Efficient Online Proportional Sampling with Applications to Smoothed Online Learning

2026-07-13 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

This research introduces a novel data structure for efficient online proportional sampling in high-dimensional domains, specifically addressing scenarios where the sampling distribution is derived from a dynamically evolving weight function over piecewise-structured partitions. The core challenge lies in managing the exponential growth of subregions, which can reach O(t^d) by round t in d dimensions. The proposed data structure supports efficient updates and sampling, effectively circumventing this complexity by leveraging axis-parallel hyperplane discontinuities. Under a σ-smoothed adaptive adversary, the authors prove a tight O(√σT) bound on the data structure's depth, and an O(log T) bound under a random-order adversary, marking the first such results for this problem class. This framework is applied to online learning with piecewise-structured rewards, yielding efficient no-regret algorithms with provable sublinear regret guarantees for both full-information and bandit feedback.

Key takeaway

For Machine Learning Engineers developing online learning systems with high-dimensional, dynamically changing reward functions, this research offers a critical advancement. You can now implement efficient proportional sampling without facing the prohibitive O(t^d) growth of subregions, thanks to the proposed data structure. This enables robust no-regret algorithms under both full-information and bandit feedback, providing provable sublinear regret guarantees. Consider integrating this structured approach to optimize performance in applications like algorithm configuration or dynamic pricing.

Key insights

A novel data structure enables efficient online proportional sampling in high-dimensional, dynamically evolving systems, achieving tight bounds.

Principles

• Exponential growth in subregions can be avoided with structured data.
• Smoothed adversaries allow for tighter bounds in online sampling.
• Piecewise structures are key to managing high-dimensional complexity.

Method

The method involves designing a data structure that supports efficient updates and proportional sampling by exploiting axis-parallel hyperplane discontinuities to avoid explicit O(t^d) subregion growth.

In practice

• Apply to principal-agent games for pricing and contracts.
• Use for algorithm configuration and parameter tuning.
• Implement no-regret online learning with structured rewards.

Topics

• Online Proportional Sampling
• Smoothed Online Learning
• High-Dimensional Data
• Data Structures
• Regret Minimization
• Algorithm Configuration

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Upper-Linearizability of Online Non-Monotone DR-Submodular Maximization over Down-Closed Convex Sets — Non-monotone DR-submodular maximization over down-closed sets is 1/e-linearizable, enabling online linear optimization reduction.
• Learning with Simulators: No Regret in a Computationally Bounded World — Given a simulator for dependent data, classical VC dimension learning guarantees can be recovered, broadening the PAC model.
• Spectral bandits for smooth graph functions with applications in recommender systems — Spectral bandits efficiently learn smooth graph functions by utilizing an "effective dimension" for scalable online recommendation.
• Revisiting Active Sequential Prediction-Powered Mean Estimation — Optimal confidence width in active sequential prediction-powered mean estimation favors a high constant query probability.
• Revisiting Decentralized Online Convex Optimization with Compressed Communication — FTRL-type algorithms can significantly improve decentralized online convex optimization with compressed communication.
• The Price of Hidden Curvature: An $\widetildeΩ (d^{5/4} qrt{T})$ Lower Bound for Bandit Convex Optimization — Stochastic bandit convex optimization is fundamentally harder than linear bandits, with a new ÕÎ(d^{5/4}√T) regret lower bound.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.