Documentation

TauCeti.MeasureTheory.OptimalTransport.MultiMarginal.Finite.Duality

Finite multi-marginal duality and complementary slackness #

For finitely many finite spaces, the multi-marginal dual problem has one real-valued potential per marginal, constrained so that their pointwise sum never exceeds the cost.

The basic identity in this file is the multi-marginal analogue of the transportation-matrix gap formula: the difference between the cost of a plan and the value of a family of potentials is the expectation of the pointwise dual gap. It gives weak duality and complementary slackness. A finite-dimensional minimax argument then proves strong duality and attainment of both problems. In particular, a feasible plan and feasible potentials are simultaneously optimal exactly when every configuration carrying mass saturates the dual constraint.

The finite coupling model and its bridge to the measure-theoretic TauCeti.MultiCoupling API live in TauCeti.MeasureTheory.OptimalTransport.MultiMarginal.Finite.Basic.

This completes the finite linear-programming part of multi-marginal duality. The index type is assumed nonempty for strong duality: with no marginals the dual has no variables and cannot represent an arbitrary cost on the one-point empty product.

References #

def TauCeti.finiteMultiDualValue {ι : Type u} {X : ι → Type v} [Fintype ι] [(i : ι) → Fintype (X i)] (μ : (i : ι) → PMF (X i)) (φ : (i : ι) → X i → ℝ) :

The value of a finite family of multi-marginal potentials: the sum of their expectations against the prescribed marginals.

Instances For
    theorem TauCeti.finiteMultiDualValue_def {ι : Type u} {X : ι → Type v} [Fintype ι] [(i : ι) → Fintype (X i)] (μ : (i : ι) → PMF (X i)) (φ : (i : ι) → X i → ℝ) :
    finiteMultiDualValue μ φ = ∑ i : ι, ∑ x : X i, ((μ i) x).toReal * φ i x

    The defining finite-sum formula for the value of multi-marginal potentials.

    theorem TauCeti.finiteMultiDualValue_add_const {ι : Type u} {X : ι → Type v} [Fintype ι] [(i : ι) → Fintype (X i)] (μ : (i : ι) → PMF (X i)) (φ : (i : ι) → X i → ℝ) (a : ι → ℝ) :
    (finiteMultiDualValue μ fun (i : ι) (x : X i) => φ i x + a i) = finiteMultiDualValue μ φ + ∑ i : ι, a i

    Adding a coordinatewise constant to the potentials adds the sum of those constants to the dual value.

    theorem TauCeti.finiteMultiDualValue_mono {ι : Type u} {X : ι → Type v} {μ : (i : ι) → PMF (X i)} [Fintype ι] [(i : ι) → Fintype (X i)] {φ ψ : (i : ι) → X i → ℝ} (h : ∀ (i : ι) (x : X i), φ i x ≤ ψ i x) :

    The finite multi-marginal dual value is monotone in every potential.

    theorem TauCeti.continuous_finiteMultiDualValue {ι : Type u} {X : ι → Type v} [Fintype ι] [(i : ι) → Fintype (X i)] (μ : (i : ι) → PMF (X i)) :
    Continuous fun (φ : (i : ι) → X i → ℝ) => finiteMultiDualValue μ φ

    The finite multi-marginal dual value is continuous in the potentials.

    def TauCeti.FiniteMultiDualFeasible {ι : Type u} {X : ι → Type v} [Fintype ι] (c : ((i : ι) → X i) → ℝ) (φ : (i : ι) → X i → ℝ) :

    A family of potentials is feasible for the finite multi-marginal dual problem when their pointwise sum never exceeds the cost.

    Instances For
      @[simp]
      theorem TauCeti.finiteMultiDualFeasible_iff {ι : Type u} {X : ι → Type v} [Fintype ι] {c : ((i : ι) → X i) → ℝ} {φ : (i : ι) → X i → ℝ} :
      FiniteMultiDualFeasible c φ ↔ ∀ (x : (i : ι) → X i), ∑ i : ι, φ i (x i) ≤ c x

      The pointwise inequality defining finite multi-marginal dual feasibility.

      theorem TauCeti.FiniteMultiDualFeasible.sum_le {ι : Type u} {X : ι → Type v} [Fintype ι] {c : ((i : ι) → X i) → ℝ} {φ : (i : ι) → X i → ℝ} (h : FiniteMultiDualFeasible c φ) (x : (i : ι) → X i) :
      ∑ i : ι, φ i (x i) ≤ c x

      Apply a feasible family of potentials to one configuration.

      theorem TauCeti.FiniteMultiDualFeasible.mono {ι : Type u} {X : ι → Type v} [Fintype ι] {c : ((i : ι) → X i) → ℝ} {φ ψ : (i : ι) → X i → ℝ} (h : FiniteMultiDualFeasible c φ) (hψ : ∀ (i : ι) (x : X i), ψ i x ≤ φ i x) :

      Lowering every potential preserves dual feasibility.

      theorem TauCeti.FiniteMultiDualFeasible.add_const {ι : Type u} {X : ι → Type v} [Fintype ι] {c : ((i : ι) → X i) → ℝ} {φ : (i : ι) → X i → ℝ} (h : FiniteMultiDualFeasible c φ) (a : ι → ℝ) (ha : ∑ i : ι, a i = 0) :
      FiniteMultiDualFeasible c fun (i : ι) (x : X i) => φ i x + a i

      Shifting each potential by a constant whose total is zero preserves dual feasibility.

      theorem TauCeti.FiniteMultiCoupling.sum_mul_mass_eq_finiteMultiDualValue {ι : Type u} {X : ι → Type v} {μ : (i : ι) → PMF (X i)} [Fintype ι] [(i : ι) → Fintype (X i)] (π : FiniteMultiCoupling μ) (φ : (i : ι) → X i → ℝ) :
      ∑ x : (i : ι) → X i, (∑ i : ι, φ i (x i)) * (↑π x).toReal = finiteMultiDualValue μ φ

      Integrating the sum of the coordinate potentials against a coupling gives their dual value. This is the finite change-of-variables identity behind multi-marginal weak duality.

      theorem TauCeti.FiniteMultiCoupling.cost_sub_finiteMultiDualValue {ι : Type u} {X : ι → Type v} {μ : (i : ι) → PMF (X i)} [Fintype ι] [(i : ι) → Fintype (X i)] (π : FiniteMultiCoupling μ) (c : ((i : ι) → X i) → ℝ) (φ : (i : ι) → X i → ℝ) :
      cost c π - finiteMultiDualValue μ φ = ∑ x : (i : ι) → X i, (c x - ∑ i : ι, φ i (x i)) * (↑π x).toReal

      The cost of a plan minus the value of a family of potentials is the mass-weighted sum of the pointwise dual gaps.

      theorem TauCeti.FiniteMultiCoupling.finiteMultiDualValue_le_cost {ι : Type u} {X : ι → Type v} {μ : (i : ι) → PMF (X i)} [Fintype ι] [(i : ι) → Fintype (X i)] (π : FiniteMultiCoupling μ) {c : ((i : ι) → X i) → ℝ} {φ : (i : ι) → X i → ℝ} (hφ : FiniteMultiDualFeasible c φ) :

      Finite multi-marginal weak duality. Every feasible family of potentials has value at most the cost of every feasible plan.

      theorem TauCeti.FiniteMultiCoupling.cost_eq_finiteMultiDualValue_iff {ι : Type u} {X : ι → Type v} {μ : (i : ι) → PMF (X i)} [Fintype ι] [(i : ι) → Fintype (X i)] (π : FiniteMultiCoupling μ) {c : ((i : ι) → X i) → ℝ} {φ : (i : ι) → X i → ℝ} (hφ : FiniteMultiDualFeasible c φ) :
      cost c π = finiteMultiDualValue μ φ ↔ ∀ (x : (i : ι) → X i), ↑π x ≠ 0 → ∑ i : ι, φ i (x i) = c x

      Finite multi-marginal complementary slackness. A feasible plan and feasible family of potentials have the same value exactly when every configuration carrying mass saturates the dual constraint.

      theorem TauCeti.FiniteMultiCoupling.forall_cost_le_and_forall_finiteMultiDualValue_le_of_eq {ι : Type u} {X : ι → Type v} {μ : (i : ι) → PMF (X i)} [Fintype ι] [(i : ι) → Fintype (X i)] (π : FiniteMultiCoupling μ) {c : ((i : ι) → X i) → ℝ} {φ : (i : ι) → X i → ℝ} (hφ : FiniteMultiDualFeasible c φ) (heq : cost c π = finiteMultiDualValue μ φ) :
      (∀ (σ : FiniteMultiCoupling μ), cost c π ≤ cost c σ) ∧ ∀ (ψ : (i : ι) → X i → ℝ), FiniteMultiDualFeasible c ψ → finiteMultiDualValue μ ψ ≤ finiteMultiDualValue μ φ

      A complementary-slackness pair is simultaneously primal- and dual-optimal among all finite multi-marginal plans and all feasible families of potentials.

      Attainment on finite spaces #

      theorem TauCeti.FiniteMultiCoupling.exists_forall_cost_le {ι : Type u} {X : ι → Type v} [Fintype ι] [(i : ι) → Fintype (X i)] (c : ((i : ι) → X i) → ℝ) (μ : (i : ι) → PMF (X i)) :
      ∃ (π : FiniteMultiCoupling μ), ∀ (σ : FiniteMultiCoupling μ), cost c π ≤ cost c σ

      Finite multi-marginal primal attainment. Every real cost has a minimizing coupling on a nonempty finite family of finite spaces.

      Dual attainment #

      Shifting the potentials by constants of total zero changes neither dual feasibility nor the dual value, so an arbitrary feasible family may first be normalised: subtract from each potential its maximum and return the accumulated total to one coordinate, after which no potential exceeds a bound M for the cost. A normalised family is then confined to a box: at a marginal point of zero weight the potential may be lowered outright, and at a point of positive weight that weight turns a lower bound for the dual value into a lower bound for the potential. Families of value at least -M therefore have representatives in a compact box, where the dual value is attained.

      theorem TauCeti.exists_forall_finiteMultiDualValue_le {ι : Type u} {X : ι → Type v} [Fintype ι] [(i : ι) → Fintype (X i)] [Nonempty ι] (c : ((i : ι) → X i) → ℝ) (μ : (i : ι) → PMF (X i)) :
      ∃ (φ : (i : ι) → X i → ℝ), FiniteMultiDualFeasible c φ ∧ ∀ (ψ : (i : ι) → X i → ℝ), FiniteMultiDualFeasible c ψ → finiteMultiDualValue μ ψ ≤ finiteMultiDualValue μ φ

      Finite multi-marginal dual attainment. Every real cost on a nonempty finite family of finite spaces has a maximizing feasible family of marginal potentials.

      Strong duality #

      theorem TauCeti.exists_cost_eq_finiteMultiDualValue {ι : Type u} {X : ι → Type v} [Fintype ι] [(i : ι) → Fintype (X i)] (c : ((i : ι) → X i) → ℝ) (μ : (i : ι) → PMF (X i)) [Nonempty ι] :

      Finite multi-marginal Kantorovich duality. For a nonempty finite family of finite spaces, the primal and dual problems have the same value and both attain it.

      theorem TauCeti.exists_isLeast_cost_isGreatest_finiteMultiDualValue {ι : Type u} {X : ι → Type v} [Fintype ι] [(i : ι) → Fintype (X i)] (c : ((i : ι) → X i) → ℝ) (μ : (i : ι) → PMF (X i)) [Nonempty ι] :
      ∃ (V : ℝ), IsLeast {r : ℝ | ∃ (π : FiniteMultiCoupling μ), FiniteMultiCoupling.cost c π = r} V ∧ IsGreatest {r : ℝ | ∃ (φ : (i : ι) → X i → ℝ), FiniteMultiDualFeasible c φ ∧ finiteMultiDualValue μ φ = r} V

      The common optimal value of the finite multi-marginal primal and dual problems.

      theorem TauCeti.FiniteMultiCoupling.forall_cost_le_and_forall_finiteMultiDualValue_le_iff {ι : Type u} {X : ι → Type v} {μ : (i : ι) → PMF (X i)} [Fintype ι] [(i : ι) → Fintype (X i)] (π : FiniteMultiCoupling μ) {c : ((i : ι) → X i) → ℝ} {φ : (i : ι) → X i → ℝ} [Nonempty ι] (hφ : FiniteMultiDualFeasible c φ) :
      ((∀ (σ : FiniteMultiCoupling μ), cost c π ≤ cost c σ) ∧ ∀ (ψ : (i : ι) → X i → ℝ), FiniteMultiDualFeasible c ψ → finiteMultiDualValue μ ψ ≤ finiteMultiDualValue μ φ) ↔ ∀ (z : (i : ι) → X i), ↑π z ≠ 0 → ∑ i : ι, φ i (z i) = c z

      A feasible finite multi-marginal plan and potential family are both optimal exactly when their pointwise dual gap vanishes on every configuration carrying mass.