Documentation

SHSLib.Gardiner.Approx.JumpStationary

Stationary distribution of the birth–death approximation of a diffusion #

Gardiner, Handbook of Stochastic Methods, §7.2.1 c), (7.2.19)–(7.2.21).

The birth–death process (7.2.16) on the lattice x = kδ, k ∈ ℕ (reflecting at x = 0), with rates t⁺(x) = A(x)/(2δ) + B(x)/(2δ²) and t⁻(x) = -A(x)/(2δ) + B(x)/(2δ²) (Gardiner.bdRate), has the detailed-balance stationary solution (7.2.19). As δ → 0 its logarithm converges to that of the stationary Fokker–Planck solution (7.2.21), P_s(x) = 𝒩 B(x)⁻¹ exp[2∫₀ˣ A/B].

def Gardiner.bdStat (A B : ℝ → ℝ) (δ P0 : ℝ) (k : ℕ) :

The stationary distribution (7.2.19) of the birth–death approximation (7.2.16), on the lattice x = kδ: P_s(kδ) = P_s(0) ∏_{j<k} [δA(jδ) + B(jδ)] / [-δA((j+1)δ) + B((j+1)δ)].

Equations
Instances For
    theorem Gardiner.bdStat_succ (A B : ℝ → ℝ) (δ P0 : ℝ) (k : ℕ) :
    bdStat A B δ P0 (k + 1) = bdStat A B δ P0 k * ((δ * A (↑k * δ) + B (↑k * δ)) / (-δ * A (↑(k + 1) * δ) + B (↑(k + 1) * δ)))
    theorem Gardiner.bdStat_flux (A B : ℝ → ℝ) {δ : ℝ} (hδ : δ ≠ 0) (P0 : ℝ) (hw : ∀ (j : ℕ), -δ * A (↑j * δ) + B (↑j * δ) ≠ 0) (k : ℕ) :
    bdRate A B δ true (↑k * δ) * bdStat A B δ P0 k = bdRate A B δ false (↑(k + 1) * δ) * bdStat A B δ P0 (k + 1)

    Detailed balance for (7.2.19): the probability flux t⁺(kδ) P_s(k) - t⁻((k+1)δ) P_s(k+1) vanishes.

    theorem Gardiner.bdStat_stationary (A B : ℝ → ℝ) {δ : ℝ} (hδ : δ ≠ 0) (P0 : ℝ) (hw : ∀ (j : ℕ), -δ * A (↑j * δ) + B (↑j * δ) ≠ 0) :
    bdRate A B δ false (↑1 * δ) * bdStat A B δ P0 1 - bdRate A B δ true (↑0 * δ) * bdStat A B δ P0 0 = 0 ∧ ∀ (k : ℕ), bdRate A B δ true (↑k * δ) * bdStat A B δ P0 k + bdRate A B δ false (↑(k + 2) * δ) * bdStat A B δ P0 (k + 2) - (bdRate A B δ true (↑(k + 1) * δ) + bdRate A B δ false (↑(k + 1) * δ)) * bdStat A B δ P0 (k + 1) = 0

    (7.2.19) solves the stationary master equation of the birth–death process (7.2.16) on ℕ (with no death out of k = 0): at k = 0, t⁻(δ)P(1) - t⁺(0)P(0) = 0, and for k ≥ 1, t⁺((k-1)δ)P(k-1) + t⁻((k+1)δ)P(k+1) - [t⁺(kδ) + t⁻(kδ)]P(k) = 0.

    theorem Gardiner.bdStat_eq_ratio (A B : ℝ → ℝ) (δ P0 : ℝ) (hB : ∀ (j : ℕ), B (↑j * δ) ≠ 0) (hu : ∀ (j : ℕ), δ * A (↑j * δ) + B (↑j * δ) ≠ 0) (hw : ∀ (j : ℕ), -δ * A (↑j * δ) + B (↑j * δ) ≠ 0) (k : ℕ) :
    bdStat A B δ P0 k = P0 * ((-δ * A (↑0 * δ) + B (↑0 * δ)) / (δ * A (↑k * δ) + B (↑k * δ))) * ∏ j ∈ Finset.range (k + 1), (1 + δ * A (↑j * δ) / B (↑j * δ)) / (1 - δ * A (↑j * δ) / B (↑j * δ))

    The second line of (7.2.19): P_s(kδ) = P_s(0) [-δA(0) + B(0)]/[δA(kδ) + B(kδ)] ∏_{j=0}^{k} (1 + δA/B)/(1 - δA/B).

    The limit (7.2.20)–(7.2.21) with a rate #

    theorem Gardiner.abs_log_one_sub_le {v : ℝ} (hv : |v| ≤ 1 / 2) :
    |Real.log (1 - v)| ≤ 2 * |v|

    |log(1 - v)| ≤ 2|v| for |v| ≤ 1/2.

    theorem Gardiner.log_ratio_cubic {v : ℝ} (hv : |v| ≤ 1 / 2) :
    |Real.log (1 + v) - Real.log (1 - v) - 2 * v| ≤ 4 * |v| ^ 3

    |log(1 + v) - log(1 - v) - 2v| ≤ 4|v|³ for |v| ≤ 1/2: the step from (7.2.19) to (7.2.20).

    theorem Gardiner.bdStat_log_rate (A B : ℝ → ℝ) {δ P0 a b : ℝ} (L : NNReal) (N : ℕ) (hδ : 0 < δ) (hP0 : 0 < P0) (hb : 0 < b) (hB : ∀ j ≤ N, b ≤ B (↑j * δ)) (hA : ∀ j ≤ N, |A (↑j * δ)| ≤ a) (hsmall : δ * a / b ≤ 1 / 2) (hg : LipschitzOnWith L (fun (z : ℝ) => A z / B z) (Set.Icc 0 (↑N * δ))) :
    |Real.log (bdStat A B δ P0 N) - (Real.log P0 + Real.log (B 0) - Real.log (B (↑N * δ)) + 2 * ∫ (z : ℝ) in 0..↑N * δ, A z / B z)| ≤ δ * (2 * ↑L * (↑N * δ) + 4 * a / b + 4 * (↑N * δ) * δ * (a / b) ^ 3)

    (7.2.20)–(7.2.21) with an explicit rate. Let δ > 0, x = Nδ, B ≥ b > 0 and |A| ≤ a at the grid points of [0, x], δa/b ≤ 1/2, and let A/B be L-Lipschitz on [0, x]. Then the stationary distribution (7.2.19) satisfies |log P_s(x) - [log P_s(0) + log B(0) - log B(x) + 2∫₀ˣ A/B]| ≤ δ (2Lx + 4a/b + 4xδ(a/b)³). In particular P_s(x) → 𝒩 B(x)⁻¹ exp[2∫₀ˣ A/B] as δ → 0, at rate O(δ).

    theorem Gardiner.bdStat_ou_rate {k D δ P0 : ℝ} (N : ℕ) (hk : 0 ≤ k) (hD : 0 < D) (hδ : 0 < δ) (hP0 : 0 < P0) (hsmall : δ * (k * (↑N * δ)) / D ≤ 1 / 2) :
    |Real.log (bdStat (fun (x : ℝ) => -k * x) (fun (x : ℝ) => D) δ P0 N) - (Real.log P0 - k * (↑N * δ) ^ 2 / D)| ≤ δ * (2 * (k / D) * (↑N * δ) + 4 * (k * (↑N * δ)) / D + 4 * (↑N * δ) * δ * (k * (↑N * δ) / D) ^ 3)

    A non-degenerate instance of Gardiner.bdStat_log_rate: the Ornstein–Uhlenbeck process A(x) = -kx, B(x) = D > 0 (for which (7.2.21) is the Gaussian exp(-kx²/D)), on [0, Nδ] with δ ≤ D/(2k(Nδ)).