Uncertainty Decomposition

Uncertainty decomposition asks why a prediction is uncertain, commonly separating aleatoric uncertainty from noisy or ambiguous observations and epistemic uncertainty from limited knowledge or model uncertainty. Methods use Bayesian posterior variance and mutual information, ensemble or dropout disagreement, evidential models, proper scoring rules, and calibration- or representation-based diagnostics; some distinguish reducible from irreducible risk or observation noise from process randomness. Applications include selective prediction and referral, out-of-distribution detection, active learning, clarification-seeking language agents, medical imaging, forecasting, and robotics, where different uncertainty sources can call for different actions. A central open issue is that these categories depend on modeling and data-acquisition assumptions: estimates can be entangled or unidentifiable, so evaluations need independent signals such as annotator disagreement, distribution shifts, or controlled noise—not just calibration or error ranking.