Skip to main content

Symplectic Stability Classes and the Efficacy of Linear Canonical Transform-Based Token Mixers

By
Sejeung Choi

Linear canonical transforms (LCTs) furnish a structured family of token-mixing operators parameterized by fractional rotation, scaling, and quadratic-phase shear through the Iwasawa decomposition. Each parameter prescription determines a symplectic matrix whose trace distinguishes elliptic from hyperbolic canonical transformations. Across separately evaluated real and imaginary parts of the sampled LCT kernel, elliptic prescriptions are associated with higher classification performance, whereas the associations of performance with kernel rank and conditioning reverse between the two parts. Under unconstrained parameter learning, elliptic initializations retain their stability class, and near-boundary hyperbolic initializations cross into the elliptic region, while farther hyperbolic initializations generally remain outside. A depth-dependent eigenvalue-amplification hypothesis does not explain the observed separation. Trace-based classification therefore provides a geometric descriptor of token-mixing efficacy that is distinct from the spectral properties of the deployed real-valued kernels.

Read on IEEE Xplore