Documentation

SHSLib.Foundations.DavisCh2StrongMarkov

The strong Markov property for arbitrary stopping times (Davis §25.5) #

SHSLib/Foundations/DavisCh2.lean proves the strong Markov property for stopping times with countably many values (Davis.pdp_strong_markov_of_countable_range). This file removes the countability restriction by the classical dyadic discretisation

τₙ := (⌊τ·2ⁿ⌋ + 1)/2ⁿ  ↓  τ,      τ < τₙ ≤ τ + 2⁻ⁿ,

each τₙ being a stopping time with countably many values, and then passing to the limit using right continuity of the paths and continuity of f and of Pₛf.

The file contains three groups of results.

References #

The n-th dyadic upper approximation of t : ℝ≥0: dyadicUpper n t = (⌊t·2ⁿ⌋ + 1)/2ⁿ, the smallest dyadic rational of level n that is strictly larger than t.

Equations
Instances For
    theorem Davis.lt_dyadicUpper (n : ℕ) (t : NNReal) :

    The dyadic upper approximation is strictly larger than the point it approximates.

    theorem Davis.dyadicUpper_le_add (n : ℕ) (t : NNReal) :
    dyadicUpper n t ≤ t + (2 ^ n)⁻¹

    The dyadic upper approximation overshoots by at most 2⁻ⁿ.

    theorem Davis.dyadicUpper_le_iff (n : ℕ) (t u : NNReal) :
    dyadicUpper n t ≤ u ↔ t < ↑⌊u * 2 ^ n⌋₊ / 2 ^ n

    Comparison of dyadicUpper n t with an arbitrary threshold u.

    theorem Davis.dyadicFloor_le (n : ℕ) (u : NNReal) :
    ↑⌊u * 2 ^ n⌋₊ / 2 ^ n ≤ u

    The dyadic lower rounding of u never exceeds u.

    theorem Davis.dyadicUpper_mem_range (n : ℕ) (t : NNReal) :
    dyadicUpper n t ∈ Set.range fun (k : ℕ) => (↑k + 1) / 2 ^ n

    dyadicUpper n takes values in the countable set of level-n dyadic rationals.

    theorem Davis.countable_dyadicRange (n : ℕ) :
    (Set.range fun (k : ℕ) => (↑k + 1) / 2 ^ n).Countable

    dyadicUpper n t → t as n → ∞.

    dyadicUpper n t → t from the right: the approximations converge to t inside Set.Ioi t.

    theorem Davis.isStoppingTime_dyadicUpper {Omega : Type u_1} [m0 : MeasurableSpace Omega] {ℱ : MeasureTheory.Filtration NNReal m0} {tau : Omega → NNReal} (htau : MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega)) (n : ℕ) :
    MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(dyadicUpper n (tau omega))

    The dyadic upper approximation of a stopping time is a stopping time.

    theorem Davis.isStoppingTime_add_const {Omega : Type u_1} [m0 : MeasurableSpace Omega] {ℱ : MeasureTheory.Filtration NNReal m0} {tau : Omega → NNReal} (htau : MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega)) (s : NNReal) :
    MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega + s)

    Shifting an ℝ≥0-valued stopping time by a constant gives a stopping time.

    theorem Davis.measurable_apply_stoppingTime_of_countable_range {Omega : Type u_1} [m0 : MeasurableSpace Omega] {ℱ : MeasureTheory.Filtration NNReal m0} {tau : Omega → NNReal} (htau : MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega)) (S : Set NNReal) (hS : S.Countable) (hrange : ∀ (omega : Omega), tau omega ∈ S) {beta : Type u_2} [MeasurableSpace beta] (h : NNReal → Omega → beta) (hmeas : ∀ (u : NNReal), Measurable (h u)) :
    Measurable fun (omega : Omega) => h (tau omega) omega

    If τ is a stopping time with countably many values and h u is ℱ u-measurable for every u, then ω ↦ h (τ ω) ω is ℱ_τ-measurable.

    theorem Davis.measurable_apply_stoppingTime_of_progMeasurable {Omega : Type u_1} {E : Type u_2} [m0 : MeasurableSpace Omega] [MeasurableSpace E] [MetricSpace E] [BorelSpace E] {ℱ : MeasureTheory.Filtration NNReal m0} {X : NNReal → Omega → E} {tau : Omega → NNReal} (hX_prog : MeasureTheory.ProgMeasurable ℱ X) (htau : MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega)) :
    Measurable fun (omega : Omega) => X (tau omega) omega

    For a progressively measurable process, the value at a stopping time is ℱ_τ-measurable.

    theorem Davis.measurable_uncurry_of_countable_range {alpha : Type u_1} {beta : Type u_2} {gamma : Type u_3} [MeasurableSpace alpha] [MeasurableSpace beta] [MeasurableSpace gamma] {r : alpha → NNReal} (hr : Measurable r) (S : Set NNReal) (hS : S.Countable) (hrange : ∀ (a : alpha), r a ∈ S) {Y : NNReal → beta → gamma} (hY : ∀ v ∈ S, Measurable (Y v)) :
    Measurable fun (p : alpha × beta) => Y (r p.1) p.2

    Countable-valued time change. If r takes countably many values and Y v is measurable for each of them, then (a, b) ↦ Y (r a) b is measurable for the product σ-algebra.

    theorem Davis.progMeasurable_of_rightContinuous {Omega : Type u_1} {E : Type u_2} [m0 : MeasurableSpace Omega] [MeasurableSpace E] [MetricSpace E] [BorelSpace E] [SecondCountableTopology E] {ℱ : MeasureTheory.Filtration NNReal m0} {X : NNReal → Omega → E} (hX_adapted : ∀ (u : NNReal), Measurable (X u)) (hX_rc : ∀ (omega : Omega) (t : NNReal), Filter.Tendsto (fun (r : NNReal) => X r omega) (nhdsWithin t (Set.Ioi t)) (nhds (X t omega))) :

    A right-continuous adapted process is progressively measurable.

    The dyadic right-endpoint approximations Xⁿ(t, ω) = X(min (dyadicUpper n t) i, ω) are measurable on Set.Iic i × Ω (they take countably many time values, all ≤ i) and converge pointwise to X(t, ω) by right continuity of the paths.

    Mathlib provides the analogous statement for continuous processes (StronglyAdapted.progMeasurable_of_continuous); this is the càdlàg version, which is the one needed for PDPs.

    theorem Davis.pdp_strong_markov_of_rightContinuous {Omega : Type u_1} {E : Type u_2} [MeasurableSpace Omega] [MeasurableSpace E] [MetricSpace E] [BorelSpace E] {P : MeasureTheory.Measure Omega} [MeasureTheory.IsProbabilityMeasure P] (pdp : PDP E) {X : NNReal → Omega → E} {ℱ : MeasureTheory.Filtration NNReal inst✝} (hX_prog : MeasureTheory.ProgMeasurable ℱ X) (hX_rc : ∀ (omega : Omega) (t : NNReal), Filter.Tendsto (fun (r : NNReal) => X r omega) (nhdsWithin t (Set.Ioi t)) (nhds (X t omega))) (hmarkov : IsPDPMarkovProcess P pdp ℱ X) {tau : Omega → NNReal} (htau : MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega)) (htau_meas : Measurable tau) {f : E → ℝ} (hf_cont : Continuous f) {C : ℝ} (hf_bdd : ∀ (x : E), |f x| ≤ C) (s : NNReal) (hg_cont : Continuous fun (x : E) => pdpSemigroup pdp s f x) (hg_bdd : ∀ (x : E), |pdpSemigroup pdp s f x| ≤ C) :
    P[fun (omega : Omega) => f (X (tau omega + s) omega) | htau.measurableSpace] =ᵐ[P] fun (omega : Omega) => pdpSemigroup pdp s f (X (tau omega) omega)

    (25.5) Strong Markov property, general stopping times.

    If X is right-continuous, progressively measurable and satisfies the simple Markov property IsPDPMarkovProcess for the PDP semigroup, then for every stopping time τ and every bounded continuous f whose semigroup image Pₛ f is continuous,

    𝔼[f(X_{τ+s}) | ℱ_τ] = (Pₛ f)(X_τ) almost surely.

    This removes the countable-range restriction in pdp_strong_markov_of_countable_range by dyadic discretisation of τ.

    theorem Davis.pdp_strong_markov_of_rightContinuous_adapted {Omega : Type u_1} {E : Type u_2} [MeasurableSpace Omega] [MeasurableSpace E] [MetricSpace E] [BorelSpace E] [SecondCountableTopology E] {P : MeasureTheory.Measure Omega} [MeasureTheory.IsProbabilityMeasure P] (pdp : PDP E) {X : NNReal → Omega → E} {ℱ : MeasureTheory.Filtration NNReal inst✝} (hX_adapted : ∀ (u : NNReal), Measurable (X u)) (hX_rc : ∀ (omega : Omega) (t : NNReal), Filter.Tendsto (fun (r : NNReal) => X r omega) (nhdsWithin t (Set.Ioi t)) (nhds (X t omega))) (hmarkov : IsPDPMarkovProcess P pdp ℱ X) {tau : Omega → NNReal} (htau : MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega)) (htau_meas : Measurable tau) {f : E → ℝ} (hf_cont : Continuous f) {C : ℝ} (hf_bdd : ∀ (x : E), |f x| ≤ C) (s : NNReal) (hg_cont : Continuous fun (x : E) => pdpSemigroup pdp s f x) (hg_bdd : ∀ (x : E), |pdpSemigroup pdp s f x| ≤ C) :
    P[fun (omega : Omega) => f (X (tau omega + s) omega) | htau.measurableSpace] =ᵐ[P] fun (omega : Omega) => pdpSemigroup pdp s f (X (tau omega) omega)

    (25.5) Strong Markov property from adaptedness and right continuity.

    Same as pdp_strong_markov_of_rightContinuous, with progressive measurability replaced by the more primitive hypothesis that X is adapted (it is then progressively measurable by progMeasurable_of_rightContinuous).

    theorem Davis.pdp_strong_markov_of_feller {Omega : Type u_1} {E : Type u_2} [MeasurableSpace Omega] [MeasurableSpace E] [MetricSpace E] [BorelSpace E] [SecondCountableTopology E] {P : MeasureTheory.Measure Omega} [MeasureTheory.IsProbabilityMeasure P] (pdp : PDPStandard E) (hrate_cont : Continuous fun (x : E) => ↑(pdp.jumpRate x)) (c : NNReal) (hrate : ∀ (y : E), pdp.jumpRate y ≤ c) (hflow : Measurable fun (q : ℝ × E) => pdp.flow q.1.toNNReal q.2) (hflow_cont : Continuous fun (q : NNReal × E) => pdp.flow q.1 q.2) (hcont : ∀ (y : E), Continuous fun (s : ℝ) => ↑(pdp.jumpRate (pdp.flow s.toNNReal y))) (hQ_cont : ∀ (f : E → ℝ), Continuous f → (∃ (C : ℝ), ∀ (x : E), |f x| ≤ C) → Continuous fun (x : E) => ∫ (y : E), f y ∂pdp.transitionMeasure x) {X : NNReal → Omega → E} {ℱ : MeasureTheory.Filtration NNReal inst✝} (hX_adapted : ∀ (u : NNReal), Measurable (X u)) (hX_rc : ∀ (omega : Omega) (t : NNReal), Filter.Tendsto (fun (r : NNReal) => X r omega) (nhdsWithin t (Set.Ioi t)) (nhds (X t omega))) (hmarkov : IsPDPMarkovProcess P pdp.toPDP ℱ X) {tau : Omega → NNReal} (htau : MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega)) (htau_meas : Measurable tau) {f : E → ℝ} (hf_cont : Continuous f) {C : ℝ} (hf_bdd : ∀ (x : E), |f x| ≤ C) (s : NNReal) :
    P[fun (omega : Omega) => f (X (tau omega + s) omega) | htau.measurableSpace] =ᵐ[P] fun (omega : Omega) => pdpSemigroup pdp.toPDP s f (X (tau omega) omega)

    (25.5) Strong Markov property for a Feller PDP.

    For a standard PDP satisfying the §27 regularity conditions (bounded continuous jump rate, continuous flow, Q preserving continuity), the continuity and boundedness of Pₛ f are automatic — they are the Feller property (pdp_feller_property) and pdpSemigroup_abs_le. So the strong Markov property holds for every stopping time and every bounded continuous f.

    theorem Davis.pdp_strong_markov_of_countable_range_adapted {Omega : Type u_1} {E : Type u_2} [MeasurableSpace Omega] [MeasurableSpace E] {P : MeasureTheory.Measure Omega} [MeasureTheory.IsProbabilityMeasure P] (pdp : PDP E) {X : NNReal → Omega → E} {ℱ : MeasureTheory.Filtration NNReal inst✝} (hX_adapted : ∀ (u : NNReal), Measurable (X u)) (hmarkov : IsPDPMarkovProcess P pdp ℱ X) {tau : Omega → NNReal} (htau : MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega)) (htau_meas : Measurable tau) (S : Set NNReal) (hS : S.Countable) (hrange : ∀ (omega : Omega), tau omega ∈ S) {f : E → ℝ} (hf : Measurable f) {C : ℝ} (hf_bdd : ∀ (x : E), |f x| ≤ C) (s : NNReal) (hg : Measurable fun (x : E) => pdpSemigroup pdp s f x) (hg_bdd : ∀ (x : E), |pdpSemigroup pdp s f x| ≤ C) :
    P[fun (omega : Omega) => f (X (tau omega + s) omega) | htau.measurableSpace] =ᵐ[P] fun (omega : Omega) => pdpSemigroup pdp s f (X (tau omega) omega)

    Countable-range strong Markov property, with the technical hypotheses discharged. A restatement of pdp_strong_markov_of_countable_range in which the ℱ_τ-measurability and integrability of (Pₛ f)(X_τ) are derived from adaptedness of X and boundedness of f (via measurable_apply_stoppingTime_of_countable_range) instead of assumed.

    theorem Davis.pdp_expectation_strong_markov {Omega : Type u_1} {E : Type u_2} [MeasurableSpace Omega] [MeasurableSpace E] [MetricSpace E] [BorelSpace E] {P : MeasureTheory.Measure Omega} [MeasureTheory.IsProbabilityMeasure P] (pdp : PDP E) {X : NNReal → Omega → E} {ℱ : MeasureTheory.Filtration NNReal inst✝} (hX_prog : MeasureTheory.ProgMeasurable ℱ X) (hX_rc : ∀ (omega : Omega) (t : NNReal), Filter.Tendsto (fun (r : NNReal) => X r omega) (nhdsWithin t (Set.Ioi t)) (nhds (X t omega))) (hmarkov : IsPDPMarkovProcess P pdp ℱ X) {tau : Omega → NNReal} (htau : MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega)) (htau_meas : Measurable tau) {f : E → ℝ} (hf_cont : Continuous f) {C : ℝ} (hf_bdd : ∀ (x : E), |f x| ≤ C) (s : NNReal) (hg_cont : Continuous fun (x : E) => pdpSemigroup pdp s f x) (hg_bdd : ∀ (x : E), |pdpSemigroup pdp s f x| ≤ C) :
    ∫ (omega : Omega), f (X (tau omega + s) omega) ∂P = ∫ (omega : Omega), pdpSemigroup pdp s f (X (tau omega) omega) ∂P

    Expectation form of the strong Markov property. Taking expectations in pdp_strong_markov_of_rightContinuous gives 𝔼[f(X_{τ+s})] = 𝔼[(Pₛ f)(X_τ)].

    A worked instance: the jump-free (deterministic) PDP #

    All the hypotheses of the general strong Markov theorem are simultaneously satisfiable: for a PDP with vanishing jump rate the semigroup is composition with the flow, the flow process is a PDP-Markov process, and the strong Markov property at an arbitrary stopping time is the flow (semigroup) identity φ(τ + s, z) = φ(s, φ(τ, z)).

    theorem Davis.pdpSurvivorFunction_of_jumpRate_zero {E : Type u_1} [MeasurableSpace E] (p : PDP E) (hzero : ∀ (x : E), p.jumpRate x = 0) (t : NNReal) (x : E) :

    With no jumps the survivor function is identically 1.

    theorem Davis.pdpSemigroupOp_of_jumpRate_zero {E : Type u_1} [MeasurableSpace E] (p : PDP E) (hzero : ∀ (x : E), p.jumpRate x = 0) (f : E → ℝ) (psi : NNReal × E → ℝ) (q : NNReal × E) :
    pdpSemigroupOp p f psi q = f (p.flow q.1 q.2)

    With no jumps the renewal operator collapses to composition with the flow.

    theorem Davis.pdpSemigroup_of_jumpRate_zero {E : Type u_1} [MeasurableSpace E] (p : PDP E) (hzero : ∀ (x : E), p.jumpRate x = 0) (t : NNReal) (f : E → ℝ) (x : E) :
    pdpSemigroup p t f x = f (p.flow t x)

    The semigroup of a jump-free PDP is composition with the flow: Pₜ f (x) = f(φ(t, x)).

    theorem Davis.flowProcess_isPDPMarkovProcess {E : Type u_1} [MeasurableSpace E] {Omega : Type u_2} [MeasurableSpace Omega] {P : MeasureTheory.Measure Omega} [MeasureTheory.IsProbabilityMeasure P] (p : PDP E) (hzero : ∀ (x : E), p.jumpRate x = 0) {ℱ : MeasureTheory.Filtration NNReal inst✝} {Z : Omega → E} (hadapt : ∀ (t : NNReal), Measurable fun (omega : Omega) => p.flow t (Z omega)) :
    IsPDPMarkovProcess P p ℱ fun (t : NNReal) (omega : Omega) => p.flow t (Z omega)

    The flow process of a jump-free PDP is a PDP-Markov process.

    Its increments are deterministic, so the conditional expectation is just the ℱ t-measurable value f(φ(s, φ(t, Z))), which is exactly (Pₛ f)(Xₜ).

    theorem Davis.flowProcess_strong_markov {E : Type u_1} [MeasurableSpace E] {Omega : Type u_2} [MeasurableSpace Omega] [MetricSpace E] [BorelSpace E] [SecondCountableTopology E] {P : MeasureTheory.Measure Omega} [MeasureTheory.IsProbabilityMeasure P] (p : PDP E) (hzero : ∀ (x : E), p.jumpRate x = 0) (hflow_time : ∀ (x : E), Continuous fun (t : NNReal) => p.flow t x) (hflow_state : ∀ (t : NNReal), Continuous (p.flow t)) {ℱ : MeasureTheory.Filtration NNReal inst✝} {Z : Omega → E} (hadapt : ∀ (t : NNReal), Measurable fun (omega : Omega) => p.flow t (Z omega)) {tau : Omega → NNReal} (htau : MeasureTheory.IsStoppingTime ℱ fun (omega : Omega) => ↑(tau omega)) (htau_meas : Measurable tau) {f : E → ℝ} (hf_cont : Continuous f) {C : ℝ} (hf_bdd : ∀ (x : E), |f x| ≤ C) (s : NNReal) :
    P[fun (omega : Omega) => f (p.flow (tau omega + s) (Z omega)) | htau.measurableSpace] =ᵐ[P] fun (omega : Omega) => f (p.flow s (p.flow (tau omega) (Z omega)))

    Strong Markov property of the deterministic flow process at an arbitrary stopping time. A concrete, hypothesis-free-in-spirit instance of pdp_strong_markov_of_rightContinuous_adapted: it shows that the hypotheses of the general theorem are simultaneously satisfiable.

    theorem Davis.isStoppingTime_const_filtration {Omega : Type u_2} [m0 : MeasurableSpace Omega] {tau : Omega → NNReal} (htau_meas : Measurable tau) :
    MeasureTheory.IsStoppingTime (MeasureTheory.Filtration.const NNReal m0 ⋯) fun (omega : Omega) => ↑(tau omega)

    Every measurable ℝ≥0-valued time is a stopping time for the constant filtration.

    A concrete jump-free PDP on ℝ: the uniform translation flow φ(t, x) = x + t, zero jump rate and (never used) unit-shift jump kernel.

    Equations
    • One or more equations did not get rendered due to their size.
    Instances For
      theorem Davis.rampPDP_strong_markov {Omega : Type u_2} [m0 : MeasurableSpace Omega] {P : MeasureTheory.Measure Omega} [MeasureTheory.IsProbabilityMeasure P] {Z : Omega → ℝ} (hZ : Measurable Z) {tau : Omega → NNReal} (htau_meas : Measurable tau) {f : ℝ → ℝ} (hf_cont : Continuous f) {C : ℝ} (hf_bdd : ∀ (x : ℝ), |f x| ≤ C) (s : NNReal) :
      P[fun (omega : Omega) => f (Z omega + ↑(tau omega + s)) | ⋯.measurableSpace] =ᵐ[P] fun (omega : Omega) => f (Z omega + ↑(tau omega) + ↑s)

      A fully concrete instance of the general strong Markov property.

      For the translation flow x ↦ x + t on ℝ, an arbitrary random initial state Z, the constant filtration and an arbitrary measurable random time τ, 𝔼[f(Z + τ + s) | ℱ_τ] = f((Z + τ) + s) almost surely. This certifies that the hypotheses of pdp_strong_markov_of_rightContinuous_adapted are simultaneously satisfiable.