comparative evaluation of Mean Squared Error (MSE) against deployment population size (M) in low￾and mid-noise environments. The Baseline policy, trained in a noise-free environment, (solid black) degrades significantly as M increases, indicating overfitting to the training physics. Policies trained with sensor-noise injection (dashed blue) exhibit robust scalability, reducing error b… view at source ↗
Figure 16
Figure 16. Figure 16: The Role of Noise for more Robust Self-Normalization. Log-log plot of total control effort vs. agent count. The Baseline policy (black), trained on a zero-noise environment, diverges from the ideal O(1/M) scaling slope at high M, indicating a loss of coordination and the onset of competitive over-actuation. The policy trained in a low noise-scenario (σz = 0.01 and σu = 0.02) maintains the correct power la… view at source ↗
Figure 17
Figure 17. Figure 17: Zero-Shot Robustness to Parametric Shifts (FKPP Tracking). The policy, trained on fixed parameters (ν = 0.005, ρ = 3), is deployed on mismatched dynamics to test generalization. (Left) The agent exhibits graceful degradation: tracking remains stable even when natural diffusion is eliminated (ν = 0.0), maintaining coherent behavior despite the expected increase in tracking error. (Right) Agents track targe… view at source ↗
Figure 18
Figure 18. Figure 18: Zero-Shot Transfer in 2D Turbulence (Stabilization). The policy, trained at ν = 5 × 10−4 , is evaluated on more turbulent flows. (Left) Vorticity fields show suppression of chaotic structures even at ν = 2 × 10−4 . (Right) Enstrophy evolution confirms exponential decay across all tested regimes. optimization epochs per rollout with a minibatch size of 200. For MATD3, the models were trained using a batch … view at source ↗
Figure 19
Figure 19. Figure 19: Scalability under Physical Shift. Residual enstrophy vs. agent count for different fluid viscosities. Performance improves monotonically with swarm size, showing that increased actuation density can compensate for the tougher dynamics of lower viscosity flows. While deploying this framework on a novel PDE currently requires task-specific empirical tuning, our experiments reveal several consistent heuristi… view at source ↗
Figure 20
Figure 20. Figure 20: Relative Scalability (Zero-Shot) of MARL-trained Operator Policies. Comparison of zero-shot scalability on the 2D Turbulence stabilization task. The plot evaluates Relative Enstrophy (%) as a function of the inference-time agent count (M), with all models trained at M = 64. DeepONet policies trained with model-free algorithms (MAPPO and MATD3) exhibit almost identical cardinality invariance properties to … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

5 extracted references · 3 canonical work pages

  1. [1]

    URL https://arxiv.org/abs/2601. 15015. GitHub: https://github.com/safe-autonomous- systems/fluidgym. Bhan, L., Shi, Y ., and Krstic, M. Neural operators for bypassing gain and control computations in PDE back- stepping.IEEE Transactions on Automatic Control, 69 (8):5310–5325, 2023. Bhan, L., Bian, Y ., Krstic, M., and Shi, Y . PDE control gym: A benchmark...

    2023

  2. [2]

    Bradbury, J., Frostig, R., Hawkins, P., Johnson, M

    URL https://proceedings.mlr.press/ v283/bousias25a.html. Bradbury, J., Frostig, R., Hawkins, P., Johnson, M. J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., and Zhang, Q. JAX: composable transformations of Python+NumPy programs, 2018. URL http://github.com/jax-ml/jax. Carrillo, J. A. and Choi, Y .-P. Mean-field l...

    2018

  3. [3]

    cc/paper_files/paper/2018/file/ 842424a1d0595b76ec4fa03c46e8d755-Paper

    URL https://proceedings.neurips. cc/paper_files/paper/2018/file/ 842424a1d0595b76ec4fa03c46e8d755-Paper. pdf. de Jong, T. O., Shukla, K., and Lazar, M. Deep Operator Neural Network Model Predictive Control.arXiv preprint arXiv:2505.18008, 2025. Demetriou, M. A. Guidance of mobile actuator-plus-sensor networks for improved control and estimation of dis- tr...

    work page doi:10.1016/j.neucom.2025.130716 2018

  4. [4]

    arXiv preprint arXiv:2006.07869 (2020)

    URL https://openreview.net/forum? id=c_E8kFWfhp0. Pan, J., Xing, J., Reiter, R., Zhai, Y ., Aljalbout, E., and Scaramuzza, D. Learning on the Fly: Rapid Policy Adap- tation via Differentiable Simulation. InIEEE Robotics and Automation Letters, 2026. Pan, Y ., Farahmand, A.-m., White, M., Nabi, S., Grover, P., and Nikovski, D. Reinforcement learning with f...

    work page arXiv 2026

  5. [5]

    maintenance mode

    URL https://openreview.net/forum? id=XopATjibyz. Theraulaz, G. and Bonabeau, E. A brief history of stigmergy. Artificial life, 5(2):97–116, 1999. Turing, A. M. The chemical basis of morphogenesis.Bul- letin of mathematical biology, 52(1):153–197, 1990. Vasanth, J., Rabault, J., Alc ´antara- ´Avila, F., Mortensen, M., and Vinuesa, R. Multi-agent reinforcem...

    work page arXiv 1999