Estimation error is policy-independent (Lemma 5.2.1)
ProvedBertsekasDP.lqg_estimation_error_policy_independentLemma 5.2.1 (the estimation error does not depend on the policy). In the linear-quadratic problem with imperfect state information, let and be any two information-feedback policies, and let , denote the state and its least-squares estimate under . Then for every stage and every outcome of positive probability,
That is, the estimation error is the same realization by realization, whichever controls were applied.
The reason is that the controls are a known function of the measurement history, so they shift the state and the measurements by a quantity that is fully known at the time of estimation and cancels in the error. The consequence is the one the source needs: in the dynamic programming recursion the quadratic estimation-error term is a function of the information alone, so it may be pulled out of the minimization over . Once that term is out of the way, what remains is the perfect-information problem with replaced by — which is exactly certainty equivalence.
Formalization Note This is the pointwise (per-outcome) form of the source's statement, which asserts the existence of a function of the information vector; the pointwise version implies it and is what the dynamic programming argument consumes. The hypothesis that the outcome has nonzero probability guarantees both conditional expectations have nonzero denominators, excluding the junk value of conditional expectation on null events.
import Mathlib import Definitions.Def_BertsekasLQGModel
namespace BertsekasDP
theorem lqg_estimation_error_policy_independent {n m q : ℕ}
{Ω₀ ΩW ΩV : Type} [Fintype Ω₀] [Fintype ΩW] [Fintype ΩV]
(M : BertsekasLQGModel n m q Ω₀ ΩW ΩV)
(π π' : List (Fin q → ℝ) → Fin m → ℝ) (k : ℕ) (hk : k ≤ M.N)
(ω : BertsekasLQGSample M) (hω : BertsekasLQGProb M ω ≠ 0) :
(BertsekasLQGTraj M π k ω).1 - BertsekasLQGEstimate M π k ω =
(BertsekasLQGTraj M π' k ω).1 - BertsekasLQGEstimate M π' k ω := by sorry
end BertsekasDPRead-back
What the Lean code literally says, in plain math · claude-fable-5
This theorem (proof sorry in the file, i.e., not supplied) asserts: for any dimensions , any finite types , any model of type BertsekasLQGModel (with all the structure's data and hypotheses as detailed in its read-back), any two policies (arbitrary functions from lists of measurement vectors in to controls in ), any time with (so ranges over inclusive — including , where no control has yet acted, and , where the conditioning measurement list ends in the noiseless ), and any sample whose product probability is nonzero, the following equality of vectors in holds:
Here is the time- state of the closed-loop trajectory recursion BertsekasLQGTraj under policy for sample (that is, and for , with measurements for ), and is the estimate of BertsekasLQGEstimate under the same policy: the -weighted average of the time- state over all samples whose full measurement list under equals that of , computed with the convention that the average is the zero vector if the total probability of the matching samples is zero (a case excluded here, since the matching set contains itself, all weights are nonnegative, and ). The corresponding quantities on the right-hand side use throughout — both the trajectory and the conditioning event are taken under . The claim is thus that the estimation error at each fixed sample of nonzero probability is the same real vector no matter which of the two policies generated the closed loop; it is stated as a pointwise (per-sample) identity, not as an identity of distributions or expectations.
Confirmed by the mission captain (proposal self-audit).