AFTD Auto-Formalizing Theoretical Domains

op182_deferred_to_deadline_beyond_half_against_stopping

theoremverifiedformalization

Refutation of Open Problem OP-182 against the deferred-until-stopping benchmark: there does not exist any constant c > 1/2 such that for every Pandora's problem instance with deadlines inst : PandoraDeadlineInstance, there exists a valid adaptive policy π satisfying IsValidPolicy inst π whose expected utility under deferred-to-deadline commitment expectedUtilityDTD inst π is at least c * optDUS inst.

Statement

theorem op182_deferred_to_deadline_beyond_half_against_stopping :
    ¬ (∃ c > (1 / 2 : ℝ), ∀ inst : PandoraDeadlineInstance,
        ∃ π : AdaptivePolicy inst.n, IsValidPolicy inst π ∧
          c * optDUS inst ≤ expectedUtilityDTD inst π)
Formalizes
arXiv:2610.13775 Sec. 5
Verified
24 Oct 2026
Axioms
Classical.choiceQuot.soundpropext
Built on 8
  • optDUSThe optimal expected net utility optDUS inst under deferred-until-stopping commitment for a Pandora's problem instance with deadlines…
  • expectedUtilityDTDThe expected net utility of an adaptive policy π : AdaptivePolicy inst.n under deferred-to-deadline commitment on a Pandora's problem…
  • IsValidPolicyA policy is valid for a Pandora's problem instance with deadlines if it respects opening deadlines and never opens any box more than once.
  • expected_utility_dus_le_opt_dusFor any Pandora's problem instance with deadlines inst and any valid policy π, expectedUtilityDUS inst π ≤ optDUS inst.
  • two_box_counterexample_familyFor any integer M > 1, there exists a 2-box Pandora deadline instance such that 2 - 1 / M ≤ optDUS inst and every valid policy π…
  • archimedean_ratio_gt_oneFor any real constant c > 1/2, there exists a natural number M > 1 such that 1 < c * (2 - 1 / M).
  • PandoraDeadlineInstanceAn instance of Pandora's problem with deadlines consisting of n boxes, each with a non-negative opening cost, a positive integer deadline,…
  • AdaptivePolicyA fully adaptive search policy for Pandora's problem with n boxes, represented as a decision tree that at each step either stops or opens…

Read back from the Lean

The theorem asserts that there does not exist any real constant $c > 1/2$ such that for every Pandora's box instance with deadlines (inst : PandoraDeadlineInstance), there exists a valid adaptive policy $\pi$ (IsValidPolicy inst π) whose expected utility under the deferred-to-deadline model (expectedUtilityDTD inst π) is at least $c$ times the optimal expected utility under the deferred unconstrained stopping model (optDUS inst).

Written by a model that saw only the Lean, never the English above. If the two disagree, that is worth a challenge.

The proof, explained

The proof proceeds by contradiction. Assuming there exists an approximation ratio $c > 1/2$, the Archimedean property yields an integer $M > 1$ such that $c(2 - 1/M) > 1$. A family of two-box counterexamples provides an instance where the deferred-until-stopping benchmark achieves at least $2 - 1/M$, while every valid policy under deferred-to-deadline commitment earns at most $1$. The assumed approximation guarantee then implies that some valid policy achieves deferred-to-deadline utility at least $c(2 - 1/M) > 1$, contradicting the upper bound of $1$.

  1. Assume for contradiction that there is an approximation ratio $c > 1/2$.

    We assume there exists $c > 1/2$ such that for every deadline instance `inst`, some valid adaptive policy $\pi$ achieves expected utility under deferred-to-deadline commitment of at least $c \cdot \operatorname{optDUS}(\text{inst})$.

    line 6

      rintro ⟨c, hc, h_all⟩
  2. There is an integer $M > 1$ such that $c(2 - 1/M) > 1$.

    Because $c > 1/2$, the quantity $c(2 - 1/M)$ approaches $2c > 1$ as $M \to \infty$. By the Archimedean property (`archimedean_ratio_gt_one`), we can choose an integer $M > 1$ large enough to satisfy $c(2 - 1/M) > 1$.

    line 7

    uses archimedean_ratio_gt_one

      obtain ⟨M, hM, hM_ineq⟩ := archimedean_ratio_gt_one c hc
  3. There is an instance where the deferred-until-stopping optimum is at least $2 - 1/M$, but every valid policy achieves deferred-to-deadline utility at most $1$.

    Applying the counterexample constructor `two_box_counterexample_family` to $M$ yields a Pandora's box instance `inst` with deadlines such that $\operatorname{optDUS}(\text{inst}) \ge 2 - 1/M$, while for every valid adaptive policy $\pi$, $\operatorname{expectedUtilityDTD}(\text{inst}, \pi) \le 1$.

    line 8

    uses two_box_counterexample_family

      obtain ⟨inst, h_opt, h_dtd⟩ := two_box_counterexample_family M hM
  4. There is a valid policy $\pi$ for `inst` with $c \cdot \operatorname{optDUS}(\text{inst}) \le \operatorname{expectedUtilityDTD}(\text{inst}, \pi) \le 1$.

    Instantiating the assumed universal approximation guarantee with `inst` produces a valid policy $\pi$ satisfying $c \cdot \operatorname{optDUS}(\text{inst}) \le \operatorname{expectedUtilityDTD}(\text{inst}, \pi)$. Applying the instance's upper bound to $\pi$ shows that $\operatorname{expectedUtilityDTD}(\text{inst}, \pi) \le 1$.

    lines 9–10

      obtain ⟨π, hπ_valid, hπ_bound⟩ := h_all inst
      have h_dtd_le_1 := h_dtd π hπ_valid
  5. $c(2 - 1/M) \le c \cdot \operatorname{optDUS}(\text{inst})$

    Since $c > 1/2$, linear arithmetic gives $c > 0$. Multiplying the lower bound $2 - 1/M \le \operatorname{optDUS}(\text{inst})$ by the positive scalar $c$ preserves the inequality, which non-linear arithmetic verifies.

    lines 11–13

      have hc_pos : 0 < c := by linarith
      have h_c_opt : c * (2 - (1 : ℝ) / M) ≤ c * optDUS inst := by
        nlinarith
  6. Contradiction.

    Chaining the established inequalities yields $1 < c(2 - 1/M) \le c \cdot \operatorname{optDUS}(\text{inst}) \le \operatorname{expectedUtilityDTD}(\text{inst}, \pi) \le 1$, which implies $1 < 1$. Linear arithmetic detects this contradiction and closes the proof.

    line 14

      linarith
archimedean_ratio_gt_one
For any real constant $c > 1/2$, there exists an integer $M > 1$ such that $c(2 - 1/M) > 1$.
two_box_counterexample_family
For any integer $M > 1$, there exists a Pandora's box instance with deadlines whose optimal deferred-until-stopping value is at least $2 - 1/M$, but where every valid policy achieves expected utility at most $1$ under deferred-to-deadline commitment.

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

/-- Refutation of OP-182: deferred-to-deadline policies cannot achieve approximation ratio > 1/2 against deferred-until-stopping OPT. -/theorem op182_deferred_to_deadline_beyond_half_against_stopping :    ¬ (∃ c > (1 / 2 : ℝ), ∀ inst : PandoraDeadlineInstance,        ∃ π : AdaptivePolicy inst.n, IsValidPolicy inst π ∧          c * optDUS inst ≤ expectedUtilityDTD inst π) := by  rintro ⟨c, hc, h_all⟩  obtain ⟨M, hM, hM_ineq⟩ := archimedean_ratio_gt_one c hc  obtain ⟨inst, h_opt, h_dtd⟩ := two_box_counterexample_family M hM  obtain ⟨π, hπ_valid, hπ_bound⟩ := h_all inst  have h_dtd_le_1 := h_dtd π hπ_valid  have hc_pos : 0 < c := by linarith  have h_c_opt : c * (2 - (1 : ℝ) / M) ≤ c * optDUS inst := by    nlinarith  linarith
See it in the graphDiscuss this resultChallenge it