theorem
Determinize.Proof.normalized_probability
{α : Type u_1}
[MeasurableSpace α]
(μ : MeasureTheory.Measure α)
[MeasureTheory.IsFiniteMeasure μ]
(positive : μ Set.univ ≠ 0)
:
theorem
Determinize.Proof.normalized_variance
(μ : MeasureTheory.Measure ℝ)
[MeasureTheory.IsFiniteMeasure μ]
(positive : μ Set.univ ≠ 0)
(moment : MeasureTheory.MemLp id 2 μ)
:
theorem
Determinize.Proof.Paper.returned_expectation_soundness
(program : Spec.Paper.Expr)
(typed : Spec.Paper.Typed [] program (Spec.Paper.Ty.float Spec.Paper.Affinity.E))
(safe : Spec.Paper.DomainSafe program)
(positive : 0 < (Spec.Paper.bigStepMeasure program) Set.univ)
(integrable : MeasureTheory.Integrable id (Spec.Paper.bigStepMeasure program))
:
0 < (Spec.Paper.bigStepMeasure program.determinize) Set.univ ∧ MeasureTheory.IsProbabilityMeasure (Spec.returnedLaw program) ∧ MeasureTheory.IsProbabilityMeasure (Spec.returnedLaw program.determinize) ∧ MeasureTheory.Integrable id (Spec.returnedLaw program) ∧ MeasureTheory.Integrable id (Spec.returnedLaw program.determinize) ∧ ∫ (x : ℝ), x ∂Spec.returnedLaw program.determinize = ∫ (x : ℝ), x ∂Spec.returnedLaw program
theorem
Determinize.Proof.Paper.conditional_variance_soundness
(program : Spec.Paper.Expr)
(typed : Spec.Paper.Typed [] program (Spec.Paper.Ty.float Spec.Paper.Affinity.E))
(safe : Spec.Paper.DomainSafe program)
(positive : 0 < (Spec.Paper.bigStepMeasure program) Set.univ)
(moment : MeasureTheory.MemLp id 2 (Spec.Paper.bigStepMeasure program))
:
0 < (Spec.Paper.bigStepMeasure program.determinize) Set.univ ∧ MeasureTheory.MemLp id 2 (Spec.Paper.bigStepMeasure program.determinize) ∧ ProbabilityTheory.variance id (Spec.returnedLaw program.determinize) ≤ ProbabilityTheory.variance id (Spec.returnedLaw program)
theorem
Determinize.Proof.Paper.conditional_extended_expectation_soundness
(program : Spec.Paper.Expr)
(typed : Spec.Paper.Typed [] program (Spec.Paper.Ty.float Spec.Paper.Affinity.E))
(safe : Spec.Paper.DomainSafe program)
(positive : 0 < (Spec.Paper.bigStepMeasure program) Set.univ)
(defined : Spec.HasExpectation (Spec.Paper.bigStepMeasure program))
:
0 < (Spec.Paper.bigStepMeasure program.determinize) Set.univ ∧ Spec.HasExpectation (Spec.Paper.bigStepMeasure program.determinize) ∧ (↑((Spec.Paper.bigStepMeasure program.determinize) Set.univ).toReal)⁻¹ * Spec.extendedExpectation (Spec.Paper.bigStepMeasure program.determinize) = (↑((Spec.Paper.bigStepMeasure program) Set.univ).toReal)⁻¹ * Spec.extendedExpectation (Spec.Paper.bigStepMeasure program)
theorem
Determinize.Proof.Paper.probability_preservation
(program : Spec.Paper.Expr)
(typed : Spec.Paper.Typed [] program (Spec.Paper.Ty.float Spec.Paper.Affinity.E))
(safe : Spec.Paper.DomainSafe program)
:
Spec.Paper.DomainSafe program.determinize ∧ (Spec.Paper.bigStepMeasure program) Set.univ + Spec.Paper.divergenceProbability program = 1 ∧ (Spec.Paper.bigStepMeasure program.determinize) Set.univ = (Spec.Paper.bigStepMeasure program) Set.univ ∧ Spec.Paper.divergenceProbability program.determinize = Spec.Paper.divergenceProbability program
Probability preservation: the target is domain-safe, the source returns or diverges, and both return, and diverge, with the same probabilities.
theorem
Determinize.Proof.Traces.conditional_variance_soundness
(program : Spec.Paper.Expr)
(typed : Spec.Paper.Typed [] program (Spec.Paper.Ty.float Spec.Paper.Affinity.E))
(safe : Spec.Paper.DomainSafe program)
(positive : 0 < (Spec.Paper.bigStepMeasure program) Set.univ)
(moment : MeasureTheory.MemLp id 2 (Spec.Paper.bigStepMeasure program))
:
MeasureTheory.Integrable
(fun (trace : Spec.Traces.Trace) =>
ProbabilityTheory.variance id ((Spec.Traces.traceAndOutputLaw program).condKernel trace))
(((Spec.Paper.bigStepMeasure program) Set.univ)⁻¹ • Spec.Traces.traceLaw program) ∧ ProbabilityTheory.variance id (Spec.returnedLaw program) = ProbabilityTheory.variance id (Spec.returnedLaw program.determinize) + ∫ (trace : Spec.Traces.Trace), ProbabilityTheory.variance id
((Spec.Traces.traceAndOutputLaw program).condKernel
trace) ∂((Spec.Paper.bigStepMeasure program) Set.univ)⁻¹ • Spec.Traces.traceLaw program
theorem
Determinize.Proof.Traces.total_masses
(program : Spec.Paper.Expr)
:
(Spec.Traces.traceLaw program) Set.univ = (Spec.Traces.traceAndOutputLaw program) Set.univ ∧ (Spec.Traces.traceAndOutputLaw program) Set.univ = (Spec.Paper.bigStepMeasure program) Set.univ
Both marginals of the joint law have the joint law's total mass, the return probability.
theorem
Determinize.Proof.Traces.lintegral_along_traces
(program : Spec.Paper.Expr)
{f : ℝ → ENNReal}
(measurable : Measurable f)
:
∫⁻ (value : ℝ), f value ∂Spec.Paper.bigStepMeasure program = ∫⁻ (trace : Spec.Traces.Trace), ∫⁻ (value : ℝ), f value ∂(Spec.Traces.traceAndOutputLaw program).condKernel trace ∂Spec.Traces.traceLaw program
A nonnegative integral over the output law, taken first over the conditional law given the trace and then over the trace law.
theorem
Determinize.Proof.Traces.returned_expectation_along_traces
(program : Spec.Paper.Expr)
:
Spec.returnedExpectation program = (↑(Spec.returnProbability program).toReal)⁻¹ * (↑(∫⁻ (trace : Spec.Traces.Trace), Spec.posPartIntegral
((Spec.Traces.traceAndOutputLaw program).condKernel trace) ∂Spec.Traces.traceLaw program) - ↑(∫⁻ (trace : Spec.Traces.Trace), Spec.negPartIntegral
((Spec.Traces.traceAndOutputLaw program).condKernel trace) ∂Spec.Traces.traceLaw program))
The expectation conditioned on returning, with the positive and negative parts of the output each integrated first over the conditional law given the trace and then over the trace law.