mutual_f_theta_concaveOn_of_weight_bound
Lemma 4 of arXiv:2610.21231: under the assumptions of Theorem 1 (g = x² f'' with g'' < 0 and g + 2(g' - η)²/(-g'') ≤ 2η + (x-1)ℓ), for full-support p_X and p_{Y|X}, r_X ↦ Θ_{p_X,p_{Y|X}}(r_X) is concave on the open probability simplex.
Statement
theorem mutual_f_theta_concaveOn_of_weight_bound {X Y : Type*} [Fintype X] [Fintype Y] (f : ℝ → ℝ) (hf1 : f 1 = 0) (hconv : ConvexOn ℝ (Set.Ioi 0) f) (hsmooth : ContDiffOn ℝ 4 f (Set.Ioi 0)) (hg2 : ∀ x > 0, deriv (deriv (subadditive_f_weight f)) x < 0) (hbound : ∃ η ℓ : ℝ, 0 < η ∧ ∀ x > 0, subadditive_f_weight f x + 2 * (deriv (subadditive_f_weight f) x - η) ^ 2 / (-deriv (deriv (subadditive_f_weight f)) x) ≤ 2 * η + (x - 1) * ℓ) (pX : X → ℝ) (PY : X → Y → ℝ) (hpX : ∀ x, 0 < pX x) (hsX : ∑ x, pX x = 1) (hPY : ∀ x y, 0 < PY x y) (hsY : ∀ x, ∑ y, PY x y = 1) : ConcaveOn ℝ {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} (mutual_f_theta f pX PY)
- Source
- arXiv:2610.21231 (On the subadditivity of mutual f-information), Appendix B, Lemma 4
- Verified
- 24 Oct 2026
- Axioms
Classical.choiceQuot.soundpropext- Built on 9
- 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…
- 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_gen_varianceThe generalized variance functional L_{u,p}(q) = E_p[g(q/p) (u - E_q u)²] of eq. (22) of arXiv:2610.21231.
- prob_simplex_concaveOn_of_hasDerivAt2A function on the open probability simplex {r : ∀x, 0 < r x, Σ r = 1} is concave if, at every point r of it and in every direction v with…
- perspective_affine_hasDerivAt_zeroFor Q0 > 0, the perspective t ↦ (Q0 + t Q1) f((C0 + t C1)/(Q0 + t Q1)) has derivative Q1 f(C0/Q0) + (C1 Q0 - C0 Q1) f'(C0/Q0)/Q0 at 0.
- hasDerivAt_comp_affine_at_zeroIf g has derivative g' at c, then t ↦ g(c + t d) has derivative g' d at 0.
- perspective_affine_deriv_hasDerivAt_zeroFor Q0 > 0, the first derivative of the affine perspective along t has derivative (C1 Q0 - C0 Q1)² f''(C0/Q0)/Q0³ at 0: the perspective's…
- subadditive_gen_variance_concaveOnInner step of Lemma 4 of arXiv:2610.21231: if g'' < 0 on (0, ∞) and g + 2(g' - η)²/(-g'') ≤ 2η + (x - 1)ℓ there, then for full-support p_X…
- subadditive_f_weight_hasDerivAt_factsIf f is C⁴ on (0, ∞), then at every x > 0, f, f', g = x² f'' and g' have derivatives f', f'', g', g'' at x.
- 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 establishes concavity of r ↦ Θ(r) on the open probability simplex via the second directional derivative criterion along zero-sum tangent vectors v. By explicit differentiation of the perspective and affine components of Θ, the second directional derivative D₂(r, v) is shown to equal ∑ᵧ Q(y) V(ry) - V(r), where ry is the posterior distribution given output y, Q is the output marginal distribution, and V is a generalized variance functional. Under the given subadditivity differential inequality on the weight function g = x² f'', V is concave on the simplex, so Jensen's inequality implies ∑ᵧ Q(y) V(ry) ≤ V(∑ᵧ Q(y) ry) = V(r), ensuring D₂(r, v) ≤ 0.
The generalized variance functional is concave on the probability simplex under the weight bound hypothesis.
We unpack the parameters η, ℓ from the weight bound assumption hb, establish differentiability of f and gw(x) = x² f''(x) using subadditive_f_weight_hasDerivAt_facts, and apply subadditive_gen_variance_concaveOn to obtain that the generalized variance functional q ↦ subadditive_gen_variance gw pX u q is concave on the open probability simplex for any test vector u.
uses
subadditive_f_weight_hasDerivAt_facts,subadditive_gen_variance_concaveOnclassical obtain ⟨η, ℓ, _, hb⟩ := hbound have hfacts := subadditive_f_weight_hasDerivAt_facts f hsmooth set f1 := deriv f set f2 := deriv (deriv f) set gw := subadditive_f_weight f have hgdef : ∀ a, gw a = a ^ 2 * f2 a := fun a => rfl have hcc := subadditive_gen_variance_concaveOn (X := X) gw (deriv gw) (deriv (deriv gw)) (fun a ha => (hfacts a ha).2.2.1) (fun a ha => (hfacts a ha).2.2.2) hg2 η ℓ hb
The directional derivative candidates D₁ and D₂ are defined, and the output marginal Q(y) is strictly positive on the simplex.
We define D₁ and D₂ representing the first and second directional derivatives along perturbation v. We verify the linearity identity hlin for the perturbed marginal ∑ₓ' (r + t • v)(x') PY(x', y) = Q(y) + t W(y), and establish that Q(y) > 0 whenever r is in the open simplex because X is nonempty (hs : ∑ r = 1) and all terms in the sum are strictly positive.
uses
Finset.sum_add_distrib,Finset.mul_sum,not_nonempty_iff,Finset.sum_pos,mul_pos,Finset.mem_univlet D1 : (X → ℝ) → (X → ℝ) → ℝ := fun r v => ∑ y, ∑ x, pX x * ((∑ x', v x' * PY x' y) * f (r x * PY x y / pX x / ∑ x', r x' * PY x' y) + (v x * PY x y / pX x * (∑ x', r x' * PY x' y) - r x * PY x y / pX x * (∑ x', v x' * PY x' y)) * f1 (r x * PY x y / pX x / ∑ x', r x' * PY x' y) / (∑ x', r x' * PY x' y)) - ∑ x, pX x * (f1 (r x / pX x) * (v x / pX x)) let D2 : (X → ℝ) → (X → ℝ) → ℝ := fun r v => ∑ y, ∑ x, pX x * ((v x * PY x y / pX x * (∑ x', r x' * PY x' y) - r x * PY x y / pX x * (∑ x', v x' * PY x' y)) ^ 2 * f2 (r x * PY x y / pX x / ∑ x', r x' * PY x' y) / (∑ x', r x' * PY x' y) ^ 3) - ∑ x, pX x * (f2 (r x / pX x) * (v x / pX x) * (v x / pX x)) have hlin : ∀ (r v : X → ℝ) (t : ℝ) (y : Y), ∑ x', (r + t • v) x' * PY x' y = ∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y := by intro r v t y simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul, add_mul, Finset.sum_add_distrib, Finset.mul_sum, mul_assoc] have hQpos : ∀ r : X → ℝ, (∀ x, 0 < r x) → ∑ x, r x = 1 → ∀ y, 0 < ∑ x', r x' * PY x' y := by intro r hr hs y have hX : Nonempty X := by by_contra h rw [not_nonempty_iff] at h simp at hs exact Finset.sum_pos (fun x _ => mul_pos (hr x) (hPY x y)) ⟨Classical.arbitrary X, Finset.mem_univ _⟩
Concavity of Θ on the open probability simplex reduces to verifying its first and second directional derivatives and showing D₂(r, v) ≤ 0 for zero-sum directions v.
We invoke prob_simplex_concaveOn_of_hasDerivAt2, which reduces concavity on the open simplex to three obligations: checking that D₁(r, v) is the first directional derivative of Θ at r along v, that D₂(r, v) is the directional derivative of D₁(·, v) at r along v, and that D₂(r, v) ≤ 0 for all zero-sum perturbation directions v.
uses
prob_simplex_concaveOn_of_hasDerivAt2apply prob_simplex_concaveOn_of_hasDerivAt2 (mutual_f_theta f pX PY) D1 D2The first directional derivative of t ↦ Θ(r + t • v) at t = 0 equals D₁(r, v).
We expand the path t ↦ Θ(r + t • v) into perspective and affine terms using hlin and algebraic rearrangement. We then differentiate term-by-term using HasDerivAt.fun_sub, HasDerivAt.fun_sum, perspective_affine_hasDerivAt_zero for the perspective summands, and hasDerivAt_comp_affine_at_zero for the affine summands.
uses
Finset.sum_congr,HasDerivAt.fun_sub,HasDerivAt.fun_sum,div_pos,mul_pos,perspective_affine_hasDerivAt_zero,HasDerivAt.const_mul,hasDerivAt_comp_affine_at_zero· intro r hr hs v _ have hfun : (fun t : ℝ => mutual_f_theta f pX PY (r + t • v)) = fun t : ℝ => ∑ y, ∑ x, pX x * ((∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y) * f ((r x * PY x y / pX x + t * (v x * PY x y / pX x)) / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y))) - ∑ x, pX x * f (r x / pX x + t * (v x / pX x)) := by funext t simp only [mutual_f_theta, hlin] congr 1 · refine Finset.sum_congr rfl (fun y _ => ?_) rw [Finset.mul_sum] refine Finset.sum_congr rfl (fun x _ => ?_) simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul] rw [show (r x + t * v x) * PY x y / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y) / pX x = (r x * PY x y / pX x + t * (v x * PY x y / pX x)) / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y) by ring] ring · refine Finset.sum_congr rfl (fun x _ => ?_) simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul] rw [show (r x + t * v x) / pX x = r x / pX x + t * (v x / pX x) by ring] rw [hfun] refine HasDerivAt.fun_sub (HasDerivAt.fun_sum (fun y _ => HasDerivAt.fun_sum (fun x _ => ?_))) (HasDerivAt.fun_sum (fun x _ => ?_)) · have hu : 0 < r x * PY x y / pX x / ∑ x', r x' * PY x' y := div_pos (div_pos (mul_pos (hr x) (hPY x y)) (hpX x)) (hQpos r hr hs y) exact (perspective_affine_hasDerivAt_zero f f1 _ (∑ x', v x' * PY x' y) _ (v x * PY x y / pX x) (hQpos r hr hs y) (hfacts _ hu).1).const_mul (pX x) · exact (hasDerivAt_comp_affine_at_zero f _ _ (v x / pX x) (hfacts _ (div_pos (hr x) (hpX x))).1).const_mul (pX x)
The second directional derivative of t ↦ D₁(r + t • v, v) at t = 0 equals D₂(r, v).
We express the map t ↦ D₁(r + t • v, v) algebraically and differentiate at t = 0. Using perspective_affine_deriv_hasDerivAt_zero for the first-derivative perspective terms and hasDerivAt_comp_affine_at_zero with HasDerivAt.mul_const for the affine terms, we find that the derivative matches D₂(r, v).
uses
HasDerivAt.fun_sub,HasDerivAt.fun_sum,div_pos,mul_pos,perspective_affine_deriv_hasDerivAt_zero,HasDerivAt.const_mul,hasDerivAt_comp_affine_at_zero,HasDerivAt.mul_const· intro r hr hs v _ have hfun : (fun t : ℝ => D1 (r + t • v) v) = fun t : ℝ => ∑ y, ∑ x, pX x * ((∑ x', v x' * PY x' y) * f ((r x * PY x y / pX x + t * (v x * PY x y / pX x)) / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y)) + (v x * PY x y / pX x * (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y) - (r x * PY x y / pX x + t * (v x * PY x y / pX x)) * (∑ x', v x' * PY x' y)) * f1 ((r x * PY x y / pX x + t * (v x * PY x y / pX x)) / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y)) / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y)) - ∑ x, pX x * (f1 (r x / pX x + t * (v x / pX x)) * (v x / pX x)) := by funext t have e1 : ∀ x y, (r + t • v) x * PY x y / pX x = r x * PY x y / pX x + t * (v x * PY x y / pX x) := by intro x y; simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul]; ring have e2 : ∀ x, (r + t • v) x / pX x = r x / pX x + t * (v x / pX x) := by intro x; simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul]; ring simp only [D1, hlin, e1, e2] rw [hfun] refine HasDerivAt.fun_sub (HasDerivAt.fun_sum (fun y _ => HasDerivAt.fun_sum (fun x _ => ?_))) (HasDerivAt.fun_sum (fun x _ => ?_)) · have hu : 0 < r x * PY x y / pX x / ∑ x', r x' * PY x' y := div_pos (div_pos (mul_pos (hr x) (hPY x y)) (hpX x)) (hQpos r hr hs y) exact (perspective_affine_deriv_hasDerivAt_zero f f1 f2 _ (∑ x', v x' * PY x' y) _ (v x * PY x y / pX x) (hQpos r hr hs y) (hfacts _ hu).1 (hfacts _ hu).2.1).const_mul (pX x) · exact ((hasDerivAt_comp_affine_at_zero f1 _ _ (v x / pX x) (hfacts _ (div_pos (hr x) (hpX x))).2.1).mul_const (v x / pX x)).const_mul (pX x)
The posterior distributions ry(·) lie in the simplex, average to r, and satisfy Jensen's inequality for the generalized variance functional.
Let Q(y) = ∑ₓ' r(x') PY(x', y), W(y) = ∑ₓ' v(x') PY(x', y), and β(x) = v(x) / r(x). We define the posterior distributions ry(x) = r(x) PY(x, y) / Q(y), check that each ry is in the simplex, and observe that ∑ᵧ Q(y) = 1 and ∑ᵧ Q(y) ry = r. By concavity of the generalized variance functional V(q) = subadditive_gen_variance gw pX β q, Jensen's inequality (ConcaveOn.le_map_sum) gives ∑ᵧ Q(y) V(ry) ≤ V(r). We also compute the posterior expectation ∑ₓ' ry(x') β(x') = W(y) / Q(y) and the prior expectation ∑ₓ' r(x') β(x') = ∑ v = 0.
uses
div_pos,mul_pos,Finset.sum_div,div_self,Finset.sum_comm,mul_one,mul_div_cancel₀,ConcaveOn.le_map_sum,Finset.sum_congr· intro r hr hs v hv set Q : Y → ℝ := fun y => ∑ x', r x' * PY x' y with hQ set W : Y → ℝ := fun y => ∑ x', v x' * PY x' y with hW set β : X → ℝ := fun x => v x / r x with hβ set ry : Y → X → ℝ := fun y x => r x * PY x y / Q y with hry have hQp : ∀ y, 0 < Q y := hQpos r hr hs have hmem : ∀ y ∈ (Finset.univ : Finset Y), ry y ∈ {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} := by intro y _ refine ⟨fun x => div_pos (mul_pos (hr x) (hPY x y)) (hQp y), ?_⟩ simp only [hry, ← Finset.sum_div] exact div_self (hQp y).ne' have hsQ : ∑ y, Q y = 1 := by simp only [hQ]; rw [Finset.sum_comm] simp only [← Finset.mul_sum, hsY, mul_one, hs] have hsum : ∑ y, Q y • ry y = r := by funext x simp only [Finset.sum_apply, Pi.smul_apply, smul_eq_mul, hry] have : ∀ y, Q y * (r x * PY x y / Q y) = r x * PY x y := fun y => mul_div_cancel₀ _ (hQp y).ne' simp only [this, ← Finset.mul_sum, hsY, mul_one] have hj := (hcc pX β hpX hsX).le_map_sum (t := Finset.univ) (w := Q) (p := ry) (fun y _ => (hQp y).le) hsQ hmem rw [hsum] at hj simp only [smul_eq_mul] at hj have hmy : ∀ y, ∑ x', ry y x' * β x' = W y / Q y := by intro y simp only [hry, hβ, hW] rw [Finset.sum_div] refine Finset.sum_congr rfl (fun x _ => ?_) have := (hr x).ne' field_simp have hm0 : ∑ x', r x' * β x' = 0 := by simp only [hβ] rw [← hv] refine Finset.sum_congr rfl (fun x _ => ?_) have := (hr x).ne' field_simp
The second directional derivative D₂(r, v) equals the nonpositive Jensen gap ∑ᵧ Q(y) V(ry) - V(r) ≤ 0.
By expanding definitions and clearing denominators using the posterior and prior expectations of β, we establish the algebraic identity D₂(r, v) = ∑ᵧ Q(y) V(ry) - V(r). Since Jensen's inequality established that ∑ᵧ Q(y) V(ry) ≤ V(r), linear arithmetic immediately yields D₂(r, v) ≤ 0, closing the final goal.
uses
Finset.sum_congr,Finset.mul_sumhave hD2 : D2 r v = ∑ y, Q y * subadditive_gen_variance gw pX β (ry y) - subadditive_gen_variance gw pX β r := by simp only [D2, subadditive_gen_variance, hmy, hm0] congr 1 · refine Finset.sum_congr rfl (fun y _ => ?_) rw [Finset.mul_sum] refine Finset.sum_congr rfl (fun x _ => ?_) simp only [hry, hgdef, hQ, hW, hβ] have hp := (hpX x).ne' have hq : (∑ x', r x' * PY x' y) ≠ 0 := (hQpos r hr hs y).ne' have hrx := (hr x).ne' rw [show r x * PY x y / (∑ x', r x' * PY x' y) / pX x = r x * PY x y / pX x / (∑ x', r x' * PY x' y) by ring] field_simp · refine Finset.sum_congr rfl (fun x _ => ?_) simp only [hgdef, hβ] have hp := (hpX x).ne' have hrx := (hr x).ne' field_simp ring rw [hD2] linarith
ConcaveOn.le_map_sum- Jensen's inequality for concave functions: a concave function evaluated at a convex combination is bounded below by the convex combination of the function values.
Finset.sum_add_distrib- Distributivity of finite sums: summing a pointwise sum of two functions equals the sum of their individual sums.
Finset.sum_comm- Swapping the order of two finite summations preserves the sum.
Finset.sum_congr- Two finite sums are equal if their index sets and summands agree pointwise.
Finset.sum_div- Dividing a finite sum by a scalar equals summing each term divided by that scalar.
Finset.mul_sum- Multiplication by a constant distributes over a finite sum.
Finset.sum_pos- A finite sum of strictly positive numbers over a nonempty finite set is strictly positive.
Finset.mem_univ- Every element of a finite type belongs to the universal finset univ.
HasDerivAt.fun_sub- The derivative of the difference of two functions is the difference of their derivatives.
HasDerivAt.fun_sum- The derivative of a finite sum of differentiable functions is the sum of their derivatives.
HasDerivAt.const_mul- The derivative of a constant times a function is the constant times the derivative of the function.
HasDerivAt.mul_const- The derivative of a function multiplied by a constant scalar is the derivative multiplied by that scalar.
div_pos- The quotient of two strictly positive real numbers is strictly positive.
div_self- Any nonzero number divided by itself is 1.
hasDerivAt_comp_affine_at_zero- The derivative of the affine composition t ↦ g(c + t * d) at t = 0 is g'(c) * d.
mul_div_cancel₀- Multiplying a nonzero denominator by a fraction with that denominator cancels the denominator.
mul_one- Multiplying any number by 1 preserves the number.
mul_pos- The product of two strictly positive real numbers is strictly positive.
not_nonempty_iff- A type is not nonempty if and only if it is empty.
perspective_affine_hasDerivAt_zero- The derivative of the perspective-affine composite t ↦ (Q₀ + t Q₁) * f((C₀ + t C₁) / (Q₀ + t Q₁)) at t = 0 is Q₁ * f(C₀ / Q₀) + (C₁ Q₀ - C₀ Q₁) * f'(C₀ / Q₀) / Q₀.
perspective_affine_deriv_hasDerivAt_zero- The derivative of the first derivative of the perspective-affine composite at t = 0 is (C₁ Q₀ - C₀ Q₁)² * f''(C₀ / Q₀) / Q₀³.
prob_simplex_concaveOn_of_hasDerivAt2- A function on the open probability simplex is concave if its first and second directional derivatives exist along zero-sum perturbation directions, and the second directional derivative is everywhere nonpositive.
subadditive_f_weight_hasDerivAt_facts- For a C⁴ function f on (0, ∞), f, f', gw, and gw' are differentiable at every positive point, where gw(x) = x² f''(x).
subadditive_gen_variance_concaveOn- Under the subadditivity differential inequality for the weight function gw, the generalized variance functional q ↦ ∑ₓ pX(x) gw(q(x) / pX(x)) (u(x) - ∑ₓ' q(x') u(x'))² is concave on the open probability simplex for any vector u.
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_concaveOn_of_weight_bound {X Y : Type*} [Fintype X] [Fintype Y] (f : ℝ → ℝ) (hf1 : f 1 = 0) (hconv : ConvexOn ℝ (Set.Ioi 0) f) (hsmooth : ContDiffOn ℝ 4 f (Set.Ioi 0)) (hg2 : ∀ x > 0, deriv (deriv (subadditive_f_weight f)) x < 0) (hbound : ∃ η ℓ : ℝ, 0 < η ∧ ∀ x > 0, subadditive_f_weight f x + 2 * (deriv (subadditive_f_weight f) x - η) ^ 2 / (-deriv (deriv (subadditive_f_weight f)) x) ≤ 2 * η + (x - 1) * ℓ) (pX : X → ℝ) (PY : X → Y → ℝ) (hpX : ∀ x, 0 < pX x) (hsX : ∑ x, pX x = 1) (hPY : ∀ x y, 0 < PY x y) (hsY : ∀ x, ∑ y, PY x y = 1) : ConcaveOn ℝ {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} (mutual_f_theta f pX PY) := by classical obtain ⟨η, ℓ, _, hb⟩ := hbound have hfacts := subadditive_f_weight_hasDerivAt_facts f hsmooth set f1 := deriv f set f2 := deriv (deriv f) set gw := subadditive_f_weight f have hgdef : ∀ a, gw a = a ^ 2 * f2 a := fun a => rfl have hcc := subadditive_gen_variance_concaveOn (X := X) gw (deriv gw) (deriv (deriv gw)) (fun a ha => (hfacts a ha).2.2.1) (fun a ha => (hfacts a ha).2.2.2) hg2 η ℓ hb let D1 : (X → ℝ) → (X → ℝ) → ℝ := fun r v => ∑ y, ∑ x, pX x * ((∑ x', v x' * PY x' y) * f (r x * PY x y / pX x / ∑ x', r x' * PY x' y) + (v x * PY x y / pX x * (∑ x', r x' * PY x' y) - r x * PY x y / pX x * (∑ x', v x' * PY x' y)) * f1 (r x * PY x y / pX x / ∑ x', r x' * PY x' y) / (∑ x', r x' * PY x' y)) - ∑ x, pX x * (f1 (r x / pX x) * (v x / pX x)) let D2 : (X → ℝ) → (X → ℝ) → ℝ := fun r v => ∑ y, ∑ x, pX x * ((v x * PY x y / pX x * (∑ x', r x' * PY x' y) - r x * PY x y / pX x * (∑ x', v x' * PY x' y)) ^ 2 * f2 (r x * PY x y / pX x / ∑ x', r x' * PY x' y) / (∑ x', r x' * PY x' y) ^ 3) - ∑ x, pX x * (f2 (r x / pX x) * (v x / pX x) * (v x / pX x)) have hlin : ∀ (r v : X → ℝ) (t : ℝ) (y : Y), ∑ x', (r + t • v) x' * PY x' y = ∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y := by intro r v t y simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul, add_mul, Finset.sum_add_distrib, Finset.mul_sum, mul_assoc] have hQpos : ∀ r : X → ℝ, (∀ x, 0 < r x) → ∑ x, r x = 1 → ∀ y, 0 < ∑ x', r x' * PY x' y := by intro r hr hs y have hX : Nonempty X := by by_contra h rw [not_nonempty_iff] at h simp at hs exact Finset.sum_pos (fun x _ => mul_pos (hr x) (hPY x y)) ⟨Classical.arbitrary X, Finset.mem_univ _⟩ apply prob_simplex_concaveOn_of_hasDerivAt2 (mutual_f_theta f pX PY) D1 D2 · intro r hr hs v _ have hfun : (fun t : ℝ => mutual_f_theta f pX PY (r + t • v)) = fun t : ℝ => ∑ y, ∑ x, pX x * ((∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y) * f ((r x * PY x y / pX x + t * (v x * PY x y / pX x)) / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y))) - ∑ x, pX x * f (r x / pX x + t * (v x / pX x)) := by funext t simp only [mutual_f_theta, hlin] congr 1 · refine Finset.sum_congr rfl (fun y _ => ?_) rw [Finset.mul_sum] refine Finset.sum_congr rfl (fun x _ => ?_) simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul] rw [show (r x + t * v x) * PY x y / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y) / pX x = (r x * PY x y / pX x + t * (v x * PY x y / pX x)) / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y) by ring] ring · refine Finset.sum_congr rfl (fun x _ => ?_) simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul] rw [show (r x + t * v x) / pX x = r x / pX x + t * (v x / pX x) by ring] rw [hfun] refine HasDerivAt.fun_sub (HasDerivAt.fun_sum (fun y _ => HasDerivAt.fun_sum (fun x _ => ?_))) (HasDerivAt.fun_sum (fun x _ => ?_)) · have hu : 0 < r x * PY x y / pX x / ∑ x', r x' * PY x' y := div_pos (div_pos (mul_pos (hr x) (hPY x y)) (hpX x)) (hQpos r hr hs y) exact (perspective_affine_hasDerivAt_zero f f1 _ (∑ x', v x' * PY x' y) _ (v x * PY x y / pX x) (hQpos r hr hs y) (hfacts _ hu).1).const_mul (pX x) · exact (hasDerivAt_comp_affine_at_zero f _ _ (v x / pX x) (hfacts _ (div_pos (hr x) (hpX x))).1).const_mul (pX x) · intro r hr hs v _ have hfun : (fun t : ℝ => D1 (r + t • v) v) = fun t : ℝ => ∑ y, ∑ x, pX x * ((∑ x', v x' * PY x' y) * f ((r x * PY x y / pX x + t * (v x * PY x y / pX x)) / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y)) + (v x * PY x y / pX x * (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y) - (r x * PY x y / pX x + t * (v x * PY x y / pX x)) * (∑ x', v x' * PY x' y)) * f1 ((r x * PY x y / pX x + t * (v x * PY x y / pX x)) / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y)) / (∑ x', r x' * PY x' y + t * ∑ x', v x' * PY x' y)) - ∑ x, pX x * (f1 (r x / pX x + t * (v x / pX x)) * (v x / pX x)) := by funext t have e1 : ∀ x y, (r + t • v) x * PY x y / pX x = r x * PY x y / pX x + t * (v x * PY x y / pX x) := by intro x y; simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul]; ring have e2 : ∀ x, (r + t • v) x / pX x = r x / pX x + t * (v x / pX x) := by intro x; simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul]; ring simp only [D1, hlin, e1, e2] rw [hfun] refine HasDerivAt.fun_sub (HasDerivAt.fun_sum (fun y _ => HasDerivAt.fun_sum (fun x _ => ?_))) (HasDerivAt.fun_sum (fun x _ => ?_)) · have hu : 0 < r x * PY x y / pX x / ∑ x', r x' * PY x' y := div_pos (div_pos (mul_pos (hr x) (hPY x y)) (hpX x)) (hQpos r hr hs y) exact (perspective_affine_deriv_hasDerivAt_zero f f1 f2 _ (∑ x', v x' * PY x' y) _ (v x * PY x y / pX x) (hQpos r hr hs y) (hfacts _ hu).1 (hfacts _ hu).2.1).const_mul (pX x) · exact ((hasDerivAt_comp_affine_at_zero f1 _ _ (v x / pX x) (hfacts _ (div_pos (hr x) (hpX x))).2.1).mul_const (v x / pX x)).const_mul (pX x) · intro r hr hs v hv set Q : Y → ℝ := fun y => ∑ x', r x' * PY x' y with hQ set W : Y → ℝ := fun y => ∑ x', v x' * PY x' y with hW set β : X → ℝ := fun x => v x / r x with hβ set ry : Y → X → ℝ := fun y x => r x * PY x y / Q y with hry have hQp : ∀ y, 0 < Q y := hQpos r hr hs have hmem : ∀ y ∈ (Finset.univ : Finset Y), ry y ∈ {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} := by intro y _ refine ⟨fun x => div_pos (mul_pos (hr x) (hPY x y)) (hQp y), ?_⟩ simp only [hry, ← Finset.sum_div] exact div_self (hQp y).ne' have hsQ : ∑ y, Q y = 1 := by simp only [hQ]; rw [Finset.sum_comm] simp only [← Finset.mul_sum, hsY, mul_one, hs] have hsum : ∑ y, Q y • ry y = r := by funext x simp only [Finset.sum_apply, Pi.smul_apply, smul_eq_mul, hry] have : ∀ y, Q y * (r x * PY x y / Q y) = r x * PY x y := fun y => mul_div_cancel₀ _ (hQp y).ne' simp only [this, ← Finset.mul_sum, hsY, mul_one] have hj := (hcc pX β hpX hsX).le_map_sum (t := Finset.univ) (w := Q) (p := ry) (fun y _ => (hQp y).le) hsQ hmem rw [hsum] at hj simp only [smul_eq_mul] at hj have hmy : ∀ y, ∑ x', ry y x' * β x' = W y / Q y := by intro y simp only [hry, hβ, hW] rw [Finset.sum_div] refine Finset.sum_congr rfl (fun x _ => ?_) have := (hr x).ne' field_simp have hm0 : ∑ x', r x' * β x' = 0 := by simp only [hβ] rw [← hv] refine Finset.sum_congr rfl (fun x _ => ?_) have := (hr x).ne' field_simp have hD2 : D2 r v = ∑ y, Q y * subadditive_gen_variance gw pX β (ry y) - subadditive_gen_variance gw pX β r := by simp only [D2, subadditive_gen_variance, hmy, hm0] congr 1 · refine Finset.sum_congr rfl (fun y _ => ?_) rw [Finset.mul_sum] refine Finset.sum_congr rfl (fun x _ => ?_) simp only [hry, hgdef, hQ, hW, hβ] have hp := (hpX x).ne' have hq : (∑ x', r x' * PY x' y) ≠ 0 := (hQpos r hr hs y).ne' have hrx := (hr x).ne' rw [show r x * PY x y / (∑ x', r x' * PY x' y) / pX x = r x * PY x y / pX x / (∑ x', r x' * PY x' y) by ring] field_simp · refine Finset.sum_congr rfl (fun x _ => ?_) simp only [hgdef, hβ] have hp := (hpX x).ne' have hrx := (hr x).ne' field_simp ring rw [hD2] linarith