AFTD Auto-Formalizing Theoretical Domains

mutual_f_information_of_independent

theoremverified

If f(1) = 0 and X, Y are independent (p_{XY}(x, y) = a(x) b(y) with a, b summing to 1), then Ĩ_f(X; Y) = 0.

Statement

theorem mutual_f_information_of_independent {X Y : Type*} [Fintype X] [Fintype Y] (f : ℝ → ℝ)
    (hf : f 1 = 0) (a : X → ℝ) (b : Y → ℝ) (ha : ∑ x, a x = 1) (hb : ∑ y, b y = 1) :
    mutual_f_information f (fun x y => a x * b y) = 0
Source
Sanity lemma for the mutual f-information, arXiv:2610.21231 (On the subadditivity of mutual f-information), Sec. I.
Verified
16 Oct 2026
Axioms
Classical.choiceQuot.soundpropext
Built on 1
  • mutual_f_informationFor a joint probability mass function p_{XY} on finite sets, the mutual f-information is Ĩ_f(X; Y) = D_f(p_{XY} ‖ p_X p_Y) = Σ_{x,y}…

The proof, explained

To show that the mutual f-information vanishes for independent distributions, the proof first evaluates the marginal sums ∑ y, a(x)b(y) = a(x) and ∑ x, a(x)b(y) = b(y) using the fact that a and b each sum to 1. This simplifies each term in the double sum defining mutual f-information to (a(x)b(y)) f((a(x)b(y))/(a(x)b(y))). A case split on whether a(x)b(y) = 0 shows that every summand is 0—either by multiplication by zero or because f(1) = 0—so the entire sum evaluates to 0.

  1. Expanding the definition of mutual f-information and evaluating the marginal distributions reduces the goal to ∑ x, ∑ y, a x * b y * f (a x * b y / (a x * b y)) = 0.

    Unfolding the definition of mutual_f_information writes it in terms of the joint distribution a(x) b(y) and its marginal sums ∑ y', a x * b y' and ∑ x', a x' * b y. Factoring out a(x) in the first marginal sum and using ∑ y, b y = 1 shows that this sum equals a(x) (proven as h1). Similarly, factoring out b(y) and using ∑ x, a x = 1 shows the second marginal sum equals b(y) (proven as h2). Rewriting the expression with these two identities simplifies the marginal product to a(x) b(y).

    lines 4–9

    uses Finset.mul_sum, mul_one, Finset.sum_mul, one_mul

      dsimp [mutual_f_information]
      have h1 : ∀ x, (∑ y', a x * b y') = a x := fun x => by
        rw [← Finset.mul_sum, hb, mul_one]
      have h2 : ∀ y, (∑ x', a x' * b y) = b y := fun y => by
        rw [← Finset.sum_mul, ha, one_mul]
      simp_rw [h1, h2]
  2. For every x and y, the summand (a x * b y) * f (a x * b y / (a x * b y)) is equal to 0.

    We perform a case analysis on whether a(x) b(y) is zero. If a(x) b(y) = 0, the term is 0 * f(...), which is 0 by zero_mul. If a(x) b(y) ≠ 0, the quotient (a(x) b(y)) / (a(x) b(y)) equals 1 by div_self; since f(1) = 0 by hypothesis hf, the term becomes (a(x) b(y)) * 0, which equals 0 by mul_zero.

    lines 10–13

    uses zero_mul, div_self, mul_zero

      have h3 : ∀ x y, (a x * b y) * f (a x * b y / (a x * b y)) = 0 := fun x y => by
        by_cases hc : a x * b y = 0
        · rw [hc, zero_mul]
        · rw [div_self hc, hf, mul_zero]
  3. The total mutual f-information is 0. automation

    Rewriting each summand with h3 transforms the double sum into a sum of zeros, which simp evaluates to 0.

    lines 14–15

      simp_rw [h3]
      simp
Finset.mul_sum
Multiplication distributes over a finite sum on the left: a * ∑ i ∈ s, f i = ∑ i ∈ s, a * f i.
mul_one
Multiplication by 1 on the right is the identity: a * 1 = a.
Finset.sum_mul
Multiplication distributes over a finite sum on the right: (∑ i ∈ s, f i) * a = ∑ i ∈ s, f i * a.
one_mul
Multiplication by 1 on the left is the identity: 1 * a = a.
zero_mul
Multiplication by 0 on the left yields 0: 0 * a = 0.
div_self
A non-zero element divided by itself equals 1: if a ≠ 0, then a / a = 1.
mul_zero
Multiplication by 0 on the right yields 0: a * 0 = 0.

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 mutual_f_information_of_independent {X Y : Type*} [Fintype X] [Fintype Y] (f : ℝ → ℝ)    (hf : f 1 = 0) (a : X → ℝ) (b : Y → ℝ) (ha : ∑ x, a x = 1) (hb : ∑ y, b y = 1) :    mutual_f_information f (fun x y => a x * b y) = 0 := by  dsimp [mutual_f_information]  have h1 : ∀ x, (∑ y', a x * b y') = a x := fun x => by    rw [← Finset.mul_sum, hb, mul_one]  have h2 : ∀ y, (∑ x', a x' * b y) = b y := fun y => by    rw [← Finset.sum_mul, ha, one_mul]  simp_rw [h1, h2]  have h3 : ∀ x y, (a x * b y) * f (a x * b y / (a x * b y)) = 0 := fun x y => by    by_cases hc : a x * b y = 0    · rw [hc, zero_mul]    · rw [div_self hc, hf, mul_zero]  simp_rw [h3]  simp
See it in the graphDiscuss this resultChallenge it