Unverified 2026

Green-Trace Cycle Regularizer

Implementation & benchmark of arXiv:2607.24857 — Arithmetic Landscape Functions of a Discrete Cat Map

Usefulness5/10

Difficulty6/10

Novelty7/10

Regularization Stability Generalization Moe routing Rnn Transformer Regularization Dynamical systems Spectral theory Linear algebra Combinatorics

Source paper: Arithmetic Landscape Functions of a Discrete Cat Map arXiv:2607.24857 ⓘ · analyzed Aug 31, 2026

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 Not peer-reviewed.

Idea description

Use the resolvent trace as a differentiable statistic that controls how strongly a learned routing or recurrent transition matrix returns to short cycles. Penalizing this quantity suppresses accidental short feedback loops, while matching a target trace can impose a desired memory profile in recurrent, graph, or mixture-of-experts architectures.

Formulas

ZT(α) = exp(Σ(n=1 to ∞) Fn α^n/n), where Fn = number of fixed points of T^n.

trace(Gα) = |X| − α d/dα log ZT(α) = |X| + Σ(n=1 to ∞) Fn α^n.

Rcycle(θ) = Σ(n=1 to K) wn α^n trace(Qθ^n), where Qθ is a row-stochastic routing matrix.

Ltotal = Ltask + λcycle Rcycle, with λcycle ≥ 0.

Mathematical statement

For a finite deterministic map T with transfer matrix P, let Gα=(I−αP)^−1 and let Fn be the number of fixed points of T^n. The dynamical zeta function is ZT(α)=exp(Σn≥1 Fn α^n/n). The Green-zeta identity is trace(Gα)=|X|−α d/dα log ZT(α)=|X|+Σn≥1 Fn α^n. For a learned row-stochastic transition matrix Qθ, replace deterministic return counts by trace(Qθ^n), and use a truncated cycle penalty Rcycle=Σn=1K wn α^n trace(Qθ^n), with nonnegative weights wn. This penalty directly measures short-return mass and is differentiable through Qθ.

Implementation notes

1. Integration point: apply the regularizer to a learned token-routing matrix, sparse mixture-of-experts router, graph transition matrix, or recurrent attention matrix. Construct Qθ from router logits with a row-wise softmax. 2. Pseudocode: Q = row_softmax(router_logits); Qpow = Q; R = 0; for n in 1..K: R += w[n]*alpha**n*trace(Qpow); Qpow = Qpow @ Q; loss = task_loss + lambda_cycle*R; backpropagate. For large N, estimate each trace with m Rademacher probes z using trace(Q^n) approximately equal to m^−1 Σi ziᵀQ^nzi. Detach no quantities if the routing matrix should receive gradients. 3. The Green-zeta identity and geometric return weighting are taken from the paper; trace(Q^n) is computed exactly for small matrices or estimated empirically for large ones. 4. First experiment: train a two-layer MoE or recurrent router on CIFAR-10 patch sequences and on synthetic modular-map prediction. Compare no regularizer, entropy regularization, and this penalty with K=4. Sweep λcycle and α. The predicted signature is a monotone reduction in measured short-return statistics trace(Q), trace(Q²), and so on, with the strongest effect at small n. The truncated penalty should approximate the measured resolvent trace with residual scaling like α^(K+1), provided α<1. An intermediate penalty is predicted to reduce recurrent loss spikes without eliminating all useful long-range returns.

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.