mutual_f_theta_self
Θ_{p_X,p_{Y|X}}(p_X) = Ĩ_f(X;Y) when f(1) = 0 (step (17)-(18) of the proof of Theorem 1 of arXiv:2610.21231).
Statement
theorem mutual_f_theta_self {X Y : Type*} [Fintype X] [Fintype Y] (f : ℝ → ℝ) (hf1 : f 1 = 0) (pX : X → ℝ) (PY : X → Y → ℝ) (hpX : ∀ x, 0 < pX x) (hsY : ∀ x, ∑ y, PY x y = 1) : mutual_f_theta f pX PY pX = mutual_f_information f (fun x y => pX x * PY x y)
- Source
- arXiv:2610.21231 (On the subadditivity of mutual f-information), Appendix B, proof of Theorem 1, eqs. (17)-(18)
- Verified
- 23 Oct 2026
- Axioms
Classical.choiceQuot.soundpropext- Built on 2
- mutual_f_thetaThe function Θ_{p_X,p_{Y|X}}(r_X) of eq. (14)/(19) of arXiv:2610.21231: Σ_y r_Y(y) D_f(r_{X|y} ‖ p_X) - D_f(r_X ‖ p_X), with r_{YX} = r_X…
- 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}…
- Used by 1
- mutual_f_information_subadditive_full_support_of_weight_boundTheorem 1 of arXiv:2610.21231, on full-support distributions. Let f with f(1) = 0 be convex and four times continuously differentiable on…
The proof, explained
The proof unfolds the definitions of Θ_{p_X, P_{Y|X}}(p_X) and Ĩ_f(p_{XY}) and establishes their equality term by term. First, the subtrahend in Θ vanishes because p_X(x)/p_X(x) = 1 and f(1) = 0, while the row sums of the transition matrix allow the marginal ∑_y p_X(x) P_{Y|X}(x, y) to simplify to p_X(x). Exchanging the order of summation on the right-hand side and distributing the marginal factor into the inner sum over X reduces the equality to pointwise agreement of the summands, which is completed by elementary ring arithmetic.
Unfolding the definitions reduces the left-hand side to its first summation, eliminating the subtrahend.
After expanding the definitions of `mutual_f_theta` and `mutual_f_information`, the subtrahend in `mutual_f_theta` is ∑ x, pX x * f (pX x / pX x). Since pX x > 0 for all x, the ratio pX x / pX x equals 1 by `div_self`. Using the hypothesis f 1 = 0, each term becomes pX x * 0 = 0 by `mul_zero`. The entire sum of zeros evaluates to 0 by `Finset.sum_const_zero`, and subtracting 0 leaves only the leading term by `sub_zero`.
uses
div_self,mul_zero,Finset.sum_const_zero,sub_zerounfold mutual_f_theta mutual_f_information have h1 : ∀ x, pX x / pX x = 1 := fun x => div_self (hpX x).ne' simp only [h1, hf1, mul_zero, Finset.sum_const_zero, sub_zero]
The marginal sum over Y of the joint distribution evaluates to pX x.
Factoring out pX x using `Finset.mul_sum` gives ∑ y, pX x * PY x y = pX x * ∑ y, PY x y. Since each row of the transition matrix sums to 1 (hsY), this evaluates to pX x * 1 = pX x by `mul_one`. Simplifying with this fact replaces the marginal factor ∑ y', pX x * PY x y' on the right-hand side with pX x.
uses
Finset.mul_sum,mul_onehave hrow : ∀ x, ∑ y, pX x * PY x y = pX x := fun x => by rw [← Finset.mul_sum, hsY, mul_one] simp only [hrow]
The order of summation on the right-hand side is swapped to sum over Y first, reducing the goal to showing equality for each y.
Applying `Finset.sum_comm` rewrites the double sum ∑ x, ∑ y on the right-hand side as ∑ y, ∑ x. By `Finset.sum_congr`, it suffices to prove that the summands for each y ∈ Y are equal.
uses
Finset.sum_comm,Finset.sum_congrrw [Finset.sum_comm] refine Finset.sum_congr rfl (fun y _ => ?_)
The outer factor on the left-hand side distributes into the inner sum over X, reducing the goal to matching each term for fixed x and y.
The factor ∑ x, pX x * PY x y is distributed into the inner sum over X using `Finset.mul_sum`. Applying `Finset.sum_congr` again reduces the equality of sums over X to proving that the inner terms agree pointwise for each x ∈ X.
uses
Finset.mul_sum,Finset.sum_congrrw [Finset.mul_sum] refine Finset.sum_congr rfl (fun x _ => ?_)
The arguments of f and their coefficients match, concluding the proof.
The division identity (a / b) / c = a / (c * b) is shown using `div_div` and `mul_comm`. Rewriting with this identity transforms the argument of f on the left, ((pX x * PY x y / ∑ x', pX x' * PY x' y) / pX x), into pX x * PY x y / (pX x * ∑ x', pX x' * PY x' y), matching the argument on the right. Ring arithmetic (`ring`) then verifies that the coefficient products (∑ x', pX x' * PY x' y) * (pX x * f(...)) and (pX x * ∑ x', pX x' * PY x' y) * f(...) are identical.
uses
div_div,mul_commhave hd : ∀ a b c : ℝ, a / b / c = a / (c * b) := fun a b c => by rw [div_div, mul_comm b c] rw [hd]; ring
div_self- For any nonzero element a of a group with zero (such as ℝ), a / a = 1.
div_div- Iterated division satisfies (a / b) / c = a / (b * c).
Finset.mul_sum- Multiplication distributes over a finite sum: a * ∑ i ∈ s, f i = ∑ i ∈ s, a * f i.
Finset.sum_congr- Two finite sums over the same set are equal if their terms are equal at every index.
Finset.sum_comm- The order of two finite summations can be exchanged: ∑ x ∈ s, ∑ y ∈ t, f x y = ∑ y ∈ t, ∑ x ∈ s, f x y.
Finset.sum_const_zero- The sum of zeros over any finite set is zero.
mul_comm- Multiplication is commutative: a * b = b * a.
mul_one- Multiplying any element by one leaves it unchanged: a * 1 = a.
mul_zero- Multiplying any element by zero yields zero: a * 0 = 0.
sub_zero- Subtracting zero from any element leaves it unchanged: a - 0 = a.
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_f_theta_self {X Y : Type*} [Fintype X] [Fintype Y] (f : ℝ → ℝ) (hf1 : f 1 = 0) (pX : X → ℝ) (PY : X → Y → ℝ) (hpX : ∀ x, 0 < pX x) (hsY : ∀ x, ∑ y, PY x y = 1) : mutual_f_theta f pX PY pX = mutual_f_information f (fun x y => pX x * PY x y) := by unfold mutual_f_theta mutual_f_information have h1 : ∀ x, pX x / pX x = 1 := fun x => div_self (hpX x).ne' simp only [h1, hf1, mul_zero, Finset.sum_const_zero, sub_zero] have hrow : ∀ x, ∑ y, pX x * PY x y = pX x := fun x => by rw [← Finset.mul_sum, hsY, mul_one] simp only [hrow] rw [Finset.sum_comm] refine Finset.sum_congr rfl (fun y _ => ?_) rw [Finset.mul_sum] refine Finset.sum_congr rfl (fun x _ => ?_) have hd : ∀ a b c : ℝ, a / b / c = a / (c * b) := fun a b c => by rw [div_div, mul_comm b c] rw [hd]; ring