[Autoplectic]

Brings dit's new discrete entropy estimators (dit PRs #295 and #305) into infoflow, benchmarks them against the old defaults, and adopts the one change that held up: Chao-Wang-Jost for the flow layers. Full results are in reports/dit_estimators.md.

Changes

• Any dit entropy estimator for the layers: infer_multiplex(estimator=...), estimate_edge, layer_statistics and crossfit_flows accept any of dit.inference.estimators.ENTROPY_ESTIMATORS. Intrinsic flow is evaluated at the hardened channel, because these estimators need integer counts (measures.evaluation_table).
• Any dit entropy estimator for selection: SkeletonSettings(entropy_estimator=...) swaps the entropies in the plug-in selection statistic. The permutation tests stay valid for any statistic.
• New default: measures.DEFAULT_LAYER_ESTIMATOR = "chao_wang_jost", replacing Miller-Madow. It's used by the functions above, the channel layer and auto's cost pilot. The selection statistic stays plug-in, and compare_networks keeps Miller-Madow. Pass estimator="miller_madow" to reproduce old results.
• Single-outcome guard: single-outcome entropy terms return 0 for dit estimators (PYM raises on them).
• dit lock: moved to 6314d14e. dit #306 renamed prng to rng, and total_correlation_ksg rejects the old name despite the deprecation alias, so infoflow now passes rng.
• Validation scripts:
  • validation/dit_estimators.py: layer accuracy on sofic processes with exact flows.
  • validation/dit_estimators_selection.py: selection on random networks with known graphs.
  • validation/dit_estimators_catfr.py: a bias probe and a selection check on catFR1 iEEG.
  • validation/new_defaults.py: configurations compared on VAR, logistic-map and Lorenz-96 networks, the motifs, NetSim and CausalRivers, with optional copula normalisation.

Findings

• Layers. Chao-Wang-Jost lowered the cross-fitted layer RMSE against Miller-Madow from 0.047 to 0.035 bits at 300 samples and from 0.019 to 0.014 at 1,000 (0.007 for both at 3,000). Most of the gain is in synergistic bias. On catFR with the source shuffled (every true layer zero), the spurious synergy per edge dropped from 0.0010 to 0.0004 bits.
• Default switch on whole networks.
  • Intrinsic-layer recall rose on every synthetic type, at unchanged precision: VAR 0.29 → 0.39, logistic maps 0.35 → 0.54, Lorenz-96 0.04 → 0.33.
  • NetSim and CausalRivers were unchanged.
  • Networks took about 1.5× as long.
• Calibration. On 40 pairs of independent AR(1) processes, false layers with Chao-Wang-Jost were 3 intrinsic, 5 synergistic and 5 shared, against 1, 5 and 13 with Miller-Madow. The uncoupled test now counts runs with any false flow (at most 1 of 8), because on one seed Chao-Wang-Jost flags both layers of the edge Miller-Madow flags once.
• Selection: not adopted.
  • Miller-Madow raised recall by 0.05 on VAR and logistic networks at no cost, but cost CausalRivers 0.06 AUROC. A decomposition run confirmed the selection change was the whole cause.
  • Chao-Wang-Jost, Wu-Yang and Panzeri-Treves raised recall more, but worsened edge-score AUROC on most networks. Panzeri-Treves is also 12-120× slower than plug-in.
• Second batch (#305): no improvement.
  • Panzeri-Treves has the lowest mean layer RMSE but is worse than Chao-Wang-Jost in most settings.
  • Quadratic extrapolation helps on catFR but not on the synthetic processes.
  • Unseen and PYM suppress the flow itself.
  • The Gaussian copula helps CausalRivers (AUROC 0.581 → 0.606) and hurts NetSim and short logistic series.
• Rejected estimators: NSB and jackknife are accurate but far too slow (about 2,500× and 180× plug-in per entropy). Shrinkage, Laplace, Bonachela and asymptotic NSB are worse than Miller-Madow.

Tests

• New tests check every estimator's layer CMI against dit's conditional_mutual_information, check the layer identities, check the selection CMI and the recovery of a chain under three estimators, and cover the single-outcome guard.
• Full suite: 277 passed, 10 skipped (optional extras not installed).

Made with Cursor