PaperlayerView exact revision on arXiv

Source-linked AI summary

Global error estimators for parametric monotone nonlinearities and neural approximations

Pablo Cortés Castillo, Wolfgang Dahmen, Jay Gopalakrishnan

arXiv:2610.05767v1math.NAcs.NAstat.ML

TL;DR

A computable error estimator is sought for any candidate function w. The paper splits the nonlinear operator into a linear part and a strongly monotone closure, using a computable surrogate for the closure term and broken test spaces for local dual-norm computations. The two estimators are globally variationally correct, with trial-norm errors proportional to corresponding residuals; the MFOSLS and MDPG formulations share discrete subspaces and identical variational-correctness constants.

  • Problem

    A computable error estimator is sought for any candidate function w.

    Source: §1 · p. 2

  • Method

    The paper splits the nonlinear operator into a linear part and a strongly monotone closure, using a computable surrogate for the closure term and broken test spaces for local dual-norm computations.

    Source: §1 · p. 2, §2.2 · p. 8

  • Results

    The two estimators are globally variationally correct, with trial-norm errors proportional to corresponding residuals; the MFOSLS and MDPG formulations share discrete subspaces and identical variational-correctness constants.

    Source: §3.1 · p. 12, §7 · p. 27

  • Takeaways & Limitations

    Because the estimators are computable for arbitrary trial functions, they can serve as loss functions for neural-network approximations of parameter-to-solution maps.

    Source: §1 · p. 1, §7 · p. 27

    The error–loss relationship depends on parameter-dependent trial norms, and whether the observed MFOSLS behavior is a locking phenomenon remains open.

    Source: §7 · pp. 27–28

Abstract

from arXiv · show

We construct computable error estimators, which double as loss functions for neural networks, for a class of parametric nonlinear partial differential equations with a monotonicity property, and prove that they are globally reliable and efficient. The value of such a loss function is bounded above and below by the squared error in the natural trial norm, for every trial function, not merely for those near the exact solution; this global property rests on monotonicity. The construction rests on splitting the nonlinear operator into a linear part and a strongly monotone closure, and on measuring the linear residual in a discrete dual norm. Since the closure contributes a dual norm that admits no closed form when the trial norm is stronger than an $L_2$ norm, the estimator is built around a computable surrogate for it, required only to satisfy a pairing bound and a Lipschitz bound. The main theorem then yields two-sided bounds with explicit constants and covers two instances. A first-order system least-squares estimator on conforming trial spaces is the first instance studied: its two-sided bound holds on the whole trial space and therefore applies to arbitrary approximations, including those not in a discrete finite element space. The second instance is a discontinuous Petrov-Galerkin estimator on trial spaces of finite element functions for each parameter value, built with broken test spaces, for which no conformity is required and the dual norm is computed by independent element-local problems. All assumptions are verified for a model class of nonlinear fluxes, with the constants tracked explicitly in terms of the parameter range. Being computable for an arbitrary input, both estimators serve as variationally correct loss functions for neural network approximations of parameter-to-solution maps.

1. Introduction

The introduction addresses the difficulty of obtaining computable, globally valid error control for nonlinear parametric PDEs. It combines monotonicity with computable residual estimators, developed as both a posteriori estimators and neural-network losses.

1. Introduction

  • Because each estimator is computable for arbitrary inputs, it can serve as a variationally correct neural-network loss for learning parameter-to-solution maps.The introduction situates this alongside data-driven regression and residual-based physics-informed approaches to operator learning.Source: §1 · pp. 1, 3
  • Nonlinear residual bounds are typically local near a regular solution, while evaluating the natural dual residual norm requires solving a global problem.The target is a computable estimator for any trial function, with two-sided bounds whose constants are independent of discretization.Source: §1 · p. 2
  • Splitting the operator into a linear part and a strongly monotone closure yields global bounds for every trial function.With broken test spaces, the linear residual’s dual norm decouples into computable element-local problems, at the cost of hybrid trial spaces and a Fortin operator.Source: §1 · p. 2
  • The closure’s dual norm may be unavailable in stronger-than-L2 trial norms, so the framework uses a computable surrogate satisfying pairing and Lipschitz bounds.The resulting theorem gives global reliability and efficiency with constants depending explicitly on closure monotonicity and Lipschitz data and the Fortin operator norm.Source: §1 · p. 2
  • MFOSLS uses conforming trial spaces and bounds error across the entire trial space, while MDPG uses a hybrid formulation with broken tests.The MFOSLS bound applies to arbitrary approximations, including ones outside discrete finite element spaces; MDPG includes a Fortin operator and an oscillation term that vanishes for piecewise constant data.Source: §1 · p. 2

2. Monotone loss functions

The section develops a globally reliable and efficient estimator for monotone nonlinear problems by splitting the operator into a linear part and a strongly monotone closure. Its abstract result yields least-squares and DPG loss functions, with different trial-space and computability requirements.

2.1. The error estimator under general assumptions.

  • The estimator splits the residual into a linear part and a nonlinear closure part, using monotonicity to obtain global rather than merely local error control.The closure dual norm can be replaced by a computable surrogate satisfying the stated bounds.Source: §2 · p. 4, §2.1 · p. 6
  • The abstract theorem gives two-sided error bounds for the loss under its assumptions, with reliability and efficiency constants controlled by the framework’s operator and surrogate bounds.A Fortin condition links the trial and discrete test spaces when the dual residual norm is measured on a strict test subspace.Source: §2.1 · pp. 5, 7
  • When the data oscillation vanishes or is comparable to best-approximation error, a loss minimizer is quasi-optimal in error.Taking the full test space with the identity Fortin map makes the oscillation term vanish.Source: §2.1 · p. 8, §2.2 · p. 8

2.2. The least-squares instance.

  • MFOSLS uses a conforming trial space and its two-sided bound holds throughout Uα, including for non-finite-element approximations.Its residual term is computable when the test-space Riesz map is explicitly invertible; for L2-type test spaces it is a squared strong-form residual integral.Source: §2.2 · p. 8

2.3. The DPG instance.

  • MDPG uses broken test spaces so the discrete dual norm of the linear residual is computed through element-local problems.Its closure term uses the canonical dual norm when that norm is computable for the chosen formulation.Source: §2.3 · pp. 8–9

3. Parametric diffusion with nonlinear fluxes

The section recasts nonlinear diffusion with a flux variable and a nonlinear closure, then realizes the estimator in broken-test-space DPG and conforming least-squares forms. The DPG closure norm has a computable L2 expression, while MFOSLS uses a computable surrogate and applies across its full trial space.

3. Parametric diffusion with nonlinear fluxes

  • The nonlinear diffusion model is recast with the flux as an additional unknown, separating linear equations from the nonlinear constitutive closure.The flux map is assumed invertible for formulations that use its inverse; the two loss functions use the flux map itself.Source: §3 · pp. 9–10

3.1. The MDPG loss function.

  • MDPG enforces the closure relation in an ultraweak formulation with broken tests, and its closure dual norm reduces to a computable L2 norm.An alternative formulation enforces the inverse closure; both formulations use the same discrete subspaces and have identical variational-correctness constants.Source: §3.1 · pp. 11–12

3.2. The MFOSLS loss function.

  • MFOSLS uses a conforming trial space without a mesh or broken spaces, and every loss term is an L2 integral.Its closure expression is generally only an upper bound for the closure dual norm because the trial norm is stronger than the corresponding weighted L2 product norm.Source: §3.2 · p. 12

4. Verification of the assumptions for a class of examples

This section verifies estimator assumptions for a class of nonlinear diffusive fluxes under flux monotonicity and parameter-uniform bounds, establishing global variational correctness for the MDPG loss and global reliability and efficiency for MFOSLS over conforming trial candidates. The MFOSLS result applies without a finite-element or mesh requirement.

4.1. A model class of nonlinear diffusive fluxes.

  • The model flux has a continuous Jacobian with spectrum in [−α^-1β2, −α^-1β1], yielding strong monotonicity and bi-Lipschitz bounds for both the flux and its inverse.The typical profile ϕ(t)=2+(1+t)^−2 satisfies (48b) with β1=1, β2=3; the invertible profile ϕ(t)=2+(1+t)^−1 has β1=2, β2=3 and an explicit inverse that enables the alternative MDPG formulation.Source: §4.1 · pp. 13–14

4.2. Variational correctness of the MDPG loss function.

  • The MDPG loss is globally reliable and efficient for every discrete trial function, with mesh-independent constants; for piecewise-constant data, its error bounds also hold without an oscillation term.The result imposes no proximity-to-solution restriction and allows arbitrary polynomial degree k, although the projection norm may depend on k.Source: §4.2 · p. 19

4.3. Variational correctness of the MFOSLS loss function.

  • For MFOSLS, the identity projection verifies Assumption 2.2, and the kernel of the linear operator is parameterized by divergence-free q and p=∇u.The supplied passage states that verification of Assumptions 2.1 and 2.3 remains; the MFOSLS verification continues in the next section part.Source: §4.2 · p. 19, §4.3 · p. 19
  • For the stated setting, Assumptions 2.1 and 2.3 hold for Bα, Cα, and Mα, including surjectivity and the specified bounds and implication.The proposition specifies constants using κ1, κ2, µ, the Friedrichs constant, and constants from Proposition 4.5.Source: §4.3 · p. 19
  • The MFOSLS loss is globally reliable and efficient for every element of the full trial space Uα = H(div) × ˚H1(Ω) × L2(Ω)d.The two-sided bound follows from the verified assumptions without using finite elements or a mesh.Source: §4.3 · pp. 20–21
  • Unlike the MDPG result, the MFOSLS bound needs no data-oscillation restriction or mesh, but requires conforming q ∈ H(div) and u ∈ ˚H1(Ω) with imposed boundary conditions.Source: §4.3 · p. 21

5. Application to neural network approximations

The paper frames neural approximation of the parameter-to-solution map as a learning problem and trains with an empirical estimator-based loss rather than solution data. The loss is variationally correct when the oscillation term vanishes, bounding squared error with constants independent of the parameter.

5.1. The learning problem.

  • The exact map α→u(α) is framed as the minimizer of population risk over a hypothesis class of functions with inputs x and α.Source: §5.1 · p. 21
  • When osc(ℓ)=0, variational correctness bounds the loss above and below by squared error with constants independent of α.Source: §5.1 · p. 21
  • Estimator-loss training uses sampled parameters without high-fidelity solution data, unlike solution-regression training, which requires such data.Source: §5.1 · p. 21
  • The experiments focus on piecewise-constant coefficients over finitely many subdomains, yielding a finite-dimensional parameter-to-solution map for interface problems.Source: §5.1 · p. 22
  • Hybrid hypotheses use spatial piecewise polynomials with parameter-dependent neural-network coefficients, simplifying boundary-condition enforcement and spatial derivatives while reserving backpropagation for parameter variables.Source: §5.1 · p. 22
  • The neural map uses leaky-ReLU layers, with a low-rank variant when r≪N; trainable weights and biases define Fθ.Source: §5.1 · pp. 22–23

6. Numerical studies

The numerical studies assess loss–error calibration and how network depth, training-set size, polynomial degree, and parameter dimension affect learned solutions. They find that performance depends on training and norm weights, and that the tested architecture reproduces key solution features in a sixteen-parameter example.

6.1. Experimental setup.

  • The experiments compare MDPG and MFOSLS losses for parameter-to-solution learning, using reference solutions only for evaluation rather than training.They are intended to illustrate responses to hypothesis-class hyperparameters and parameter-range effects, not validate the theory or demonstrate large-scale capability.Source: §6 · p. 23, §6.1 · p. 23

6.2. Loss–error calibration.

  • Loss–error ratios vary by about 70-fold for MDPG and 500-fold for MFOSLS across boxes with the same coefficient contrast, as coefficient scale changes the norm weights.The replacement of exact states by refined-mesh references is checked only on eight test parameters, so the calibration assessment is a plausibility check.Source: §6.2 · pp. 24–25

6.3. Network depth.

  • At the tested training budget, depths 1 and 3 finish well above depths 5 and 7; extending depth 5 training to 2 × 10^4 epochs yields only about another factor of two.The depth-5, 2 × 10^4-epoch choice is presented as a practical accuracy–cost compromise, not evidence that training has converged.Source: §6.3 · p. 25, Figure 1 · p. 25

6.4. Training and spatial accuracy.

  • Increasing the training set from 128 to 4096 substantially reduces test loss and flux error, although component errors need not decrease monotonically.Even the largest training set leaves a gap to the FEM benchmark; with fixed batch size, the gain reflects both more data and more optimization steps.Source: §6.4 · pp. 25–26, Table 2 · p. 26
  • Raising polynomial degree from 0 to 1 sharply improves network predictions, while from k = 1 to k = 2 FEM losses fall much faster than network losses.Enriching the trial space reduces discretization error only, so learning error must also fall to benefit from a richer space.Source: §6.4 · pp. 26–27, Table 3 · p. 26
  • For the lowest-order MDPG scalar error, network and FEM values are 0.576 and 0.737, versus 0.101 and 0.0981 for MFOSLS.Because the MDPG FEM result also has poor scalar accuracy, the discrepancy is not attributable solely to network training; it is greatly reduced at k = 1.Source: §6.4 · p. 27

6.5. A sixteen-parameter example.

  • In the sixteen-parameter example, both networks reproduce the displayed sample’s main scalar and flux features, including abrupt flux changes across subdomain interfaces.The experiment uses four times as many parameters and three times the contrast as preceding experiments, without retuning architecture or optimizer settings; one sample does not quantify accuracy.Source: §6.5 · p. 27

7. Conclusions

The two loss functions are globally variationally correct, with an error–loss relation that depends on parameter-dependent trial norms. Experiments find that learning error dominates discretization error and that MFOSLS is slightly more accurate in most experiments.

7. Conclusions

  • Both MFOSLS and MDPG are globally variationally correct, with monotonicity providing global stability in parameter-dependent trial norms.MFOSLS uses L2-type test spaces; MDPG uses DPG concepts because its test space is not self-dual.Source: §7 · p. 27
  • Sufficiently deep networks are needed, trained-network error is dominated by learning error rather than finite element discretization, and MFOSLS is slightly more accurate in most experiments.The experiments also show that the error–loss relationship depends on the parameter.Source: §7 · p. 27
  • Figure 2 pairs coefficient samples on 2 × 2 and 4 × 4 partitions with Qmix reference and network solutions displayed on common row-wise color scales.The reference uses h = 0.05; the networks use h = 0.1.Source: Figure 2 · p. 28
  • At lowest order, MDPG has a large scalar error already present in its finite element solution, unlike MFOSLS; whether this is locking remains unresolved.The authors contrast this with linear diffusion, where DPG was more robust to parameter variation, and leave nonlinear renormalization ideas for future work.Source: §7 · pp. 27–28

Statements and Declarations

  • The work reports NSF and German Research Foundation support, declares no competing interests, and makes Section 6 data available from the corresponding author on reasonable request.Source: Statements and Declarations · p. 28

Related papers

Loading 2610.05767v1…