Equivariant Critical-Mode Branch Seeding
Implementation & benchmark of arXiv:2608.11101 — Local and Global Equivariant Bifurcation for Periodic Weyl and Riesz Fractional Equations
Source paper: Local and Global Equivariant Bifurcation for Periodic Weyl and Riesz Fractional Equations arXiv:2608.11101 ⓘ · analyzed Sep 1, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
When a symmetry-frequency block becomes critical, initialize or perturb the network specifically along its critical representation rather than injecting isotropic noise into all hidden channels. This creates trainable branches for the symmetry patterns predicted by the bifurcation calculation and can expose useful periodic solutions that ordinary symmetry-preserving training fails to reach.
Formulas
Mathematical statement
Let Pσ project hidden states onto the isotypical component associated with irreducible representation σ, and let Qk project a periodic trajectory onto temporal Fourier mode k. Define Jσ,k = QkPσMθPσQk, where Mθ is the one-period return Jacobian. A critical vector vσ,k satisfies Jσ,kvσ,k = λσ,kvσ,k with |1 − λσ,k| ≤ ε. A periodic perturbation seed is δhₜ = a Re(exp(2πikt/T)vσ,k), where a is the seed amplitude and T is the period. If a subgroup H should remain invariant, apply its projector ΠH = |H|⁻¹ Σg∈H ρ(g). The transferred bifurcation mechanism is that critical representations determine admissible symmetry types; the neural test is whether those subspaces produce distinct useful attractors or trajectory branches.
Implementation notes
1. Integration point: add a branch-seeding stage to a periodic RNN or neural ODE used for trajectory prediction. First train a base model to a periodic orbit or recurrent state h*. Then estimate its one-period return Jacobian and identify the lowest-margin symmetry-frequency block. 2. Pseudocode: compute Jacobian-vector products with automatic differentiation; project them into every Pσ sector; use an FFT over the period to separate k; run Arnoldi on each block; select vσ,k whose eigenvalue is closest to one; generate seeds with amplitudes a in {0.01, 0.03, 0.1} and both signs; apply ΠH; roll out every seed; continue training the branches using task loss, periodic closure loss, and the diversity penalty in the formula. Reject branches whose closure residual exceeds a fixed tolerance. 3. Taken from the paper: spatial-isotypic and temporal-mode decomposition and the role of critical representations in determining possible symmetry types. Estimated empirically: Jacobians, eigenvectors, nonlinear branch quality, basin sizes, and whether branches differ beyond a trivial time phase shift. 4. First cheap experiment: use a four-node cyclic ring with a shared equivariant GRU to learn trajectories containing traveling and standing waves. Compare isotropic Gaussian seeds, random Fourier seeds, and critical-mode seeds at equal compute. The prediction is that the first successful nontrivial branch lies in the block with smallest |1 − λσ,k|. Vary seed amplitude and measure a sharp basin transition: below a critical amplitude no branch leaves the base attractor, while above it at least one branch has low periodic-closure error. Report sector agreement and closure-error transition, not only test loss.
Verification
This idea has not been verified yet.
Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.
Artifacts
Artifacts unavailable.