arXiv:2608.19835 2026 Dynamics 1 ideas extracted · analyzed Sep 1, 2026
The paper identifies a concrete mechanism for stable, norm-preserving evolution: symmetric channel-coupling matrices combined with divergence-free spatial coefficients make the transport operator skew-adjoint in an L2 Hilbert space. This implies conservation of feature energy and, under the stated regularity conditions, uniqueness of the Cauchy evolution rather than merely formal stability. A direct neural-network transfer is an energy-preserving spatial or sequence-mixing block implemented with a skew-symmetric discretized generator and an exponential or Cayley update, providing a principled alternative to unconstrained residual layers.
Replace an unconstrained spatial residual block by a discretized transport evolution whose generator is skew-adjoint. Symmetric channel matrices and divergence-free spatial coefficients make the continuous operator energy-preserving, while a matrix exponential or Cayley transform gives an exactly norm-preserving discrete layer.