Systematic characterization of surrogate mismatch

Investigate systematically how mismatch between a differentiable surrogate’s target condition and the actual rare event affects estimation accuracy for more complex events.

Background

Iterative Unalignment uses a differentiable surrogate to guide proposal optimization toward rare events, while the binary event indicator defines the event being estimated. The surrogate may target a necessary condition, a broader condition, or only part of the event, creating undercoverage or overcoverage relative to the true event region.

The paper reports preliminary experiments in which a proposal trained to amplify one token is evaluated on events that add alternative required or optional tokens. Estimation error increases as the surrogate and event diverge, but the paper does not systematically characterize this tradeoff for more complex events. Such a characterization would clarify when surrogate mismatch remains tolerable and when it causes unacceptable importance-weight instability or estimation error.