✓✓ Beats tuned baseline 2026

Holonomy-Twisted Message Passing

Implementation & benchmark of arXiv:2608.01596 — Beyond $K$-Theory: Geometry and Holomorphy in Hyperbolic Band Theory

Usefulness7/10

Difficulty5/10

Novelty6/10

Architecture Accuracy Generalization Graph nn Embedding Representation theory Linear algebra Spectral theory

Source paper: Beyond $K$-Theory: Geometry and Holomorphy in Hyperbolic Band Theory arXiv:2608.01596 ⓘ · analyzed Aug 31, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Replace an ordinary graph-neural-network edge message by a message transported through a unitary representation of the edge's fundamental-group label. The layer can distinguish globally different holonomy sectors even when the underlying bundles or ordinary graph topology are identical, while inverse edge labels enforce a Hermitian and unitary consistency constraint.

Formulas

$$H_{\rho}=\sum_{v\in V}|v\rangle\langle v|\otimes V_v\otimes I_r+\sum_{e\in\mathcal{E}^+}\left(|t(e)\rangle\langle s(e)|\otimes T_e\otimes\rho(\gamma_e)+|s(e)\rangle\langle t(e)|\otimes T_e^*\otimes\rho(\gamma_e)^*\right)$$

$$\gamma_{\bar e}=\gamma_e^{-1},\qquad T_{\bar e}=T_e^*,\qquad \rho(\gamma_e)\in U(r)$$

$$h_v^{\ell+1}=\phi_\ell\left(W_0^{(\ell)}h_v^\ell+\sum_{e:t(e)=v}W_e^{(\ell)}\,\rho(\gamma_e)h_{s(e)}^\ell\right)$$

Mathematical statement

For a finite directed quotient graph Q=(V,E), each oriented edge e has source s(e), target t(e), a group label gamma_e in Gamma, and a reverse edge bar(e) satisfying gamma_bar(e)=gamma_e^{-1}. The paper assigns a hopping matrix T_e with T_bar(e)=T_e^*, and a unitary representation rho: Gamma -> U(r). Its twisted Hamiltonian has the structure H_rho = sum_v |v><v| tensor V_v tensor I_r + sum_{e in E+} (|t(e)><s(e)| tensor T_e tensor rho(gamma_e) + |s(e)><t(e)| tensor T_e^* tensor rho(gamma_e)^*). Here V_v is an on-site Hermitian matrix, E+ contains one orientation of each geometric edge, m is the orbital width, and r is the representation dimension. The ML adaptation uses the same transport operator in a residual message-passing update. For the abelian case, rho is a character chi: Gamma -> U(1), so each edge contributes a phase chi(gamma_e); these phases are not recoverable from the ordinary K^0 class even though they change spectra.

Implementation notes

1. Integration point: modify the aggregation operation of a message-passing GNN, graph transformer, or molecular GNN. Every directed edge must carry a label gamma_e, represented either by a word in a supplied group or by a precomputed unitary matrix U_e. Keep node states in complex width d, or represent each complex vector by concatenated real and imaginary parts. For the first MVP use scalar phases U_e=exp(i theta_e). 2. Pseudocode: initialize node states h_v. For every layer ell, set m_v=0. For every directed edge e:u->v, compute transported state z_e=U_e h_u, compute message q_e=MLP_ell([Re(z_e), Im(z_e), edge_features_e]), and add q_e to m_v. Update h_v=LayerNorm(h_v+MLP_update([h_v,m_v])). Add reverse edges with U_bar_e=conjugate(U_e). 3. The paper-derived constraints are unitary transport and inverse/reverse consistency; neural weights, nonlinearities, and normalization are learned. If matrices are learned, parameterize U_e by a matrix exponential of a skew-Hermitian matrix or regularize ||U_e^*U_e-I||_F^2. 4. First cheap experiment: generate small quotient graphs with surface-group edge labels and define targets from spectra of the corresponding twisted Hamiltonian. Construct pairs with identical adjacency and ordinary topological summaries but different cycle holonomies. Compare a GCN or GAT baseline against the twisted layer at equal parameter count. Success is higher accuracy on labels determined by cycle products, lower spectral-prediction error on unseen holonomy assignments, and no degradation when all U_e equal the identity.

Verification

Beats tuned baseline

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: Significant win ✓

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 9/10

Built a scalar-phase holonomy-twisted message-passing/Hamiltonian MVP with inverse-edge transport and a controlled holonomy classification test. The mechanism checks passed: unitary norm error was 1.78e-15, Hermitian error was 0, eigenvalue imaginary parts were 1.41e-16, the flux trace sweep matched 2n(cos(Phi)-1) within 1.42e-13, and equal-holonomy spectra differed by at most 8.88e-16. In the controlled task, twisted features achieved 1.0 accuracy versus 0.5 for an ordinary-graph constant baseline; this demonstrates the promised information access, but not a broad neural-network performance win.

Agent confidence: 9/10

Baseline: 0.50 accuracy on 200 identical-adjacency graphs with constant ordinary features
Idea: 1.00 accuracy using tr(H^8) from unitary phase transport; unit norm max error 1.78e-15; Hermitian max error 0; max flux-trace prediction error 1.42e-13; same-holonomy spectral spread 8.88e-16

Limitations: Only scalar U(1) phases and small cycle graphs were tested. The classification comparison uses an exact spectral invariant rather than a trained multi-layer GNN, and the baseline is intentionally unable to distinguish identical ordinary inputs; no molecular, noisy, non-cycle, non-abelian, or compute-matched benchmark was evaluated.

How to run: python3 holonomy_mvp.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.01596) Sep 2, 2026

Beats the tuned baseline by 100.0% (p=0.0081, wins 8 / 8 paired seeds; metric: lower is better).

Benchmark: Holonomy_cycle_sector
Model: matched_cycle_mlp
Paired seeds: 8
Baseline mean: 0.5
Idea mean: 0
Effect (Δ): -0.5 (−100.0%; negative = idea better)
Wins: 8 / 8 paired seeds
p-value: 0.0081 (permutation test, 20 000 shuffles)
Smallest detectable effect: ±0.0%
Mechanism: Confirmed ✓
Practical effect: Helps
Baseline tuning: swept over 3 configs

Limitations:

Only scalar U(1) phases and a synthetic fixed 8-cycle classification task were tested. The model uses a compact cycle transport representation rather than a general multi-layer node/message-passing GNN; non-abelian representations, varying graph topology, noise, compute overhead, and real-world graph transfer were not evaluated.

How to run: python3 stage2_bench.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Aug 31, 2026 latest Worked ✓

Agent confidence: 9/10

The registered custom graph track contains the required edge-labeled cycle holonomy structure. The end-to-end holonomy-twisted system achieved 0.000 test error versus 0.500 for the matched identity-transport baseline on all 8 paired seeds; the permutation p-value was 0.0081, yielding the harness verdict idea better (significant). The trained-model signature quantitatively confirmed unitary norm preservation, although the result is limited to a small synthetic U(1) cycle task.

Baseline: mean test error 0.500, std 0.000, 8 seeds; baseline lr sweep {0.003, 0.01, 0.03} all mean 0.500
Idea: mean test error 0.000, std 0.000, 8 seeds; shared lr sweep {0.003, 0.01, 0.03} all mean 0.000
P value: 0.0081
Paired delta: -0.500

Benchmark result