AFTD Auto-Formalizing Theoretical Domains

relu_continuous_flag_two_step_erasure

theoremverified

Suppose a real-valued flag update G is continuous in its test input at an initial flag p, with G(0,p)=p and G(1,p)=p-c, and a true test clears the flag c: G(1,c)=0. If c<p≤2c, there is an intermediate test value e in [0,1] for which applying the e test and then the true test clears the flag to zero.

Statement

theorem relu_continuous_flag_two_step_erasure (G : ℝ → ℝ → ℝ) (c p : ℝ) (hp : c < p) (hp2 : p ≤ 2 * c) (hcont : ContinuousOn (fun e => G e p) (Set.Icc 0 1)) (hzero : G 0 p = p) (hone : G 1 p = p - c) (hclear : G 1 c = 0) : ∃ e ∈ Set.Icc (0 : ℝ) 1, G 1 (G e p) = 0
Source
Intermediate-value obstruction for the clause-flag approach to OP-75, arXiv:2610.06092, Remark 5.15; this concerns a general class of candidate updates, not the open problem itself.
Verified
15 Nov 2026
Axioms
Classical.choiceQuot.soundpropext

The proof, explained

To clear the flag in two steps, it suffices to find an intermediate test value e ∈ [0, 1] that sets the flag to c, since a subsequent true test clears c to 0. Under the bounds c < p ≤ 2c, the target value c lies between the endpoint values G(1, p) = p - c and G(0, p) = p. The Intermediate Value Theorem applied to the continuous update map e ↦ G(e, p) on [0, 1] guarantees the existence of such an e.

  1. c lies in the interval [G(1, p), G(0, p)] = [p - c, p].

    Rewriting the interval endpoints using G(0, p) = p and G(1, p) = p - c reduces the membership condition to p - c ≤ c and c ≤ p. The inequality p - c ≤ c follows from p ≤ 2c, and c ≤ p follows from c < p, both verified by linear arithmetic.

    lines 2–4

    uses Set.mem_Icc

      have hmem : c ∈ Set.Icc ((fun e => G e p) 1) ((fun e => G e p) 0) := by
        simp only [hone, hzero, Set.mem_Icc]
        constructor <;> linarith
  2. There exists an input e ∈ [0, 1] such that G(e, p) = c.

    Since 0 ≤ 1 and the function e ↦ G(e, p) is continuous on [0, 1], the Intermediate Value Theorem (intermediate_value_Icc') guarantees that the image of [0, 1] covers the interval [G(1, p), G(0, p)]. Because c belongs to this interval, there exists a point e ∈ [0, 1] with G(e, p) = c.

    lines 5–6

    uses intermediate_value_Icc'

      obtain ⟨e, he, heq⟩ := intermediate_value_Icc' (by norm_num : (0 : ℝ) ≤ 1) hcont hmem
      change G e p = c at heq
  3. The witness e ∈ [0, 1] clears the flag after the second test: G(1, G(e, p)) = 0.

    Substituting G(e, p) = c into G(1, G(e, p)) yields G(1, c), which equals 0 by the assumption hclear. Thus, e satisfies all required conditions.

    line 7

      exact ⟨e, he, by rw [heq, hclear]⟩
Set.mem_Icc
A point x belongs to the closed interval [a, b] if and only if a ≤ x and x ≤ b.
intermediate_value_Icc'
If a function f is continuous on [a, b] with a ≤ b, then the closed interval [f(b), f(a)] is contained in the image f([a, b]).

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

theorem relu_continuous_flag_two_step_erasure (G : ℝ → ℝ → ℝ) (c p : ℝ) (hp : c < p) (hp2 : p ≤ 2 * c) (hcont : ContinuousOn (fun e => G e p) (Set.Icc 0 1)) (hzero : G 0 p = p) (hone : G 1 p = p - c) (hclear : G 1 c = 0) : ∃ e ∈ Set.Icc (0 : ℝ) 1, G 1 (G e p) = 0 := by  have hmem : c ∈ Set.Icc ((fun e => G e p) 1) ((fun e => G e p) 0) := by    simp only [hone, hzero, Set.mem_Icc]    constructor <;> linarith  obtain ⟨e, he, heq⟩ := intermediate_value_Icc' (by norm_num : (0 : ℝ) ≤ 1) hcont hmem  change G e p = c at heq  exact ⟨e, he, by rw [heq, hclear]⟩
See it in the graphDiscuss this resultChallenge it