Feature Learning for the High Dimensional Stationary Sch\"odinger Equation with Deep Ritz Method

Gilad Lerman; Yao Yao; Yulong Lu

A restricted neural network provably solves a high-dimensional Schrödinger equation in logarithmic time, and its features align with the source.

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 · deepseek-v4-flash

2026-08-02 08:15 UTC pith:GTSZ3WKZ

load-bearing objection Abstract-only: genuinely new convergence-rate and feature-emergence results in a restricted setting, but the 'approximate global minimum' claim overstates what the stated bound actually proves. the 3 major comments →

arxiv 2607.06369 v2 pith:GTSZ3WKZ submitted 2026-07-07 math.OC stat.ML

Feature Learning for the High Dimensional Stationary Sch\"odinger Equation with Deep Ritz Method

Yao Yao , Yulong Lu , Gilad Lerman This is my paper

classification math.OC stat.ML

keywords deep Ritz methodstationary Schrödinger equationfeature learningsingle-index modelRiemannian gradient descentloss landscaperegularized Ritz energyhigh-dimensional PDE

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.

This paper studies feature learning in the deep Ritz method, a neural-network approach to the stationary Schrödinger equation with Neumann boundary conditions. It proves that when the hypothesis class is limited to single-index models, Riemannian gradient descent on the Ritz energy loss reaches within epsilon of a constant multiple of the optimal loss in O(log(1/epsilon)) iterations. It then analyzes the loss landscape under the assumption that the PDE source itself is single-index: the minimizer of the Ritz energy is attained at the feature vector aligned with the source. For a two-neuron architecture with the first feature fixed along the source, the regularized Ritz loss exhibits the emergence of a second feature as the regularization parameter varies. These results provide a first theoretical account of how features emerge in the deep Ritz method, though the landscape results depend on the source being single-index.

Core claim

The central discovery is twofold. First, in an agnostic setting where the hypothesis is a single-index model but the PDE solution is arbitrary, Riemannian gradient descent on the Ritz loss converges to an approximate global minimum: after T = O(log(1/epsilon)) iterations, the energy is within epsilon of a constant multiple of the optimal energy in that class. Second, when the source term of the Schrödinger equation is itself single-index, the Ritz loss landscape has a structural property: the global minimizer of the energy is attained at the feature vector aligned with the source. For a two-neuron multi-index model, with the first neuron's feature already aligned to the source, the regulariz

What carries the argument

The analysis rests on two objects: the single-index model, a hypothesis function that depends on the input only through one linear feature w, and the Ritz energy, the variational loss functional of the stationary Schrödinger operator that the deep Ritz method minimizes. The convergence proof uses Riemannian gradient descent on the unit sphere of feature directions, exploiting the geometry of the single-index parameter space. For the two-neuron case, the regularization parameter acts as the control knob that drives the emergence of a second feature, and its role is characterized by analyzing the landscape of the regularized Ritz loss.

Load-bearing premise

The landscape and feature-emergence theorems assume the PDE source term is a single-index function; real Schrödinger sources such as Coulomb potentials are not single-index, so these results do not transfer directly to standard physical settings, and the agnostic convergence result only certifies a restricted hypothesis class up to an unspecified multiplicative constant.

What would settle it

For a strictly nonlinear single-index source, f(x) = g(<w*,x>), compute the single-index Ritz energy landscape over unit feature vectors on a bounded domain with Neumann boundary conditions. If the global minimizer is not parallel to w*, the feature-alignment theorem is false. This numerical calculation directly tests the claimed landscape property.

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

If this is right

In the single-index hypothesis class, the Ritz method attains near-optimal energy in a number of iterations logarithmic in 1/epsilon, independent of the ambient dimension.
When the source is single-index, the energy-minimizing feature is exactly the source feature, implying the learned representation recovers a physically meaningful direction in the potential.
In the two-neuron architecture, decreasing the regularization parameter causes a second feature to emerge, providing a provable mechanism for feature learning in a shallow neural network.
The numerical experiments confirm the predicted feature-emergence threshold in the two-neuron setting, validating the theoretical landscape analysis.
These results establish a theoretical baseline for why the deep Ritz method can learn low-dimensional structure in high-dimensional Schrödinger problems.

Where Pith is reading between the lines

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

If the logarithmic convergence extends beyond single-index hypotheses, the deep Ritz method could be certified for high-dimensional Schrödinger problems where mesh-based methods fail, but the current proof is limited to a narrow class.
The single-index assumption on the source is a strong idealization; real Coulomb-like potentials are not single-index, so the alignment and emergence theorems do not directly apply to standard physical settings.
The unspecified multiplicative constant in the agnostic convergence bound invites a follow-up analysis to pin it down or show it is dimension-independent, which would make the bound practically useful.
The two-neuron analysis might generalize to deeper multi-index models where layers of features emerge hierarchically under progressive regularization, though such an extension is not demonstrated here.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Abstract-only: genuinely new convergence-rate and feature-emergence results in a restricted setting, but the 'approximate global minimum' claim overstates what the stated bound actually proves.

read the letter

You asked for my read on arXiv:2607.06369. Since we only have the abstract, I'll keep this appropriately provisional, but there are two things worth saying up front.

What's genuinely new here is the combination: Riemannian gradient descent on the Ritz loss with a single-index hypothesis, a logarithmic iteration bound in an agnostic setting, and a landscape-level explanation for how a second feature emerges under regularization in a two-neuron model. That's not a restatement of earlier results as far as I can tell from the abstract. If the proofs hold, this is a useful step for the neural-PDE-solver subfield, even though the architectures are toy models relative to practice.

The paper also deserves credit for stating its two caveats honestly: the "agnostic" result only certifies the restricted hypothesis class, and the landscape/emergence results assume the source term is single-index. Those limits don't sink the paper; they just define its scope.

The real soft spot is the convergence claim. The abstract says gradient descent reaches an "approximate global minimum" because after T = O(log(1/ε)) iterations the loss is "within ε of a constant multiple of the optimal loss." That's not what "approximate global minimum" normally means. Unless that constant is 1+o(1) or a numerical factor close to 1, a bound of L(w_T) ≤ C L* + ε with C unspecified says almost nothing about how close the algorithm gets to the optimum. It could be a factor of ten away and still satisfy the stated bound. This may be a presentation problem rather than a mathematical error, but as written it's an overclaim. The authors need to either state the constant explicitly or reframe the result as a constant-factor approximation.

The other weaknesses are inherent to reviewing an abstract. I can't check the proofs, the numerical experiments, or whether the cited prior work is fairly represented. The single-index-source limitation is real: actual Schrödinger potentials like Coulomb are not single-index, so the feature-emergence story doesn't transfer to the main applications. But that's a scope issue, not a flaw within the model.

Overall: this looks like a serious piece of work that needs a referee to check the constant in the convergence theorem and confirm that the experiments actually validate the emergence theory. The central ideas are coherent and unlikely to be circular. It deserves a proper peer review, not a desk reject, provided the authors tighten the language around the global-minimum claim.

I'd bring it to a reading group if we had the full text. For now, I'd cite it only if I needed to point to the logarithmic-rate result in single-index Ritz — and I'd caution whoever reads it about the constant.

Recommendation: send it to review.

Referee Report

3 major / 3 minor

Summary. The paper studies the deep Ritz method for the stationary Schrödinger equation with Neumann boundary conditions, focusing on feature learning in restricted architectures. It claims (i) an agnostic convergence result: Riemannian gradient descent over a single-index hypothesis class reaches a loss within ε of a constant multiple of the optimal loss in O(log(1/ε)) iterations; (ii) a landscape analysis under a single-index source assumption, showing that the minimum Ritz energy is attained at a feature aligned with the source in the single-index case and characterizing second-feature emergence in the two-neuron case as the regularization parameter varies; and (iii) numerical experiments supporting the emergence theory.

Significance. If the convergence result holds with a constant close to one, it would constitute a rare rigorous performance guarantee for a neural PDE solver on a restricted hypothesis class, and the alignment/emergence analysis would be a novel contribution to understanding feature learning in physics-informed deep learning. The paper is careful to separate agnostic and source-dependent settings, and the numerical experiments are a positive addition. However, the abstract-level statement of the convergence result is currently too weak to support the 'approximate global minimum' claim without an explicit bound on the multiplicative constant, and the landscape results are explicitly conditional on a source model that does not cover the Schrödinger potentials of primary physical interest.

major comments (3)

[Abstract, convergence claim] The abstract states that after T = O(log(1/ε)) iterations 'the loss is within ε of a constant multiple of the optimal loss,' and calls this an approximate global minimum. As stated, this implication is invalid unless the constant is near 1 (e.g., C = 1 + o(1)). With an unspecified C, a bound of the form L(w_T) ≤ C L* + ε is vacuous for large C; it only shows that the algorithm reaches a point in an unknown level set. Please either state the constant explicitly, prove that it can be taken as 1+o(1), or rephrase the claim as convergence to a constant-factor approximation rather than an approximate global minimum. This is the central advertised result and needs to be precise.
[Abstract, landscape results] The alignment and emergence theorems assume the source term of the PDE follows a single-index model, and the two-neuron analysis further assumes the first feature is already aligned with the source. Real Schrödinger sources, such as Coulomb-type or other external potentials, are generally not single-index, so these landscape results do not transfer to the actual PDEs the method is used for. The agnostic convergence theorem avoids the source assumption, but it suffers from the unquantified constant issue. Please clarify the intended domain of applicability and discuss whether the single-index assumption can be relaxed or is a genuine restriction.
[Abstract, two-neuron case] The statement 'given that the first feature is aligned with the source' is a strong initialization assumption. If the feature-emergence characterization holds only in this favorable regime, it does not establish that gradient descent generically discovers such alignment from arbitrary initializations. The paper should state explicitly whether the emergence result is for this preconditioned setting only, and if so, discuss what happens without this assumption.

minor comments (3)

[Abstract, notation and definitions] The phrase 'Riemannian gradient descent' is not specified in the abstract. Please define the Riemannian metric and the manifold structure on the feature space in the theorem statements.
[Abstract, quantitative statement] Consider replacing the phrase 'within ε of a constant multiple of the optimal loss' with a concrete inequality, e.g., L(w_T) ≤ C L* + ε, with C explicitly defined and its dependence on problem parameters stated.
[Abstract, terminology] The term 'approximate global minimum' is standardly used for additive or multiplicative (1+ε) approximation. Using it for a constant-factor bound with arbitrary C is misleading; choose terminology that matches the actual guarantee.

Circularity Check

0 steps flagged

No circularity detected in abstract-only review; claims are mathematical results with stated assumptions.

full rationale

This review is based solely on the abstract; no derivation chain, equations, fitted parameters, self-citations, or imported uniqueness theorems are available to examine. The convergence claim is presented as a theorem about Riemannian gradient descent on the Ritz loss, the landscape claim is conditional on a single-index source, and the emergence claim is explicitly conditional on an aligned first feature. None of these statements, as stated, defines its target in terms of an input or presents a fitted quantity as a prediction. The abstract's 'constant multiple' language is a potential correctness or precision concern, not a circularity concern, because it does not show that any result is equivalent to its assumptions by construction. Without quotable evidence of a specific reduction, the hard rules require a non-finding. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

All results are conditional on low-dimensional structure assumptions (single-index source, restricted hypothesis classes); no new physical entities or fitted potentials are introduced. The main unstated costs are the theorem constants and regularity conditions, which the abstract does not report.

free parameters (2)

multiplicative constant C in convergence bound = unspecified in abstract
The headline convergence claim is "within ε of a constant multiple of the optimal loss"; the constant's value (or whether it is near 1, making the bound meaningful) is not stated in the abstract, and the claim's strength depends directly on it.
regularization parameter = unspecified in abstract; varies
The second-feature emergence is characterized as this parameter varies; it is a method hyperparameter, not fitted to data, but the abstract gives no value or range, so its role in the claimed transition cannot be checked from the abstract alone.

axioms (5)

domain assumption Hypothesis restricted to a single-index model in the agnostic convergence analysis
The convergence and optimal-loss statements are for this restricted class, not for the full deep Ritz ansatz; stated in the abstract's first result.
domain assumption PDE source term follows a single-index model
The feature-alignment and second-feature emergence results are conditional on this structure; stated in the abstract's second result.
domain assumption Two-neuron multi-index hypothesis with first feature aligned with the source
The emergence characterization is proved only in this setting and conditions on the first feature already being aligned; stated in the abstract's third result.
domain assumption Stationary Schrödinger equation with Neumann boundary conditions
The entire analysis is for this PDE class and boundary condition; the abstract gives no indication the results extend beyond it.
standard math Standard smoothness/regularity conditions on the Ritz loss for Riemannian gradient descent convergence
Any O(log(1/ε)) convergence theorem requires smoothness/compactness assumptions; the abstract does not state them, so they are assumed in the full text.

pith-pipeline@v1.3.0-alltime-deepseek · 511 in / 12287 out tokens · 102259 ms · 2026-08-02T08:15:41.676130+00:00 · methodology

0 comments

read the original abstract

This paper investigates feature learning within the framework of the deep Ritz method for solving the stationary Schr\"odinger equation with Neumann boundary conditions. We first analyze the convergence of Riemannian gradient descent in an agnostic setting, where the hypothesis function is restricted to a single-index model while the PDE solution is arbitrary. We prove that gradient descent reaches an approximate global minimum: after T = O(log(1/\epsilon)) iterations, the loss is within \epsilon of a constant multiple of the optimal loss. We then examine the loss landscape when the source term of the PDE itself follows a single-index model, considering hypothesis functions given by either a single-index model or a two-neuron multi-index model. In the single-index case, we show that the minimum Ritz energy is attained at the feature vector aligned with that of the source term. In the two-neuron case, we study the landscape of regularized Ritz losses and characterize how a second feature emerges, given that the first feature is aligned with the source, as the regularization parameter varies. Finally, numerical experiments are presented to validate the feature emergence theory in the two-neuron setting.