Intrinsic optimality of the linear evaluation complexity measure

Determine whether the instance-dependent quantity \(\rho^\star\), defined as the minimum over mixtures of target policies of the maximum quadratic uncertainty of their feature means under the mixture’s feature second-moment matrix, captures the intrinsic difficulty of simultaneous policy evaluation in stochastic contextual bandits with linear reward structure.

Background

The paper introduces ρ⋆\rho^\star as an instance-dependent measure of the difficulty of constructing a shared sampling mixture for evaluating multiple target policies. PRISMA achieves an upper bound on the number of context draws and reward observations that scales, up to logarithmic factors, with (1∨ρ⋆)/ε2(1\vee\rho^\star)/\varepsilon^2.

The analysis does not establish that this quantity is necessary for policy evaluation. The authors identify matching lower bounds as a way to determine whether mixtures of target policies are optimally efficient and whether more general sampling strategies could achieve lower evaluation costs.

References

An open question is whether $\rho\opt$ captures the intrinsic difficulty of evaluation in the linear setting. Our analysis provides an upper bound in terms of this quantity, but does not establish its necessity. Matching lower bounds would help determine when mixtures of target policies achieve the optimal evaluation cost and whether more general sampling strategies can improve upon them.

— Explore Once, Evaluate Many: Multiple Policy Evaluation for Stochastic Linear Bandits  (2610.22252 - Dragazis et al., 15 Oct 2026) in Section 5, Discussion