Skip to content

v0.12.0

Choose a tag to compare

@tjkessler tjkessler released this 05 Aug 17:35
· 64 commits to main since this release
bd82ebd

Scale-and-honesty release on the 0.11 stack: Kronecker-sum exact spectrum
for eligible networked graph operators, and docs-only topology-criticality
honesty. Homogeneous scientific defaults and the linear latent operator
contract are unchanged.

Added

  • Exact Kronecker-sum spectrum path for eligible discrete and continuous
    graph operators (koopman_graph.operators.kronecker_spectrum;
    GraphKoopmanOperator.spectrum /
    ContinuousGraphKoopmanOperator.spectrum). Eligibility requires a
    shared self factor, adjacency in {"symmetric", "random_walk"}, and
    sparsity in {"dense", "block_diagonal"}. Eigenvalues and ambient
    eigenvectors match the dense assembly oracle within float tolerance on
    path/ring fixtures.
  • Three-path spectrum contract: distributed leading-modulus Arnoldi
    (discrete / multiplex hetero), Kronecker-sum exact reduction when
    eligible, otherwise dense assembly. Continuous distributed has no
    Arnoldi spectrum path and uses dense L_eff.
  • Sphinx Scale / architecture / FAQ prose distinguishing the three
    spectrum paths and residual dense ceilings for exact inverse, Jacobi /
    Richardson, and eig-regularization hinges. Tutorial
    examples/29_large_graph_block_diagonal.ipynb records a dense vs
    Kronecker spectrum timing probe at the existing probe (N).
  • Topology-criticality honesty in limitations (cross-linked from
    architecture / capabilities), with paper.bib entry Ghosh2025.
    Tutorial examples/18_networked_koopman_dynamic_topology.ipynb notes
    that dynamic topology remains mechanical (no spectral-degeneracy /
    early-warning diagnostic API).

Notes

  • Kronecker reduction applies to spectrum only. Exact inverse /
    Jacobi / Richardson and eig-regularization hinges remain dense
    ((N\cdot d)) ceilings.
  • No Kronecker-style spectrum for hetero / hypergraph ambient operators
    in this release.
  • Homogeneous training defaults are unchanged relative to 0.11.0.

Full changelog: see CHANGELOG.md ([0.12.0] - 2026-08-05).