Classical fractons with cosmological fixed points
arXiv:2608.07672
2026
Dynamics
1 ideas extracted · analyzed Sep 1, 2026
What the math gives to ML
The paper provides a nonstandard attractor mechanism: a conservative Hamiltonian system can lack attractors in full phase space while its projected scale and shape variables converge to stable fixed-point configurations. The transferable asset is the separation of global scale from normalized shape, together with central-configuration equations that predict which parameter distributions are attracting. A neural-network implementation can augment parameters with momentum-like variables, decompose weights into norm and direction, and stabilize only the directional shape dynamics while allowing compensating motion in hidden momentum variables. This gives a falsifiable prediction: normalized parameter configurations should converge at a measurable rate even if the complete augmented optimizer remains non-dissipative.
Ideas from this paper
Unverified
2026
Represent each trainable parameter block as a global scale multiplied by a normalized shape, and evolve the shape through a projected Hamiltonian optimizer. The optimizer is designed so that normalized weights can approach a stable central configuration while auxiliary momenta retain phase-space volume that prevents ordinary Hamiltonian dynamics from having a full-space attractor.
Useful6/10
Difficulty6/10
Novelty8/10