Van Kampen's system size expansion #
Gardiner, Handbook of Stochastic Methods, §7.2.3, equations (7.2.28)–(7.2.40) and (7.2.59)–(7.2.61).
The master equation has finitely many jumps a → a + r_j with extensive rates
W(a; r_j) = Ω ψ_j(a/Ω) (7.2.29), so that the jump moments are αₙ(a) = Ω α̃ₙ(a/Ω)
(7.2.31). In the fluctuation variable z of (7.2.30), a = Ωφ(t) + Ω^{1/2} z, the
backward generator of the master equation becomes Gardiner.vkGen
(Gardiner.finJumpGen_fluct, Gardiner.hasDerivAt_fluct_frame).
Main results:
Gardiner.vkGen_sub_lnaGen_le— the rescaled master-equation generator is withinO(Ω^{-1/2})of the linear-noise operator on smooth test functions, with the explicit boundΩ^{-1/2} [K z² |f'|/2 + L |z| |f''|/2 + ‖f'''‖ β̃₃(φ + Ω^{-1/2}z)/6], whereKboundsα̃₁'',Lis a Lipschitz constant ofα̃₂andβ̃₃ = Σ_j |ψ_j| |r_j|³is the third absolute jump moment;Gardiner.vkGen_tendsto_lnaGenis the limitΩ → ∞.- Erratum (7.2.35), (7.2.40). The limiting operator is
α̃₁'(φ) z ∂_z f + ½ α̃₂(φ) ∂_z² f, i.e. the Fokker–Planck equation is∂ₜP = -α̃₁'(φ(t)) ∂_z(zP) + ½ α̃₂(φ(t)) ∂_z² Pand the SDE isdz = α̃₁'(φ(t)) z dt + √α̃₂(φ(t)) dW. The text printsα̃₁in place of its derivativeα̃₁'in both (7.2.35) and (7.2.40); its own example (7.2.47) usesα̃₁' = -k₁.Gardiner.lnaFPis the corrected right-hand side. - §7.2.3 a), comparison with Kramers–Moyal:
Gardiner.km_change_of_variablesis (7.2.36)–(7.2.37), which is the small-noise equation (6.3.1) withε² = 1/Ω(7.2.38);Gardiner.vkGen_sub_kmGen_leandGardiner.kmGen_sub_lnaGen_leshow that the master-equation and the Kramers–Moyal generators have the same limit and differ byO(Ω^{-1/2}). - (7.2.33)–(7.2.34): the coefficient
vkCoeff mofΩ^{-(m-2)/2}(backward form);vkCoeff 1is cancelled by the transport term exactly whenφ' = α̃₁(φ)(7.2.33), andvkCoeff 2is the linear-noise operator. - (7.2.59)–(7.2.60):
Gardiner.vk_moment_hierarchy, from integration by parts (Gardiner.integral_pow_mul_tower). - The moments of the linear-noise equation, (4.4.84)–(4.4.85) in one dimension:
Gardiner.lna_mean_ode,Gardiner.lna_second_moment_ode,Gardiner.linear_ode_eq.
(7.2.31): the intensive jump moments α̃ₙ(x) = Σ_j ψ_j(x) r_jⁿ.
Equations
- Gardiner.intMoment s r ψ n x = Gardiner.finJumpMoment s r ψ x n
Instances For
(7.2.32), the transport term: along the moving frame a = Ωφ(t) + Ω^{1/2} z, a function
of z changes at rate -Ω^{1/2} φ'(t) ∂_z f.
The backward generator of the fluctuation variable z (7.2.30)–(7.2.32): the
master-equation generator in the moving frame plus the transport term,
Σ_j Ω ψ_j(φ + Ω^{-1/2} z) [f(z + Ω^{-1/2} r_j) - f(z)] - Ω^{1/2} φ' f'(z).
The test function is given by its derivative tower D.
Equations
Instances For
The Kramers–Moyal route (7.2.36)–(7.2.39): the generator of the Kramers–Moyal
Fokker–Planck equation (7.2.37) in the fluctuation variable. It is the small-noise
generator snGen of (6.3.8) with A = α̃₁, B = α̃₂, ε = Ω^{-1/2} (7.2.38).
Equations
- Gardiner.kmGen' s r ψ Ω phi D z = Gardiner.snGen (Gardiner.intMoment s r ψ 1) (Gardiner.intMoment s r ψ 2) (√Ω)⁻¹ phi z (D 1 z) (D 2 z)
Instances For
The master equation and its Kramers–Moyal truncation have the same system-size
limit: in the fluctuation variable, with φ' = α̃₁(φ) (7.2.33), the two generators differ
by at most Ω^{-1/2} ‖f'''‖ β̃₃(φ + Ω^{-1/2}z)/6.
The Kramers–Moyal route to the linear noise approximation (7.2.37)–(7.2.39): the
generator of the Kramers–Moyal equation in the fluctuation variable is within
Ω^{-1/2} (K z² |f'| + L |z| |f''|)/2 of the linear-noise operator
α̃₁'(φ) z f' + ½ α̃₂(φ) f''. Here E is a derivative tower of α̃₁ with |α̃₁''| ≤ K
between φ and φ + Ω^{-1/2} z, and L controls the variation of α̃₂.
Van Kampen's expansion with an explicit error term. On a test function f with
|f'''| ≤ M, the generator of the fluctuation variable (7.2.30)–(7.2.32), with φ solving
the deterministic equation (7.2.33), differs from the linear-noise operator
α̃₁'(φ) z f'(z) + ½ α̃₂(φ) f''(z) (the corrected (7.2.35)) by at most
Ω^{-1/2} [K z² |f'(z)|/2 + L |z| |f''(z)|/2 + M β̃₃(φ + Ω^{-1/2} z)/6].
The system-size limit (7.2.35), corrected: as Ω → ∞ the generator of the
fluctuation variable converges to the linear-noise operator on every smooth test function
with bounded third derivative, provided α̃₁'' is bounded, α̃₂ is Lipschitz at φ, and the
rates are continuous at φ.
(7.2.33)–(7.2.34): the coefficients of the expansion #
The coefficient of Ω^{-(m-2)/2} in (7.2.34), in backward (generator) form:
Σ_{n=1}^{m} α̃ₙ^{(m-n)}(φ) z^{m-n} f⁽ⁿ⁾(z) / (n! (m-n)!), where Dα n k = α̃ₙ^{(k)}(φ) and
Df n = f⁽ⁿ⁾(z). (The forward form has (-∂_z)ⁿ z^{m-n} in place of z^{m-n} f⁽ⁿ⁾.)
Equations
Instances For
The corrected linear-noise Fokker–Planck equation (7.2.35) #
(7.2.35), corrected: the right-hand side
-α̃₁'(φ) ∂_z(z P) + ½ α̃₂(φ) ∂_z² P of the linear-noise Fokker–Planck equation, with
a₁ = α̃₁'(φ(t)) and a₂ = α̃₂(φ(t)). (The text prints α̃₁[φ(t)] for a₁.) The
corresponding SDE (7.2.40) is dz = a₁ z dt + √a₂ dW(t).
Equations
Instances For
The linear-noise operator is the generator of the corrected (7.2.35): for a test
function f and a density P, ∫ f · lnaFP P = ∫ (a₁ z f' + ½ a₂ f'') P
(integration by parts; surface terms vanish by integrability).
The moments of the corrected linear-noise equation: ∫ z · lnaFP P = a₁ ∫ z P and
∫ z² · lnaFP P = 2 a₁ ∫ z² P + a₂ ∫ P (the m = 2 truncation of (7.2.60)).
The mean of the linear-noise approximation (7.2.48), (4.4.84): if P(t, ·) solves
the corrected (7.2.35) with coefficient a₁(t) and d/dt commutes with ∫ dz (hswap),
then d⟨z⟩/dt = a₁(t) ⟨z⟩.
The second moment of the linear-noise approximation (7.2.49), (4.4.85): under the
same hypotheses, d⟨z²⟩/dt = 2 a₁(t) ⟨z²⟩ + a₂(t) ∫ P.
(7.2.36)–(7.2.37): the Kramers–Moyal equation in intensive variables #
(7.2.36) ⟹ (7.2.37): if αₙ(a) = Ω α̃ₙ(a/Ω) (7.2.31), then the density
P_x(x) = Ω P_a(Ωx) of x = a/Ω satisfies ∂ₜ P_x = -∂ₓ[α̃₁ P_x] + (1/2Ω) ∂ₓ²[α̃₂ P_x], i.e.
the small-noise equation (6.3.1) with ε² = 1/Ω (7.2.38). Pointwise:
Ω · RHS(7.2.36)(Ωx) = RHS(6.3.1)(x).
(7.2.59)–(7.2.61): the moment hierarchy #
(7.2.60): suppose ∂ₜ P = Σ_{(m,n) ∈ T} c_{m,n} (-∂_z)ⁿ [z^{m-n} P] (the right-hand
side of (7.2.34), truncated to a finite set T of indices; c_{m,n} is
Ω^{-(m-2)/2} α̃ₙ^{(m-n)}(φ)/(n!(m-n)!)), where G (m,n) is a derivative tower of
z^{m-n} P. If d/dt commutes with ∫ dz (hswap), then
d⟨zᵏ⟩/dt = Σ_{(m,n) ∈ T} c_{m,n} k!/(k-n)! ⟨z^{(k-n)+(m-n)}⟩, which is (7.2.60): the
falling factorial k!/(k-n)! vanishes for n > k, so the inner sum runs to min(m, k).