mutual_power_divergence_information_eq_sum_rpow
For β ∈ (0, 1) and a joint probability mass function p_{XY}, the mutual f_β-information equals (Σ_{x,y} p(x, y)^β (p_X(x) p_Y(y))^{1−β} − 1)/(β(β − 1)).
Statement
theorem mutual_power_divergence_information_eq_sum_rpow {X Y : Type*} [Fintype X] [Fintype Y] (β : ℝ) (hβ0 : 0 < β) (hβ1 : β < 1) (p : X → Y → ℝ) (hp : ∀ x y, 0 ≤ p x y) (hp1 : ∑ x, ∑ y, p x y = 1) : mutual_f_information (power_divergence_generator β) p = (∑ x, ∑ y, p x y ^ β * ((∑ y', p x y') * (∑ x', p x' y)) ^ (1 - β) - 1) / (β * (β - 1))
- Source
- Lemma for mutual_power_divergence_information_subadditive_of_two_thirds_le; standard closed form of power divergences, arXiv:2610.21231 (On the subadditivity of mutual f-information), Sec. I.
- Verified
- 16 Oct 2026
- Axioms
Classical.choiceQuot.soundpropext- Built on 2
- 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}…
- power_divergence_generatorThe power-divergence generator f_β(x) = (x^β − βx + β − 1)/(β(β − 1)), for β ∉ {0, 1} and x ≥ 0 (β = 1/2 gives 2(√x − 1)², the squared…
- Used by 1
- power_divergence_exists_binary_counterexample_of_lt_golden_ratioFor every exponent beta in (0, (sqrt 5 - 1)/2), there exist strictly positive binary probability mass functions pX, PY, PZ forming a…
The proof, explained
The proof unfolds the definition of mutual f-information with the power-divergence generator f_β(u) = (u^β − β u + (β − 1))/(β(β − 1)) evaluated at u = p(x, y)/q(x, y), where q(x, y) = p_X(x) p_Y(y) is the product of the marginal probabilities. A pointwise expansion of each summand reduces the numerator to p(x, y)^β q(x, y)^(1−β) − β p(x, y) + (β − 1) q(x, y), handling the boundary case q(x, y) = 0 by showing p(x, y) must also vanish. Summing over all x and y, the linear terms evaluate to −β and β − 1 using the normalization of the joint and marginal probabilities, which sum to −1 and yield the desired formula.
To establish the pointwise expansion `h_term`, it suffices to show that for every pair (x, y), q * ((p x y / q) ^ β - β * (p x y / q) + (β - 1)) = p x y ^ β * q ^ (1 - β) - β * p x y + (β - 1) * q, where q = (∑ y', p x y') * (∑ x', p x' y).
We state an auxiliary lemma `h_term` computing each summand of the mutual f-information. For fixed x and y, let q denote the product of the marginal probabilities p_X(x) and p_Y(y). By definition, `power_divergence_generator β (p x y / q)` equals ((p x y / q) ^ β - β * (p x y / q) + (β - 1)) / (β * (β - 1)). Rewriting multiplication across division with `mul_div_assoc` and canceling the common denominator β * (β - 1) reduces the goal to simplifying the numerator.
uses
mul_div_assochave _ := hβ1 have h_term : ∀ (x : X) (y : Y), ((∑ y', p x y') * ∑ x', p x' y) * power_divergence_generator β (p x y / ((∑ y', p x y') * ∑ x', p x' y)) = (p x y ^ β * ((∑ y', p x y') * ∑ x', p x' y) ^ (1 - β) - β * p x y + (β - 1) * ((∑ y', p x y') * ∑ x', p x' y)) / (β * (β - 1)) := by intro x y set q := (∑ y', p x y') * ∑ x', p x' y change q * (((p x y / q) ^ β - β * (p x y / q) + (β - 1)) / (β * (β - 1))) = _ rw [← mul_div_assoc] congr 1
0 ≤ q
Because every p x y is non-negative, the finite sums defining the marginals ∑ y', p x y' and ∑ x', p x' y are non-negative by `Finset.sum_nonneg`. Their product q is therefore non-negative by `mul_nonneg`.
uses
mul_nonneg,Finset.sum_nonneghave hq_nonneg : 0 ≤ q := by apply mul_nonneg · exact Finset.sum_nonneg (fun y' _ => hp x y') · exact Finset.sum_nonneg (fun x' _ => hp x' y)
When q = 0, both sides of the numerator identity vanish.
We split into cases using `eq_or_lt_of_le`. If q = 0, the product of the two marginal sums is zero, so by `mul_eq_zero.mp` at least one marginal sum is zero. Since each is a sum of non-negative terms, all terms in that sum must be zero (`Finset.sum_eq_zero_iff_of_nonneg`), which forces p x y = 0. Substituting q = 0 and p x y = 0 into both sides, 0 ^ β = 0 because β > 0 (`Real.zero_rpow`), and the identity reduces to 0 = 0.
uses
eq_or_lt_of_le,mul_eq_zero.mp,Finset.sum_eq_zero_iff_of_nonneg,Finset.mem_univ,zero_mul,Real.zero_rpow,ne_of_gtrcases eq_or_lt_of_le hq_nonneg with hq_zero | hq_pos · have hq0 : q = 0 := hq_zero.symm have hp_zero : p x y = 0 := by cases mul_eq_zero.mp hq0 with | inl hx => have hx_all : ∀ y', p x y' = 0 := by rw [Finset.sum_eq_zero_iff_of_nonneg (fun y' _ => hp x y')] at hx exact fun y' => hx y' (Finset.mem_univ y') exact hx_all y | inr hy => have hy_all : ∀ x', p x' y = 0 := by rw [Finset.sum_eq_zero_iff_of_nonneg (fun x' _ => hp x' y)] at hy exact fun x' => hy x' (Finset.mem_univ x') exact hy_all x rw [hq0, hp_zero] rw [zero_mul, Real.zero_rpow (ne_of_gt hβ0)] ring
When q > 0, the numerator expands to p x y ^ β * q ^ (1 - β) - β * p x y + (β - 1) * q.
When q > 0, we distribute the exponent over the division via `Real.div_rpow` to get (p x y / q) ^ β = p x y ^ β / q ^ β. Using `Real.rpow_sub` and `Real.rpow_one`, we write q ^ (1 - β) as q / q ^ β. Multiplying q through the three terms, the first term becomes p x y ^ β * (q / q ^ β) = p x y ^ β * q ^ (1 - β), the middle term simplifies to β * p x y via `mul_div_cancel₀`, and the constant term becomes (β - 1) * q.
uses
ne_of_gt,Real.div_rpow,le_of_lt,Real.rpow_sub,Real.rpow_one,mul_div_cancel₀· have hq_ne : q ≠ 0 := ne_of_gt hq_pos rw [Real.div_rpow (hp x y) (le_of_lt hq_pos)] have hq_pow : q ^ (1 - β) = q / q ^ β := by rw [Real.rpow_sub hq_pos, Real.rpow_one] rw [hq_pow] have h1 : q * (p x y ^ β / q ^ β) = p x y ^ β * (q / q ^ β) := by ring have h2 : q * (β * (p x y / q)) = β * p x y := by calc q * (β * (p x y / q)) = β * (q * (p x y / q)) := by ring _ = β * p x y := by rw [mul_div_cancel₀ (p x y) hq_ne] calc q * (p x y ^ β / q ^ β - β * (p x y / q) + (β - 1)) _ = q * (p x y ^ β / q ^ β) - q * (β * (p x y / q)) + (β - 1) * q := by ring _ = p x y ^ β * (q / q ^ β) - β * p x y + (β - 1) * q := by rw [h1, h2]
The mutual f-information can be written as (∑ x, ∑ y, p x y ^ β * q(x, y) ^ (1 - β) - ∑ x, ∑ y, β * p x y + ∑ x, ∑ y, (β - 1) * q(x, y)) / (β * (β - 1)).
We unfold `mutual_f_information` into a double sum over x and y and substitute the pointwise formula `h_term` into each summand. Pulling the division by β * (β - 1) outside the summations via `Finset.sum_div`, we eliminate the denominator on both sides. We then distribute the summation over addition and subtraction using `Finset.sum_add_distrib` and `Finset.sum_sub_distrib`.
uses
Finset.sum_div,Finset.sum_add_distrib,Finset.sum_sub_distribunfold mutual_f_information simp_rw [h_term] simp_rw [← Finset.sum_div] congr 1 simp_rw [Finset.sum_add_distrib, Finset.sum_sub_distrib]
∑ x, ∑ y, β * p x y = β
We factor β out of the double sum using `Finset.mul_sum`. Since the total probability mass ∑ x, ∑ y, p x y is 1 by assumption `hp1`, this leaves β * 1 = β by `mul_one`.
uses
Finset.mul_sum,mul_onehave h_beta : ∑ x, ∑ y, β * p x y = β := by have h : (∑ x, ∑ y, β * p x y) = β * ∑ x, ∑ y, p x y := by simp_rw [← Finset.mul_sum] rw [h, hp1, mul_one]
∑ x, ∑ y, ((∑ y', p x y') * ∑ x', p x' y) = 1
For a fixed x, the marginal sum ∑ y', p x y' is independent of y and factors out of the sum over y using `Finset.mul_sum`. The inner sum over y and x' equals the total sum 1 by swapping the summation order (`Finset.sum_comm`) and using `hp1`. Summing the remaining marginal ∑ y', p x y' over x then equals 1 by `hp1`.
uses
Finset.mul_sum,Finset.sum_comm,mul_onehave hq : ∑ x, ∑ y, ((∑ y', p x y') * ∑ x', p x' y) = 1 := by have h_inner : ∀ x, ∑ y, ((∑ y', p x y') * ∑ x', p x' y) = ∑ y', p x y' := by intro x rw [← Finset.mul_sum] have h_comm : ∑ y, ∑ x', p x' y = ∑ x', ∑ y, p x' y := Finset.sum_comm rw [h_comm, hp1, mul_one] calc ∑ x, ∑ y, ((∑ y', p x y') * ∑ x', p x' y) _ = ∑ x, ∑ y', p x y' := by simp_rw [h_inner] _ = 1 := hp1
∑ x, ∑ y, (β - 1) * ((∑ y', p x y') * ∑ x', p x' y) = β - 1
Factoring (β - 1) out of the double sum with `Finset.mul_sum` and substituting the previous result that the double sum of the product of marginals equals 1 gives (β - 1) * 1 = β - 1 by `mul_one`.
uses
Finset.mul_sum,mul_onehave h_q_term : ∑ x, ∑ y, (β - 1) * ((∑ y', p x y') * ∑ x', p x' y) = β - 1 := by have h : (∑ x, ∑ y, (β - 1) * ((∑ y', p x y') * ∑ x', p x' y)) = (β - 1) * ∑ x, ∑ y, ((∑ y', p x y') * ∑ x', p x' y) := by simp_rw [← Finset.mul_sum] rw [h, hq, mul_one]
The numerator simplifies to ∑ x, ∑ y, p x y ^ β * q(x, y) ^ (1 - β) - 1. automation
Substituting the evaluated sums `h_beta` (giving β) and `h_q_term` (giving β - 1) into the expression leaves -β + (β - 1) = -1, which `ring` simplifies to complete the proof.
rw [h_beta, h_q_term] ring
mul_div_assoc- Multiplying a fraction by a factor is equal to multiplying the numerator by that factor: a * (b / c) = (a * b) / c.
mul_nonneg- The product of two non-negative numbers is non-negative.
Finset.sum_nonneg- The sum of non-negative terms over a finite set is non-negative.
eq_or_lt_of_le- If a ≤ b, then either a = b or a < b.
mul_eq_zero.mp- If the product of two real numbers is zero, at least one of them must be zero.
Finset.sum_eq_zero_iff_of_nonneg- A sum of non-negative terms over a finite set is zero if and only if each term in the sum is zero.
Finset.mem_univ- Every element of a finite type belongs to the universal finset `univ`.
zero_mul- Multiplying zero by any number yields zero: 0 * a = 0.
Real.zero_rpow- Zero raised to any non-zero real power is zero: 0 ^ x = 0 when x ≠ 0.
ne_of_gt- If b < a, then a ≠ b.
Real.div_rpow- For non-negative real numbers x and y, the power of their quotient is the quotient of their powers: (x / y) ^ z = x ^ z / y ^ z.
le_of_lt- A strict inequality a < b implies the non-strict inequality a ≤ b.
Real.rpow_sub- For a positive real number x, x ^ (y - z) = x ^ y / x ^ z.
Real.rpow_one- Any real number raised to the power 1 equals itself: x ^ 1 = x.
mul_div_cancel₀- For b ≠ 0, multiplying b by a / b cancels to a.
Finset.sum_div- Dividing a finite sum by a scalar equals the sum of the divided terms: (∑ i, f i) / a = ∑ i, (f i / a).
Finset.sum_add_distrib- The sum of sums of two functions equals the sum of each function's values: ∑ x, (f x + g x) = ∑ x, f x + ∑ x, g x.
Finset.sum_sub_distrib- The sum of differences of two functions equals the difference of the sums: ∑ x, (f x - g x) = ∑ x, f x - ∑ x, g x.
Finset.mul_sum- A scalar factor can be factored out of a finite sum: a * ∑ i, f i = ∑ i, a * f i.
Finset.sum_comm- The order of two independent finite summations can be interchanged: ∑ x, ∑ y, f x y = ∑ y, ∑ x, f x y.
mul_one- Multiplying any number by 1 leaves it unchanged: a * 1 = a.
- The proof opens with an unused hypothesis `have _ := hβ1` (line 6) recording `β < 1`, which is not referenced thereafter.
- Several algebraic identities and cancellations (e.g., lines 35, 41, 46, 70) are solved automatically by the `ring` tactic.
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
Choose a tactic — a marked line — to see the goals Lean had before and after it, as in Lean’s infoview. Step with ← → or j k.
theorem mutual_power_divergence_information_eq_sum_rpow {X Y : Type*} [Fintype X] [Fintype Y] (β : ℝ) (hβ0 : 0 < β) (hβ1 : β < 1) (p : X → Y → ℝ) (hp : ∀ x y, 0 ≤ p x y) (hp1 : ∑ x, ∑ y, p x y = 1) : mutual_f_information (power_divergence_generator β) p = (∑ x, ∑ y, p x y ^ β * ((∑ y', p x y') * (∑ x', p x' y)) ^ (1 - β) - 1) / (β * (β - 1)) := by have _ := hβ1 have h_term : ∀ (x : X) (y : Y), ((∑ y', p x y') * ∑ x', p x' y) * power_divergence_generator β (p x y / ((∑ y', p x y') * ∑ x', p x' y)) = (p x y ^ β * ((∑ y', p x y') * ∑ x', p x' y) ^ (1 - β) - β * p x y + (β - 1) * ((∑ y', p x y') * ∑ x', p x' y)) / (β * (β - 1)) := by intro x y set q := (∑ y', p x y') * ∑ x', p x' y change q * (((p x y / q) ^ β - β * (p x y / q) + (β - 1)) / (β * (β - 1))) = _ rw [← mul_div_assoc] congr 1 have hq_nonneg : 0 ≤ q := by apply mul_nonneg · exact Finset.sum_nonneg (fun y' _ => hp x y') · exact Finset.sum_nonneg (fun x' _ => hp x' y) rcases eq_or_lt_of_le hq_nonneg with hq_zero | hq_pos · have hq0 : q = 0 := hq_zero.symm have hp_zero : p x y = 0 := by cases mul_eq_zero.mp hq0 with | inl hx => have hx_all : ∀ y', p x y' = 0 := by rw [Finset.sum_eq_zero_iff_of_nonneg (fun y' _ => hp x y')] at hx exact fun y' => hx y' (Finset.mem_univ y') exact hx_all y | inr hy => have hy_all : ∀ x', p x' y = 0 := by rw [Finset.sum_eq_zero_iff_of_nonneg (fun x' _ => hp x' y)] at hy exact fun x' => hy x' (Finset.mem_univ x') exact hy_all x rw [hq0, hp_zero] rw [zero_mul, Real.zero_rpow (ne_of_gt hβ0)] ring · have hq_ne : q ≠ 0 := ne_of_gt hq_pos rw [Real.div_rpow (hp x y) (le_of_lt hq_pos)] have hq_pow : q ^ (1 - β) = q / q ^ β := by rw [Real.rpow_sub hq_pos, Real.rpow_one] rw [hq_pow] have h1 : q * (p x y ^ β / q ^ β) = p x y ^ β * (q / q ^ β) := by ring have h2 : q * (β * (p x y / q)) = β * p x y := by calc q * (β * (p x y / q)) = β * (q * (p x y / q)) := by ring _ = β * p x y := by rw [mul_div_cancel₀ (p x y) hq_ne] calc q * (p x y ^ β / q ^ β - β * (p x y / q) + (β - 1)) _ = q * (p x y ^ β / q ^ β) - q * (β * (p x y / q)) + (β - 1) * q := by ring _ = p x y ^ β * (q / q ^ β) - β * p x y + (β - 1) * q := by rw [h1, h2] unfold mutual_f_information simp_rw [h_term] simp_rw [← Finset.sum_div] congr 1 simp_rw [Finset.sum_add_distrib, Finset.sum_sub_distrib] have h_beta : ∑ x, ∑ y, β * p x y = β := by have h : (∑ x, ∑ y, β * p x y) = β * ∑ x, ∑ y, p x y := by simp_rw [← Finset.mul_sum] rw [h, hp1, mul_one] have hq : ∑ x, ∑ y, ((∑ y', p x y') * ∑ x', p x' y) = 1 := by have h_inner : ∀ x, ∑ y, ((∑ y', p x y') * ∑ x', p x' y) = ∑ y', p x y' := by intro x rw [← Finset.mul_sum] have h_comm : ∑ y, ∑ x', p x' y = ∑ x', ∑ y, p x' y := Finset.sum_comm rw [h_comm, hp1, mul_one] calc ∑ x, ∑ y, ((∑ y', p x y') * ∑ x', p x' y) _ = ∑ x, ∑ y', p x y' := by simp_rw [h_inner] _ = 1 := hp1 have h_q_term : ∑ x, ∑ y, (β - 1) * ((∑ y', p x y') * ∑ x', p x' y) = β - 1 := by have h : (∑ x, ∑ y, (β - 1) * ((∑ y', p x y') * ∑ x', p x' y)) = (β - 1) * ∑ x, ∑ y, ((∑ y', p x y') * ∑ x', p x' y) := by simp_rw [← Finset.mul_sum] rw [h, hq, mul_one] rw [h_beta, h_q_term] ring