Optimal Deterministic Multicalibration and Omniprediction

2026-06-18 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Mathematics & Computational Sciences · Depth: Expert, quick

Summary

A new algorithm, published on 2026-06-18, achieves minimax-optimal Õ(ε⁻³) sample complexity for ε-multicalibration using a deterministic predictor. This resolves an open problem regarding the necessity of randomization for optimal sample complexity, explicitly asked by [CLNR26]. Previously, all predictors attaining this optimal rate were randomized, while deterministic counterparts exhibited substantially worse sample complexity. The algorithm further generalizes to produce optimal deterministic predictors satisfying outcome indistinguishability (OI) for finite or finitely covered test collections. This advancement also yields deterministic omnipredictors and panpredictors with optimal sample complexity, addressing open problems posed by [OKK25] and [BHHLZ25].

Key takeaway

For research scientists developing trustworthy machine learning models, this work demonstrates that optimal multicalibration and related fairness properties no longer require randomized predictors. You can now achieve minimax-optimal Õ(ε⁻³) sample complexity with deterministic algorithms, simplifying model deployment and analysis. Consider integrating these deterministic approaches when designing or evaluating systems requiring strong calibration and outcome indistinguishability guarantees.

Key insights

Deterministic multicalibration can achieve minimax-optimal sample complexity, resolving a key open problem in trustworthy machine learning.

Principles

• Multicalibration ensures unbiased predictions across reweighted contexts.
• Outcome indistinguishability extends calibration to broader test collections.
• Optimal sample complexity is achievable deterministically.

Method

The algorithm first provides optimal deterministic multicalibration, then generalizes to produce optimal deterministic predictors satisfying outcome indistinguishability for various test collections.

In practice

• Apply deterministic predictors for fairness in machine learning.
• Use for optimal omniprediction and panprediction tasks.

Topics

• Machine Learning
• Multicalibration
• Deterministic Algorithms
• Sample Complexity
• Omniprediction
• Outcome Indistinguishability

Best for: AI Scientist, Research Scientist

Related on AIssential

• Optimal Deterministic Multicalibration and Omniprediction — Deterministic predictors can achieve minimax-optimal sample complexity for multicalibration and omniprediction, resolving a key open problem.
• PAC-Bayesian Generalization Guarantees for Fairness on Stochastic and Deterministic Classifiers — PAC-Bayesian bounds can guarantee fairness for both stochastic and deterministic classifiers.
• Minimax Optimal Two-Sample Testing under Local Differential Privacy — Minimax optimal two-sample tests under LDP are developed for multinomial and continuous data.
• Quantifying Sensitivity for Tree Ensembles: A symbolic and compositional approach — The paper introduces XCount, a compositional, ADD-based algorithm for quantifying DTE sensitivity with certified approximate counts, outperforming prior methods.
• UnpredictaBench: A Benchmark for Evaluating Distributional Randomness in LLMs — LLMs struggle significantly with generating samples calibrated to true underlying distributions, hindering their use in simulations.
• Partial Causal Structure Learning for Valid Selective Conformal Inference under Interventions — Selective conformal prediction under interventions can be robustly maintained by learning partial causal structures.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.