Documentation

SHSLib.Gardiner.Backward

The backward differential Chapman–Kolmogorov equation (Gardiner §3.6) #

The backward differential Chapman–Kolmogorov equation: §3.6 of Gardiner, "Equations for Time Development in Initial Time — Backward Equations", equations (3.6.1)–(3.6.4).

Notation. G s z stands for p(x, t | z, s), the transition density with a fixed final condition (x, t), regarded as a function of the initial point z and the initial time s. k ds y z stands for p(z, t' + ds | y, t'), the transition density over the short time ds starting from the initial point y at the initial time t', and W z y is the jump rate W(z | y, t') of (3.4.1).

The derivation follows §3.6: the Chapman–Kolmogorov equation turns the difference quotient (3.6.1) into (3.6.2); the continuity assumption of the text replaces the inner time t' + Δt' by t', giving (3.6.3); and then the argument of §3.4.1 — Taylor expansion inside the ball |z - y| < ε, conditions (ii) and (iii) for the first two moments, and condition (i) outside the ball — produces (3.6.4).

References #

The quantities of the backward derivation #

def Gardiner.bwdQuot {n : ℕ} (k : ℝ → Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (ds : ℝ) :

The difference quotient of (3.6.3), tested against a function g: (1/Δt') ∫ dz p(z, t' + Δt' | y, t') [g(z) - g(y)]. For g = p(x, t | ·, t') this is the negative of the quotient appearing in (3.6.3).

Equations
Instances For
    def Gardiner.bQ12 {n : ℕ} (k : ℝ → Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (ds ε : ℝ) :

    The first and second order Taylor contributions of bwdQuot on the ball |z - y| < ε; the analogue of lines 1 and 2 of (3.4.12).

    Equations
    Instances For
      def Gardiner.bQ3 {n : ℕ} (k : ℝ → Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (ds ε : ℝ) :

      The Taylor remainder contribution of bwdQuot on the ball |z - y| < ε; the analogue of line 3 of (3.4.12).

      Equations
      Instances For
        def Gardiner.bQjump {n : ℕ} (k : ℝ → Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (ds ε : ℝ) :

        The contribution of the region |z - y| ≥ ε; the analogue of lines 4–6 of (3.4.12).

        Equations
        Instances For
          def Gardiner.bQsq {n : ℕ} (k : ℝ → Sp n → Sp n → ℝ) (y : Sp n) (ds ε : ℝ) :

          The second moment over the ball, controlling the remainder as in (3.4.14).

          Equations
          Instances For
            def Gardiner.bwdGen {n : ℕ} (A : Sp n → Fin n → ℝ) (B : Sp n → Fin n → Fin n → ℝ) (g : Sp n → ℝ) (y : Sp n) :

            The drift and diffusion part of the backward generator: the first two terms of (3.6.4), with the opposite sign.

            Equations
            Instances For
              def Gardiner.bJumpPre {n : ℕ} (W : Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (ε : ℝ) :

              The jump part of the backward generator, over the region |z - y| ≥ ε.

              Equations
              Instances For
                def Gardiner.bwdJump {n : ℕ} (W : Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (ε : ℝ) :

                The jump term of (3.6.4) itself: ∫ dz W(z | y, t') [g(y) - g(z)], over the region |z - y| ≥ ε.

                Equations
                Instances For
                  theorem Gardiner.bJumpPre_eq_neg_bwdJump {n : ℕ} (W : Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (ε : ℝ) :
                  bJumpPre W g y ε = -bwdJump W g y ε

                  From (3.6.1) to (3.6.2) #

                  theorem Gardiner.backward_ck_step {n : ℕ} (G : ℝ → Sp n → ℝ) (k : ℝ → Sp n → Sp n → ℝ) (y : Sp n) (t' ds : ℝ) (hnorm : ∫ (z : Sp n), k ds y z = 1) (hck : G t' y = ∫ (z : Sp n), k ds y z * G (t' + ds) z) (hkint : MeasureTheory.Integrable (fun (z : Sp n) => k ds y z) MeasureTheory.volume) (hint : MeasureTheory.Integrable (fun (z : Sp n) => k ds y z * G (t' + ds) z) MeasureTheory.volume) :
                  G (t' + ds) y - G t' y = ∫ (z : Sp n), k ds y z * (G (t' + ds) y - G (t' + ds) z)

                  The step (3.6.1) → (3.6.2): by the Chapman–Kolmogorov equation applied to the second term and the normalisation of the transition density in the first, p(x,t|y,t'+Δt') - p(x,t|y,t') is the integral of p(z,t'+Δt'|y,t') [p(x,t|y,t'+Δt') - p(x,t|z,t'+Δt')].

                  The analogue of (3.4.12): splitting into the two regions #

                  theorem Gardiner.bwdQuot_split {n : ℕ} (k : ℝ → Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (ds ε : ℝ) (hInear : MeasureTheory.IntegrableOn (fun (z : Sp n) => taylorPoly g y z * k ds y z) (Metric.ball y ε) MeasureTheory.volume) (hIrem : MeasureTheory.IntegrableOn (fun (z : Sp n) => taylorRem g y z * k ds y z) (Metric.ball y ε) MeasureTheory.volume) (hIfar : MeasureTheory.IntegrableOn (fun (z : Sp n) => (g z - g y) * k ds y z) (Metric.ball y ε)ᶜ MeasureTheory.volume) :
                  bwdQuot k g y ds = bQ12 k g y ds ε + bQ3 k g y ds ε + bQjump k g y ds ε

                  The decomposition of the difference quotient into the Taylor contributions on |z - y| < ε and the jump contribution on |z - y| ≥ ε.

                  The analogue of (3.4.14): the remainder is negligible #

                  theorem Gardiner.bQ3_le {n : ℕ} (k : ℝ → Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (ds ε δ : ℝ) (hds : 0 < ds) (hk0 : ∀ (z : Sp n), 0 ≤ k ds y z) (hrem : ∀ (z : Sp n), ‖z - y‖ < ε → |taylorRem g y z| ≤ δ * ‖z - y‖ ^ 2) (hIrem : MeasureTheory.IntegrableOn (fun (z : Sp n) => taylorRem g y z * k ds y z) (Metric.ball y ε) MeasureTheory.volume) (hIsq : MeasureTheory.IntegrableOn (fun (z : Sp n) => ‖z - y‖ ^ 2 * k ds y z) (Metric.ball y ε) MeasureTheory.volume) :
                  |bQ3 k g y ds ε| ≤ δ * bQsq k y ds ε

                  The remainder contribution is bounded by the second moment times the supremum of the Taylor remainder on the ball.

                  theorem Gardiner.bQsq_nonneg {n : ℕ} (k : ℝ → Sp n → Sp n → ℝ) (y : Sp n) (ds ε : ℝ) (hds : 0 < ds) (hk0 : ∀ (z : Sp n), 0 ≤ k ds y z) :
                  0 ≤ bQsq k y ds ε

                  The first two moments: conditions (ii) and (iii) give the drift and diffusion #

                  theorem Gardiner.bQ12_eq {n : ℕ} (k : ℝ → Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (ds ε : ℝ) (hint1 : ∀ (i : Fin n), MeasureTheory.IntegrableOn (fun (z : Sp n) => (z.ofLp i - y.ofLp i) * k ds y z) (Metric.ball y ε) MeasureTheory.volume) (hint2 : ∀ (i j : Fin n), MeasureTheory.IntegrableOn (fun (z : Sp n) => (z.ofLp i - y.ofLp i) * (z.ofLp j - y.ofLp j) * k ds y z) (Metric.ball y ε) MeasureTheory.volume) :
                  bQ12 k g y ds ε = ∑ i : Fin n, pd g i y * (1 / ds * ∫ (z : Sp n) in Metric.ball y ε, (z.ofLp i - y.ofLp i) * k ds y z) + 1 / 2 * ∑ i : Fin n, ∑ j : Fin n, pd2 g i j y * (1 / ds * ∫ (z : Sp n) in Metric.ball y ε, (z.ofLp i - y.ofLp i) * (z.ofLp j - y.ofLp j) * k ds y z)

                  The Taylor contribution bQ12 is the corresponding combination of the first and second moments of the transition density over the ball.

                  theorem Gardiner.bQ12_tendsto {n : ℕ} {k : ℝ → Sp n → Sp n → ℝ} {A : Sp n → Fin n → ℝ} {B : Sp n → Fin n → Fin n → ℝ} (hA : CondDrift k A) (hB : CondDiffusion k B) (g : Sp n → ℝ) (y : Sp n) (hint1 : ∀ (ds ε : ℝ) (i : Fin n), MeasureTheory.IntegrableOn (fun (z : Sp n) => (z.ofLp i - y.ofLp i) * k ds y z) (Metric.ball y ε) MeasureTheory.volume) (hint2 : ∀ (ds ε : ℝ) (i j : Fin n), MeasureTheory.IntegrableOn (fun (z : Sp n) => (z.ofLp i - y.ofLp i) * (z.ofLp j - y.ofLp j) * k ds y z) (Metric.ball y ε) MeasureTheory.volume) :
                  ∃ (K : ℝ), ∀ ε > 0, ∃ (L : ℝ), Filter.Tendsto (fun (ds : ℝ) => bQ12 k g y ds ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds L) ∧ |L - bwdGen A B g y| ≤ K * ε

                  Conditions (ii) and (iii) of §3.4 give the limit of the Taylor contribution bQ12: it converges, for each ε, to the drift/diffusion part bwdGen of the backward generator up to an error O(ε).

                  The backward generator: the analogue of (3.4.16) #

                  theorem Gardiner.backward_generator {n : ℕ} (k : ℝ → Sp n → Sp n → ℝ) (A : Sp n → Fin n → ℝ) (B : Sp n → Fin n → Fin n → ℝ) (W : Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (L J : ℝ) (e M : ℝ → ℝ) (C : ℝ) (hg : ContDiff ℝ 2 g) (hgs : HasCompactSupport g) (hsplit : ∀ ds > 0, ∀ ε > 0, bwdQuot k g y ds = bQ12 k g y ds ε + bQ3 k g y ds ε + bQjump k g y ds ε) (hQ3le : ∀ ds > 0, ∀ ε > 0, ∀ δ ≥ 0, (∀ (z : Sp n), ‖z - y‖ < ε → |taylorRem g y z| ≤ δ * ‖z - y‖ ^ 2) → |bQ3 k g y ds ε| ≤ δ * bQsq k y ds ε) (hD : Filter.Tendsto (bwdQuot k g y) (nhdsWithin 0 (Set.Ioi 0)) (nhds L)) (h12 : ∀ ε > 0, Filter.Tendsto (fun (ds : ℝ) => bQ12 k g y ds ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds (bwdGen A B g y + e ε))) (he : Filter.Tendsto e (nhdsWithin 0 (Set.Ioi 0)) (nhds 0)) (hsq : ∀ ε > 0, Filter.Tendsto (fun (ds : ℝ) => bQsq k y ds ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds (M ε))) (hM : ∀ ε > 0, |M ε| ≤ C) (hjump : ∀ ε > 0, Filter.Tendsto (fun (ds : ℝ) => bQjump k g y ds ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds (bJumpPre W g y ε))) (hPV : Filter.Tendsto (fun (ε : ℝ) => bJumpPre W g y ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds J)) :
                  L = bwdGen A B g y + J

                  The generator form of the backward equation. The limit L of the difference quotient (3.6.3), tested against a compactly supported C² function g, is the sum of the drift/diffusion part bwdGen and the principal value J of the jump integral.

                  As in §3.4.1, the steps that the text justifies informally appear as hypotheses: hsplit is the splitting into the two regions, hQ3le the estimate of the Taylor remainder, h12 and he the limit of the first two Taylor terms up to O(ε), hsq and hM the boundedness of the second moment, hjump the limit of the contribution of |z - y| ≥ ε given by condition (i), and hPV the existence of the principal value integral.

                  The derivation, assembled #

                  theorem Gardiner.integrableOn_ball_mul_of_bounded {n : ℕ} {f h : Sp n → ℝ} {y : Sp n} {ε C : ℝ} (hf : MeasureTheory.Integrable f MeasureTheory.volume) (hh : Continuous h) (hb : ∀ z ∈ Metric.ball y ε, |h z| ≤ C) :
                  theorem Gardiner.integrableOn_compl_ball_mul_of_bounded {n : ℕ} {f h : Sp n → ℝ} {y : Sp n} {ε C : ℝ} (hf : MeasureTheory.Integrable f MeasureTheory.volume) (hh : Continuous h) (hb : ∀ (z : Sp n), |h z| ≤ C) :
                  theorem Gardiner.backward_generator_of_kernel {n : ℕ} {k : ℝ → Sp n → Sp n → ℝ} {A : Sp n → Fin n → ℝ} {B : Sp n → Fin n → Fin n → ℝ} (W : Sp n → Sp n → ℝ) (g : Sp n → ℝ) (y : Sp n) (L J : ℝ) (M : ℝ → ℝ) (C : ℝ) (hg : ContDiff ℝ 2 g) (hgs : HasCompactSupport g) (hk0 : ∀ (ds : ℝ) (z : Sp n), 0 ≤ k ds y z) (hkint : ∀ (ds : ℝ), MeasureTheory.Integrable (fun (z : Sp n) => k ds y z) MeasureTheory.volume) (hA : CondDrift k A) (hB : CondDiffusion k B) (hD : Filter.Tendsto (bwdQuot k g y) (nhdsWithin 0 (Set.Ioi 0)) (nhds L)) (hsq : ∀ ε > 0, Filter.Tendsto (fun (ds : ℝ) => bQsq k y ds ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds (M ε))) (hM : ∀ ε > 0, |M ε| ≤ C) (hjump : ∀ ε > 0, Filter.Tendsto (fun (ds : ℝ) => bQjump k g y ds ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds (bJumpPre W g y ε))) (hPV : Filter.Tendsto (fun (ε : ℝ) => bJumpPre W g y ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds J)) :
                  L = bwdGen A B g y + J

                  The generator form of the backward equation, assembled from the conditions (ii) and (iii) of §3.4 and the limits taken in the text. Compared with backward_generator, the splitting into the two regions, the estimate of the Taylor remainder and the limit of the first two Taylor terms are now proved, from the non-negativity and integrability of the transition density and from conditions (ii) and (iii).

                  (3.6.4): the backward differential Chapman–Kolmogorov equation #

                  theorem Gardiner.backwardDifferentialCK {n : ℕ} {k : ℝ → Sp n → Sp n → ℝ} {A : Sp n → Fin n → ℝ} {B : Sp n → Fin n → Fin n → ℝ} (G : ℝ → Sp n → ℝ) (W : Sp n → Sp n → ℝ) (y : Sp n) (t' L Jb : ℝ) (M : ℝ → ℝ) (C : ℝ) (hg : ContDiff ℝ 2 (G t')) (hgs : HasCompactSupport (G t')) (hk0 : ∀ (ds : ℝ) (z : Sp n), 0 ≤ k ds y z) (hkint : ∀ (ds : ℝ), MeasureTheory.Integrable (fun (z : Sp n) => k ds y z) MeasureTheory.volume) (hA : CondDrift k A) (hB : CondDiffusion k B) (hD : Filter.Tendsto (bwdQuot k (G t') y) (nhdsWithin 0 (Set.Ioi 0)) (nhds L)) (hsq : ∀ ε > 0, Filter.Tendsto (fun (ds : ℝ) => bQsq k y ds ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds (M ε))) (hM : ∀ ε > 0, |M ε| ≤ C) (hjump : ∀ ε > 0, Filter.Tendsto (fun (ds : ℝ) => bQjump k (G t') y ds ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds (bJumpPre W (G t') y ε))) (hPV : Filter.Tendsto (fun (ε : ℝ) => bwdJump W (G t') y ε) (nhdsWithin 0 (Set.Ioi 0)) (nhds Jb)) (hnorm : ∀ ds > 0, ∫ (z : Sp n), k ds y z = 1) (hck : ∀ ds > 0, G t' y = ∫ (z : Sp n), k ds y z * G (t' + ds) z) (hGint : ∀ ds > 0, MeasureTheory.Integrable (fun (z : Sp n) => k ds y z * G (t' + ds) z) MeasureTheory.volume) (h363 : Filter.Tendsto (fun (ds : ℝ) => (1 / ds * ∫ (z : Sp n), k ds y z * (G (t' + ds) y - G (t' + ds) z)) - 1 / ds * ∫ (z : Sp n), k ds y z * (G t' y - G t' z)) (nhdsWithin 0 (Set.Ioi 0)) (nhds 0)) :
                  Filter.Tendsto (fun (ds : ℝ) => (G (t' + ds) y - G t' y) / ds) (nhdsWithin 0 (Set.Ioi 0)) (nhds (-∑ i : Fin n, A y i * pd (G t') i y - 1 / 2 * ∑ i : Fin n, ∑ j : Fin n, B y i j * pd2 (G t') i j y + Jb))

                  The backward differential Chapman–Kolmogorov equation, (3.6.4).

                  G s z is p(x, t | z, s) and k ds y z is p(z, t' + ds | y, t'). The conclusion is the derivative in the initial time t' of p(x, t | y, t'):

                  ∂p/∂t' = - ∑ᵢ Aᵢ(y,t') ∂p/∂yᵢ - ½ ∑ᵢⱼ Bᵢⱼ(y,t') ∂²p/∂yᵢ∂yⱼ
                             + ∫ dz W(z|y,t') [p(x,t|y,t') - p(x,t|z,t')].
                  

                  The hypotheses hnorm, hck, hGint give the passage (3.6.1) → (3.6.2), h363 is the continuity step (3.6.2) → (3.6.3) of the text (the two difference quotients have the same limit), and the remaining hypotheses are those of backward_generator_of_kernel.