pith. sign in

arxiv: 2511.16062 · v2 · pith:DWISEW37new · submitted 2025-11-20 · 💻 cs.LG

Gauge-Equivariant Graph Networks via Self-Interference Cancellation

Yoonhyuk Choi , Jiho Choi , Jiwoo Kang This is my paper

Pith reviewed 2026-05-21 17:44 UTC · model grok-4.3

classification 💻 cs.LG
keywords gauge-equivariant GNNself-interference cancellationoversmoothingheterophilygraph neural networksU(1) gaugerank-1 projectionmessage passing
0
0 comments X

The pith

A projection-based mechanism cancels self-interference to reduce oversmoothing in gauge-equivariant graph networks.

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

Graph neural networks often struggle with heterophilous graphs because signals become self-reinforcing and phase-inconsistent during message passing. The paper shows that existing gauge-equivariant GNNs lack mechanisms to handle self-interference from redundant low-frequency components, which drives oversmoothing under gauge transport. GESC addresses this by using a U(1) phase connection followed by a rank-1 projection that suppresses self-parallel components before attention, plus a sign-aware gate for negatively aligned neighbors. This replaces standard additive aggregation with an explicit interference-aware approach. Readers should care if they work with graphs where standard methods fail to propagate useful information across dissimilar nodes.

Core claim

The paper claims that the absence of interference handling in gauge-based GNNs is a primary driver of oversmoothing, and introduces GESC to explicitly model and cancel self-interference from redundant low-frequency components using a projection mechanism, leading to consistent outperformance on diverse graph benchmarks.

What carries the argument

A rank-1 projection applied after a U(1) phase connection that suppresses self-parallel components in the message passing process.

If this is right

  • Absence of interference handling drives oversmoothing in gauge-based GNNs.
  • GESC outperforms recent state-of-the-art models across diverse graph benchmarks.
  • The approach provides a unified interference-aware view of message passing.
  • Replacing additive aggregation with projection suppresses self-parallel components effectively.

Where Pith is reading between the lines

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

  • This could inspire interference-aware designs in non-gauge GNNs to address oversmoothing.
  • Connections to phase inconsistency in signals might link to other heterophily solutions.
  • Scalability tests on larger graphs would show if the projection remains efficient.

Load-bearing premise

Self-interference from redundant low-frequency components is the dominant cause of oversmoothing in prior gauge-equivariant GNNs.

What would settle it

Running the same benchmarks with and without the rank-1 projection and sign-aware gate to check if performance gains disappear and oversmoothing metrics increase when interference handling is removed.

Figures

Figures reproduced from arXiv: 2511.16062 by Jiho Choi, Jiwoo Kang, Yoonhyuk Choi.

Figure 1
Figure 1. Figure 1: (Left) While traditional additive aggregation uni￾formly accumulates neighbor messages, (right) our wave￾interference mechanism employs magnetic transport, SIC, and sign-aware gating to align phases and cancel redundant components, effectively suppressing detrimental neighbors while preserving informative signals. be faithfully captured by additive aggregation. Such phe￾nomena naturally arise in systems go… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of GESC. (Left) Neighbor messages are parallel transported and self-parallel components are cancelled via rank-1 projection, yielding gauge-invariant complex scores. (Middle) A sign-aware residual gate softly interpolates messages before hybrid magnitude-phase attention. (Right) Aggregated features undergo NodeNorm and modReLU, followed by classification with cross-entropy and JS consistency. The … view at source ↗
Figure 4
Figure 4. Figure 4: (Q4) Node classification accuracy versus network [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: (Q3) Sensitivity of node classification accuracy [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation on phase-aware attention variants. We compare the baseline hybrid attention in Eq. [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Consistency of GESC under random U(1) phase perturbations on the Cora dataset. We inject synthetic phase noise into node features and edge transports (hi ← e iϕihi , Uji ← e i(ϕj−ϕi)Uji) and sweep the perturbation scale α ∈ {0, 0.25, 0.5, 0.75, 1.0}. (a) Prediction agreement between original and perturbed runs, (b) logit-level ℓ2 deviation, and (c) attention distribution KL divergence. The full GESC remain… view at source ↗
read the original abstract

Graph Neural Networks (GNNs) excel on homophilous graphs but often fail under heterophily due to self-reinforcing and phase-inconsistent signals. We propose a \textbf{G}auge-\textbf{E}quivariant Graph Network with \textbf{S}elf-Interference \textbf{C}ancellation (GESC), which replaces additive aggregation with a projection-based interference mechanism. Unlike prior magnetic or gauge-equivariant GNNs that rely on additive message mixing, GESC explicitly models self-interference arising from redundant low-frequency components. We show that the absence of interference handling in existing gauge-based GNNs is a primary driver of oversmoothing under gauge transport. We introduce a $\mathrm{U}(1)$ phase connection followed by a rank-1 projection that suppresses self-parallel components before attention, and a sign-aware gate that regulates negatively aligned neighbors. Across diverse graph benchmarks, GESC consistently outperforms recent state-of-the-art models while offering a unified, interference-aware view of message passing. Our code is available at https://github.com/ChoiYoonHyuk/GESC.

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

2 major / 2 minor

Summary. The manuscript introduces Gauge-Equivariant Graph Network with Self-Interference Cancellation (GESC). It replaces additive aggregation in gauge-equivariant GNNs with a projection-based mechanism consisting of a U(1) phase connection, a rank-1 projection that suppresses self-parallel low-frequency components before attention, and a sign-aware gate for negatively aligned neighbors. The central claim is that the absence of interference handling in prior gauge-based models is a primary driver of oversmoothing under gauge transport; the paper reports consistent outperformance over recent state-of-the-art models on diverse graph benchmarks and releases code.

Significance. If the attribution of gains specifically to self-interference cancellation holds, the work supplies a concrete mechanism for mitigating oversmoothing in heterophilous and gauge-transport settings and unifies message passing under an interference-aware lens. The public code release supports reproducibility.

major comments (2)
  1. [Experimental results / ablation studies] The central claim that redundant low-frequency self-interference (rather than attention dilution, normalization, or the sign-aware gate) is the dominant oversmoothing driver requires isolation. The reported gains compare the full GESC model against external baselines, but no ablation is described that holds the U(1) phase connection and sign-aware gate fixed while removing only the rank-1 projection. This is load-bearing for the claim in the abstract that 'the absence of interference handling ... is a primary driver of oversmoothing under gauge transport.'
  2. [Method section (projection formula)] The exact definition and derivation of the rank-1 projection operator (including how it is applied after the U(1) phase connection and before attention) must be stated with explicit equations. Without this, it is not possible to verify that the projection indeed cancels self-parallel components by construction rather than through additional learned parameters.
minor comments (2)
  1. [Notation and preliminaries] Notation for the sign-aware gate threshold or scaling factor should be introduced once and used consistently; the abstract refers to it only descriptively.
  2. [Figures] Figure captions should explicitly state which baselines are gauge-equivariant versus non-equivariant to aid comparison with the interference-cancellation narrative.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major point below and will revise the manuscript accordingly to strengthen the presentation and empirical support for our claims.

read point-by-point responses

  1. Referee: [Experimental results / ablation studies] The central claim that redundant low-frequency self-interference (rather than attention dilution, normalization, or the sign-aware gate) is the dominant oversmoothing driver requires isolation. The reported gains compare the full GESC model against external baselines, but no ablation is described that holds the U(1) phase connection and sign-aware gate fixed while removing only the rank-1 projection. This is load-bearing for the claim in the abstract that 'the absence of interference handling ... is a primary driver of oversmoothing under gauge transport.'

    Authors: We agree that an internal ablation isolating the rank-1 projection is necessary to substantiate the central claim. In the revised manuscript we will add an ablation that fixes the U(1) phase connection and sign-aware gate while comparing performance with and without the rank-1 projection. This will directly test whether the projection, rather than the other components, is responsible for mitigating self-interference and oversmoothing. revision: yes

  2. Referee: [Method section (projection formula)] The exact definition and derivation of the rank-1 projection operator (including how it is applied after the U(1) phase connection and before attention) must be stated with explicit equations. Without this, it is not possible to verify that the projection indeed cancels self-parallel components by construction rather than through additional learned parameters.

    Authors: We acknowledge that the current exposition would benefit from greater mathematical precision. In the revised manuscript we will add explicit equations in the Method section defining the rank-1 projection operator, deriving its action on the phase-adjusted features, and showing its application immediately after the U(1) phase connection and before the attention step. The derivation will demonstrate that the operator projects out the self-parallel component by construction. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical claims and independent architectural components

full rationale

The paper defines GESC via explicit new components (U(1) phase connection, rank-1 projection for self-interference suppression, sign-aware gate) and reports benchmark outperformance as empirical outcomes. The statement that absence of interference handling drives oversmoothing is framed as an observation from prior models lacking the mechanism, not as a quantity derived by construction from the paper's own fitted parameters or equations. No self-definitional loops, fitted-input predictions, or load-bearing self-citations appear; the derivation chain remains self-contained against external benchmarks and code release.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the modeling choice that self-interference is the dominant source of oversmoothing and that a rank-1 projection after a U(1) phase connection is sufficient to cancel it; no new physical entities are introduced.

free parameters (1)
  • sign-aware gate threshold or scaling factor
    The gate that regulates negatively aligned neighbors is introduced without a derivation from first principles and is therefore treated as a tunable component.
axioms (1)
  • domain assumption U(1) phase connection can be attached to existing gauge-equivariant message passing without breaking equivariance
    Invoked when the authors state they introduce a U(1) phase connection followed by the projection.

pith-pipeline@v0.9.0 · 5723 in / 1395 out tokens · 34119 ms · 2026-05-21T17:44:30.892476+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

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.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

  1. [1]

    52 yields X j∈N(i) α(m) ji bm(m) j→i 2 ≤ ∥W (m)∥2 ·max j∈N(i) ∥h(t) j ∥2,(62) which proves the claim

    By the triangle inequality and convexity (a weighted average is bounded by the maximum term), one can introduce: X j∈N(i) α(m) ji bm(m) j→i 2 ≤ X j∈N(i) α(m) ji ∥bm(m) j→i∥2 ≤max j∈N(i) ∥bm(m) j→i∥2 ≤max j∈N(i) ∥˜h(m) j→i∥2.(61) Combining with Eq. 52 yields X j∈N(i) α(m) ji bm(m) j→i 2 ≤ ∥W (m)∥2 ·max j∈N(i) ∥h(t) j ∥2,(62) which proves the claim. B.3. Ga...

    work page 2017

  2. [2]

    introduces magnetic Laplacians for directional structure, and GCNII (Chen et al., 2020) incorporates identity mapping to counter over-smoothing. • Adaptive and structure-enhanced models:ACM-GCN (Luan et al., 2022) uses channel mixing, GloGNN (Li et al., 2022a) adds global nodes to enhance long-range propagation, Auto-HeG (Zheng et al., 2023) automates het...

    work page 2020