Documentation

SHSLib.Gardiner.Approx.SmallNoiseStationary

Asymptotic method for stationary distributions #

Gardiner, Handbook of Stochastic Methods, §6.3.3, (6.3.52)–(6.3.56).

For the Fokker–Planck equation (6.3.52) with small diffusion ε² B, the ansatz (6.3.53) p_s(x) = exp[-φ(x)/ε²] turns the stationary equation into (6.3.54), whose leading order in ε is the Hamilton–Jacobi equation (6.3.56).

Erratum. The printed (6.3.54) reads … + ε²[-Σ ∂ᵢAᵢ + Σ ∂ᵢBᵢⱼ ∂ⱼφ + Σ ∂ᵢ∂ⱼBᵢⱼ] = 0; the correct equation (formalized here) has the coefficients -½(∂ᵢBᵢⱼ∂ⱼφ + ∂ⱼBᵢⱼ∂ᵢφ) (= -Σ ∂ᵢBᵢⱼ∂ⱼφ for symmetric B), an additional term -½ Σ Bᵢⱼ ∂ᵢ∂ⱼφ at order ε², and the term ½ Σ ∂ᵢ∂ⱼBᵢⱼ at order ε⁴, not ε². The leading order (6.3.56) is unaffected.

Calculus for pd #

theorem Gardiner.pd_add_fun {n : ℕ} (u v : Sp n → ℝ) (i : Fin n) (z : Sp n) (hu : DifferentiableAt ℝ u z) (hv : DifferentiableAt ℝ v z) :
pd (fun (w : Sp n) => u w + v w) i z = pd u i z + pd v i z
theorem Gardiner.pd_const_mul_fun {n : ℕ} (a : ℝ) (u : Sp n → ℝ) (i : Fin n) (z : Sp n) (hu : DifferentiableAt ℝ u z) :
pd (fun (w : Sp n) => a * u w) i z = a * pd u i z
theorem Gardiner.pd_exp_div {n : ℕ} (phi : Sp n → ℝ) (ε : ℝ) (i : Fin n) (z : Sp n) (hphi : DifferentiableAt ℝ phi z) :
pd (fun (w : Sp n) => Real.exp (-phi w / ε ^ 2)) i z = -pd phi i z / ε ^ 2 * Real.exp (-phi z / ε ^ 2)

The stationary equation (6.3.54) #

def Gardiner.snHJ {n : ℕ} (A : Sp n → Fin n → ℝ) (B : Sp n → Fin n → Fin n → ℝ) (phi : Sp n → ℝ) (z : Sp n) :

The Hamilton–Jacobi operator Σ Aᵢ∂ᵢφ + ½ Σ Bᵢⱼ ∂ᵢφ ∂ⱼφ, the left-hand side of (6.3.56).

Equations
Instances For
    def Gardiner.snFirst {n : ℕ} (A : Sp n → Fin n → ℝ) (B : Sp n → Fin n → Fin n → ℝ) (phi : Sp n → ℝ) (z : Sp n) :

    The ε² bracket of the corrected (6.3.54).

    Equations
    • One or more equations did not get rendered due to their size.
    Instances For
      def Gardiner.snSecond {n : ℕ} (B : Sp n → Fin n → Fin n → ℝ) (z : Sp n) :

      The ε⁴ bracket of the corrected (6.3.54).

      Equations
      Instances For
        def Gardiner.snStatEq {n : ℕ} (A : Sp n → Fin n → ℝ) (B : Sp n → Fin n → Fin n → ℝ) (phi : Sp n → ℝ) (ε : ℝ) (z : Sp n) :

        The left-hand side of (6.3.54) (corrected).

        Equations
        Instances For
          theorem Gardiner.fpRHS_exp_eq {n : ℕ} (A : Sp n → Fin n → ℝ) (B : Sp n → Fin n → Fin n → ℝ) (phi : Sp n → ℝ) (hA : ∀ (i : Fin n), ContDiff ℝ 1 fun (w : Sp n) => A w i) (hB : ∀ (i j : Fin n), ContDiff ℝ 2 fun (w : Sp n) => B w i j) (hphi : ContDiff ℝ 2 phi) {ε : ℝ} (hε : ε ≠ 0) (t : ℝ) (z : Sp n) :
          fpRHS (fun (x : ℝ) (w : Sp n) => Real.exp (-phi w / ε ^ 2)) A (fun (w : Sp n) (i j : Fin n) => ε ^ 2 * B w i j) t z = Real.exp (-phi z / ε ^ 2) / ε ^ 2 * snStatEq A B phi ε z

          (6.3.53)–(6.3.54). For C¹ drift, C² diffusion and C² potential φ, the Fokker–Planck operator of (6.3.52) (diffusion ε²B) applied to exp(-φ/ε²) equals ε⁻² exp(-φ/ε²) times the left-hand side of the corrected (6.3.54).

          theorem Gardiner.stationary_exp_iff {n : ℕ} (A : Sp n → Fin n → ℝ) (B : Sp n → Fin n → Fin n → ℝ) (phi : Sp n → ℝ) (hA : ∀ (i : Fin n), ContDiff ℝ 1 fun (w : Sp n) => A w i) (hB : ∀ (i j : Fin n), ContDiff ℝ 2 fun (w : Sp n) => B w i j) (hphi : ContDiff ℝ 2 phi) {ε : ℝ} (hε : ε ≠ 0) (t : ℝ) :
          (∀ (z : Sp n), fpRHS (fun (x : ℝ) (w : Sp n) => Real.exp (-phi w / ε ^ 2)) A (fun (w : Sp n) (i j : Fin n) => ε ^ 2 * B w i j) t z = 0) ↔ ∀ (z : Sp n), snStatEq A B phi ε z = 0

          (6.3.53) is a stationary solution of (6.3.52) iff the corrected (6.3.54) holds.

          theorem Gardiner.hamiltonJacobi_of_stationary {n : ℕ} (A : Sp n → Fin n → ℝ) (B : Sp n → Fin n → Fin n → ℝ) (phi0 phi1 : Sp n → ℝ) (hA : ∀ (i : Fin n), ContDiff ℝ 1 fun (w : Sp n) => A w i) (hB : ∀ (i j : Fin n), ContDiff ℝ 2 fun (w : Sp n) => B w i j) (hphi0 : ContDiff ℝ 2 phi0) (hphi1 : ContDiff ℝ 2 phi1) (t : ℝ) (hstat : ∃ᶠ (ε : ℝ) in nhdsWithin 0 {0}ᶜ, ∀ (z : Sp n), fpRHS (fun (x : ℝ) (w : Sp n) => Real.exp (-(phi0 w + ε * phi1 w) / ε ^ 2)) A (fun (w : Sp n) (i j : Fin n) => ε ^ 2 * B w i j) t z = 0) (z : Sp n) :
          snHJ A B phi0 z = 0

          (6.3.55)–(6.3.56): the Hamilton–Jacobi equation. Let φ_ε = φ₀ + εφ₁ (the expansion (6.3.55) to first order, with C² coefficients). If exp(-φ_ε/ε²) is a stationary solution of (6.3.52) for arbitrarily small ε ≠ 0, then φ₀ satisfies Σ Aᵢ∂ᵢφ₀ + ½ Σ Bᵢⱼ ∂ᵢφ₀ ∂ⱼφ₀ = 0.

          Non-degenerate instance: the Ornstein–Uhlenbeck process #

          theorem Gardiner.pd_coord_self {n : ℕ} (i : Fin n) (z : Sp n) :
          pd (fun (w : Sp n) => w.ofLp i) i z = 1
          theorem Gardiner.pd_ou_potential (k : ℝ) (z : Sp 1) :
          pd (fun (w : Sp 1) => k * w.ofLp 0 ^ 2) 0 z = 2 * k * z.ofLp 0
          theorem Gardiner.pd_ou_drift (k : ℝ) (z : Sp 1) :
          pd (fun (w : Sp 1) => -k * w.ofLp 0) 0 z = -k
          theorem Gardiner.pd2_ou_potential (k : ℝ) (z : Sp 1) :
          pd2 (fun (w : Sp 1) => k * w.ofLp 0 ^ 2) 0 0 z = 2 * k
          theorem Gardiner.pd_const_fun {n : ℕ} (c : ℝ) (i : Fin n) (z : Sp n) :
          pd (fun (x : Sp n) => c) i z = 0
          theorem Gardiner.ou_snStatEq (k ε : ℝ) (z : Sp 1) :
          snStatEq (fun (w : Sp 1) (x : Fin 1) => -k * w.ofLp 0) (fun (x : Sp 1) (x_1 x_2 : Fin 1) => 1) (fun (w : Sp 1) => k * w.ofLp 0 ^ 2) ε z = 0
          theorem Gardiner.ou_stationary_exp (k : ℝ) {ε : ℝ} (hε : ε ≠ 0) (t : ℝ) (z : Sp 1) :
          fpRHS (fun (x : ℝ) (w : Sp 1) => Real.exp (-(k * w.ofLp 0 ^ 2) / ε ^ 2)) (fun (w : Sp 1) (x : Fin 1) => -k * w.ofLp 0) (fun (x : Sp 1) (x_1 x_2 : Fin 1) => ε ^ 2 * 1) t z = 0

          Non-degenerate instance of (6.3.53)–(6.3.54). For the Ornstein–Uhlenbeck process A(x) = -kx, B = 1 (so that (6.3.52) is ∂ₜp = ∂ₓ(kxp) + ½ε² ∂ₓ²p), the function exp(-kx²/ε²) is an exact stationary solution for every ε ≠ 0.

          theorem Gardiner.ou_hamiltonJacobi (k t : ℝ) (z : Sp 1) :
          snHJ (fun (w : Sp 1) (x : Fin 1) => -k * w.ofLp 0) (fun (x : Sp 1) (x_1 x_2 : Fin 1) => 1) (fun (w : Sp 1) => k * w.ofLp 0 ^ 2) z = 0

          Non-degenerate instance of (6.3.55)–(6.3.56). The hypotheses of hamiltonJacobi_of_stationary hold for the Ornstein–Uhlenbeck process (φ₀ = kx², φ₁ = 0), and its conclusion is the Hamilton–Jacobi equation -kx · 2kx + ½ (2kx)² = 0.