prob_simplex_concaveOn_of_hasDerivAt2
A 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 Σ v = 0, t ↦ F(r + t v) has derivative D1 r v at 0, t ↦ D1(r + t v) v has derivative D2 r v at 0, and D2 r v ≤ 0 (second-derivative test along segments).
Statement
theorem prob_simplex_concaveOn_of_hasDerivAt2 {X : Type*} [Fintype X] (Φ : (X → ℝ) → ℝ) (D1 D2 : (X → ℝ) → (X → ℝ) → ℝ) (h1 : ∀ r : X → ℝ, (∀ x, 0 < r x) → ∑ x, r x = 1 → ∀ v : X → ℝ, ∑ x, v x = 0 → HasDerivAt (fun t : ℝ => Φ (r + t • v)) (D1 r v) 0) (h2 : ∀ r : X → ℝ, (∀ x, 0 < r x) → ∑ x, r x = 1 → ∀ v : X → ℝ, ∑ x, v x = 0 → HasDerivAt (fun t : ℝ => D1 (r + t • v) v) (D2 r v) 0) (h3 : ∀ r : X → ℝ, (∀ x, 0 < r x) → ∑ x, r x = 1 → ∀ v : X → ℝ, ∑ x, v x = 0 → D2 r v ≤ 0) : ConcaveOn ℝ {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} Φ
- Source
- arXiv:2610.21231 (On the subadditivity of mutual f-information), Appendix B, proof of Lemma 4
- Verified
- 24 Oct 2026
- Axioms
Classical.choiceQuot.soundpropext- Built on 4
- ProbabilisticTheory.instArchimedeanOrderUnitSpaceRealThe real numbers form an Archimedean order-unit space.
- ProbabilisticTheory.OrderedVectorSpaceAn ordered real vector space.
- ProbabilisticTheory.OrderUnitSpaceAn ordered real vector space whose distinguished element
1is an order unit: it is nonnegative, and every element is bounded above by a… - ProbabilisticTheory.ArchimedeanOrderUnitSpaceAn order-unit space whose distinguished order unit is Archimedean.
- Used by 2
- mutual_f_theta_concaveOn_of_weight_boundLemma 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…
- 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…
The proof, explained
The proof establishes concavity of Φ on the open probability simplex by reducing it to single-variable concavity along line segments. For any two distributions a and b in the simplex, the displacement v = b - a has zero sum, so the entire segment connecting them stays within the open simplex. Using a chain-rule shift lemma, the restriction φ(t) = Φ(a + t • v) is shown to be twice differentiable on [0, 1] with second derivative D2(a + t • v, v) ≤ 0. The standard second-derivative test on [0, 1] then yields concavity of φ, which specializes at the endpoints to the concavity inequality for Φ.
The open probability simplex S = {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} is convex.
Given any two points a, b in the simplex and convex weights α, β ≥ 0 with α + β = 1, we verify the two defining conditions for α • a + β • b. For strict positivity at any coordinate x, if α > 0, the sum α a(x) + β b(x) is strictly positive by positivity reasoning; if α = 0, then β = 1 and the combination reduces to b(x) > 0. For the sum condition, linearity of summation distributes α • a + β • b into α ∑ a(x) + β ∑ b(x) = α · 1 + β · 1 = 1.
uses
Pi.add_apply,Pi.smul_apply,smul_eq_mul,Finset.sum_add_distrib,Finset.mul_sumhave hS : Convex ℝ {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} := by intro a ha b hb α β hα hβ hαβ refine ⟨fun x => ?_, ?_⟩ · simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul] rcases hα.lt_or_eq with h | h · have := ha.1 x; have := hb.1 x; positivity · subst h have : β = 1 := by linarith subst this; simp [hb.1 x] · simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul, Finset.sum_add_distrib, ← Finset.mul_sum, ha.2, hb.2] linarith
For any points a, b in the simplex, the direction v = b - a has coordinate sum 0, and the segment a + t • v lies in the simplex for all t ∈ [0, 1].
To establish concavity of Φ on the simplex, it suffices to take arbitrary points a, b in the simplex and convex weights α, β ≥ 0 with α + β = 1, and show α Φ(a) + β Φ(b) ≤ Φ(α • a + β • b). Setting v = b - a, its coordinates sum to ∑ b(x) - ∑ a(x) = 1 - 1 = 0. For each t ∈ [0, 1], the point a + t • v rewrites as (1 - t) • a + t • b, which is a convex combination of a and b (since 0 ≤ 1 - t, 0 ≤ t, and (1 - t) + t = 1), hence lies in the simplex by its convexity.
uses
Finset.sum_sub_distribrefine ⟨hS, ?_⟩ intro a ha b hb α β hα hβ hαβ set v : X → ℝ := b - a with hvdef have hv : ∑ x, v x = 0 := by simp [v, Finset.sum_sub_distrib, ha.2, hb.2] have hseg : ∀ t ∈ Set.Icc (0:ℝ) 1, a + t • v ∈ {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} := by intro t ht have : a + t • v = (1 - t) • a + t • b := by funext x; simp [v]; ring rw [this]; exact hS ha hb (by linarith [ht.2]) ht.1 (by ring)
Directional derivatives along the segment can be translated from 0 to any parameter t.
Define φ(t) = Φ(a + t • v) and ψ(t) = D1(a + t • v, v). A general shift lemma is proven: for any functional Gv, if s ↦ Gv((a + t • v) + s • v) has derivative d at s = 0, then s ↦ Gv(a + s • v) has derivative d at s = t. This follows from the chain rule applied to the composition with s ↦ s - t (which has derivative 1 at t), noting that (a + t • v) + (s - t) • v = a + s • v everywhere.
uses
sub_self,hasDerivAt_id',HasDerivAt.comp,mul_one,Filter.Eventually.of_forall,Function.comp,Pi.add_apply,Pi.smul_apply,smul_eq_mulset φ : ℝ → ℝ := fun t => Φ (a + t • v) with hφ set ψ : ℝ → ℝ := fun t => D1 (a + t • v) v with hψ have shift : ∀ (Gv : (X → ℝ) → ℝ) (d t : ℝ), HasDerivAt (fun s : ℝ => Gv ((a + t • v) + s • v)) d 0 → HasDerivAt (fun s : ℝ => Gv (a + s • v)) d t := by intro Gv d t h have hg : HasDerivAt (fun s : ℝ => Gv ((a + t • v) + s • v)) d (t - t) := by rw [sub_self]; exact h have hh : HasDerivAt (fun s : ℝ => s - t) 1 t := (hasDerivAt_id' t).sub_const t have := HasDerivAt.comp (h := fun s : ℝ => s - t) t hg hh rw [mul_one] at this refine this.congr_of_eventuallyEq (Filter.Eventually.of_forall fun s => ?_) simp only [Function.comp] congr 1 funext x; simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul]; ring
For each t ∈ [0, 1], φ has derivative ψ(t) at t, and ψ has derivative D2(a + t • v, v) at t.
For any t ∈ [0, 1], the segment point a + t • v belongs to the simplex and v has sum 0. Applying hypothesis h1 at this base point gives that s ↦ Φ((a + t • v) + s • v) has derivative D1(a + t • v, v) = ψ(t) at s = 0, so the shift lemma implies that φ'(t) = ψ(t). Similarly, applying hypothesis h2 gives that s ↦ D1((a + t • v) + s • v, v) has derivative D2(a + t • v, v) at s = 0, so by the shift lemma, ψ'(t) = D2(a + t • v, v).
have hd1 : ∀ t ∈ Set.Icc (0:ℝ) 1, HasDerivAt φ (ψ t) t := by intro t ht obtain ⟨hp, hs⟩ := hseg t ht exact shift Φ _ t (h1 _ hp hs v hv) have hd2 : ∀ t ∈ Set.Icc (0:ℝ) 1, HasDerivAt ψ (D2 (a + t • v) v) t := by intro t ht obtain ⟨hp, hs⟩ := hseg t ht exact shift (fun r => D1 r v) _ t (h2 _ hp hs v hv)
The one-dimensional restriction φ is concave on the closed interval [0, 1].
We apply the single-variable second-derivative test (`concaveOn_of_deriv2_nonpos`) on [0, 1], whose interior is (0, 1). The function φ is continuous on [0, 1] and differentiable on (0, 1) because it has derivative ψ(t) at every point in [0, 1]. Near any t ∈ (0, 1), the derivative deriv φ agrees with ψ on the open neighborhood (0, 1), so deriv φ is differentiable on (0, 1) with derivative D2(a + t • v, v). Finally, hypothesis h3 ensures D2(a + t • v, v) ≤ 0 everywhere on (0, 1) since each segment point is in the simplex and ∑ v(x) = 0.
uses
interior_Icc,concaveOn_of_deriv2_nonpos,convex_Icc,Set.Ioo_subset_Icc_self,Ioo_mem_nhds,Function.iterate_succ,Function.iterate_one,Function.comphave hint : interior (Set.Icc (0:ℝ) 1) = Set.Ioo 0 1 := interior_Icc have hcc : ConcaveOn ℝ (Set.Icc (0:ℝ) 1) φ := by refine concaveOn_of_deriv2_nonpos (convex_Icc 0 1) (fun t ht => (hd1 t ht).continuousAt.continuousWithinAt) ?_ ?_ ?_ · rw [hint]; intro t ht exact (hd1 t (Set.Ioo_subset_Icc_self ht)).differentiableAt.differentiableWithinAt · rw [hint]; intro t ht have heq : deriv φ =ᶠ[nhds t] ψ := by filter_upwards [Ioo_mem_nhds ht.1 ht.2] with s hs exact (hd1 s (Set.Ioo_subset_Icc_self hs)).deriv exact ((hd2 t (Set.Ioo_subset_Icc_self ht)).congr_of_eventuallyEq heq).differentiableAt.differentiableWithinAt · rw [hint]; intro t ht have heq : deriv φ =ᶠ[nhds t] ψ := by filter_upwards [Ioo_mem_nhds ht.1 ht.2] with s hs exact (hd1 s (Set.Ioo_subset_Icc_self hs)).deriv rw [Function.iterate_succ, Function.iterate_one, Function.comp, ((hd2 t (Set.Ioo_subset_Icc_self ht)).congr_of_eventuallyEq heq).deriv] obtain ⟨hp, hs⟩ := hseg t (Set.Ioo_subset_Icc_self ht) exact h3 _ hp hs v hv
The concavity inequality α Φ(a) + β Φ(b) ≤ Φ(α • a + β • b) holds.
Evaluating the concavity of φ on [0, 1] at 0 and 1 with weights α and β gives α φ(0) + β φ(1) ≤ φ(α · 0 + β · 1) = φ(β). By definition of φ, φ(0) = Φ(a), φ(1) = Φ(a + v) = Φ(b), and φ(β) = Φ(a + β • v) = Φ(α • a + β • b) (using α = 1 - β). Substituting these equalities yields the concavity inequality for Φ.
uses
zero_smul,add_zero,one_smul,smul_eq_mul,mul_zero,mul_one,zero_addhave := hcc.2 (x := 0) (y := 1) (by simp) (by simp) hα hβ hαβ simp only [φ, zero_smul, add_zero, one_smul, smul_eq_mul, mul_zero, mul_one, zero_add] at this have e1 : a + v = b := by funext x; simp [v] have e2 : a + β • v = α • a + β • b := by funext x; simp [v]; rw [show α = 1 - β by linarith]; ring rw [e1, e2] at this exact this
Pi.add_apply- Pointwise addition of functions: (f + g)(x) = f(x) + g(x).
Pi.smul_apply- Pointwise scalar multiplication of functions: (c • f)(x) = c • f(x).
smul_eq_mul- Scalar multiplication in ℝ coincides with standard real multiplication: a • b = a * b.
Finset.sum_add_distrib- Finite summation distributes over addition: ∑ (f(x) + g(x)) = ∑ f(x) + ∑ g(x).
Finset.mul_sum- Multiplication by a constant distributes over a finite sum: a * ∑ f(x) = ∑ (a * f(x)).
Finset.sum_sub_distrib- Finite summation distributes over subtraction: ∑ (f(x) - g(x)) = ∑ f(x) - ∑ g(x).
sub_self- Subtracting an element from itself yields zero: a - a = 0.
hasDerivAt_id'- The identity function x ↦ x has derivative 1 at every point.
HasDerivAt.comp- The chain rule for differentiation: if h has derivative h' at x and g has derivative g' at h(x), then g ∘ h has derivative g' * h' at x.
mul_one- Multiplication by one is the identity: a * 1 = a.
Filter.Eventually.of_forall- A property that holds everywhere holds eventually in any filter.
Function.comp- Function composition: (g ∘ f)(x) = g(f(x)).
interior_Icc- The interior of a closed interval [a, b] in ℝ is the open interval (a, b).
concaveOn_of_deriv2_nonpos- A continuous function on a convex set in ℝ that is twice differentiable on the interior with non-positive second derivative is concave.
convex_Icc- Any closed interval [a, b] in ℝ is a convex set.
Set.Ioo_subset_Icc_self- The open interval (a, b) is a subset of the closed interval [a, b].
Ioo_mem_nhds- If a < x < b, then the open interval (a, b) is an open neighborhood of x.
Function.iterate_succ- The (n+1)-th iterate of a function satisfies f^(n+1) = f^n ∘ f.
Function.iterate_one- The first iterate of a function is the function itself: f^1 = f.
zero_smul- Scaling any vector by the scalar zero yields zero: 0 • m = 0.
add_zero- Adding zero to an element leaves it unchanged: a + 0 = a.
one_smul- Scaling an element by the scalar one leaves it unchanged: 1 • b = b.
mul_zero- Multiplying any element by zero yields zero: a * 0 = 0.
zero_add- Adding zero on the left leaves an element unchanged: 0 + a = a.
- The argument relies on reducing multidimensional concavity to one dimension along the segment between two arbitrary points; formally, moving the base point of the derivative from 0 to an arbitrary parameter t along the segment is handled via a custom translation lemma (`shift`) combining the chain rule with filter-based congruence.
- In Mathlib, `concaveOn_of_deriv2_nonpos` is formulated in terms of iterated derivatives `deriv^[2]`, which requires verifying that `deriv φ` coincides with `ψ` on a neighborhood of each interior point `t ∈ (0, 1)`.
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 prob_simplex_concaveOn_of_hasDerivAt2 {X : Type*} [Fintype X] (Φ : (X → ℝ) → ℝ) (D1 D2 : (X → ℝ) → (X → ℝ) → ℝ) (h1 : ∀ r : X → ℝ, (∀ x, 0 < r x) → ∑ x, r x = 1 → ∀ v : X → ℝ, ∑ x, v x = 0 → HasDerivAt (fun t : ℝ => Φ (r + t • v)) (D1 r v) 0) (h2 : ∀ r : X → ℝ, (∀ x, 0 < r x) → ∑ x, r x = 1 → ∀ v : X → ℝ, ∑ x, v x = 0 → HasDerivAt (fun t : ℝ => D1 (r + t • v) v) (D2 r v) 0) (h3 : ∀ r : X → ℝ, (∀ x, 0 < r x) → ∑ x, r x = 1 → ∀ v : X → ℝ, ∑ x, v x = 0 → D2 r v ≤ 0) : ConcaveOn ℝ {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} Φ := by have hS : Convex ℝ {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} := by intro a ha b hb α β hα hβ hαβ refine ⟨fun x => ?_, ?_⟩ · simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul] rcases hα.lt_or_eq with h | h · have := ha.1 x; have := hb.1 x; positivity · subst h have : β = 1 := by linarith subst this; simp [hb.1 x] · simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul, Finset.sum_add_distrib, ← Finset.mul_sum, ha.2, hb.2] linarith refine ⟨hS, ?_⟩ intro a ha b hb α β hα hβ hαβ set v : X → ℝ := b - a with hvdef have hv : ∑ x, v x = 0 := by simp [v, Finset.sum_sub_distrib, ha.2, hb.2] have hseg : ∀ t ∈ Set.Icc (0:ℝ) 1, a + t • v ∈ {r : X → ℝ | (∀ x, 0 < r x) ∧ ∑ x, r x = 1} := by intro t ht have : a + t • v = (1 - t) • a + t • b := by funext x; simp [v]; ring rw [this]; exact hS ha hb (by linarith [ht.2]) ht.1 (by ring) set φ : ℝ → ℝ := fun t => Φ (a + t • v) with hφ set ψ : ℝ → ℝ := fun t => D1 (a + t • v) v with hψ have shift : ∀ (Gv : (X → ℝ) → ℝ) (d t : ℝ), HasDerivAt (fun s : ℝ => Gv ((a + t • v) + s • v)) d 0 → HasDerivAt (fun s : ℝ => Gv (a + s • v)) d t := by intro Gv d t h have hg : HasDerivAt (fun s : ℝ => Gv ((a + t • v) + s • v)) d (t - t) := by rw [sub_self]; exact h have hh : HasDerivAt (fun s : ℝ => s - t) 1 t := (hasDerivAt_id' t).sub_const t have := HasDerivAt.comp (h := fun s : ℝ => s - t) t hg hh rw [mul_one] at this refine this.congr_of_eventuallyEq (Filter.Eventually.of_forall fun s => ?_) simp only [Function.comp] congr 1 funext x; simp only [Pi.add_apply, Pi.smul_apply, smul_eq_mul]; ring have hd1 : ∀ t ∈ Set.Icc (0:ℝ) 1, HasDerivAt φ (ψ t) t := by intro t ht obtain ⟨hp, hs⟩ := hseg t ht exact shift Φ _ t (h1 _ hp hs v hv) have hd2 : ∀ t ∈ Set.Icc (0:ℝ) 1, HasDerivAt ψ (D2 (a + t • v) v) t := by intro t ht obtain ⟨hp, hs⟩ := hseg t ht exact shift (fun r => D1 r v) _ t (h2 _ hp hs v hv) have hint : interior (Set.Icc (0:ℝ) 1) = Set.Ioo 0 1 := interior_Icc have hcc : ConcaveOn ℝ (Set.Icc (0:ℝ) 1) φ := by refine concaveOn_of_deriv2_nonpos (convex_Icc 0 1) (fun t ht => (hd1 t ht).continuousAt.continuousWithinAt) ?_ ?_ ?_ · rw [hint]; intro t ht exact (hd1 t (Set.Ioo_subset_Icc_self ht)).differentiableAt.differentiableWithinAt · rw [hint]; intro t ht have heq : deriv φ =ᶠ[nhds t] ψ := by filter_upwards [Ioo_mem_nhds ht.1 ht.2] with s hs exact (hd1 s (Set.Ioo_subset_Icc_self hs)).deriv exact ((hd2 t (Set.Ioo_subset_Icc_self ht)).congr_of_eventuallyEq heq).differentiableAt.differentiableWithinAt · rw [hint]; intro t ht have heq : deriv φ =ᶠ[nhds t] ψ := by filter_upwards [Ioo_mem_nhds ht.1 ht.2] with s hs exact (hd1 s (Set.Ioo_subset_Icc_self hs)).deriv rw [Function.iterate_succ, Function.iterate_one, Function.comp, ((hd2 t (Set.Ioo_subset_Icc_self ht)).congr_of_eventuallyEq heq).deriv] obtain ⟨hp, hs⟩ := hseg t (Set.Ioo_subset_Icc_self ht) exact h3 _ hp hs v hv have := hcc.2 (x := 0) (y := 1) (by simp) (by simp) hα hβ hαβ simp only [φ, zero_smul, add_zero, one_smul, smul_eq_mul, mul_zero, mul_one, zero_add] at this have e1 : a + v = b := by funext x; simp [v] have e2 : a + β • v = α • a + β • b := by funext x; simp [v]; rw [show α = 1 - β by linarith]; ring rw [e1, e2] at this exact this