Complex dynamics perspective for birational maps of the plane arising from cluster algebra mutations
arXiv:2607.08125
2026
Architecture
2 ideas extracted · analyzed Aug 30, 2026
What the math gives to ML
The transferable asset is the functorial tropicalization of birational cluster mutations: complicated rational updates become explicitly computable, homogeneous, piecewise-linear maps on exponent or log-magnitude coordinates. These maps are compositional, exactly invertible, and have analytically known expanding and contracting directions through their tropical eigenvalues. A promising neural-network use is a reversible, low-cost state-mixing block for normalizing flows, reversible transformers, or memory-constrained recurrent networks, with the dynamical degree providing a direct knob for controlling long-horizon expansion.
Ideas from this paper
△ Mechanism confirmed, baseline not beaten
2026
Replace a conventional two-layer channel mixer in a reversible architecture with the tropicalization of two cluster mutations. For every pair of channels, the block applies sign-dependent integer shears and reflections, giving a cheap piecewise-linear transformation that is exactly invertible and requires no stored activations during backpropagation. Continuous trainable affine scale and mixing parameters can be placed around the fixed tropical core.
Useful7/10
Difficulty4/10
Novelty8/10
Unverified
2026
Use the tropical dynamical degree as an analytic expansion budget for repeated neural blocks. Layers with $pq>4$ deliberately expand along a known tropical eigendirection, while layers with $pq\leq4$ avoid exponential asymptotic growth; a schedule can therefore increase representational mixing without allowing hidden-state norms to explode.
Useful5/10
Difficulty3/10
Novelty7/10