AFTD Auto-Formalizing Theoretical Domains

tsallis_three_qa_weight_score_not_concave

theoremverifiedoriginal

Without convex exposure the QA weight-score need not be concave: for the cubic Tsallis rule, forecasts (0,0,1), (1/4,3/4,0), outcome 1, weights v = (3/10,7/10), w = (0,1) and t = 2/3, the score of the pool at t v + (1-t) w is strictly less than t times the score at v plus (1-t) times the score at w.

Statement

theorem tsallis_three_qa_weight_score_not_concave :
    ∃ (ps : Fin 2 → Fin 3 → ℝ) (v w : Fin 2 → ℝ) (t : ℝ) (xv xw xt : Fin 3 → ℝ) (j : Fin 3),
      (∀ i, ps i ∈ stdSimplex ℝ (Fin 3)) ∧ v ∈ stdSimplex ℝ (Fin 2) ∧
      w ∈ stdSimplex ℝ (Fin 2) ∧ 0 ≤ t ∧ t ≤ 1 ∧
      is_qa_pool (tsallis_expected_reward 3) (tsallis_exposure 3) ps v xv ∧
      is_qa_pool (tsallis_expected_reward 3) (tsallis_exposure 3) ps w xw ∧
      is_qa_pool (tsallis_expected_reward 3) (tsallis_exposure 3) ps (t • v + (1 - t) • w) xt ∧
      proper_score_of_exposure (tsallis_expected_reward 3) (tsallis_exposure 3) xt j <
        t * proper_score_of_exposure (tsallis_expected_reward 3) (tsallis_exposure 3) xv j +
          (1 - t) * proper_score_of_exposure (tsallis_expected_reward 3) (tsallis_exposure 3) xw j
Related work
operator (Claude), 2026-10-03: answer to arXiv:2102.07081 v2, Sec. 7 (does Theorem 5.1, concavity of the QA weight-score in the weights, hold without convex exposure?): no, the cubic Tsallis rule with experts (0,0,1), (1/4,3/4,0) and weights (0,1), (1/5,4/5), (3/10,7/10) is a counterexample
Verified
20 Oct 2026
Axioms
Classical.choiceQuot.soundpropext
Built on 6
  • is_qa_poolx is the quasi-arithmetic pool of forecasts p_i with weights w_i: x lies in the simplex and minimizes G(y) - <y, sum_i w_i g(p_i)> over…
  • proper_score_of_exposureThe score of forecast x on outcome j for the proper scoring rule with expected reward G and exposure g: G(x) + <g(x), e_j - x>.
  • qa_counterexample_forecastsThe two expert forecasts (0, 0, 1) and (1/4, 3/4, 0) on three outcomes.
  • qa_pool_tsallis_three_of_kktFor the two counterexample forecasts and any weights, a simplex point meeting the KKT conditions is the cubic Tsallis QA pool.
  • tsallis_expected_rewardThe Tsallis expected reward with natural parameter gamma: the sum of x_k^gamma.
  • tsallis_exposureThe Tsallis exposure (gradient of the expected reward): gamma x_k^(gamma - 1).

Lean source view module on GitHub

/-- arXiv:2102.07081 (Sec. 7) asks whether Theorem 5.1 (the weight-score of the QA pool is concave in the weights) holds without convex exposure. It does not: for the cubic Tsallis rule, three outcomes, experts `(0,0,1)` and `(1/4,3/4,0)`, outcome 1 and weights `(0,1)`, `(3/10,7/10)` and their `1/3 : 2/3` mixture `(1/5,4/5)`, the score at the mixture is strictly below the mixture of the scores. -/theorem tsallis_three_qa_weight_score_not_concave :    ∃ (ps : Fin 2 → Fin 3 → ℝ) (v w : Fin 2 → ℝ) (t : ℝ) (xv xw xt : Fin 3 → ℝ) (j : Fin 3),      (∀ i, ps i ∈ stdSimplex ℝ (Fin 3)) ∧ v ∈ stdSimplex ℝ (Fin 2) ∧      w ∈ stdSimplex ℝ (Fin 2) ∧ 0 ≤ t ∧ t ≤ 1 ∧      is_qa_pool (tsallis_expected_reward 3) (tsallis_exposure 3) ps v xv ∧      is_qa_pool (tsallis_expected_reward 3) (tsallis_exposure 3) ps w xw ∧      is_qa_pool (tsallis_expected_reward 3) (tsallis_exposure 3) ps (t • v + (1 - t) • w) xt ∧      proper_score_of_exposure (tsallis_expected_reward 3) (tsallis_exposure 3) xt j <        t * proper_score_of_exposure (tsallis_expected_reward 3) (tsallis_exposure 3) xv j +          (1 - t) * proper_score_of_exposure (tsallis_expected_reward 3) (tsallis_exposure 3) xw j := by  refine ⟨qa_counterexample_forecasts, ![3/10, 7/10], ![0, 1], 2/3, ![0, 35/64, 29/64],    ![1/4, 3/4, 0], ![0, 5/8, 3/8], 0, ?_, ?_, ?_, by norm_num, by norm_num, ?_, ?_, ?_, ?_⟩  · intro i; fin_cases i <;>      exact ⟨fun k => by fin_cases k <;> norm_num [qa_counterexample_forecasts],        by simp [qa_counterexample_forecasts, Fin.sum_univ_three, Matrix.cons_val_two, Matrix.tail_cons, Matrix.head_cons] <;> norm_num⟩  · exact ⟨fun k => by fin_cases k <;> simp <;> norm_num, by simp [Fin.sum_univ_two]; norm_num⟩  · exact ⟨fun k => by fin_cases k <;> simp, by simp [Fin.sum_univ_two]⟩  · refine qa_pool_tsallis_three_of_kkt _ _ (-5817/20480)      ⟨fun k => by fin_cases k <;> simp <;> norm_num, by simp [Fin.sum_univ_three]; norm_num⟩ ?_    intro k    fin_cases k <;>      simp [Fin.sum_univ_two, tsallis_exposure, qa_counterexample_forecasts] <;> norm_num  · refine qa_pool_tsallis_three_of_kkt _ _ 0      ⟨fun k => by fin_cases k <;> simp <;> norm_num, by simp [Fin.sum_univ_three]; norm_num⟩ ?_    intro k    fin_cases k <;>      simp [Fin.sum_univ_two, tsallis_exposure, qa_counterexample_forecasts] <;> norm_num  · refine qa_pool_tsallis_three_of_kkt _ _ (-57/320)      ⟨fun k => by fin_cases k <;> simp <;> norm_num, by simp [Fin.sum_univ_three]; norm_num⟩ ?_    intro k    fin_cases k <;>      simp [Fin.sum_univ_two, tsallis_exposure, qa_counterexample_forecasts] <;> norm_num  · simp [proper_score_of_exposure, tsallis_expected_reward, tsallis_exposure, Fin.sum_univ_three]    norm_num
See it in the graphDiscuss this resultChallenge it