subadditive_f_generator_satisfies_weight_bound
First claim of Corollary 1 of arXiv:2610.21231: let s be twice continuously differentiable and positive on (0, ∞) with s'' < 0, let K_s(x) = s(x) + 2(s'(x))²/(-s''(x)) be concave on (0, ∞), and α ≥ K_s(1). Then the generator f(x) = α x log x + ∫_1^x (x - t) s(t)/t² dt satisfies the hypotheses of Theorem 1: f is convex and C⁴ on (0, ∞), f(1) = 0, g = x² f'' = α x + s has g'' < 0, and g(x) + 2(g'(x) - η)²/(-g''(x)) ≤ 2η + (x - 1)ℓ for all x > 0, with η = α and ℓ = α + (a supergradient of K_s at 1).
Statement
theorem subadditive_f_generator_satisfies_weight_bound (s : ℝ → ℝ) (α : ℝ)
(hs_smooth : ContDiffOn ℝ 2 s (Set.Ioi 0))
(hs_pos : ∀ x > 0, 0 < s x)
(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) (subadditive_f_generator α s) ∧
subadditive_f_generator α s 1 = 0 ∧
ContDiffOn ℝ 4 (subadditive_f_generator α s) (Set.Ioi 0) ∧
(∀ x > 0, deriv (deriv (subadditive_f_weight (subadditive_f_generator α s))) x < 0) ∧
∃ η ℓ : ℝ, 0 < η ∧ ∀ x > 0,
subadditive_f_weight (subadditive_f_generator α s) x +
2 * (deriv (subadditive_f_weight (subadditive_f_generator α s)) x - η) ^ 2 /
(-deriv (deriv (subadditive_f_weight (subadditive_f_generator α s))) x) ≤
2 * η + (x - 1) * ℓ- Formalizes
- arXiv:2610.21231 (On the subadditivity of mutual f-information), Corollary 1 (the claim that g satisfies the assumptions of Theorem 1, Appendix C). Differs: s is assumed C² on (0, ∞), which the paper leaves implicit (it needs s'' and Theorem 1 needs f ∈ C⁴); ℓ uses a one-sided derivative of the concave K_s at 1 instead of K_s'(1).
- Verified
- 23 Oct 2026
- Axioms
Classical.choiceQuot.soundpropext- Built on 10
- subadditive_f_generatorThe generator function f associated to parameter α and function s is f(x) = α x log x + ∫_1^x (x - t) (s(t) / t^2) dt.
- subadditive_f_ksFor a twice differentiable function s, the associated K_s function is K_s(x) = s(x) + 2(s'(x))^2/(-s''(x)).
- subadditive_f_weightThe weight g(x) = x² f''(x) attached to a generator f in Theorem 1 of arXiv:2610.21231 (f'' is the iterated derivative
deriv (deriv f)). - subadditive_ks_identityFor real numbers α and x and functions s, g : ℝ → ℝ, if g x = α * x + s x, deriv g x - α = deriv s x, and deriv (deriv g) x = deriv (deriv…
- subadditive_tangent_bound_stepIf K_1 ≤ η_g, ℓ_g = η_g + K'_1, and K_x ≤ K_1 + K'_1 (x - 1), then η_g x + K_x ≤ 2 η_g + (x - 1) ℓ_g.
- concave_on_ioi_exists_supergradient_at_oneA function K concave on (0, ∞) has a supergradient at 1: there is ℓ with K(x) ≤ K(1) + ℓ (x - 1) for all x > 0 (the first-order inequality…
- subadditive_f_generator_hasDerivAtFor s continuous on (0, ∞) and y > 0, the generator f(x) = α x log x + ∫_1^x (x - t) s(t)/t² dt has derivative f'(y) = α (log y + 1) +…
- subadditive_f_generator_deriv2For s continuous on (0, ∞) and y > 0, the derivative of the generator f of equation (4) has derivative α/y + s(y)/y² at y, i.e. f''(y) =…
- subadditive_f_generator_contDiffOnIf s is twice continuously differentiable on (0, ∞), then the generator f of equation (4) is four times continuously differentiable on (0,…
- subadditive_f_generator_weight_eventuallyEqFor s continuous on (0, ∞) and x > 0, the weight g(y) = y² f''(y) of the generator f of equation (4) agrees with α y + s(y) on a…
- Used by 1
- subadditive_mutual_f_information_family_full_supportCorollary 1 of arXiv:2610.21231, on full-support distributions: let s be twice continuously differentiable, positive and concave on (0, ∞)…
The proof, explained
The theorem verifies that the generator f(x) = α x log x + ∫₁ˣ (x - t) s(t)/t² dt satisfies the five conditions required by Theorem 1: convexity, f(1) = 0, C⁴ smoothness on (0, ∞), strict concavity of the weight function g(x) = x² f''(x) = α x + s(x), and the quadratic-concavity weight bound. Smoothness, normalization, and convexity follow from the derivative formulas for f together with α > 0 and s > 0, while g'' < 0 reduces directly to s'' < 0. The weight bound is established by setting η = α and ℓ = α + ℓ_K, where ℓ_K is a supergradient of the concave function K_s at 1, which reduces the bound to the supergradient inequality for K_s via an algebraic identity.
The parameter α is strictly positive.
By definition, K_s(1) = s(1) + 2(s'(1))² / (-s''(1)). Since s(1) > 0, (s'(1))² ≥ 0, and -s''(1) > 0 by the assumption s'' < 0, K_s(1) is the sum of a positive number and a non-negative quotient, hence K_s(1) > 0. Since α ≥ K_s(1), we obtain α > 0.
uses
lt_of_lt_of_le,one_pos,div_nonneg,neg_pos.mprhave hc : ContinuousOn s (Set.Ioi 0) := hs_smooth.continuousOn have hα_pos : 0 < α := by refine lt_of_lt_of_le ?_ hα unfold subadditive_f_ks have h1 := hs_pos 1 one_pos have h2 := div_nonneg (by positivity : (0 : ℝ) ≤ 2 * deriv s 1 ^ 2) (neg_pos.mpr (hs_deriv2_neg 1 one_pos)).le linarith
The function s is differentiable at every point y > 0.
Because s is C² on (0, ∞) and (0, ∞) is an open neighborhood of y, s is differentiable at y.
uses
isOpen_Ioi.mem_nhdshave hsd : ∀ y > 0, DifferentiableAt ℝ s y := fun y hy => (hs_smooth.differentiableOn (by norm_num) y hy).differentiableAt (isOpen_Ioi.mem_nhds hy)
For any x > 0, the derivative of the weight function g agrees locally around x with y ↦ α + s'(y).
In a neighborhood of x, the weight function g(y) = y² f''(y) equals α y + s(y) by subadditive_f_generator_weight_eventuallyEq. Differentiating α y + s(y) term by term gives α + s'(y).
uses
lt_mem_nhds,subadditive_f_generator_weight_eventuallyEq,hasDerivAt_id-- derivatives of the weight g = x² f'' = α x + s x have hg1 : ∀ x > 0, deriv (subadditive_f_weight (subadditive_f_generator α s)) =ᶠ[nhds x] fun y => α + deriv s y := by intro x hx filter_upwards [lt_mem_nhds hx] with y hy rw [(subadditive_f_generator_weight_eventuallyEq α s hc hy).deriv_eq] exact (((hasDerivAt_id y).const_mul α).add (hsd y hy).hasDerivAt).deriv.trans (by simp)
For every x > 0, the second derivative of the weight function satisfies g''(x) = s''(x).
Differentiating g'(y) = α + s'(y) at x eliminates the constant α using deriv_const_add, giving g''(x) = s''(x).
uses
deriv_const_addhave hg2 : ∀ x > 0, deriv (deriv (subadditive_f_weight (subadditive_f_generator α s))) x = deriv (deriv s) x := by intro x hx rw [(hg1 x hx).deriv_eq] exact deriv_const_add α
The generator satisfies f(1) = 0, is C⁴ on (0, ∞), and has strictly concave weight g''(x) < 0 for all x > 0.
Evaluating f(1) yields 0 directly because log 1 = 0 and the integral vanishes. The C⁴ smoothness of f on (0, ∞) is provided by subadditive_f_generator_contDiffOn. Finally, g''(x) = s''(x) < 0 follows from the hypothesis hs_deriv2_neg and the second derivative computation.
uses
subadditive_f_generator_contDiffOnrefine ⟨?_, by simp [subadditive_f_generator], subadditive_f_generator_contDiffOn α s hs_smooth, fun x hx => (hg2 x hx).symm ▸ hs_deriv2_neg x hx, ?_⟩
The generator f is convex on (0, ∞).
On the convex domain (0, ∞), by convexOn_of_deriv2_nonneg' it suffices to verify that f''(y) ≥ 0 for all y > 0. By subadditive_f_generator_deriv2, f''(y) = α / y + s(y) / y². Since α > 0, y > 0, and s(y) > 0, this expression is strictly positive.
uses
convexOn_of_deriv2_nonneg',convex_Ioi,subadditive_f_generator_hasDerivAt,subadditive_f_generator_deriv2,Function.iterate_succ_apply,Function.iterate_one· refine convexOn_of_deriv2_nonneg' (convex_Ioi 0) (fun y hy => (subadditive_f_generator_hasDerivAt α s hc hy).differentiableAt.differentiableWithinAt) (fun y hy => (subadditive_f_generator_deriv2 α s hc hy).differentiableAt.differentiableWithinAt) (fun y hy => ?_) have hy' : (0 : ℝ) < y := hy rw [Function.iterate_succ_apply, Function.iterate_one, (subadditive_f_generator_deriv2 α s hc hy').deriv] have := hs_pos y hy' positivity
There exist parameters η > 0 and ℓ ∈ ℝ satisfying the weight bound for all x > 0.
Because K_s is concave on (0, ∞), it admits a supergradient ℓ_K at 1, so K_s(x) ≤ K_s(1) + ℓ_K (x - 1). Choosing η = α (which is positive) and ℓ = α + ℓ_K, the algebraic identity subadditive_ks_identity reduces the left-hand side of the weight bound to α x + K_s(x). Applying subadditive_tangent_bound_step using the supergradient bound and K_s(1) ≤ α yields α x + K_s(x) ≤ 2α + (x - 1)(α + ℓ_K), which establishes the claim.
uses
concave_on_ioi_exists_supergradient_at_one,subadditive_ks_identity,subadditive_f_generator_weight_eventuallyEq,subadditive_tangent_bound_step· obtain ⟨ℓK, hℓK⟩ := concave_on_ioi_exists_supergradient_at_one hKs_concave refine ⟨α, α + ℓK, hα_pos, fun x hx => ?_⟩ have hid := subadditive_ks_identity α x s (subadditive_f_weight (subadditive_f_generator α s)) (subadditive_f_generator_weight_eventuallyEq α s hc hx).eq_of_nhds (by rw [(hg1 x hx).eq_of_nhds]; ring) (hg2 x hx) rw [hid] exact subadditive_tangent_bound_step α (α + ℓK) _ ℓK x _ rfl hα (by linarith [hℓK x hx])
lt_of_lt_of_le- If a < b and b ≤ c, then a < c.
one_pos- The real number 1 is strictly positive.
div_nonneg- The quotient of two non-negative numbers is non-negative.
neg_pos.mpr- The negation of a strictly negative number is strictly positive.
isOpen_Ioi.mem_nhds- The open interval (a, ∞) is a neighborhood of each of its points.
lt_mem_nhds- If a < b, then a < x holds for all x in a neighborhood of b.
subadditive_f_generator_weight_eventuallyEq- For s continuous on (0, ∞) and x > 0, the weight function g(y) = y² f''(y) equals α y + s(y) in a neighborhood of x.
hasDerivAt_id- The identity function has derivative 1 everywhere.
deriv_const_add- Adding a constant to a function does not change its derivative.
subadditive_f_generator_contDiffOn- If s is C² on (0, ∞), then the generator f is C⁴ on (0, ∞).
convexOn_of_deriv2_nonneg'- A twice-differentiable function on a convex set is convex if its second derivative is everywhere non-negative.
convex_Ioi- The open half-line (r, ∞) is a convex set.
subadditive_f_generator_hasDerivAt- The first derivative of the generator f at y > 0 is α(log y + 1) + ∫₁ʸ s(t)/t² dt.
subadditive_f_generator_deriv2- The derivative of f' (the second derivative f'') at y > 0 is α/y + s(y)/y².
Function.iterate_succ_apply- The (n + 1)-fold iterate of a function f at x equals the n-fold iterate applied to f(x).
Function.iterate_one- The 1-fold iterate of a function is the function itself.
concave_on_ioi_exists_supergradient_at_one- A concave function on (0, ∞) admits a supergradient ℓ at 1.
subadditive_ks_identity- If g(x) = α x + s(x), g'(x) - α = s'(x), and g''(x) = s''(x), then g(x) + 2(g'(x) - α)² / (-g''(x)) = α x + K_s(x).
subadditive_tangent_bound_step- If ℓ_g = η_g + K'_1, K_1 ≤ η_g, and K(x) ≤ K_1 + K'_1(x - 1), then η_g x + K(x) ≤ 2 η_g + (x - 1) ℓ_g.
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 subadditive_f_generator_satisfies_weight_bound (s : ℝ → ℝ) (α : ℝ) (hs_smooth : ContDiffOn ℝ 2 s (Set.Ioi 0)) (hs_pos : ∀ x > 0, 0 < s x) (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) (subadditive_f_generator α s) ∧ subadditive_f_generator α s 1 = 0 ∧ ContDiffOn ℝ 4 (subadditive_f_generator α s) (Set.Ioi 0) ∧ (∀ x > 0, deriv (deriv (subadditive_f_weight (subadditive_f_generator α s))) x < 0) ∧ ∃ η ℓ : ℝ, 0 < η ∧ ∀ x > 0, subadditive_f_weight (subadditive_f_generator α s) x + 2 * (deriv (subadditive_f_weight (subadditive_f_generator α s)) x - η) ^ 2 / (-deriv (deriv (subadditive_f_weight (subadditive_f_generator α s))) x) ≤ 2 * η + (x - 1) * ℓ := by have hc : ContinuousOn s (Set.Ioi 0) := hs_smooth.continuousOn have hα_pos : 0 < α := by refine lt_of_lt_of_le ?_ hα unfold subadditive_f_ks have h1 := hs_pos 1 one_pos have h2 := div_nonneg (by positivity : (0 : ℝ) ≤ 2 * deriv s 1 ^ 2) (neg_pos.mpr (hs_deriv2_neg 1 one_pos)).le linarith have hsd : ∀ y > 0, DifferentiableAt ℝ s y := fun y hy => (hs_smooth.differentiableOn (by norm_num) y hy).differentiableAt (isOpen_Ioi.mem_nhds hy) -- derivatives of the weight g = x² f'' = α x + s x have hg1 : ∀ x > 0, deriv (subadditive_f_weight (subadditive_f_generator α s)) =ᶠ[nhds x] fun y => α + deriv s y := by intro x hx filter_upwards [lt_mem_nhds hx] with y hy rw [(subadditive_f_generator_weight_eventuallyEq α s hc hy).deriv_eq] exact (((hasDerivAt_id y).const_mul α).add (hsd y hy).hasDerivAt).deriv.trans (by simp) have hg2 : ∀ x > 0, deriv (deriv (subadditive_f_weight (subadditive_f_generator α s))) x = deriv (deriv s) x := by intro x hx rw [(hg1 x hx).deriv_eq] exact deriv_const_add α refine ⟨?_, by simp [subadditive_f_generator], subadditive_f_generator_contDiffOn α s hs_smooth, fun x hx => (hg2 x hx).symm ▸ hs_deriv2_neg x hx, ?_⟩ · refine convexOn_of_deriv2_nonneg' (convex_Ioi 0) (fun y hy => (subadditive_f_generator_hasDerivAt α s hc hy).differentiableAt.differentiableWithinAt) (fun y hy => (subadditive_f_generator_deriv2 α s hc hy).differentiableAt.differentiableWithinAt) (fun y hy => ?_) have hy' : (0 : ℝ) < y := hy rw [Function.iterate_succ_apply, Function.iterate_one, (subadditive_f_generator_deriv2 α s hc hy').deriv] have := hs_pos y hy' positivity · obtain ⟨ℓK, hℓK⟩ := concave_on_ioi_exists_supergradient_at_one hKs_concave refine ⟨α, α + ℓK, hα_pos, fun x hx => ?_⟩ have hid := subadditive_ks_identity α x s (subadditive_f_weight (subadditive_f_generator α s)) (subadditive_f_generator_weight_eventuallyEq α s hc hx).eq_of_nhds (by rw [(hg1 x hx).eq_of_nhds]; ring) (hg2 x hx) rw [hid] exact subadditive_tangent_bound_step α (α + ℓK) _ ℓK x _ rfl hα (by linarith [hℓK x hx])