A model reduction method based on nonlinear optimization for multiscale stochastic optimal control problems

Jingyi Zhang; Lingling Ma; Qiuqi Li

arxiv: 2510.11325 · v2 · pith:NJL5TZ4Enew · submitted 2025-10-13 · 🧮 math.OC

A model reduction method based on nonlinear optimization for multiscale stochastic optimal control problems

Lingling Ma , Jingyi Zhang , Qiuqi Li This is my paper

Pith reviewed 2026-05-21 20:42 UTC · model grok-4.3

classification 🧮 math.OC

keywords model reductionstochastic optimal controldata-driven methodsmultiscale problemsgradient-based optimizationPDE constraintsreduced-order modeling

0 comments

The pith

A data-driven reduced-order model for stochastic optimal control is built by minimizing output error with gradient optimization on input-output data alone.

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

The paper establishes a non-intrusive framework for approximating the parameter-to-output mapping in stochastic optimal control problems with PDE constraints. It constructs an L2-optimal reduced-order model by minimizing output error through gradient-based optimization that uses only input-output pairs. Training data comes from offline GMsFEM solves for multiscale cases, and the resulting model keeps computational cost independent of the full PDE dimension. A sympathetic reader would care because this enables accurate real-time control in settings where direct simulation of the original system is prohibitive.

Core claim

An L2-optimal reduced-order model is constructed to directly approximate the parameter-to-output mapping. The model is obtained by minimizing the L2 norm of the output error via gradient-based optimization, requiring only input-output data without access to the full-order system matrices or state variables, with high-fidelity training data generated by GMsFEM.

What carries the argument

The L2-optimal reduced-order model obtained by gradient-based minimization of output error, which directly approximates the parameter-to-output mapping from input-output data.

If this is right

The reduced model keeps computational complexity independent of the original PDE dimension.
Accuracy is preserved specifically for control-relevant outputs in the stochastic setting.
The approach applies directly to stochastic diffusion and advection-diffusion equations.
The framework supports real-time applications because its cost does not scale with the full system size.

Where Pith is reading between the lines

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

The same data-driven optimization could be tested on control problems with time-dependent or nonlinear PDE constraints if suitable input-output data can be generated.
Connecting this reduction step to existing stochastic optimization solvers might further reduce online computation time beyond what is shown.
The method's performance on parameters with different correlation lengths or higher-dimensional random fields remains open for direct numerical checks.

Load-bearing premise

The offline input-output data must be representative enough for the gradient optimization to produce a reduced model that accurately captures control-relevant outputs across the stochastic parameter space.

What would settle it

Compute control outputs with the reduced model on a fresh set of stochastic parameters outside the training data and compare them to full-order GMsFEM results; large deviation in output error would falsify the approximation claim.

read the original abstract

This paper proposes a non-intrusive, data-driven reduced-order modeling framework for stochastic optimal control problems governed by partial differential equations. The control problem is formulated with a quadratic cost functional and stochastic PDE constraints, and an L2-optimal reduced-order model is constructed to directly approximate the parameter-to-output mapping. The model is obtained by minimizing the L2 norm of the output error via gradient-based optimization, requiring only input-output data without access to the full-order system matrices or state variables. To efficiently generate high-fidelity training data for multiscale problems, the Generalized Multiscale Finite Element Method (GMsFEM) is employed as an offline solver. The proposed framework ensures accuracy in control-relevant outputs while maintaining computational complexity independent of the original PDE dimension, making it suitable for real-time applications. Numerical experiments on stochastic diffusion and advection-diffusion equations demonstrate the accuracy, efficiency, and robustness of the method.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

A data-driven non-intrusive ROM for stochastic multiscale OCP via GMsFEM data and gradient L2 optimization, with thin theoretical support.

read the letter

This paper gives a practical non-intrusive way to build reduced models for stochastic optimal control of multiscale PDEs. They generate input-output pairs offline with GMsFEM, then fit a reduced model by minimizing the L2 output error through gradient-based nonlinear optimization, all without needing the full system matrices or states. The online phase stays cheap and independent of the original dimension, which suits real-time control use cases. The experiments on stochastic diffusion and advection-diffusion problems indicate the approach can preserve control-relevant accuracy at lower cost. That combination of GMsFEM data generation with direct L2-optimal fitting is the concrete extension they contribute. It builds on existing data-driven reduction ideas without claiming a first-principles breakthrough. The main soft spot is the lack of supporting analysis. The abstract and description provide no quantitative error bounds, convergence results for the optimization, or checks on how training data coverage affects the outcome. Gradient descent on a non-convex loss from multiscale stochastic outputs can land in poor local minima, and nothing in the reported work addresses whether that happens or how to mitigate it. The claim that the reduced model captures the parameter-to-output map reliably therefore rests mostly on the numerical examples shown. This is aimed at researchers in computational optimal control who need implementable reduction tools for high-dimensional stochastic problems. A reader working on data-driven methods for PDE-constrained optimization would find the setup and results useful to examine. It deserves a serious referee because the method is concrete, the experiments are there to evaluate, and the gaps are fixable with added analysis rather than fatal to the core idea.

Referee Report

2 major / 2 minor

Summary. This paper proposes a non-intrusive, data-driven reduced-order modeling framework for multiscale stochastic optimal control problems governed by PDEs. An L2-optimal reduced model is constructed to approximate the parameter-to-output mapping by minimizing the L2 norm of the output error via gradient-based nonlinear optimization, using only input-output data generated offline by GMsFEM without access to full-order matrices or states. The approach is tested numerically on stochastic diffusion and advection-diffusion equations, with claims of accuracy, efficiency, and suitability for real-time applications.

Significance. If the optimization reliably produces accurate reduced models that preserve control performance, the non-intrusive character and independence from full-order system details would represent a practical advance for high-dimensional stochastic PDE-constrained optimization. The combination of GMsFEM data generation with direct output-error minimization addresses a relevant computational bottleneck, but the lack of supporting analysis for global optimality or error control reduces the immediate significance.

major comments (2)

[Numerical optimization procedure (around the description of gradient-based minimization)] The central claim that gradient-based minimization produces an L2-optimal reduced-order model is load-bearing for the entire framework. The optimization problem for fitting a parametric reduced model to multiscale stochastic outputs is expected to be non-convex; the manuscript provides no analysis, multiple random initializations, or comparisons against global optimization methods to show that reported solutions avoid poor local minima whose error differs substantially from the true L2 minimizer.
[Numerical experiments section] No quantitative a-posteriori error bounds, convergence rates, or explicit comparison of the reduced-model control cost against the full-order GMsFEM solution are supplied. The numerical experiments are described as demonstrating accuracy, but without these metrics it is impossible to verify that the reduced model captures control-relevant outputs uniformly across the stochastic parameter space.

minor comments (2)

[Abstract] The abstract states that the framework 'ensures accuracy'; this should be rephrased to 'demonstrates accuracy in the reported experiments' to reflect the numerical character of the evidence.
[Model reduction formulation] Notation for the reduced model parameters and the precise form of the L2 loss functional should be introduced with explicit equations rather than descriptive text only.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point by point below and indicate the revisions we will make to strengthen the presentation and empirical support for the method.

read point-by-point responses

Referee: The central claim that gradient-based minimization produces an L2-optimal reduced-order model is load-bearing for the entire framework. The optimization problem for fitting a parametric reduced model to multiscale stochastic outputs is expected to be non-convex; the manuscript provides no analysis, multiple random initializations, or comparisons against global optimization methods to show that reported solutions avoid poor local minima whose error differs substantially from the true L2 minimizer.

Authors: We agree that the underlying optimization problem is non-convex and that gradient-based minimization does not guarantee a global L2 minimizer. The manuscript emphasizes the practical utility of the resulting reduced models rather than providing theoretical global optimality guarantees, which would require substantial additional analysis beyond the scope of the current work. In the revised manuscript, we will add numerical results from multiple random initializations across the test cases to demonstrate consistency of the obtained reduced models and their output errors. We will also include a comparison against a global optimization method (e.g., differential evolution) on a simplified low-dimensional instance to show that the local solutions achieve errors close to those from global search. These additions will provide empirical support for the reliability of the reported results without claiming theoretical global optimality. revision: yes
Referee: No quantitative a-posteriori error bounds, convergence rates, or explicit comparison of the reduced-model control cost against the full-order GMsFEM solution are supplied. The numerical experiments are described as demonstrating accuracy, but without these metrics it is impossible to verify that the reduced model captures control-relevant outputs uniformly across the stochastic parameter space.

Authors: We acknowledge that the current numerical section relies primarily on relative output errors and visual comparisons rather than explicit control-cost metrics or a-posteriori bounds. In the revised manuscript we will add direct comparisons of the optimal control costs obtained with the reduced-order model versus the full-order GMsFEM solver for the reported test cases. Deriving rigorous a-posteriori error bounds or convergence rates is challenging in this non-intrusive, data-driven setting because the method deliberately avoids access to full-order operators or states; we will clarify this limitation in the text. The added control-cost comparisons will provide quantitative evidence that the reduced model preserves control performance across the stochastic parameter space. revision: partial

Circularity Check

0 steps flagged

No circularity: derivation is an explicit data-driven optimization on externally generated inputs

full rationale

The paper constructs the reduced-order model by explicitly minimizing an L2 output-error objective over input-output pairs generated offline by GMsFEM. This is a standard supervised fitting procedure whose output is defined to be the minimizer of that external loss; it does not reduce by construction to any internal definition, self-citation chain, or renamed known result. The central claim therefore remains an independent numerical statement about the quality of the obtained minimizer rather than a tautology.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Abstract-only review prevents exhaustive identification; the framework implicitly relies on standard assumptions about existence and uniqueness of solutions to the stochastic PDE optimal control problem and on the ability of GMsFEM to produce sufficiently accurate training data.

axioms (2)

domain assumption Existence of solutions to the stochastic PDE-constrained optimal control problem with quadratic cost.
Required for the control problem formulation to be well-posed before reduction.
domain assumption GMsFEM offline solver produces high-fidelity input-output pairs representative of the full-order mapping.
Central to generating training data without full-order matrices.

pith-pipeline@v0.9.0 · 5682 in / 1467 out tokens · 31452 ms · 2026-05-21T20:42:17.322588+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

minimizing the L2 norm of the output error via gradient-based optimization, requiring only input-output data without access to the full-order system matrices or state variables
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

L2-optimal reduced-order model ... parameter-separable structure

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.