Implementation of Milstein Schemes for Stochastic Delay-Differential Equations with Arbitrary Fixed Delays
Mitchell Griggs, Kevin Burrage, Pamela Burrage · 2025
Read Implementation of Milstein Schemes for Stochastic Delay-Differential Equations with Arbitrary Fixed Delays by Mitchell Griggs, Kevin Burrage, Pamela Burrage (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-15365.txt.
Introduction
Mitchell Griggs, Kevin Burrage, and Pamela Burrage built ways to run noise-driven delay models whose time lags refuse to line up, in a paper published in 2025. Such models mix present values with values from fixed times in the past. Neat numerical schemes exist, but they also need past readings at combined lags, which fall off any even grid. Earlier work coped only when the lags divided the step size cleanly.
For the half-order scheme, they keep a fixed step and fill past readings by linear interpolation. For the full-order scheme, they grow the grid to hold each needed past time, which makes step sizes vary. A related method grows the paired noise sums on that uneven grid as well. Worked examples confirm that both schemes reach their claimed rates. The tools free delay models from the divisible-lag restriction.
Catalog record
| CREATOR | Mitchell Griggs, Kevin Burrage, Pamela Burrage |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 87 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-15365
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Implementation of Milstein Schemes for Stochastic Delay-Differential Equations with Arbitrary Fixed Delays Mitchell Griggs, Kevin Burrage, Pamela Burrage Abstract This paper develops methods for numerically solving stochastic delay-differential equations (SDDEs) with multiple fixed delays that do not align with a uniform time mesh. We focus on numerical schemes of strong convergence orders 1 / 2 1/2 and 1 1 , such as the Euler–Maruyama and Milstein schemes, respectively. Although numerical schemes for SDDEs with delays τ 1 , … , τ K \tau_{1},\ldots,\tau_{K} are theoretically established, their implementations require evaluations at both present times such as t n t_{n} , and also at delayed times such as t n − τ k t_{n}-\tau_{k} and t n − τ l − τ k t_{n}-\tau_{l}-\tau_{k} . As a result, previous simulations of these schemes have been largely restricted to the case of divisible delays. We develop simulation techniques for the general case of indivisible delays where delayed times such as t n − τ k t_{n}-\tau_{k} are not restricted to a uniform time mesh. To achieve order of convergence (OoC) 1 / 2 1/2 , we implement the schemes with a fixed step size while using linear interpolation to approximate delayed scheme values. To achieve OoC 1 1 , we construct an augmented time mesh that includes all time points required to evaluate the schemes, which necessitates using a varying step size. We also introduce a technique to simulate delayed iterated stochastic integrals on the augmented time mesh, by extending an established method from the divisible-delays setting. We then confirm that the numerical schemes achieve their theoretical convergence orders with computational examples. 2020 MSC 60H35 , 65C30 , 65L20 1 Introduction In this paper, we develop practical methods to produce numerical solutions for Itô SDDEs of the form \displaystyle\mathrm{d}X(t) \displaystyle=[A_{0}X(t)+f(t,X(t),X(t-\tau_{1}),\ldots,X(t-\tau_{K}))]\,\mathrm{d}t \displaystyle\quad+\sum_{j=1}^{m}[A_{j}X(t)+g_{j}(t,X(t),X(t-\tau_{1}),\ldots,X(t-\tau_{K}))]\,\mathrm{d}W_{j}(t),\quad t\in[0,T], (1) \displaystyle X(t) \displaystyle=\phi(t),\quad t\in[-\tau,0]. Equation ( 1 ) includes d -dimensional solution process X=(X(t))_{t\in[0,T]} , constant matrices A_{0},\ldots,A_{m}\in\mathbb{R}^{d\times d} , functions f:[0,T]\times\mathbb{R}^{d}\times(\mathbb{R}^{d})^{K}\to\mathbb{R}^{d} and g_{1},\ldots,g_{m}:[0,T]\times\mathbb{R}^{d}\times(\mathbb{R}^{d})^{K}\to\mathbb{R}^{d} , constant delays \tau_{1},\ldots,\tau_{K}>0 , history process \phi , and independent Wiener processes W_{1},\ldots,W_{m} . The greatest delay is \tau=\max\{\tau_{1},\ldots,\tau_{K}\} , the history process satisfies the moment condition \sup_{t\in[-\tau,0]}\mathbb{E}[\left|\phi(t)\right|^{4}]<\infty , and each of f,g_{1},\ldots,g_{m} is C^{1} in the first component ( t ) while also being C^{2} in all other components ( X(t),X(t-\tau_{1}),\ldots,X(t-\tau_{K}) ). Without loss of generality, we suppose that one of the delays divides T . That is, for some k , there exists N_{k}\in\mathbb{N} such that N_{k}\tau_{k}=T . If this is not the case then we simply include another delay, \tau_{K+1}=T , and include X(t-\tau_{K+1}) in the arguments for f,g_{1},\ldots,g_{m} . Equation ( 1 ) arises in models involving both randomness and time delays. These models have recently found applications in economics [ 1 , 2 ] , medicine [ 3 , 4 ] , ecology [ 5 ] , and other fields. In these and other models, properties of the solution trajectories depend on the delay values. For example, in the biomechanical model by Boulet et al., it is experimentally found that certain delay parameters lead to solution blowup (see the stability diagram in [ 3 , Figure 8] ). In our work, we allow for general delays, that may or may not align with a uniform time mesh. We develop simulation techniques to numerically solve ( 1 ), with convergence orders 1/2 and 1 . The standard method for numerically solving (non-delayed) stochastic differential equations (SDEs) is the Euler–Maruyama (EM) approximation by Maruyama [ 6 ] , which has OoC 1/2 . The EM scheme is easily extendable to ( 1 ), and only includes additional calculations for delayed scheme values. Achieving OoC 1 with SDE schemes requires alternative techniques such as the method by Milstein [ 7 ] , which includes the iterated stochastic integrals I_{ij}(t_{n},t_{n+1})=\int_{t_{n}}^{t_{n+1}}\int_{t_{n}}^{s}\,\mathrm{d}W_{i}(u)\,\mathrm{d}W_{j}(s),\quad i,j=1,\ldots,m. (2) The Milstein scheme is further complicated when delays are included (except for the case of delay-free noise, where g_{j}(t,X(t),X(t-\tau_{1}),\ldots,X(t-\tau_{K}))=g_{j}(t,X(t)) , j=1,\ldots,m ), as this includes the delayed iterated stochastic integrals I_{ij}^{\tau_{k}}(t_{n},t_{n+1})=\int_{t_{n}}^{t_{n+1}}\int_{t_{n}}^{s}\,\mathrm{d}W_{i}(u-\tau_{k})\,\mathrm{d}W_{j}(s)=\int_{t_{n}}^{t_{n+1}}\int_{t_{n}-\tau_{k}}^{s-\tau_{k}}\,\mathrm{d}W_{i}(u)\,\mathrm{d}W_{j}(s),\quad i,j=1,\ldots,m,\quad k=1,\ldots,K. (3) The SDDE Milstein scheme is given by Küchler and Platen [ 8 ] when K=1 , but simulation methods for ( 2 ) were still in development at the time of this research, so examples from this time only include delay-free noise. Hofmann and Müller-Gronbach [ 9 ] give an SDDE Milstein scheme for K delays, using the approximations I_{ij}(t_{n},t_{n+1})\approx\frac{(W_{i}(t_{n+1})-W_{i}(t_{n}))\,(W_{j}(t_{n+1})-W_{j}(t_{n}))}{2} (4) and I_{ij}^{\tau_{k}}(t_{n},t_{n+1})\approx\frac{(W_{i}(t_{n+1}-\tau_{k})-W_{i}(t_{n}-\tau_{k}))\,(W_{j}(t_{n+1})-W_{j}(t_{n}))}{2} (5) to simulate the iterated integrals. This produces an efficient Milstein approximation of ( 1 ), but with OoC 1/2 . One of our objectives in this paper is to improve the approximation ( 5 ) of ( 3 ), to improve the SDDE Milstein scheme so as to successfully achieve demonstrable OoC 1 . To produce an improved approximation for ( 3 ), we find inspiration from techniques that have been developed for the non-delayed case, which we extend to the setting with delays. One method to approximate ( 2 ) is the trapezoidal method by Milstein [ 10 ] . Cao, Zhang, and Karniadakis [ 11 ] extend the trapezoidal method to simulate ( 3 ), for the SDDE Milstein scheme in the case of a single delay that divides T . We extend this approximation further, to the case of multiple delays, including indivisible ones where some delay \tau_{k} does not divide T . Our focus is on implementation techniques, to complement existing theory. For example, there exists a continuous Milstein scheme for ( 1 ), by Kloeden and Shardlow [ 12 ] , but this scheme requires simulation of I_{ij}^{\tau_{k}}(t_{n},t) for every time t . However, it is not possible to simulate such a scheme on a continuum of values, due to the fluctuating iterated stochastic integrals. Therefore we restrict simulations to discrete mesh points, simulating the numerical schemes at every point needed for the Milstein scheme. This requires a varying step size when the delays are indivisible. We give the construction of this time mesh specifically for the Milstein scheme, but our implementation method is readily applicable for other schemes, which we demonstrate with the SDDE Magnus–Milstein (MM) method, by Griggs, Burrage, and Burrage [ 13 ] . This paper is written to give instructions for simulating numerical SDDE schemes with strong OoC 1 , and is presented in the following structure. In Section 2 , we present our selected schemes, which are later used while constructing the augmented time mesh. We discuss implementation complications for these schemes, in Section 3 . We simulate these schemes with OoC 1/2 in Section 4 , by approximating delayed scheme values with linear interpolants. We begin improving the convergence orders of these schemes in Section 5 , by augmenting the simulation time mesh. We then extend the trapezoidal approximation (for delayed iterated stochastic integrals) to this augmented mesh, in Section 6 . We provide example simulations of the schemes on the augmented mesh, in Section 7 . In Section 8 , we conclude this work with discussion of the size of the augmented mesh, the applicability of our technique for higher-order schemes, and also the significance of our method. 2 Selected numerical schemes In this section, we define our selected numerical schemes used to approximate solutions of ( 1 ). We begin