Theorem 21.2 — finite-horizon Bayesian sequential testing (printed statement)
DisprovedPeskirShiryaev.SequentialTesting.theorem_21_2Consider the finite-horizon Bayesian test of the two fixed drift hypotheses for , with prior , positive error losses , and horizon . Let , , and let be the contact boundaries of the variational value. The infinite-horizon threshold pair satisfies its root equations, and lies in the continuous monotone class before , with both endpoints .
The theorem asserts that
is admissible and attains the posterior value. There is an admissible optimal binary decision whose values satisfy the printed cases: when the stopped posterior equals , or when it equals . Its Bayesian risk equals and is no larger than that of any admissible bounded test.
With and , the boundaries solve
for and , and are the unique solution in the stated boundary class. The theorem also asserts the two Gaussian expectation formulas (21.2.47) and from (21.2.48), for both and all admissible elapsed times. If , then on .
This is the chapter's complete finite-horizon boundary characterization, including its explicit numerical kernels.
Formalization Note This faithful anchor intentionally retains the two documented printed defects: the equality-only decision graph does not cover immediate stopping outside the continuation interval (F1), and the signed transition formula fails for (F2). Consequently the anchor is expected to be disproved. At zero elapsed time the otherwise singular transition expression is assigned its deterministic point-mass value; Gaussian integration uses the standard normal measure. The finite-horizon restriction is the setting of §21.2 despite the theorem's citation of the earlier unrestricted equations.
import Definitions.Def_PeskirShiryaev_SequentialTesting_BoundaryClass open MeasureTheory ProbabilityTheory Set open scoped ENNReal NNReal
namespace PeskirShiryaev.SequentialTesting
/-- Theorem 21.2 as printed, (21.2.44)-(21.2.48). F1 and F2 are deliberately
retained: the two equality-only decision branches and the signed Gaussian
CDF coefficient. This anchor is expected to be false; see the separate variant. -/
theorem theorem_21_2 {Ω : Type*} [MeasurableSpace Ω]
(P0 P1 : Measure Ω) [IsProbabilityMeasure P0] [IsProbabilityMeasure P1]
(θ : Ω → Bool) (hθ : Measurable θ)
(hθ0 : ∀ᵐ ω ∂P0, θ ω = false) (hθ1 : ∀ᵐ ω ∂P1, θ ω = true)
(W : ℝ≥0 → Ω → ℝ) (hWmeas : ∀ u, Measurable (W u))
(hWcont : ∀ ω, Continuous (fun u => W u ω)) (hW0 : ∀ ω, W 0 ω = 0)
(hB0 : IsBrownianReal W P0) (hB1 : IsBrownianReal W P1)
(μ σ a b T : ℝ) (hμ : μ ≠ 0) (hσ : σ ≠ 0)
(ha : 0 < a) (hb : 0 < b) (hT : 0 < T) :
let E : Model Ω :=
{ P0 := P0, P1 := P1, prob0 := inferInstance, prob1 := inferInstance,
θ := θ, θ_measurable := hθ, θ_zero := hθ0, θ_one := hθ1,
W := W, W_measurable := hWmeas, W_continuous := hWcont, W_zero := hW0,
brownian0 := hB0, brownian1 := hB1 }
let C : Parameters :=
{ μ := μ, σ := σ, a := a, b := b, T := T, drift_ne := hμ,
noise_ne := hσ, a_pos := ha, b_pos := hb, horizon_pos := hT }
let g := optimalBoundary E C
ThresholdPair C (lowerThreshold C) (upperThreshold C) ∧
BoundaryClass C g ∧ SolvesIntegralEquations E C g ∧
(∀ h : Fin 2 → ℝ → ℝ, BoundaryClass C h →
SolvesIntegralEquations E C h → ∀ i, EqOn (h i) (g i) (Icc 0 C.T)) ∧
(∀ p ∈ Icc (0 : ℝ) 1,
let τ := exitTime E C g 0 p
Admissible E C C.T τ ∧ stoppingCost E C 0 p τ = value E C 0 p ∧
∃ d : Ω → Bool, AdmissibleDecision E C τ d ∧
PrintedDecision E C g p τ d ∧ bayesRisk E C p τ d = value E C 0 p ∧
∀ (s : Ω → ℝ≥0) (e : Ω → Bool), Admissible E C C.T s →
AdmissibleDecision E C s e → bayesRisk E C p τ d ≤ bayesRisk E C p s e) ∧
(∀ t ∈ Icc (0 : ℝ) C.T, ∀ i : Fin 2,
terminalMean E C (g i t) (C.T - t) = gaussianTerminalMean C (g i t) (C.T - t)) ∧
(∀ t ∈ Icc (0 : ℝ) C.T, ∀ i j : Fin 2, ∀ u ∈ Icc (0 : ℝ) (C.T - t),
transitionCDF E C (g i t) u (g j (t + u)) =
gaussianTransitionCDF (C.μ / C.σ) (g i t) u (g j (t + u))) ∧
(C.a = C.b → ∀ t ∈ Icc (0 : ℝ) C.T, g 1 t = 1 - g 0 t) := by sorry
end PeskirShiryaev.SequentialTesting
Read-back
What the Lean code literally says, in plain math · claude-opus-5-5
The statement assumes the following data and hypotheses:
- Sample space and measures. is a measurable space carrying two probability measures, and .
- Hypothesis indicator. is measurable. It equals false -almost surely and true -almost surely.
- Process. is a real-valued process indexed by the nonnegative reals . Each is measurable, and every path is continuous. Also for every , not just almost every .
- Brownian hypotheses. satisfies the predicate
IsBrownianRealunder and also under . That predicate comes from the imported module and is not shown here. By its name it says is a real Brownian motion under each measure, but its exact content cannot be read from this code. - Real parameters. are real numbers with , , , and .
From these the statement builds a model (the measures, , and the hypotheses above) and a parameter record . It then sets
which is a pair of real functions , indexed by . None of the named notions used below is defined in the visible code; they all come from the imported module. These names are: ThresholdPair, lowerThreshold, upperThreshold, BoundaryClass, SolvesIntegralEquations, exitTime, Admissible, AdmissibleDecision, stoppingCost, value, bayesRisk, PrintedDecision, terminalMean, gaussianTerminalMean, transitionCDF and gaussianTransitionCDF. What follows reports only how they are combined.
The statement asserts that all of the following hold together.
- Thresholds. The pair satisfies .
- Boundary class. belongs to .
- Integral equations. satisfies .
- Uniqueness. Take any pair of real functions on that is in and satisfies . Then for each , for all . Nothing is claimed about outside .
- Optimality. For every , let . Then:
- is .
- .
- There exists a decision with all of the following properties:
- is an .
- satisfies .
- .
- For every random time that is , and every decision that is :
- Terminal mean. For all and :
- Transition CDF. For all , all and all :
- Symmetry. If , then for all .
The declaration's attached comment says it reproduces "Theorem 21.2 as printed, (21.2.44)–(21.2.48)". The comment says it deliberately keeps "two equality-only decision branches and the signed Gaussian CDF coefficient", and states that the declaration "is expected to be false". That remark is the author's comment, not part of the assertion.
Degenerate cases. The hypotheses rule out a zero horizon and zero or , so is an ordinary real number and no division-by-zero default arises in item 7. The interval is never empty. cannot be empty, because it carries probability measures.
- Endpoints of . The boundary values and are included in item 5. At those values, item 5 still demands an admissible , a decision attaining the value, and minimality over all admissible pairs.
- Endpoint . At , item 6 compares the two terminal-mean quantities with remaining time . Item 7 then allows only , comparing the two CDFs at time increment with and .
- Increment in general. For any , is allowed in item 7, which compares with the Gaussian formula at increment .
- Outside . and are constrained off only through whatever requires.
- Junk values inside imported definitions. Whether any imported definition hits a total-function default cannot be determined from this code. Examples would be an integral of a non-integrable function, a supremum of an empty or unbounded set, or an infimum defining an exit time that never occurs.
- Joint satisfiability. The hypotheses do not obviously contradict each other. They require to separate from almost surely, while satisfies the same Brownian predicate under both measures.
- When . Item 8 holds vacuously.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.