pith. sign in

arxiv: 2605.03542 · v2 · pith:WW65ATP3new · submitted 2026-05-05 · 🧮 math.NA · cs.LG· cs.NA

Random test functions, H⁻¹ norm equivalence, and stochastic variational physics-informed neural networks

Diego Marcondes This is my paper

Pith reviewed 2026-05-07 14:49 UTC · model grok-4.3

classification 🧮 math.NA cs.LGcs.NA
keywords H^{-1} norm equivalencerandom test functionsstochastic variational PINNsweak solutionsphysics-informed neural networkselliptic PDEsSobolev spaces
0
0 comments X

The pith

The H^{-1} norm of any functional is equivalent to its expected squared evaluation against a random test function whose distribution depends only on the domain.

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

The paper proves that the H^{-1} norm of any functional can be recovered exactly by taking the expectation of its squared evaluation on samples from a random test function distribution that depends solely on the domain. This holds even though the test functions have negative Sobolev regularity in dimensions two and higher. The equivalence defines stochastically weak solutions that are the same as classical weak solutions. It leads to a new class of neural network methods for solving PDEs by minimizing an empirical stochastic norm of the residual instead of a deterministic one.

Core claim

The central discovery is that the H^{-1} norm of any functional is equivalent to its expected squared evaluation against a random test function whose distribution depends only on the domain. Realisations of this random test function have negative Sobolev regularity for d greater than or equal to 2, yet this roughness does not prevent averaging over the distribution from exactly recovering the correct weak topology, independently of the differential operator. This equivalence introduces stochastically weak solutions, which coincide with classical weak solutions, and motivates stochastic variational physics-informed neural networks trained by minimising an empirical approximation of the stocha

What carries the argument

A domain-dependent probability distribution over test functions such that the expected value of the squared pairing with any functional recovers the H^{-1} norm and the associated weak topology.

Load-bearing premise

There exists a distribution over random test functions, depending only on the domain, for which the expected squared evaluation exactly equals the H^{-1} norm of the functional for any functional.

What would settle it

A concrete counterexample consisting of a domain, a specific functional, and its H^{-1} norm that does not match the computed expectation over samples from the proposed random test function distribution.

Figures

Figures reproduced from arXiv: 2605.03542 by Diego Marcondes.

Figure 1
Figure 1. Figure 1: Examples of realisations of Φ. (a) Sampled by (4.10) with τ “ 10 for d “ 1, 2, 3 in a grid with, respectively, 1024, 10242 and 5123 points. For d “ 3 we present the slice φp0.5, ¨, ¨q of the sampled realisation. (b) Sampled via the truncated sum (4.12) with M “ 4, 096, τ “ 0.1 and 22, 500 points sampled uniformly. (c) Sampled by truncating the sum (4.13) with M “ 1, 024, τ “ 0.1 and 7, 105 points in a grid… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Solutions of the boundary value problems in Experiment 1 (6.3) (solid line in red) with the approximation by the SV-PINNs with DAFF trained with L-BFGS for 5, 000 steps (dashed line in blue), and the pointwise error between them. (b) Average L 2 relative error with ˘ one standard error over the 3 repetitions at all training steps for each method, optimiser and architecture features for solving the boun… view at source ↗
Figure 3
Figure 3. Figure 3: Solution of the boundary value problems in Experiment 2 (6.4), the approximation by an SV-PINN trained with L-BFGS for 5, 000 steps and its point￾wise error view at source ↗
Figure 4
Figure 4. Figure 4: Average L 2 relative error with ˘ one standard error over the 3 repeti￾tions at all training steps for each method and optimiser for solving the boundary value problems in Experiment 2 (6.4). training steps. The performance of SV-PINNs was more stable, i.e., smaller standard deviations, over the three repetitions than that of PINNs. The SV-PINNs converged with significantly fewer steps than PINNs, at￾taini… view at source ↗
Figure 5
Figure 5. Figure 5: (a) Solution of the boundary value problem in Experiment 3 (6.5), the approximation by an SV-PINN trained with L-BFGS for 5, 000 steps and its pointwise error. (b) Average L 2 relative error with ˘ one standard error over the 3 repetitions at all training steps for each method and optimiser. 6.4. Experiment 4: Helmholtz equation in 2D. To further increase the complexity, we consider the two-dimensional Hel… view at source ↗
Figure 6
Figure 6. Figure 6: (a) Solution of the boundary value problem in Experiment 4 (6.6), the approximation by an SV-PINN trained with L-BFGS for 5, 000 steps and its pointwise error. (b) Average L 2 relative error with ˘ one standard error over the 3 repetitions at all training steps for each method and optimiser. As in the previous experiments, the SV-PINNs trained with L-BFGS were able to recover the solution with an L 2 relat… view at source ↗
Figure 7
Figure 7. Figure 7: Solution of the boundary value problems in Experiment 5 (6.7), the approximation by an SV-PINN trained with L-BFGS for 5, 000 steps and its point￾wise error view at source ↗
Figure 8
Figure 8. Figure 8: Average L 2 relative error with ˘ one standard error over the 3 repeti￾tions at all training steps for each method and optimiser for solving the boundary value problems in Experiment 5 (6.7). these frequencies are more compatible with the solution for higher values of k. As with the other examples, the convergence of the SV-PINNs trained by L-BFGS is significantly faster, achieving an L 2 relative error lo… view at source ↗
Figure 9
Figure 9. Figure 9: Solution of the boundary value problems in Experiment 6 (6.8), the approximation by an SV-PINN trained with L-BFGS for 5, 000 steps and its point￾wise error. We consider a slice of the respective functions with the first variable x “ 0.5. As expected, the PINNs perform particularly well, especially for higher values of k, attaining L 2 relative errors on average between 6.7 ˆ 10´3 and 8.5 ˆ 10´3 , outperfo… view at source ↗
Figure 10
Figure 10. Figure 10: Average L 2 relative error with ˘ one standard error over the 3 repeti￾tions at all training steps for each method and optimiser for solving the boundary value problems in Experiment 6 (6.8) view at source ↗
Figure 11
Figure 11. Figure 11: Solution of the boundary value problems in Experiment 7 (7.1), the approximation by an SV-PINN trained with L-BFGS for 5, 000 steps and its point￾wise error. We consider an architecture with 64 DAFF and one hidden layer with 1, 024 nodes, τ “ 0.1 for numerical stability, the collocation points in a grid with 7, 105 points, and the same training methods as in the previous sections. In order to compute the … view at source ↗
Figure 12
Figure 12. Figure 12: Average L 2 relative error with ˘ one standard error over the 3 repeti￾tions at all training steps for each method and optimiser for solving the boundary value problems in Experiment 7 (7.1) view at source ↗
Figure 13
Figure 13. Figure 13: Solution of the boundary value problems in Experiment 8 (8.1), the approximation by an SV-PINN trained with L-BFGS for 5, 000 steps and its point￾wise error. norm applicable to broader classes of PDEs. We emphasise that the follow￾ing extensions are presented at a conceptual level and detailed analytical and numerical investigations are left for future work. 9.1. High-order elliptic operators and poly-har… view at source ↗
Figure 14
Figure 14. Figure 14: Average L 2 relative error with ˘ one standard error over the 3 repeti￾tions at all training steps for each method and optimiser for solving the boundary value problems in Experiment 8 (8.1). in which B iu Bν i is the i-th normal derivative. For instance, B 0u Bν 0 “ u|BΩ and B 1u Bν 1 “ ∇u ¨ ν in which ν is the normal vector. To characterise the weak solutions of (9.1), we consider the eigenvalues and ei… view at source ↗
read the original abstract

The dual norm characterisation of weak solutions of second-order linear elliptic partial differential equations is mathematically natural but computationally intractable: evaluating the $H^{-1}$ norm of the residual requires a supremum over an infinite-dimensional test space. We prove that the $H^{-1}$ norm of any functional is equivalent to its expected squared evaluation against a random test function whose probability distribution depends only on the domain. Crucially, realisations of this random test function have negative Sobolev regularity for $d \geq 2$, yet this roughness is not an obstacle: averaging over the distribution exactly recovers the correct weak topology, independently of the differential operator, and no supremum evaluation is necessary. This equivalence introduces the notion of stochastically weak solutions, which coincide with classical weak solutions, and motivates stochastic variational physics-informed neural networks (SV-PINNs): neural networks trained by minimising an empirical approximation of the stochastic norm of the PDE residual. Although instantiated here with neural networks, the underlying principle is independent of the trial space and suggests a broader paradigm for numerical methods based on stochastic rather than deterministic test spaces. The framework extends naturally to higher-order elliptic, parabolic and hyperbolic equations and to abstract operator equations on Hilbert spaces. As a proof of concept, we present numerical experiments on eight challenging second-order linear elliptic problems spanning high-frequency and multi-scale solutions, indefinite operators, variable coefficients, and non-standard domains, in which SV-PINNs consistently and significantly outperform standard PINNs, recovering solutions to within one percent relative error in hundreds of L-BFGS steps.

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.

Referee Report

0 major / 2 minor

Summary. The manuscript proves that the H^{-1} norm of any functional is equivalent to the square root of its expected squared pairing against a random test function whose law depends only on the domain. This equivalence is independent of the differential operator, yields stochastically weak solutions that coincide with classical weak solutions, and motivates SV-PINNs: neural networks trained by minimizing an empirical Monte-Carlo approximation to the stochastic residual norm. The framework is illustrated on eight second-order linear elliptic problems (high-frequency, multi-scale, indefinite, variable-coefficient, non-standard domains) where SV-PINNs recover solutions to within 1% relative error in a few hundred L-BFGS iterations and consistently outperform standard PINNs.

Significance. If the equivalence is rigorously established, the work supplies a computationally tractable, operator-independent surrogate for the dual norm that removes the need to solve auxiliary supremum problems. This is a genuine advance for variational PINNs and related mesh-free methods. The numerical evidence on a deliberately diverse test suite, together with the claim that negative Sobolev regularity of the random fields does not obstruct recovery of the H^{-1} topology, strengthens the practical case. The extension sketched to higher-order, parabolic, hyperbolic and abstract Hilbert-space problems further increases potential impact.

minor comments (2)
  1. [Abstract] Abstract: the statement that solutions are recovered “to within one percent relative error” should specify the norm (L^2, H^1, etc.) in which the error is measured and indicate whether the figure is an average or worst-case value over the eight test problems.
  2. [Numerical experiments] Implementation details for sampling the random test functions (discretization of the covariance operator, number of Monte-Carlo samples per loss evaluation, and handling of the negative-order regularity for d ≥ 2) are only sketched; a short algorithmic box or pseudocode would improve reproducibility.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive and encouraging report, which accurately summarizes the main theoretical result on the H^{-1} equivalence and its use in SV-PINNs, as well as the numerical experiments on diverse elliptic problems. We appreciate the assessment of significance and the recommendation for minor revision.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central result is a functional-analytic existence proof: there exists a random test function distribution (depending only on the domain) such that the expected squared duality pairing recovers the H^{-1} norm exactly. This is achieved by taking the covariance operator of the random field to be the inverse Riesz map of the H^1_0 inner product, which is a standard, operator-independent construction on the domain. The equivalence therefore holds by the definition of covariance and the Riesz representation theorem, but this constitutes a valid mathematical demonstration rather than a self-referential loop or a fitted quantity renamed as a prediction. No load-bearing step in the derivation reduces to its own inputs by construction; the subsequent SV-PINN training procedure follows from the established equivalence without circularity. The argument is self-contained against external functional-analytic benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard functional-analysis properties of Sobolev spaces together with the paper-specific assertion that a domain-dependent random distribution recovers the dual norm via expectation; no free parameters or new physical entities are introduced.

axioms (2)
  • standard math Standard properties of Sobolev spaces H^1 and its dual H^{-1} for second-order elliptic operators
    Invoked to define weak solutions and the dual-norm characterisation of residuals.
  • ad hoc to paper Existence of a random test-function distribution depending only on the domain such that expectation of squared evaluations recovers the H^{-1} norm independently of the operator
    This is the load-bearing statement proved in the paper and required for the stochastic formulation to be equivalent to the classical weak form.

pith-pipeline@v0.9.0 · 5579 in / 1511 out tokens · 158275 ms · 2026-05-07T14:49:20.046692+00:00 · methodology

Review history (2 revisions) →

discussion (0)

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