AFTD Auto-Formalizing Theoretical Domains

subadditive_mutual_f_information_family_full_support

theoremverifiedformalization

Corollary 1 of arXiv:2610.21231, on full-support distributions: let s be twice continuously differentiable, positive and concave on (0, ∞) with s'' < 0, let K_s(x) = s(x) + 2(s'(x))²/(-s''(x)) be concave on (0, ∞), and let α ≥ K_s(1). Then f(x) = α x log x + ∫_1^x (x - t) s(t)/t² dt is convex on (0, ∞), f(1) = 0, and Ĩ_f(X; YZ) ≤ Ĩ_f(X; Y) + Ĩ_f(X; Z) for every full-support Markov chain Y → X → Z on finite sets.

Statement

theorem subadditive_mutual_f_information_family_full_support (s : ℝ → ℝ) (α : ℝ)
    (hs_smooth : ContDiffOn ℝ 2 s (Set.Ioi 0))
    (hs_pos : ∀ x > 0, 0 < s x)
    (hs_concave : ConcaveOn ℝ (Set.Ioi (0 : ℝ)) s)
    (hs_deriv2_neg : ∀ x > 0, deriv (deriv s) x < 0)
    (hKs_concave : ConcaveOn ℝ (Set.Ioi (0 : ℝ)) (subadditive_f_ks s))
    (hα : subadditive_f_ks s 1 ≤ α) :
    ConvexOn ℝ (Set.Ioi (0 : ℝ)) (subadditive_f_generator α s) ∧
      subadditive_f_generator α s 1 = 0 ∧
      mutual_f_information_subadditive_full_support (subadditive_f_generator α s)
Formalizes
arXiv:2610.21231 (On the subadditivity of mutual f-information), Corollary 1. Differs: subadditivity is stated for full-support distributions; on all distributions it cannot be stated with a real-valued f, since f(0) = ∫_0^1 s(t)/t dt = +∞ whenever s(0+) > 0 (Lean's junk value 0 there makes the all-distributions statement false, e.g. for s = 1 + √x, α = 4: experiment #3380). s is assumed C² on (0, ∞), implicit in the paper.
Verified
24 Oct 2026
Axioms
Classical.choiceQuot.soundpropext
Built on 6

The proof, explained

The proof establishes that the generator f(x) satisfies the general differential condition for subadditivity of mutual f-information. First, the lemma subadditive_f_generator_satisfies_weight_bound shows that the assumptions on s and α ensure f is convex, satisfies f(1) = 0, is C⁴, and satisfies an affine bound on its weight function. Then, mutual_f_information_subadditive_full_support_of_weight_bound uses these verified properties to conclude that mutual f-information is subadditive for all full-support Markov chains.

  1. The generator f is convex on (0, ∞), vanishes at 1, is C⁴, has strictly concave weight function g, and satisfies an affine differential bound on g.

    By applying subadditive_f_generator_satisfies_weight_bound to the hypotheses on s (C² smoothness, positivity, negative second derivative, concavity of K_s) and α ≥ K_s(1), we obtain five key properties for f: convexity on (0, ∞), f(1) = 0, C⁴ regularity on (0, ∞), strict concavity of the weight function g = subadditive_f_weight f (i.e. g'' < 0), and the existence of constants η > 0 and ℓ ∈ ℝ such that g(x) + 2(g'(x) - η)² / (-g''(x)) ≤ 2η + (x - 1)ℓ for all x > 0.

    lines 11–12

    uses subadditive_f_generator_satisfies_weight_bound

      obtain ⟨hconv, hf1, hsmooth4, hg2, hbound⟩ :=
        subadditive_f_generator_satisfies_weight_bound s α hs_smooth hs_pos hs_deriv2_neg hKs_concave hα
  2. f is convex on (0, ∞), satisfies f(1) = 0, and induces a subadditive mutual f-information on full-support Markov chains.

    The three required conclusions are combined into a conjunction: the convexity of f and f(1) = 0 were obtained in the previous step, and the subadditivity of mutual f-information follows by invoking mutual_f_information_subadditive_full_support_of_weight_bound with the properties (f(1) = 0, convexity, C⁴ smoothness, g'' < 0, and the affine weight bound) verified above.

    lines 13–15

    uses mutual_f_information_subadditive_full_support_of_weight_bound

      exact ⟨hconv, hf1,
        mutual_f_information_subadditive_full_support_of_weight_bound _
          hf1 hconv hsmooth4 hg2 hbound⟩
subadditive_f_generator_satisfies_weight_bound
For a C² positive function s on (0, ∞) with s'' < 0, concave K_s, and α ≥ K_s(1), the generator f(x) = α x log x + ∫_1^x (x-t) s(t)/t² dt is convex on (0, ∞), satisfies f(1) = 0, is C⁴ on (0, ∞), has weight function g with g'' < 0, and satisfies an affine weight bound g(x) + 2(g'(x) - η)² / (-g''(x)) ≤ 2η + (x - 1)ℓ for some η > 0 and ℓ ∈ ℝ.
mutual_f_information_subadditive_full_support_of_weight_bound
If a function f satisfies f(1) = 0, is convex and C⁴ on (0, ∞), has weight function g with g'' < 0, and satisfies the differential bound g(x) + 2(g'(x) - η)² / (-g''(x)) ≤ 2η + (x - 1)ℓ for some η > 0 and ℓ ∈ ℝ, then the mutual f-information is subadditive for every full-support Markov chain Y → X → Z.

Written by a model (gemini-3.8-flash) from the Lean proof and the goal states Lean reported at each step. Lean verified the proof; this explanation is a reading aid and is not itself machine-checked.

Lean source view module on GitHub

theorem subadditive_mutual_f_information_family_full_support (s : ℝ → ℝ) (α : ℝ)    (hs_smooth : ContDiffOn ℝ 2 s (Set.Ioi 0))    (hs_pos : ∀ x > 0, 0 < s x)    (hs_concave : ConcaveOn ℝ (Set.Ioi (0 : ℝ)) s)    (hs_deriv2_neg : ∀ x > 0, deriv (deriv s) x < 0)    (hKs_concave : ConcaveOn ℝ (Set.Ioi (0 : ℝ)) (subadditive_f_ks s))    (hα : subadditive_f_ks s 1 ≤ α) :    ConvexOn ℝ (Set.Ioi (0 : ℝ)) (subadditive_f_generator α s) ∧      subadditive_f_generator α s 1 = 0 ∧      mutual_f_information_subadditive_full_support (subadditive_f_generator α s) := by  obtain ⟨hconv, hf1, hsmooth4, hg2, hbound⟩ :=    subadditive_f_generator_satisfies_weight_bound s α hs_smooth hs_pos hs_deriv2_neg hKs_concave hα  exact ⟨hconv, hf1,    mutual_f_information_subadditive_full_support_of_weight_bound _      hf1 hconv hsmooth4 hg2 hbound⟩
See it in the graphDiscuss this resultChallenge it