arxiv: 2603.16849 · v3 · submitted 2026-03-17 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

GIST: Gauge-Invariant Spectral Transformers for Scalable Graph Neural Operators

Mattia Rigotti , Nicholas Thumiger , Thomas Frick

Authors on Pith no claims yet

Pith reviewed 2026-05-15 09:22 UTC · model grok-4.3

classification 💻 cs.LG

keywords gauge invariancegraph neural operatorsspectral transformersdiscretization invariancescalable GNNmesh neural networksneural operators

0 comments

The pith

GIST computes graph neural operator attention via inner products of approximate spectral embeddings to achieve exact gauge invariance at linear complexity.

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

Neural operators on irregular meshes face a tension where full spectral encodings require expensive eigendecomposition and break gauge invariance through solver artifacts, while fast approximations sacrifice symmetry by design. GIST resolves this by restricting attention to pairwise inner products of efficient approximate spectral embeddings. The paper proves these inner products estimate an exactly gauge-invariant graph kernel with end-to-end linear complexity and formally connects gauge invariance to discretization-invariant learning with bounded mismatch error. If correct, this enables models that reliably transfer across mesh resolutions of the same domain or across related graphs in inductive settings. Empirically the approach reaches state-of-the-art accuracy on large mesh benchmarks while matching strong baselines on standard graph tasks.

Core claim

GIST is a scalable graph neural operator that restricts attention to pairwise inner products of efficient approximate spectral embeddings; these inner products estimate an exactly gauge-invariant graph kernel at O(N) complexity, and gauge invariance implies discretization-invariant learning with bounded mismatch error.

What carries the argument

Pairwise inner products of efficient approximate spectral embeddings, which estimate an exactly gauge-invariant graph kernel.

If this is right

Models transfer across different mesh resolutions of the same physical domain with only bounded error.
Processing remains linear in the number of nodes, enabling graphs with hundreds of thousands of points.
State-of-the-art accuracy holds on AirfRANS, ShapeNet-Car, DrivAerNet and DrivAerNet++ benchmarks.
Performance matches strong baselines on standard graph tasks such as PPI node classification.

Where Pith is reading between the lines

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

The same inner-product construction could be applied to other spectral methods outside graphs to enforce invariance without cubic costs.
Direct measurement of the kernel mismatch error on a sequence of successively refined meshes would provide a practical test of the bound.
The approach opens a route to neural operators that remain stable under adaptive mesh refinement in time-dependent simulations.

Load-bearing premise

The inner products of the approximate embeddings estimate an exactly gauge-invariant kernel without the approximations introducing significant gauge-breaking artifacts or loss of expressivity.

What would settle it

Training GIST on one mesh resolution and testing on a significantly different resolution of the same domain, then measuring whether prediction error exceeds the claimed bounded mismatch.

read the original abstract

Neural operators on irregular meshes face a fundamental tension. Spectral positional encodings, the natural choice for capturing geometry, require cubic-complexity eigendecomposition and inadvertently break gauge invariance through numerical solver artifacts; existing efficient approximations sacrifice gauge symmetry by design. Both failure modes break discretization invariance: models fail to transfer across mesh resolutions of the same domain, and similarly across different graphs of related structure in inductive settings. We propose GIST (Gauge-Invariant Spectral Transformer), a scalable neural operator that resolves this tension by restricting attention to pairwise inner products of efficient approximate spectral embeddings. We prove these inner products estimate an exactly gauge-invariant graph kernel at end-to-end $\mathcal{O}(N)$ complexity, and establish a formal connection between gauge invariance and discretization-invariant learning with bounded mismatch error. To our knowledge, GIST is the first scalable graph neural operator with a provable discretization-mismatch bound. Empirically, GIST sets state-of-the-art on the AirfRANS, ShapeNet-Car, DrivAerNet, and DrivAerNet++ mesh benchmarks (up to 750K nodes), and additionally matches strong baselines on standard graph benchmarks (e.g., 99.50% micro-F1 on PPI).

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

GIST uses inner products of approximate spectral embeddings to claim exact gauge invariance and a discretization bound at linear cost, but the approximation step needs checking to confirm the proof holds.

read the letter

The main point is that this paper gives a construction for spectral graph neural operators that avoids breaking gauge invariance by design. Instead of feeding approximate eigenvectors directly into the model, it restricts attention to their pairwise inner products, which the authors prove estimates a gauge-invariant kernel at O(N) cost while adding a bound on discretization mismatch error. That combination looks new relative to prior spectral work on irregular domains, and the empirical side shows SOTA numbers on large mesh tasks like AirfRANS and DrivAerNet up to 750k nodes plus solid results on PPI. The framing of the tension between spectral encodings and invariance is clear and useful for anyone working on physics-informed operators on meshes. The bound itself, if it survives scrutiny, is the kind of formal handle that could matter for transfer across resolutions. The soft spot sits in the proof that the inner products remain exactly gauge-invariant after approximation. The stress-test concern lands here: standard efficient spectral approximations can introduce numerical artifacts that break exact symmetry, and it is not obvious from the abstract whether the derivation uses a symmetry-preserving method or simply bounds the resulting error. If the latter, the “exactly” claim and the tightness of the mismatch bound both weaken. Experiments would also benefit from explicit ablations isolating the invariance contribution versus other modeling choices. This paper is aimed at people building scalable neural operators for scientific simulation on irregular data. A reader focused on discretization-invariant models or gauge-equivariant architectures would get concrete ideas to try, even if the full derivations require verification. I would send it to peer review because the core construction is specific enough to be tested and the empirical claims are on relevant benchmarks.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces GIST, a scalable graph neural operator that resolves the tension between spectral positional encodings' cubic complexity and their tendency to break gauge invariance. By restricting attention to pairwise inner products of efficient approximate spectral embeddings, the authors claim to prove that these inner products estimate an exactly gauge-invariant graph kernel at end-to-end O(N) complexity. They further establish a formal connection between gauge invariance and discretization-invariant learning with a bounded mismatch error, positioning GIST as the first scalable graph neural operator with a provable discretization-mismatch bound. Empirically, GIST achieves state-of-the-art results on large mesh benchmarks including AirfRANS, ShapeNet-Car, DrivAerNet, and DrivAerNet++ (up to 750K nodes) while matching strong baselines on standard graph tasks such as PPI (99.50% micro-F1).

Significance. If the central theoretical claims hold, this work would be significant for neural operators on irregular domains. It offers a practical O(N) method that preserves gauge invariance exactly in the kernel construction, directly linking this property to improved discretization invariance and transfer across mesh resolutions or related graphs. The provision of a provable mismatch bound and strong empirical performance on large-scale benchmarks (hundreds of thousands of nodes) would represent a meaningful advance over prior spectral and approximation-based approaches.

major comments (2)

[Abstract / Theoretical Analysis] Abstract and theoretical section on gauge invariance: The central claim that 'pairwise inner products of efficient approximate spectral embeddings estimate an exactly gauge-invariant graph kernel' requires the chosen approximation to preserve gauge symmetry exactly in the inner-product step. Standard efficient spectral approximations (e.g., Nyström, random features, or Lanczos) typically introduce gauge-breaking artifacts; if the proof only bounds error after the fact rather than constructing an exactly equivariant approximation, both the 'exactly gauge-invariant' statement and the subsequent discretization-mismatch bound are at risk. Please identify the specific approximation technique and show explicitly (via the relevant theorem or lemma) that it maintains exact invariance.
[Theoretical Analysis] Discretization-mismatch bound: The manuscript asserts a 'provable discretization-mismatch bound' connecting gauge invariance to discretization-invariant learning. The bound, its derivation, and the precise conditions (including any dependence on mesh regularity or approximation rank) must be stated explicitly; without this, it is unclear whether the bound is non-vacuous or holds under the same approximations used for scalability.

minor comments (2)

[Abstract] The abstract reports SOTA results on multiple mesh benchmarks but does not indicate whether ablation studies isolate the contribution of the gauge-invariant inner-product construction versus other architectural choices.
[Method] Notation for the approximate spectral embeddings and the inner-product kernel should be introduced with explicit definitions early in the method section to aid readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful review and constructive feedback on our work. We address the major comments point by point below, providing clarifications on the theoretical claims and indicating revisions to the manuscript.

read point-by-point responses

Referee: [Abstract / Theoretical Analysis] Abstract and theoretical section on gauge invariance: The central claim that 'pairwise inner products of efficient approximate spectral embeddings estimate an exactly gauge-invariant graph kernel' requires the chosen approximation to preserve gauge symmetry exactly in the inner-product step. Standard efficient spectral approximations (e.g., Nyström, random features, or Lanczos) typically introduce gauge-breaking artifacts; if the proof only bounds error after the fact rather than constructing an exactly equivariant approximation, both the 'exactly gauge-invariant' statement and the subsequent discretization-mismatch bound are at risk. Please identify the specific approximation technique and show explicitly (via the relevant theorem or lemma) that it maintains exact invariance.

Authors: The approximation technique employed is a scalable randomized spectral embedding method based on random projections for approximating the top eigenvectors of the graph Laplacian (detailed in Section 3.2). Crucially, Theorem 3.4 proves that the pairwise inner products of these embeddings exactly estimate the gauge-invariant kernel. This is because any gauge transformation corresponds to a unitary transformation on the embedding space, and the inner product is invariant under unitary transformations: for embeddings Φ and gauge-transformed Φ' = U Φ where U is unitary, <Φ_i, Φ_j> = <Φ'_i, Φ'_j>. The approximation is designed such that this invariance holds exactly in the inner-product computation, independent of the approximation error in the embeddings themselves. We have revised the manuscript to include an explicit statement of this lemma in the main theoretical section and added a dedicated proof in the appendix for clarity. revision: yes
Referee: [Theoretical Analysis] Discretization-mismatch bound: The manuscript asserts a 'provable discretization-mismatch bound' connecting gauge invariance to discretization-invariant learning. The bound, its derivation, and the precise conditions (including any dependence on mesh regularity or approximation rank) must be stated explicitly; without this, it is unclear whether the bound is non-vacuous or holds under the same approximations used for scalability.

Authors: The discretization-mismatch bound is formalized in Theorem 5.1, which states that the error in discretization invariance is bounded by O(δ + ε), where δ is the mesh regularity parameter (assuming quasi-uniform meshes as per standard assumptions in the field) and ε is the spectral approximation error depending on the rank k of the embedding. The derivation proceeds by showing that gauge invariance implies the kernel is discretization-invariant up to the approximation error, with the proof relying on the stability of the inner-product kernel under mesh refinements. We have expanded the statement of Theorem 5.1 in the revised version to explicitly list all conditions, including the dependence on mesh regularity and the approximation rank k, and included a detailed derivation in the main text rather than solely in the appendix. revision: yes

Circularity Check

0 steps flagged

No circularity: proof of gauge-invariant kernel approximation stands independently

full rationale

The paper's central claim is a proof that pairwise inner products of efficient approximate spectral embeddings estimate an exactly gauge-invariant graph kernel with O(N) complexity and bounded discretization mismatch. No equations or steps in the abstract or description reduce this by construction to fitted parameters, self-citations, or ansatzes; the approximation is treated as given and the invariance is derived formally rather than assumed via definition or prior self-work. This is the common case of a self-contained theoretical result with external empirical validation on mesh benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that pairwise inner products of approximate spectral embeddings preserve exact gauge invariance up to bounded error; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (1)

domain assumption Inner products of efficient approximate spectral embeddings estimate an exactly gauge-invariant graph kernel with bounded mismatch error
This is the load-bearing mathematical claim stated in the abstract that enables both the O(N) complexity and the discretization-invariance guarantee.

pith-pipeline@v0.9.0 · 5516 in / 1273 out tokens · 46457 ms · 2026-05-15T09:22:26.962811+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 (J uniqueness and J(x)=J(1/x) symmetry) echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

their inner products are unbiased estimators of the graph kernel K=f(P)f(P)^T, which depends only on the graph operator P and is therefore exactly gauge-invariant... K(i, j) is a function of P alone and therefore exactly gauge-invariant: it is unchanged by sign flips, eigenspace rotations, and numerical solver choices
IndisputableMonolith/Foundation/ArithmeticFromLogic.lean embed_injective / embed_strictMono_of_one_lt (orbit embeddings preserve structure under symmetry) refines

?

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

gauge invariance guarantees discretization-invariant learning with bounded discretization mismatch error

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.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Faster by Design: Interactive Aerodynamics via Neural Surrogates Trained on Expert-Validated CFD
cs.LG 2026-04 unverdicted novelty 7.0

A graph-based neural operator trained on expert-validated race-car CFD data reaches accuracy levels usable for early-stage interactive aerodynamic design exploration.