PaperlayerView exact revision on arXiv

Source-linked AI summary

Stable time-stepping via residual minimization: finite-element and neural-network approximations for transient parabolic problems

Judit Muñoz-Matute, Sergio Rojas

arXiv:2609.27548v1math.NAcs.NA

TL;DR

Transient coercive variational problems call for stable minimum-residual time stepping, but the time-step operator's norm couples temporal and spatial scales. This paper uses a steady spatial test norm scaled by the time step and derives estimates for conforming, broken-test, finite-element, and neural-network settings. Experiments report expected finite-element convergence and improved error versus spatial complexity under residual-based marking compared with uniform refinement.

  • Problem

    Using the full Backward Euler operator to define the test norm couples the L2 and elliptic scales, making estimates in standard discrete parabolic energy quantities less direct.

    Source: §1 · p. 2

  • Method

    The framework represents each Backward Euler residual in the steady spatial test norm scaled by the time step, then adds a complementary residual contribution for finite-dimensional test-space components that are missed.

    Source: §5.2 · pp. 15–16, §7 · p. 23

  • Results

    Residual-based marking improves error versus spatial complexity compared with uniform refinement, while the augmented indicator more consistently localizes reference error along adaptive sequences.

    Source: §7 · p. 23

  • Takeaways & Limitations

    The framework supplies stability estimates and fully discrete reliability estimates for conforming and broken-test formulations, with computable residual decompositions also applicable to neural-network trials.

    Source: §7 · p. 23

    Efficiency estimates are not uniform reverse bounds as τ → 0, broken-setting efficiency includes the skeleton error, and adaptive computations use one spatial mesh over the whole time interval per cycle.

    Source: §7 · p. 23

Abstract

from arXiv · show

We propose a time-stepping minimum-residual (MinRes) framework for transient coercive variational problems, applicable to finite-element and neural-network trial approximations. For the Backward Euler scheme, we measure the residual at each time level in the dual norm induced by the steady test-space norm, scaled by the time step. This choice yields stability estimates controlling a discrete parabolic energy quantity, up to the time-discretization defect. We develop conforming and broken-test formulations and specialize the analysis to diffusion-advection-reaction problems. For finite-dimensional test spaces, we derive fully discrete reliability estimates by supplementing the computable discrete residual with a complementary contribution that accounts for residual components the test space does not resolve. We then extend the framework to neural-network trial classes and obtain computable residual decompositions for conforming and broken polynomial test spaces. Numerical experiments confirm the expected finite-element convergence rates, examine the effect of test-space refinement for a global space-time neural approximation, and demonstrate residual-driven spatial refinement for a transient problem with a moving localized feature.

1 Introduction

The introduction motivates a time-stepping minimum-residual framework that uses a steady spatial test norm to analyze transient problems. It extends this approach to finite-element and neural-network trials, with residual contributions that account for components missed by finite-dimensional test spaces.

1 Introduction

  • Using the steady spatial test-space inner product scaled by the time step keeps the residual norm independent of the time-step size and supports a classical Backward Euler energy argument.The alternative time-step-operator norm couples the L2 and elliptic scales, making estimates in standard discrete parabolic energy quantities less direct.Source: §1 · p. 2
  • The framework separates representation-independent time-stepping stability analysis from the later discretization of the residual norm.The analysis applies to arbitrary trial sequences, not only finite-dimensional linear spaces or a particular minimum-residual algorithm.Source: §1 · p. 2
  • For neural trials, residual evaluation at discrete time levels uses spatial quadratures rather than integration over the full space-time domain.This retains a global space-time neural approximation while keeping numerical integration d-dimensional, a relevant distinction given quadrature cost and accuracy [20] [4].Source: §1 · p. 2
  • Fully discrete reliability requires supplementing the detected residual with a complementary contribution for residual components that the finite-dimensional test space misses.A small discrete residual alone need not control the error when the test space does not detect the full residual.Source: §1 · p. 3
  • The paper establishes conforming and broken-test stability and efficiency estimates, fully discrete residual comparisons, and computable local contributions for diffusion–advection–reaction problems.It also extends the framework to global space-time neural-network trial classes.Source: §1 · p. 3

2 Preliminaries

The preliminaries set out an abstract coercive parabolic problem and its diffusion–advection–reaction realization. They then specify conforming and broken-test variational settings and the assumptions used for the latter.

2.1 Abstract problem and notation

  • The abstract problem seeks a time-dependent solution with prescribed initial data, driven by a time-independent continuous and coercive bilinear form.The setting uses a bounded Lipschitz domain in dimensions one through three and a Hilbert space V continuously embedded in L2(Ω).Source: §2.1 · p. 3
  • The diffusion–advection–reaction operator combines diffusion, advection, and reaction with time-independent coefficients; the heat equation is recovered by setting ϵ = 1, β = 0, and γ = 0.The stated coefficients satisfy ϵ > 0, β ∈ W 1,∞(Ω)d, and γ ∈ L∞(Ω).Source: §2.1 · pp. 3–4

2.2 Conforming and broken variational settings

  • The conforming setting uses the abstract variational problem, while the broken-test setting uses elementwise H1 tests and a skeleton variable in its variational formulation.The broken formulation assumes boundedness, a mesh-independent inf–sup condition, and field coercivity; its bilinear form includes a normal-trace pairing.Source: §2.2 · pp. 4–5
  • For diffusion–advection–reaction problems, the broken bilinear form includes the elementwise diffusion and advection-reaction terms together with a skeleton normal-trace term.For compatible solutions, elementwise integration by parts links the broken formulation to the physical diffusive normal trace.Source: §2.2 · p. 5

3 Semidiscrete Riesz representatives

The semidiscrete analysis represents the Backward Euler residual for arbitrary trial sequences and establishes stability and efficiency estimates in both conforming and broken-test settings. The stability bounds account for the time-discretization defect, while the efficiency estimates are not uniform reverse bounds as the time step tends to zero.

3 Semidiscrete Riesz representatives

  • The semidiscrete framework associates a residual Riesz representative with arbitrary trial approximations at discrete time levels and does not require a particular trial-space representation or minimum-residual algorithm.The analysis is developed for Backward Euler, and the residual-scaling strategy is stated to adapt to other implicit schemes with scheme-dependent stability arguments and consistency terms.Source: §3 · p. 5
  • The consistency contribution to the squared stability estimates is of order τ^2 under the stated additional regularity assumption on the exact solution.The defect is the Backward Euler consistency defect; the consistency analysis assumes ∂ttu ∈ L2(0, T; L2(Ω)).Source: §3 · p. 5

3.1 Conforming formulation

  • In the conforming formulation, the semidiscrete residual representatives satisfy a stability estimate and a separate efficiency estimate for the error quantities.The stability argument uses the error equation, the L2 polarization identity, and coercivity of the spatial bilinear form.Source: §3.1 · p. 6

3.2 Broken-test formulation

  • The broken-test formulation likewise provides semidiscrete stability and efficiency estimates under its stated assumptions.The efficiency bound is not a τ-uniform reverse bound for the stability quantities; it involves the full trial-space error, including the skeleton component, whereas stability controls the field component's parabolic energy.Source: §3.2 · pp. 7–8

4 Fully discrete FE-based MinRes formulations

The fully discrete FE formulations minimize a time-level residual in a finite-dimensional test space, for both conforming and broken tests. Reliability compares that computable residual with the continuous residual, accounting for components the test space may miss.

4 Fully discrete FE-based MinRes formulations

  • At each time level, the fully discrete conforming and broken formulations minimize the discrete residual in a dual norm induced by a finite-dimensional test space.The resulting minimization has a mixed saddle-point characterization involving a discrete residual representative.Source: §4 · p. 8, §4.1 · pp. 8–9

4.1 Conforming formulation

  • The conforming formulation assumes uniformly equivalent test-space and spatial norms and an injective discrete time-step operator; under these conditions, the trial-space dimension cannot exceed the test-space dimension.The finite-dimensional spaces satisfy Xh ⊂ V and Yh ⊂ V.Source: §4.1 · p. 8
  • When the conforming trial and test spaces coincide, the standard Backward Euler Galerkin solution is the unique MinRes solution because its discrete residual vanishes.The general fully discrete error relation is restricted to the discrete test space, so it cannot directly be tested with the full error.Source: §4.1 · p. 9
  • A conforming fully discrete reliability estimate compares the computable discrete residual with the continuous residual by adding a complementary contribution for what the test space misses.That complementary residual measures the part invisible to the discrete test space.Source: §4.1 · pp. 9–10

4.2 Broken-test formulation

  • For the heat equation, the broken primal DPG field matches the standard Backward Euler Galerkin approximation under the stated compatibility assumptions and a time-step-dependent test inner product.The identity need not hold for a general discrete test inner product.Source: §4.2 · p. 11
  • The broken-test formulation also has a fully discrete reliability result comparing the discrete and continuous residuals, with the complementary broken residual accounting for unresolved components.The theorem is stated for finite-dimensional trial and broken test spaces.Source: §4.2 · pp. 10–11
  • In the broken formulation, elementwise residual localization uses element and flux residuals; with local approximation properties and the physical diffusive normal trace, the skeleton contribution vanishes.The stated localization also assumes H2 regularity of the trial approximation on each element.Source: §4.2 · p. 12

5 Neural-network MinRes approximations

The framework extends residual-based stability analysis to neural-network trial classes while retaining finite-dimensional test spaces. Bubble-enriched conforming and broken polynomial tests yield computable residual comparisons, with the broken formulation requiring an augmented estimator for reliable error control.

5 Neural-network MinRes approximations

  • The neural-network extension applies the representation-independent stability analysis to arbitrary trial sequences and relates the continuous residual representative to its computable discrete counterpart.The analysis does not require the trial class to be linear.Source: §5 · p. 12

5.1 Bubble-enriched conforming test space

  • The bubble-enriched conforming test space supports a neural-network reliability estimate by bounding the residual component not detected by the discrete tests.The stated bound has a constant depending on mesh shape regularity and the constants in the bubble quasi-interpolation estimates.Source: §5.1 · pp. 13–14

5.2 Broken polynomial test space

  • For broken polynomial tests, the residual comparison separates the component detected by elementwise averages from a mesh-weighted complementary intra-element component.This comparison remains valid for degree-zero tests, used in the numerical experiments.Source: §5.2 · p. 15
  • In the broken formulation, Jη is the accumulated discrete MinRes objective, whereas Jaug additionally accounts for continuous-residual components missed by the finite-dimensional test space.Jaug provides reliable error estimation up to the time-discretization defect; the analysis does not generally establish the same reliability for Jη alone.Source: §5.2 · pp. 15–16

6 Numerical experiments

The experiments verify finite-element convergence, isolate how test-space refinement resolves a fixed neural-network residual, and test residual-driven spatial refinement for a transient problem with a moving localized feature.

6.1 Smooth heat equation: finite-element verification

  • For the smooth heat problem, MinRes and standard finite-element full discrete-in-time errors and ηh converge at approximately first order, while final-time L2 errors converge at approximately second order.On the last refinement, the observed rates are 1.01, 1.00, and 1.01 for the MinRes full error, finite-element full error, and ηh, respectively, and 2.00 and 1.99 for final-time L2 errors.Source: §6.1 · p. 17, Figure 1 · p. 17
  • The residual representative tracks the MinRes error closely: ηh/Eh increases only from approximately 1.14 on the coarsest mesh to 1.20 on the finest.Here ηh is used diagnostically without the complementary residual; that contribution is examined in the neural-network and adaptive experiments.Source: §6.1 · p. 17

6.2 Smooth heat equation: global space–time neural approximation

  • With the trained network and time partition fixed, the resolved fraction rises from approximately 0.12 to 0.89 as nx increases from 4 to 32, while the unresolved fraction falls from approximately 0.88 to 0.11.The weighted complementary-to-detected contribution ratio also decreases from approximately 0.96 to 0.015; the simultaneous decrease in unweighted unresolved residual shows improved resolution, not merely an effect of mesh weighting.Source: §6.2 · p. 19, Figure 2 · p. 19

6.3 Diffusion–advection–reaction problem: estimator-driven spatial marking

  • Edar is 7.62 × 10−4 with 1068 cells for η-marking and 7.55 × 10−4 with 1022 cells for augmented marking, versus 8.36 × 10−4 with 2048 cells for uniform refinement.Both residual-driven strategies achieve slightly lower errors using about half as many cells as the finest uniform mesh.Source: §6.3 · pp. 21–22, Figure 4 · p. 22
  • On the final adaptive meshes, A(η, E) ≈ 0.835 and A(ξ, E) ≈ 0.910; the augmented indicator remains more closely aligned with the manufactured local error along the adaptive sequences.The augmented indicator is comparatively stable on intermediate meshes where the η-based distribution temporarily loses alignment.Source: §6.3 · p. 22
  • Both adaptive strategies concentrate refinement along the moving feature’s trajectory and produce meshes with comparable overall complexity.The experiment uses fixed coefficients and does not claim parameter robustness as ϵ → 0.Source: §6.3 · pp. 20, 22, Figure 5 · p. 22

7 Conclusions

The framework supports stable time-stepping MinRes for coercive transient problems, with computable formulations for finite-element and neural-network trials; experiments demonstrate convergence and adaptive-refinement results alongside stated efficiency and mesh limitations.

7 Conclusions

  • The framework provides stability estimates and fully discrete reliability estimates for conforming and broken-test formulations, with computable residual decompositions for neural-network trials.Finite-dimensional test spaces supplement the computable residual with a complementary contribution; neural decompositions use spatial quadratures at discrete time levels.Source: §7 · p. 23
  • Experiments show expected convergence for the smooth heat problem, improved error versus spatial complexity under residual-based marking, and more consistent error localization with the augmented indicator.For neural approximation, test-space refinement resolves an increasing fraction of a fixed residual while the relative complementary contribution decreases.Source: §7 · p. 23
  • Efficiency estimates are not uniform reverse bounds as τ →0; broken-setting estimates involve the full trial-space error, including its skeleton component.Adaptive computations use one spatial mesh over the entire time interval at each refinement cycle.Source: §7 · p. 23
Loading 2609.27548v1…