Proof of Theorem 4, p.4 — for ,
ProvedPhysReaudit.multiplier_ge_tp2o-gran-per-chapterp2o-plan-paperp2o-v1physical-auditingsequential-testingstatistics
Let and , and let be i.i.d. Bernoulli with and . Then for every , almost surely,
So has multipliers at most one under every harmful parameter, which is what makes Ville's inequality bound false acceptance of the SPRT.
Differences from the source:
- encoding: q₀ = (b − ε)/(κp), t = b/(κp) with 0 < ε < b < κp → any reals 0 < q₀ < t < 1; reason: exactly the pairs that arise (κ = a = 1, b = t, ε = t − q₀ realise any such pair), so the same instances
- encoding: the SPRT's target observations Y_1, Y_2, … and F_n = σ(Y_1, …, Y_n) → the coordinates y 0, y 1, … of the i.i.d. Bernoulli(q) target tape (Measure.infinitePi of bern q) and pastFiltration n (coordinates < n); reason: the SPRT reads the target tape in order, so its i-th observation is target entry i − 1
Preamble
import Mathlib import Definitions.Def_PhysReaudit_Delta import Definitions.Def_PhysReaudit_lawQ import Definitions.Def_PhysReaudit_sprtS open MeasureTheory
Formal statement
namespace PhysReaudit
/-- **The likelihood-ratio multiplier for `q ≥ t`** (Q. Li, *When Can Physical Evidence Survive a World-Model Update?*, review revision of 3 November 2026 (sha256 a20c437b…), proof of Theorem 4, p.4 (PDF p.5), display
"Eq[e^{Sn+1−Sn} | Fn] = 1 + (q − t)(q0 − t)/(t(1 − t)) ≤ 1").
Let `0 < q₀ < t < 1` and `q ∈ [t, 1]`. Under the i.i.d. Bernoulli(q) target tape, for every `n`,
the conditional expectation of `e^{S_{n+1} − S_n}` given `F_n = σ(Y_1, …, Y_n)` is almost surely the constant
`1 + (q − t)(q₀ − t)/(t(1 − t))`, and this constant is at most `1`. -/
theorem multiplier_ge_t {q₀ t : ℝ} (h0 : 0 < q₀) (h1 : q₀ < t) (h2 : t < 1) {q : ℝ}
(hq : q ∈ Set.Icc t 1) (n : ℕ) :
(Measure.infinitePi fun _ : ℕ => bern q)[fun y => Real.exp (sprtS q₀ t y (n + 1) - sprtS q₀ t y n)
| pastFiltration n]
=ᵐ[Measure.infinitePi fun _ : ℕ => bern q] (fun _ => 1 + (q - t) * (q₀ - t) / (t * (1 - t))) ∧
1 + (q - t) * (q₀ - t) / (t * (1 - t)) ≤ 1 := by sorry
end PhysReaudit
Source
Q. Li, When Can Physical Evidence Survive a World-Model Update? Finite-cost re-auditing and reusable policy certificates, review revision of 3 November 2026 (STOC 2027 submission, sha256 a20c437b34115b84982339e9a6fa319cb3a66f66d9d4c474465cbb954c79093d), proof of Theorem 4, p.4 (PDF p.5), display "Eq[e^{Sn+1−Sn} | Fn] = 1 + (q − t)(q0 − t)/(t(1 − t)) ≤ 1"
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.