PaperlayerView exact revision on arXiv

Source-linked AI summary

Explicit domain preserving numerical schemes for a class of stochastic differential equations

Charles-Edouard Bréhier, David Cohen

arXiv:2608.25685v1math.NAmath.PR

TL;DR

The paper addresses the difficulty that standard numerical methods may leave bounded domains when approximating constrained stochastic differential equations. It develops explicit domain-preserving schemes for arbitrary-dimensional hypercubes and proves convergence, including a strong-order-one variant for one-dimensional Brownian motion. The schemes preserve the domain for every time-step size, while the general method has strong order 1/2 and weak order 1.

  • Problem

    Standard numerical methods such as Euler–Maruyama and Milstein almost surely do not preserve bounded domains for arbitrary time-step sizes.

  • Method

    The paper constructs explicit schemes using auxiliary equations, coefficient freezing, flow splitting, and integrator conditions designed to preserve hypercubes for arbitrary dimensions and noise dimensions.

  • Results

    The schemes preserve the hypercube almost surely for any time-step size; the general scheme has strong order 1/2 and weak order 1, while a one-dimensional-noise variant has strong order 1.

  • Takeaways & Limitations

    The framework provides explicit numerical approximations that retain the prescribed hypercube domain while offering the stated strong and weak convergence guarantees.

  • Takeaways & Limitations

    The presentation simplifies the domain to [-1,1]^d, with arbitrary coordinate intervals requiring minor modifications, and domain preservation depends on imposed integrator conditions.

Abstract

from arXiv · show

We construct and analyze numerical schemes for systems of stochastic differential equations, which preserve almost surely a given hypercube of arbitrary dimension. We propose a new general class of explicit schemes, such that for any choice of the time-step size the numerical solution takes values in the hypercube. We prove strong and weak convergence results for this general class of domain preserving numerical schemes, with strong order $1/2$ and weak order $1$ in general. We also construct a variant of the scheme which achieves strong order $1$ when the stochastic differential equation is driven by a one-dimensional Brownian motion. The convergence results are illustrated with numerical experiments.

1. Introduction

The paper introduces explicit numerical schemes that preserve stochastic differential equation solutions within hypercubes for any time-step size. It establishes convergence rates for general schemes and a strong-order-one variant under one-dimensional Brownian noise.

  • Motivation: Standard Euler–Maruyama and Milstein schemes almost surely fail to preserve bounded domains for arbitrary time-step sizes.This motivates domain-preserving methods for applications whose state variables must remain bounded.
  • Contribution: The proposed explicit schemes preserve the hypercube D = [-1,1]^d almost surely for any time-step size, including arbitrary dimensions, multidimensional noise, and drift.The schemes are primarily formulated for Itô noise, with an adaptation for Stratonovich noise.
  • Construction: The construction freezes coefficient factors and uses Stratonovich reformulation with Lie–Trotter splitting of associated flows.The variable factor σ is retained because replacing it directly with standard explicit Euler–Maruyama would lose domain preservation.
  • Convergence: The general scheme has strong order 1/2 and weak order 1 under appropriate regularity conditions.The convergence statements are given through strong and weak error estimates.
  • Higher-order variant: A corrected variant achieves strong order 1 for scalar equations driven by one-dimensional Brownian motion.Its correction term is analogous to the correction distinguishing Milstein from Euler–Maruyama.
  • Experiments: Numerical experiments illustrate the results as system dimension, Brownian-motion dimension, and drift components vary.The paper also discusses transfer of the analysis to Stratonovich noise through the equivalent Itô formulation.

2. Setting

The setting is a class of Itô stochastic differential equations evolving almost surely in the hypercube D = [-1,1]^d. Boundary-vanishing vector-field components enable a coordinatewise factorization that underpins the auxiliary-equation formulation.

  • Domain: The domain is the d-dimensional hypercube D = [-1,1]^d, with interior (-1,1)^d and boundary ∂D.The presentation fixes all coordinate bounds to -1 and 1 for simplicity, while allowing arbitrary intervals with minor modifications.
  • Assumptions: Each vector-field component vanishes when its corresponding coordinate is on either boundary face, and sufficiently regular coefficients yield a unique solution remaining in D almost surely.The stated baseline regularity is C2, with stronger conditions imposed for convergence results.
  • Stochastic model: The model is an Itô stochastic differential equation with drift g_0 and diffusion vector fields g_1,...,g_M driven by an R^M-valued Brownian motion.The Brownian components are independent standard real-valued Brownian motions.
  • Regularity: The solution satisfies the temporal regularity estimate E[||X(t)-X(s)||^2] ≤ C(T)|t-s| on finite time intervals.This estimate is used in the subsequent error analysis.
  • Equivalent formulation: Boundary-vanishing diffusion coefficients factor coordinatewise as g = fσ with σ(y) = (y-1)(y+1), producing bounded auxiliary scalar stochastic and ordinary differential equations.The factorization follows from the endpoint conditions and preserves the boundary zeros explicitly.
  • Equivalent formulation: The equivalent auxiliary formulation is the fundamental ingredient for constructing domain-preserving schemes and for analyzing their errors.The auxiliary equations separate the stochastic and ordinary differential components associated with each coordinate.

3. Domain preserving integrators

The paper constructs explicit integrators that preserve the hypercube domain for every time-step size. The general scheme achieves strong order 1/2 and weak order 1, while a one-dimensional-noise variant achieves strong order 1.

  • The scheme updates each component using frozen vector-field coefficients while retaining a variable σ in the stochastic equation, unlike explicit Euler–Maruyama.Keeping σ variable is identified as a key ingredient for domain preservation.
  • The proposed class combines exact ordinary-flow integration with numerical integrators Φ for the auxiliary stochastic equation.The ordinary differential equation can be solved exactly, whereas the auxiliary stochastic equation requires approximation.
  • The flow ϕ maps every y ∈ [-1, 1] back into [-1, 1] for every real time s, providing a basic domain-preserving building block.
  • The example integrators are domain preserving and converge with strong order 1 for the auxiliary stochastic differential equation.
  • Φ(s, γ, y) ∈ [-1, 1] for all s ≥ 0, γ ∈ R, and y ∈ [-1, 1] is the condition ensuring domain preservation.The integrators introduced in Examples 1 and 2 satisfy this condition.
  • The general scheme converges with strong order 1/2 and weak order 1, while a variant reaches strong order 1 when driven by one-dimensional Brownian motion.The numerical experiments indicate that the general convergence orders are optimal.

4. Main results and numerical illustrations

The proposed schemes preserve the hypercube domain almost surely, while achieving strong order 1/2 and weak order 1 generally; a variant reaches strong order 1 for one-dimensional noise.

  • 4.1. Domain preservation: The numerical experiments compare DP(0.5;1) and DP(0.5;2) with Euler–Maruyama, and DP(1.0;1) and DP(1.0;2) with Milstein.The reported proportions validate domain preservation for the proposed schemes, unlike the corresponding standard methods.
  • 4.1. Domain preservation: The schemes (17) and (33) are domain preserving, keeping X_n in D almost surely for any admissible time-step size.This follows by recursion from the boundary-preserving condition on the mapping Φ.
  • 4.2. Strong error estimates: Strong order 1/2 is established for scheme (17) across varied dimensions, Brownian dimensions, drift settings, and numerical experiments.Figures 2 and 3 show that the order 1/2 is optimal for the tested cases.
  • 4.3. Weak error estimates: Weak order 1 is established for scheme (17) for C3 test functions and illustrated with different functions θ in dimensions 1 and 2.Figures 4 and 5 indicate that the weak order 1 is optimal in the reported experiments.

5. Domain preserving schemes for Stratonovich stochastic differential equations

The paper adapts explicit domain-preserving schemes to Stratonovich SDEs by converting them to an equivalent Itô formulation. The resulting schemes preserve the hypercube for every time-step size and retain the stated convergence orders.

  • Itô reformulation: The Stratonovich SDE is rewritten in equivalent Itô form, adding the correction 1/2 Σ_m (g_m)'g_m to the drift.The correction is essential for constructing consistent domain-preserving schemes.
  • Assumptions: The construction assumes the drift and diffusion vector fields satisfy stronger boundary conditions, including vanishing components on the hypercube boundary.These conditions ensure the transformed coefficients retain the structure required by the domain-preserving construction.
  • Domain preservation: For any time-step size, the resulting numerical solution remains in D = [-1,1]^d almost surely when initialized in D.Domain preservation follows from applying the original scheme to the equivalent Itô equation.
  • Higher-order variant: A second-order variant achieves strong order 1 for one-dimensional Brownian motion while preserving the domain.The order-1 result applies to schemes of the specified second-order type.
  • Convergence: The Stratonovich-adapted schemes inherit strong order 1/2 and weak order 1 convergence under the stated regularity and integrator assumptions.The convergence results follow directly from the corresponding Itô-scheme theorems.
  • Consistency: The naive frozen-flow scheme is generally inconsistent because it omits the correction term required when f is nonconstant.The numerical experiments illustrate this inconsistency and motivate the equivalent-Itô construction.

6. Proof of Theorem 7

The proof establishes strong order 1/2 by embedding the discrete scheme in auxiliary processes and decomposing the error into controllable drift, diffusion, and interpolation terms. Domain preservation and bounded coefficients provide the estimates needed for the final error bound.

  • Proof objective: The proof targets mean-square convergence of order 1/2 under regularity assumptions on the integrator and coefficient mappings.The argument assumes Assumptions 2–3 and finite ||Φ||_{3,β}.
  • Auxiliary processes: The auxiliary process rZ_k is introduced to represent within-step stochastic evolution and provide increment bounds.Its increments combine bounded coefficient terms with Brownian increments.
  • Domain preservation: The auxiliary process remains in [-1,1], and its grid-point values coincide with the numerical iterates because the integrator is domain preserving.This connects the continuous auxiliary construction to the discrete scheme without leaving the domain.
  • Auxiliary processes: The auxiliary process rX_k satisfies an interval-wise SDE and remains close to the grid value X_n,k with mean-square error O(τ).The proof uses Itô’s formula, Cauchy–Schwarz, and the Itô isometry.
  • Error decomposition: The global strong error is decomposed into several drift, diffusion, and auxiliary-process error terms before each contribution is bounded.The decomposition is introduced in equations (68), (72), and (76).
  • Conclusion: Combining the error-term estimates yields the theorem’s mean-square strong convergence bound of order 1/2.The final estimate is obtained after grouping terms controlled by the auxiliary-process bounds.

7. Proof of Theorem 9

The weak-error proof uses the Kolmogorov equation, auxiliary processes, and a telescoping decomposition. Regularity and increment bounds control the resulting error terms, which sum to the weak error estimate.

  • Assumptions: The proof assumes regularity and growth control for Φ, ensuring bounds on B_γΦ, ψ, and their first and second derivatives.The condition ||Φ||_4,β < ∞ supplies the required control.
  • Weak-error representation: The Kolmogorov solution u converts the terminal weak error into a difference between u evaluated at the numerical and exact solutions.Specifically, E[θ(X_N)] − E[θ(X(T))] becomes E[u(0,X_N)] − E[u(T,X_0)].
  • Error decomposition: A telescoping sum decomposes the weak error into one-step terms ε_n, which are represented using the auxiliary process matching X_n at grid points.The identity rX(t_n)=X_n enables the replacement of grid values by the auxiliary process.
  • Conclusion: Summing the auxiliary error bounds over all time intervals yields the weak error estimate and completes Theorem 9.The theorem follows after the estimates established in Lemmas 19, 20, and 21 are combined.
  • Auxiliary estimates: Auxiliary-process estimates bound increments and derivative-related error terms on each time interval.Lemma 22 controls Δ_n,k, while Lemma 23 bounds δ^1_n,k,ℓ using regularity of u and increments of rX.

8. Proof of Theorem 10

The proof of Theorem 10 analyzes a more delicate one-dimensional-Brownian-motion error decomposition. Auxiliary processes remain domain-preserving and satisfy increment estimates that lead, via discrete Grönwall, to the strong error estimate.

  • Objective: The proof targets first-order mean-square convergence for the domain-preserving scheme when the system is driven by one-dimensional Brownian motion.The analysis is stated to require more effort and a more delicate error decomposition than the general case.
  • Auxiliary process: The auxiliary process pX_k remains in [-1,1], agrees with X_n,k at grid points, and is continuous across grid intervals.These properties follow from the domain-preserving integrator Φ and the scheme definition.
  • Auxiliary estimates: Moment and exponential bounds for pZ_k support the control of auxiliary processes and error terms.Lemma 24 provides increment and moment bounds, while Lemma 25 controls exponential expressions for sufficiently small τ.
  • Auxiliary estimates: The auxiliary process satisfies the increment bound E[|pX_k(t) − X_n,k|^2] ≤ C_p(T)τ.The bound holds uniformly over components, time steps, and t within each interval.
  • Error conclusion: After decomposing the component errors, a discrete Grönwall argument establishes the strong error estimates and completes Theorem 10.The component errors are defined relative to the exact solution at grid times.

Appendix A. Proof of Lemmas 3 and 4

The appendix verifies that the proposed integrators satisfy the assumptions needed for domain preservation and convergence. It uses flow properties, smoothness, endpoint invariance, and derivative-growth bounds.

  • Flow properties: The mappings ϕ and φ are flows of ordinary differential equations whose vector fields satisfy the required smoothness and endpoint conditions.The vector fields are class C^8 and satisfy σ(-1)=σ(1)=0.
  • Domain preservation: The flow maps preserve [-1,1] for every initial value in the interval and every real flow time.This property is stated for both ϕ and φ.
  • Integrator assumptions: The integrators Φ from Examples 1 and 2 satisfy the domain-preservation assumptions through the listed conditions.The appendix checks conditions (28)–(34), including condition (31) as a consequence of condition (34).
  • Integrator assumptions: The integrator from Example 2 maps admissible inputs into [-1,1], satisfying condition (28).The stated verification holds for s ≥ 0, γ ∈ R, and y ∈ [-1,1].
  • Regularity: Compositions of the flows and σ_1 give integrators Φ of class C^8 with controlled derivatives and finite ||Φ||_p,β for needed orders.The appendix identifies p=2, 3, and 4 as the orders needed for Theorems 7, 9, and 10.
Loading 2608.25685v1…