# Harmonic Global Latent Channels

- ID: 167
- Canonical URL: https://synthcore.org/idea/167/harmonic-global-latent-channels
- API JSON: https://synthcore.org/api/idea/167.json
- API Markdown: https://synthcore.org/api/idea/167.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2608.25266](https://arxiv.org/abs/2608.25266)
- Category: architecture
- Solves: accuracy, generalization, sample-efficiency
- ML areas: graph-nn, world-model, embedding
- Math tags: geometry, topology, linear-algebra, pde
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Augment a mesh or graph neural network with an explicit low-dimensional channel for topological circulation or flux modes. The network predicts a local gauge-fixed field u and global coefficients a, then reconstructs the physical field as y = u + Ha, so local message passing does not need to synthesize global modes through many layers.

## Mathematical statement

The paper's central decomposition is y = u + h, with u perpendicular to the harmonic space H^k and h in H^k, where y is the physical k-form field, u is the gauge-fixed local state, and H^k is the harmonic subspace. Its period map is Pi: H^k -> R^{b_k}, where b_k is the k-th Betti number and Pi(h) records global circulation or flux coordinates. In a discrete mesh, let B_k and B_{k+1} be incidence matrices on k- and (k+1)-simplices, let M_k be the positive-definite k-form mass matrix, and define the discrete codifferential delta_k = M_{k-1}^{-1} B_k^T M_k. Construct an M_k-orthonormal harmonic basis H satisfying B_{k+1}H = 0, delta_k H = 0, and H^T M_k H = I. If P is a discrete period operator obtained by integrating basis k-forms over a chosen homology-cycle basis, use Pi(y) = Py as the global observable. The adapted neural decomposition is y_theta = u_theta + H a_theta, with the gauge condition H^T M_k u_theta = 0, trained using both the task loss and a period loss.

## Key formulas

- $$y=u+h,\qquad u\perp\mathfrak{H}^{k},\qquad h\in\mathfrak{H}^{k}.$$
- $$\bm{\Pi}:\mathfrak{H}^{k}\longrightarrow\mathbb{R}^{b_{k}}.$$
- $$B_{k+1}H=0,\qquad \delta_kH=0,\qquad H^{\mathsf T}M_kH=I,\qquad \delta_k=M_{k-1}^{-1}B_k^{\mathsf T}M_k.$$
- $$y_\theta=u_\theta+Ha_\theta,\qquad H^{\mathsf T}M_ku_\theta=0,\qquad \mathcal{L}=\mathcal{L}_{\mathrm{task}}(y_\theta)+\lambda_{\Pi}\lVert Py_\theta-\pi^*\rVert_2^2+\lambda_{\mathrm{g}}\lVert H^{\mathsf T}M_ku_\theta\rVert_2^2.$$

## Implementation notes

(1) Exact integration point: use this as the output head of a mesh neural operator or graph neural network whose output is a discrete differential k-form, such as edge values for k = 1 or face fluxes for k = 2. Precompute the mesh incidence matrices B_k and B_{k+1} and mass matrices M_k and M_{k-1}. Compute a basis H for the discrete harmonic subspace by solving the nullspace constraints B_{k+1}H = 0 and M_{k-1}^{-1}B_k^T M_k H = 0, then M_k-orthonormalize the columns. (2) Pseudocode: `z = GNN(features, mesh); u_raw = local_head(z); coeff = H.T @ M_k @ u_raw; u = u_raw - H @ coeff; a = global_pool(z, cycle_features); y = u + H @ a; loss = task_loss(y, target) + lambda_period * norm(P @ y - period_target)^2 + lambda_gauge * norm(H.T @ M_k @ u)^2`. For forecasting, concatenate the current global coefficient a_t to every local node or edge feature and predict a_{t+1}. (3) Compute H, the weighted projection H(H^T M_k input), and P exactly from the mesh; estimate only the neural mappings from input features to the local field and coefficients. If period labels are unavailable, use P @ target as the target period. For changing meshes, recompute H and P or transport coefficients through a consistent homology-cycle basis, fixing basis signs and ordering to prevent label permutations. (4) First cheap experiment: use an annulus or square-with-hole mesh, generate Poisson or Maxwell fields with randomized circulation, and compare a four-layer message-passing baseline with this harmonic-head model at equal parameter count. Measure field L2 error, period error, and performance after mesh refinement. Success means lower circulation error at the same depth, particularly when the hole is farther apart than the baseline receptive field, plus improved zero-shot transfer to refined meshes.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: no_effect
- Verdict: Built a weighted discrete harmonic decomposition and a four-layer ring message-passing model with an explicit harmonic latent head. The numerical check verified the harmonic constraints, weighted orthonormality, gauge projection, reconstruction, and period consistency to approximately 1e-16. On the toy global-circulation task, the harmonic head improved field RMSE, period RMSE, and far-edge RMSE, though it used more parameters and the task favors global circulation.

### Mechanism check

- Verdict: Built a weighted discrete harmonic decomposition and a four-layer ring message-passing model with an explicit harmonic latent head. The numerical check verified the harmonic constraints, weighted orthonormality, gauge projection, reconstruction, and period consistency to approximately 1e-16. On the toy global-circulation task, the harmonic head improved field RMSE, period RMSE, and far-edge RMSE, though it used more parameters and the task favors global circulation.
- Confidence: 8/10
- Limitations: Only a 24-edge weighted cycle with one harmonic mode was tested. No 2D mesh, mesh-refinement transfer, equal-parameter matching, FLOP or speed measurement, period-supervised loss, or multiple random seeds were tested. The harmonic model has 3122 parameters versus 2497 for the baseline.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 8/8
- Benchmark verdict: idea better (significant)

## Artifacts

- [bench_report.json](https://synthcore.org/code/1193/bench_report.json)
- [custom_harmonic_track.py](https://synthcore.org/code/1193/custom_harmonic_track.py)
- [experiment.py](https://synthcore.org/code/1193/experiment.py)
- [report.md](https://synthcore.org/code/1193/report.md)
- [report_bench_2026-09-03T031807.md](https://synthcore.org/code/1193/report_bench_2026-09-03T031807.md)
- [results.json](https://synthcore.org/code/1193/results.json)
- [run_bench.py](https://synthcore.org/code/1193/run_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/1193)

## Disclaimer

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