Documentation

Determinize.Proof.Traces.Fibers

Fiber soundness #

FiberSound fiber source target is the invariant of the lockstep argument: the source joint law is the target joint law pushed through the fibers (fiberLift), and almost every target point carries a good fiber (FiberHasMean: a probability measure with a finite mean equal to the point's output). The invariant survives pushing traces forward, mixing over a parameter and countable sums, and it yields the trace factorization (FiberSound.factorization).

Replace the output of a point by a draw from fiber at its trace.

Equations
Instances For
    theorem Determinize.Proof.StepTraces.fiberLift_apply {α : Type u_1} [MeasurableSpace α] (fiber : Paper.SFiniteKernel α ℝ) (point : α × ℝ) :
    (fiberLift fiber).kernel point = MeasureTheory.Measure.map (fun (x : ℝ) => (point.1, x)) (fiber.kernel point.1)

    The fiber at a point is a probability measure with a finite mean, and that mean is the point's output.

    Equations
    Instances For

      source is target pushed through the fibers, and almost every target point carries a good fiber.

      Equations
      • One or more equations did not get rendered due to their size.
      Instances For
        theorem Determinize.Proof.StepTraces.FiberSound.terminal {α : Type u_1} [MeasurableSpace α] (fiber : Paper.SFiniteKernel α ℝ) (trace : α) (output : ℝ) (good : FiberHasMean fiber.kernel (trace, output)) :
        FiberSound fiber (MeasureTheory.Measure.map (fun (value : ℝ) => (trace, value)) (fiber.kernel trace)) (MeasureTheory.Measure.dirac (trace, output))
        def Determinize.Proof.StepTraces.mapTraceOutput {α : Type u_1} {β : Type u_2} (f : α → β) (point : α × ℝ) :
        β × ℝ

        Apply f to the trace component of a point.

        Equations
        Instances For
          theorem Determinize.Proof.StepTraces.FiberSound.mapTrace_ae {α : Type u_1} {β : Type u_2} [MeasurableSpace α] [MeasurableSpace β] (fiber : Paper.SFiniteKernel α ℝ) (nextFiber : Paper.SFiniteKernel β ℝ) (source target : MeasureTheory.Measure (α × ℝ)) (sound : FiberSound fiber source target) (f : α → β) (hf : Measurable f) (sameFiber : ∀ᵐ (point : α × ℝ) ∂target, nextFiber.kernel (f point.1) = fiber.kernel point.1) :
          theorem Determinize.Proof.StepTraces.FiberSound.mapTrace {α : Type u_1} {β : Type u_2} [MeasurableSpace α] [MeasurableSpace β] (fiber : Paper.SFiniteKernel α ℝ) (nextFiber : Paper.SFiniteKernel β ℝ) (source target : MeasureTheory.Measure (α × ℝ)) (sound : FiberSound fiber source target) (f : α → β) (hf : Measurable f) (sameFiber : ∀ (trace : α), nextFiber.kernel (f trace) = fiber.kernel trace) :

          mapTrace_ae with the fibers agreeing everywhere.

          theorem Determinize.Proof.StepTraces.FiberSound.mix {α : Type u_1} {β : Type u_2} [MeasurableSpace α] [MeasurableSpace β] (fiber : Paper.SFiniteKernel α ℝ) (outer : MeasureTheory.Measure β) (source target : ProbabilityTheory.Kernel β (α × ℝ)) (sound : ∀ᵐ (parameter : β) ∂outer, FiberSound fiber (source parameter) (target parameter)) :
          FiberSound fiber (outer.bind ⇑source) (outer.bind ⇑target)
          theorem Determinize.Proof.StepTraces.FiberSound.sum {α : Type u_1} [MeasurableSpace α] {ι : Type u_4} [Countable ι] (fiber : Paper.SFiniteKernel α ℝ) (source target : ι → MeasureTheory.Measure (α × ℝ)) (sound : ∀ (i : ι), FiberSound fiber (source i) (target i)) :
          theorem Determinize.Proof.StepTraces.compProd_eq_bind_pair {α : Type u_1} {β : Type u_2} [MeasurableSpace α] [MeasurableSpace β] (ν : MeasureTheory.Measure α) [MeasureTheory.SFinite ν] (κ : ProbabilityTheory.Kernel α β) [ProbabilityTheory.IsSFiniteKernel κ] :
          ν.compProd κ = ν.bind fun (point : α) => MeasureTheory.Measure.map (fun (value : β) => (point, value)) (κ point)

          Composing a measure with a kernel pairs each point with its draw: Mathlib's μ ⊗ₘ κ = (Kernel.id ×ₖ κ) ∘ₘ μ, pointwise.

          theorem Determinize.Proof.StepTraces.FiberSound.factorization {α : Type u_4} [MeasurableSpace α] {source target : MeasureTheory.Measure (α × ℝ)} {fiber : Paper.SFiniteKernel α ℝ} [ProbabilityTheory.IsMarkovKernel fiber.kernel] (sound : FiberSound fiber source target) (massTarget : target Set.univ ≤ 1) :
          have ν := MeasureTheory.Measure.map Prod.fst target; have output := fun (trace : α) => ∫ (value : ℝ), value ∂fiber.kernel trace; ν Set.univ ≤ 1 ∧ Measurable output ∧ source = ν.compProd fiber.kernel ∧ target = MeasureTheory.Measure.map (fun (t : α) => (t, output t)) ν ∧ ∀ᵐ (t : α) ∂ν, MeasureTheory.Integrable id (fiber.kernel t) ∧ output t = ∫ (r : ℝ), r ∂fiber.kernel t