Active Bellman Fixed Point and Value Iteration
Abstract
A contractive active Bellman operator has one value and geometric iteration.
Theorem 1.1 (The active Bellman value is unique and reached geometrically).
Proof. Machine-checked in Lean as D5/S3/Observer/DynamicProgramming/BellmanFixedPointIteration.bellman_contraction_unique_fixed_point_and_iteration_bound (✓ std3). ∎
Source. Repository-derived.
Commentary.
The value carrier is an arbitrary complete metric semilattice. For bounded belief-value functions this lattice infimum is pointwise minimum, so T(v) is constructed as the minimum of the stopping value G and the least continuation Q_i(v).
The displayed Lipschitz premise is the contraction estimate established immediately before the source theorem: every continuation changes future value by at most the discount factor. Together with gamma strictly below one it makes the constructed T a strict contraction.
Mathlib’s canonical contraction fixed point supplies v star and proves that every other fixed value equals it. Iterating the Lipschitz estimate and using that every iterate fixes v star gives exactly the stated gamma-to-n distance bound for every initial value.
References
- Truth anchor:
D5/S3/Observer/DynamicProgramming/BellmanFixedPointIteration.bellman_contraction_unique_fixed_point_and_iteration_bound