Central Limit Theorems for Stochastic Gradient Descent Quantile Estimators
This paper introduces a novel theoretical framework for understanding quantile estimation using stochastic gradient descent (SGD) with constant learning rates. The authors treat the SGD iteration as a Markov chain, demonstrating its convergence to a stationary distribution regardless of initialization. They prove a central limit theorem for the quantile SGD estimator, providing the first theoretical guarantees for this method in non-smooth and non-strongly convex settings. Additionally, a recursive algorithm is proposed for constructing confidence intervals, with numerical studies validating the approach. AI
IMPACT Provides foundational theoretical guarantees for quantile estimation in SGD, potentially improving the reliability of ML models in non-standard settings.