InformationTheory

InformationTheory.Shannon.Hoeffding.MinimizerAttainment

source

Hoeffding I-projection minimizer attainment — IsHoeffdingTiltMinimal discharge #

Hoeffding/Lagrange.lean discharged the mem half of IsHoeffdingLagrangeHyp from the IVT and reduced the realises half to the strictly-primitive predicate

  `IsHoeffdingTiltMinimal P₁ P₂ α λ`
    := `IsMinOn (klDivPmf · P₂) (hoeffdingConstraintSet P₁ α) (hoeffdingTilt P₁ P₂ λ)`,

the Csiszár I-projection minimality of the exponential tilt. This file discharges that minimality fully, for the interior regime 0 < λ ≤ 1 with the IVT constraint-match klDivPmf (tilt) P₁ = α.

Approach #

The discharge is the exponential-family Pythagorean identity, derived from the log-linearity of the tilt and a per-coordinate algebraic identity that holds even at zero atoms (so the minimality extends to the whole constraint set K, boundary included).

The key per-coordinate fact, valid for R > 0 and all Q ≥ 0 (including Q = 0, where 0 · log 0 = 0):

  `R · klFun (Q / R) = Q · log Q - Q · log R + (R - Q)`.

Summing this against the three references P₂, P₁, T = hoeffdingTilt P₁ P₂ λ with weights λ, 1 - λ, -1 makes the Q log Q terms cancel (coefficients λ + (1-λ) - 1 = 0) and the Q log R terms collapse via the tilt's constant-log-ratio identity log T - (1-λ) log P₁ - λ log P₂ = -log Z (hoeffdingTilt_log_ratio_const) into the flat term Q · (-log Z). Summing over the simplex (∑ Q = 1) gives the master identity

  `λ · D(Q‖P₂) + (1-λ) · D(Q‖P₁) - D(Q‖T) = -log Z`        (∀ Q ∈ stdSimplex).

Specialising to Q = T (D(T‖T)=0) and subtracting yields the Pythagorean difference

  `λ (D(Q‖P₂) - D(T‖P₂)) + (1-λ)(D(Q‖P₁) - D(T‖P₁)) = D(Q‖T) ≥ 0`.

For Q ∈ K the constraint gives D(Q‖P₁) ≤ α = D(T‖P₁), so the (1-λ)·(…) term is ≤ 0 (using 1-λ ≥ 0); hence λ (D(Q‖P₂) - D(T‖P₂)) ≥ 0, and with λ > 0 we get D(T‖P₂) ≤ D(Q‖P₂) — exactly IsMinOn.

What this file publishes #

Per-coordinate klFun identity (valid at zero atoms) #

theorem

InformationTheory.Shannon.HoeffdingMinimizerAttainment.klFun_ref_mul

source
{R Q : } (hR : 0 < R) (hQ : 0 Q) :
R * klFun (Q / R) = Q * Real.log Q - Q * Real.log R + (R - Q)

Per-coordinate klFun identity: for a positive reference R and any Q ≥ 0 (including Q = 0, where 0 · log 0 = 0), R · klFun (Q / R) = Q · log Q - Q · log R + (R - Q).

Used by

    klDivPmf cross-entropy sum form (no full support) #

    theorem

    InformationTheory.Shannon.HoeffdingMinimizerAttainment.klDivPmf_eq_entropyCross_sum

    source
    {α : Type u_1} [Fintype α] [Nonempty α] [MeasurableSpace α] [MeasurableSingletonClass α] {Q R : α} (hR_pos : ∀ (a : α), 0 < R a) (hQ_nn : ∀ (a : α), 0 Q a) :
    CsiszarProjection.klDivPmf Q R = a : α, (Q a * Real.log (Q a) - Q a * Real.log (R a) + (R a - Q a))

    Cross-entropy sum form of klDivPmf, valid for any Q ∈ stdSimplex (full support not required) and full-support reference R: klDivPmf Q R = ∑ a, (Q a · log (Q a) - Q a · log (R a) + (R a - Q a)).

    Used by

      Master exponential-family identity #

      theorem

      InformationTheory.Shannon.HoeffdingMinimizerAttainment.hoeffdingTilt_kl_master

      source
      {α : Type u_1} [Fintype α] [Nonempty α] [MeasurableSpace α] [MeasurableSingletonClass α] (P₁ P₂ : α) (hP₁_pos : ∀ (a : α), 0 < P₁ a) (hP₂_pos : ∀ (a : α), 0 < P₂ a) (hP₁_sum : a : α, P₁ a = 1) (hP₂_sum : a : α, P₂ a = 1) (lam : ) {Q : α} (hQ : Q stdSimplex α) :

      Master identity: for every Q ∈ stdSimplex the log-linear weighting of the three KL divergences against P₂, P₁, and the tilt collapses to the flat -log Z:

      `λ · klDivPmf Q P₂ + (1-λ) · klDivPmf Q P₁ - klDivPmf Q (tilt) = -log Z(λ)`.
      

      This holds with no full-support hypothesis on Q.

      Used by

        Pythagorean difference identity #

        theorem

        InformationTheory.Shannon.HoeffdingMinimizerAttainment.hoeffdingTilt_kl_pythagoras_diff

        source
        {α : Type u_1} [Fintype α] [Nonempty α] [MeasurableSpace α] [MeasurableSingletonClass α] (P₁ P₂ : α) (hP₁_pos : ∀ (a : α), 0 < P₁ a) (hP₂_pos : ∀ (a : α), 0 < P₂ a) (hP₁_sum : a : α, P₁ a = 1) (hP₂_sum : a : α, P₂ a = 1) (lam : ) {Q : α} (hQ : Q stdSimplex α) :

        Pythagorean difference identity (Csiszár): subtracting the master identity at Q and at the tilt T (where klDivPmf T T = 0):

        `λ (klDivPmf Q P₂ - klDivPmf T P₂) + (1-λ)(klDivPmf Q P₁ - klDivPmf T P₁)
           = klDivPmf Q T`. 
        
        Used by

          IsHoeffdingTiltMinimal discharge #

          theorem

          InformationTheory.Shannon.HoeffdingMinimizerAttainment.isHoeffdingTiltMinimal_of_constraint_eq

          source
          {α : Type u_1} [Fintype α] [Nonempty α] [MeasurableSpace α] [MeasurableSingletonClass α] (P₁ P₂ : α) (hP₁_pos : ∀ (a : α), 0 < P₁ a) (hP₂_pos : ∀ (a : α), 0 < P₂ a) (hP₁_sum : a : α, P₁ a = 1) (hP₂_sum : a : α, P₂ a = 1) {alpha lam : } (h_lam_pos : 0 < lam) (h_lam_le : lam 1) (h_kl : CsiszarProjection.klDivPmf (HoeffdingTilt.hoeffdingTilt P₁ P₂ lam) P₁ = alpha) :

          I-projection minimality discharge: for 0 < λ ≤ 1 and the IVT constraint-match klDivPmf (tilt) P₁ = α, the tilt minimises klDivPmf · P₂ over the constraint set K(α). This fully discharges the primitive IsHoeffdingTiltMinimal (whole constraint set, boundary included).

          Used by

            Constructive IsHoeffdingLagrangeHyp (both halves) #

            theorem

            InformationTheory.Shannon.HoeffdingMinimizerAttainment.isHoeffdingLagrangeHyp_of_constraint_eq

            source
            {α : Type u_1} [Fintype α] [Nonempty α] [MeasurableSpace α] [MeasurableSingletonClass α] (P₁ P₂ : α) (hP₁_pos : ∀ (a : α), 0 < P₁ a) (hP₂_pos : ∀ (a : α), 0 < P₂ a) (hP₁_sum : a : α, P₁ a = 1) (hP₂_sum : a : α, P₂ a = 1) {alpha lam : } (h_lam_pos : 0 < lam) (h_lam_le : lam 1) (h_kl : CsiszarProjection.klDivPmf (HoeffdingTilt.hoeffdingTilt P₁ P₂ lam) P₁ = alpha) :

            Fully constructive Lagrange hypothesis: from 0 < λ ≤ 1 and the IVT constraint-match, build IsHoeffdingLagrangeHyp with both mem and realises constructive — no minimality hypothesis carried.

            Used by

              Interior existence (IVT + discharged minimality) #

              theorem

              InformationTheory.Shannon.HoeffdingMinimizerAttainment.exists_isHoeffdingLagrangeHyp_interior

              source
              {α : Type u_1} [Fintype α] [Nonempty α] [MeasurableSpace α] [MeasurableSingletonClass α] (P₁ P₂ : α) (hP₁_pos : ∀ (a : α), 0 < P₁ a) (hP₂_pos : ∀ (a : α), 0 < P₂ a) (hP₁_sum : a : α, P₁ a = 1) (hP₂_sum : a : α, P₂ a = 1) {alpha : } (h_alpha_pos : 0 < alpha) (h_alpha_le : alpha CsiszarProjection.klDivPmf P₂ P₁) :
              lamSet.Ioc 0 1, HoeffdingTilt.IsHoeffdingLagrangeHyp P₁ P₂ alpha lam

              Interior existence: for interior 0 < α ≤ klDivPmf P₂ P₁, the IVT supplies a λ ∈ (0,1] whose tilt matches the constraint, and the in-file minimality discharge upgrades it to a fully constructive IsHoeffdingLagrangeHyp — no external minimality hypothesis.

              Used by