Time-Aligned Multichannel Coherence Estimation under Intermittent Observations

Abstract

Time misalignment and intermittent readout can bias multichannel dependence estimates by mixing timing error with missing data. Time-Aligned Multichannel Coherence Estimation (TAMCE) reports a causal network coherence statistic G together with an explicit observation-support state Q using normalized recursive channel states, independent timing calibration, and confidence-weighted correlation. A non-saturated benchmark with 16 independent seeds compares coincidence, zero filling, pairwise-complete correlation, causal hold-last completion, non-causal interpolation, diagonal intermittent-observation Kalman filtering, and TAMCE. At channel-correlated missingness 0.60, TAMCE obtains AUC 0.749 and latent-target RMSE 0.554, versus 0.723 and 0.713 for aligned zero filling and 0.826 and 0.418 for Kalman filtering. The TAMCE-zero-fill AUC difference is not statistically significant (p = 0.087). Same-record statistic ablations and selective risk-coverage tests show that Q is useful for abstention but carries essentially the same availability information as direct mask-derived pair support. Heavy-tail and drift stresses do not reverse the Kalman fidelity advantage. TAMCE is therefore supported as a model-light coherence-and-provenance interface for online monitoring and data-quality logic, not as a universally superior estimator.

23 pages, 13 figures, 2 tables. Added a 16-seed non-saturated benchmark, stronger missing-data baselines, latent-target fidelity, risk-coverage analysis, network-statistic ablations, and robustness tests. Claims are explicitly bounded; no detector-level validation or estimator superiority is claimed