Euler–Maruyama one-step diffusion error: and
OpenAppliedSDE.SarkkaSolin.euler_maruyama_diffusion_local_errorSetting. Let be a Brownian motion on a probability space with a filtration to which is adapted and whose future increments are independent of the past. Let be globally Lipschitz with a common constant , and let . Let be a process with almost surely continuous paths, with for , such that is jointly measurable, progressively measurable and square integrable on for every , and such that for every , almost surely,
where the last term is the platform's Itô integral stochasticIntegralBrownian. For each let be the Euler–Maruyama scheme on the grid with step : and
Write (the binder I with hI names this process) and .
Claim. There is a constant , independent of and , such that for all and the martingale part of the one-step error
satisfies
Proof idea. is the Itô integral of , so the Itô isometry gives , and the increment bound finishes the estimate. The orthogonality holds because is a martingale increment over and is square integrable and almost surely equal to an -measurable variable.
import Mathlib import Definitions.Def_KLSModular_LevyStochCalc_Brownian_Filtered import Definitions.Def_KLSModular_LevyStochCalc_Brownian_ItoL2CompletionMaster open MeasureTheory ProbabilityTheory Filter LevyStochCalc.Brownian open scoped ENNReal NNReal Topology BigOperators
theorem AppliedSDE.SarkkaSolin.euler_maruyama_diffusion_local_error
{Ω : Type*} [MeasurableSpace Ω]
(P : Measure Ω) [IsProbabilityMeasure P]
(W : LevyStochCalc.Brownian.BrownianMotion P)
(ℱ : Filtration ℝ ‹MeasurableSpace Ω›)
(hW : LevyStochCalc.Brownian.IsBrownianFiltration W ℱ)
(T x₀ : ℝ) (hT : 0 < T)
(b σ : ℝ → ℝ)
(hLip : ∃ L : ℝ, 0 ≤ L ∧
(∀ x y : ℝ, |b x - b y| ≤ L * |x - y|) ∧
(∀ x y : ℝ, |σ x - σ y| ≤ L * |x - y|))
(X : ℝ → Ω → ℝ)
(hXcont : ∀ᵐ ω ∂P, Continuous (fun t : ℝ => X t ω))
(hXmom : ∀ t : ℝ, t ∈ Set.Icc (0 : ℝ) T →
Integrable (fun ω : Ω => (X t ω) ^ 2) P)
(H : Ω → ℝ → ℝ)
(hH : ∀ (ω : Ω) (s : ℝ), H ω s = σ (X s ω))
(hHmeas : Measurable (Function.uncurry H))
(hHprog : LevyStochCalc.Probability.ProgressivelyMeasurable ℱ H)
(hHsq : ∀ (t : ℝ), 0 < t →
∫⁻ ω, ∫⁻ s in Set.Icc (0 : ℝ) t,
(‖H ω s‖₊ : ℝ≥0∞) ^ 2 ∂volume ∂P < ⊤)
(hXeq : ∀ t : ℝ, 0 ≤ t → t ≤ T →
(fun ω : Ω => X t ω) =ᵐ[P] fun ω : Ω =>
x₀ + (∫ s in (0 : ℝ)..t, b (X s ω)) +
LevyStochCalc.Brownian.Ito.stochasticIntegralBrownian
W ℱ hW H hHmeas hHprog hHsq t ω)
(I : ℝ → Ω → ℝ)
(hI : ∀ (t : ℝ) (ω : Ω), I t ω =
LevyStochCalc.Brownian.Ito.stochasticIntegralBrownian
W ℱ hW H hHmeas hHprog hHsq t ω)
(E : (n : ℕ) → Fin (n + 1) → Ω → ℝ)
(hE0 : ∀ (n : ℕ) (ω : Ω), E n 0 ω = x₀)
(hErec : ∀ (n : ℕ), 0 < n →
∀ (k : Fin n) (ω : Ω),
E n k.succ ω = E n k.castSucc ω +
b (E n k.castSucc ω) * (T / n) +
σ (E n k.castSucc ω) *
(W.W ((((k.val + 1 : ℕ) : ℝ) * T / n)) ω -
W.W (((k.val : ℝ) * T / n)) ω)) :
∃ c : ℝ, 0 ≤ c ∧ ∀ (n : ℕ), 0 < n → ∀ k : Fin n,
(∫ ω, ((I (((k.val + 1 : ℕ) : ℝ) * T / n) ω - I ((k.val : ℝ) * T / n) ω) -
σ (E n k.castSucc ω) *
(W.W (((k.val + 1 : ℕ) : ℝ) * T / n) ω - W.W ((k.val : ℝ) * T / n) ω)) ^ 2 ∂P
≤ c * (T / n) * ((∫ ω, (X ((k.val : ℝ) * T / n) ω - E n k.castSucc ω) ^ 2 ∂P) + T / n)) ∧
(∫ ω, (X ((k.val : ℝ) * T / n) ω - E n k.castSucc ω) *
((I (((k.val + 1 : ℕ) : ℝ) * T / n) ω - I ((k.val : ℝ) * T / n) ω) -
σ (E n k.castSucc ω) *
(W.W (((k.val + 1 : ℕ) : ℝ) * T / n) ω - W.W ((k.val : ℝ) * T / n) ω)) ∂P = 0) := by
sorry