Theorem 2.1 (Strong Duality)
ProvedInfoRelax.IdealPenalty.theorem_2_1_strong_dualitydualitydynamic-programminginformation-relaxationp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let be a relaxation of the natural filtration , let the feasible set be nonempty, and assume exists for every policy . Then
Furthermore, if the primal value on the left is finite (the primal problem is bounded), the dual problem on the right has an optimal solution whose dual bound equals the primal value.
This is the analogue of the strong duality theorem of linear programming: there is no gap between the primal DP and its information-relaxation dual.
Formalization Note Both sides are in the extended reals, so (7) is asserted also when the primal value is . "Bounded" means the primal value is not .
Preamble
import Mathlib import Definitions.Def_InfoRelax_IdealPenalty_Framework open MeasureTheory
Formal statement
namespace InfoRelax.IdealPenalty
/-- **Theorem 2.1 (Strong Duality)**, Brown, Smith & Sun (2010), Oper. Res., DOI
10.1287/opre.1090.0796, p. 4, eq. (7).
Let `𝔾` be a relaxation of `𝔽`. Then
`sup_{α_F ∈ 𝒜_𝔽} 𝔼[r(α_F)] = inf_{z ∈ 𝒵_𝔽} sup_{α_G ∈ 𝒜_𝔾} 𝔼[r(α_G) − z(α_G)]`.
Furthermore, if the primal problem is bounded, the dual problem has an optimal solution
`z* ∈ 𝒵_𝔽` that achieves this bound.
**Formalization Note.** General framework of §2.1–§2.2. Values in `EReal`; (7) is stated also when
the primal value is `+∞`. "The primal problem is bounded" is `primalValue ≠ ⊤`; "achieves this
bound" is `dualBound 𝔾 z* = primalValue` (which, by (7), is the dual value). `A.Nonempty` is the
page's tacit assumption; `hr` is the integrability convention. -/
theorem theorem_2_1_strong_duality {Ω X : Type*} [mΩ : MeasurableSpace Ω] [MeasurableSpace X]
{T : ℕ} (μ : Measure Ω) [IsProbabilityMeasure μ] (A : Set (Fin (T + 1) → X))
(hA : A.Nonempty) (𝔽 𝔾 : Filtration (Fin (T + 1)) mΩ) (r : (Fin (T + 1) → X) → Ω → ℝ)
(hr : RewardIntegrable μ A r) (h𝔾 : IsRelaxation 𝔽 𝔾) :
primalValue μ A 𝔽 r = dualValue μ A 𝔽 𝔾 r ∧
(primalValue μ A 𝔽 r ≠ ⊤ →
∃ z ∈ dualFeasible μ A 𝔽, dualBound μ A 𝔾 r z = primalValue μ A 𝔽 r) := by sorry
end InfoRelax.IdealPenalty
Source
Brown, Smith & Sun, Information Relaxations and Duality in Stochastic Dynamic Programs, Oper. Res. (Articles in Advance, 2010), DOI 10.1287/opre.1090.0796, p. 4, Theorem 2.1, eq. (7)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.