Core Concepts
The authors derive minimax sample complexity lower bounds as well as upper bounds for estimating risk measures such as variance, Value-at-Risk (VaR), and Conditional Value-at-Risk (CVaR) in discounted Markov cost processes.
Abstract
The key points of the content are:
Lower bounds:
The authors derive a minimax sample complexity lower bound of Ω(1/ε^2) for estimating VaR, CVaR, and variance in two types of Markov cost process (MCP) problem instances: one with deterministic costs and the other with stochastic costs.
The lower bound proofs involve novel techniques, including solving a constrained optimization problem for the deterministic costs case.
The lower bound for mean estimation also improves upon the existing Ω(1/ε) bound.
Upper bounds:
Using a truncation scheme, the authors derive an upper bound of Õ(1/ε^2) for CVaR and variance estimation, matching the corresponding lower bounds up to logarithmic factors.
They also propose an extension of the estimation scheme to cover a broader class of Lipschitz-continuous risk measures, such as spectral risk measures and utility-based shortfall risk.
Significance:
To the best of the authors' knowledge, their work is the first to provide lower and upper bounds for estimating any risk measure beyond the mean within a Markovian setting.
The lower bounds establish the sample complexity requirements for accurately estimating various risk measures in discounted Markov cost processes.
The upper bounds provide practical estimation schemes that achieve the optimal sample complexity, up to logarithmic factors.