Preference-Class Gap and Pathwise Distortion in Dynamic Decision Models
Аннотация
An analyst may optimise an alternative objective while judging its policy by a benchmark cost. We give a sufficient certificate that the resulting expected loss lies within a stated allowance. For finite-horizon additive rewards on a common feasible system, the Preference-Class Gap is a uniform distance between canonical representatives modulo common positive scale and component constants. The Pathwise Distortion Index is benchmark regret normalised by the sum of component ranges. A component-oscillation bound implies $$\textrm{PDI}\le \min \{1,2N\textrm{PPG}\}$$ PDI ≤ min { 1 , 2 N PPG } , and a certified optimisation error enters additively. These certificates do not require the benchmark optimum; an inconclusive certificate calls for a separate audit. We distinguish net regret from coupling-dependent gross diagnostics. An inventory study includes a hand-derived certificate for a nearby objective and a fixed-cost-free surrogate with a proved base-stock recursion. In a synthetic inventory study, a declared benchmark loss allowance distinguishes certification before benchmark optimisation from adoption or rejection after a value audit. A surrogate without fixed ordering costs admits a base-stock recursion on the specified finite grid, reducing the work required to solve the alternative model.