Convergence rate in the splitting-up method for rough differential equations
Peter H.C. Pang · 2024
Read Convergence rate in the splitting-up method for rough differential equations by Peter H.C. Pang (2024) 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-2412-00432.txt.
Introduction
Peter H C Pang works out how fast a split-step scheme nears true paths for rough-driven curves in a 2024 paper. The curves are pushed by a jagged driver too rough for smooth maths. A split-up scheme builds paths step by step, and its match to true paths needs proof.
Pang builds the paths with that scheme and shows the match in a fresh way. A rate for the match comes with the proof, not just the match itself. The note is short and tight, aimed at the core rate claim for this rough class.
Catalog record
| CREATOR | Peter H.C. Pang |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-00432
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Convergence rate in the splitting-up method for rough differential equations Thanks: The author is grateful for the support of the Norwegian Research Council via the project INICE (301538). Peter H.C. Pang Abstract . In this note we construct solutions to rough differential equations d Y = f ( Y ) d X \mathrm{d}Y=f(Y)\,\mathrm{d}X with a driver X ∈ C α ( [ 0 , T ] , ℝ d ) X\in C^{\alpha}([0,T];\mathbb{R}^{d}) , 1 3 < α ⩽ 1 2 \frac{1}{3}<\alpha\leqslant\frac{1}{2} , using a splitting-up scheme. We show convergence of our scheme to solutions in the sense of Davie by a new argument and give a rate of convergence. 2020 Mathematics Subject Classification 60L20, 60H35, 65C30 1. Introduction 1.1. Rough differential equations Suppose X\in C^{\alpha}([0,T];\mathbb{R}^{d}) for a fixed \alpha\in(1/3,1/2] and consider the rough differential equation \displaystyle\mathrm{d}Y=f(Y)\,\mathrm{d}X, (1) where f is a fixed function in C^{\gamma-1}_{b}(\mathbb{R}^{d}) , and \gamma>\frac{1}{\alpha}\geqslant 2 . Standard rough paths theory [ 3 , Ch. 4, 8] informs us that in order to make sense of this equation, we must include additional information in the form of the area \int_{s}^{t}X_{s,r}\,\mathrm{d}X_{r}:=\mathbb{X}_{s,t}, satisfying \sup_{s\not=t}|\mathbb{X}_{s,t}|/|t-s|^{2\alpha}<\infty and Chen’s relation. In this way, the augmented problem \mathrm{d}Y=f(Y)\,\mathrm{d}{\bf X},\qquad{\bf X}:=(X,\mathbb{X}) can be understood via a sewing lemma to be asking for Y such that Y_{s,t}=\int_{s}^{t}f(Y_{r})\,\mathrm{d}{\bf X}_{r}:=\lim_{\left\|P\right\|\downarrow 0}\sum_{t_{i}\in P}f(Y_{t_{i}})X_{t_{i},t_{i+1}}+f(Y_{t_{i}})f^{\prime}(Y_{t_{i}})\mathbb{X}_{t_{i},t_{i+1}}, for every s,t\in[0,T] , and P are finite partitions of [s,t] with mesh size \leqslant\left\|P\right\| . The problem being thus understood, it can be solved via Picard iterations. In this note, we revisit the problem of existence of solutions to the RDE ( 1 ) by considering a computational way of coming to assign a meaning and a solution to ( 1 ), similar to the classical paper [ 1 ] (see also [ 6 , Section 4] ). 1.2. Operator splitting One fruitful way of constructing numerical schemes for (partial) differential equations is known as the operator splitting , or fractional steps method. A careful introduction to this in the context of deterministic PDEs can be found in [ 9 ] . Consider a problem of the form \frac{\mathrm{d}Y}{\mathrm{d}t}=\mathcal{A}(Y), where \mathcal{A} is an operator that can be deomposed into a sum of simpler operators \mathcal{A}=\mathcal{A}_{1}+\mathcal{A}_{2} , (often) exhibiting qualitatively different behaviours. Let \mathcal{S}_{t} be the corresponding solution operator so that a solution can be written as Y(t)=\mathcal{S}_{t}\overline{Y} for an initial condition \overline{Y} . For j=1,2 , let \mathcal{S}_{t}^{j} be the solution operator for the equation \displaystyle\frac{\mathrm{d}Y^{j}}{\mathrm{d}t}=\mathcal{A}_{j}(Y^{j}). It stands to reason that the solution Y can be approximated by propagating the initial condition \overline{Y} repeatedly using \mathcal{S}_{{h}}^{2}\mathcal{S}_{{h}}^{1} for small time steps {h} , culminating in a Trotter formula of the form: \displaystyle\mathcal{S}_{t}=\lim_{n\to\infty}\Big(\mathcal{S}_{t/n}^{2}\mathcal{S}_{t/n}^{1}\Big)^{n}. Similar to a sewing lemma, such a formula tells us that on small time steps, the integral \mathcal{S}_{{h}}\overline{Y} of the full problem can be well approximated by the composed integral \mathcal{S}_{{h}}^{2}\mathcal{S}_{{h}}^{1}\overline{Y} . Let N=[T/h] be a large integer. For 0\leqslant j\leqslant N , set t_{j}:=j{h} . Turning again to the differential equation ( 1 ), we can consider the following approximating system which contains the second order information Z as an interlacing step: \displaystyle Y^{h,1}_{t} \displaystyle=Y^{h,2}_{t_{j}^{-}}+f(Y^{h,1}_{t_{j}^{-}})X_{t_{j},t},\quad t\in(t_{j},t_{j+1}], (2) \displaystyle Y^{h,2}_{t} \displaystyle=Y^{h,1}_{t_{j+1}}+Z(Y^{h,1}_{t_{j+1}})_{t_{j},t},\quad t\in[t_{j},t_{j+1}), and appended with the initial condition Y^{h,2}_{0}=\overline{Y}\in\mathbb{R}^{d} . (Since the standard notation in rough paths theory is to put two temporal subscripts to indicate the difference X_{t,s}=X(t)-X(s) , as well as for more general functions taking in two temporal arguments, we emphasise that Z takes in three arguments, and Z(Y_{t})_{r,s} does not mean a difference Z(Y_{t})_{s}-Z(Y_{t})_{r} .) For a functon g continuous on (t_{j},t_{j+1}) , let g_{t_{j}^{+}} denote the limit to t_{j} from above. From the continuity of X , we find that \displaystyle Y^{h,1}_{t_{j}^{+}} \displaystyle=Y^{h,2}_{t_{j}^{-}},\qquad\text{and similarly, }\qquad Y^{h,2}_{t_{j}}=Y^{h,1}_{t_{j+1}}, assuming Z(\cdot)_{t,t}\equiv 0 . Associated with Y^{h,i} , i=1,2 , is an auxiliary solution Y^{h} that runs at twice the rate, and has the advantage of being continuous at the grid points t_{j} (at least, if Z is continuous in its third argument). For t\in(t_{j},t_{j+1}] , we define: \displaystyle Y^{h}_{t}:=\begin{cases}Y^{h,1}_{t_{j}+2(t-t_{j})}&t\in(t_{j},t_{j+1/2}]\\ Y^{h,2}_{t_{j}+2(t-t_{j+1/2})}&t\in(t_{j+1/2},t_{j+1}]\end{cases}, (3) where t_{j+1/2}:=t_{j}+\frac{1}{2}{h} . We have Y^{h}_{t_{j}}=Y^{h,1}_{t_{j}^{+}}=Y^{h,2}_{t_{j}^{-}} . From the first equation of ( 2 ), Y^{h,1} is as smooth as X in the interior of each interval of the partition. We can therefore hope that Y^{h} converges to a function in C^{\alpha}([0,T];\mathbb{R}^{d}) that can be understood as a solution to ( 1 ). Following [ 1 ] , we define solutions as follows: Definition 1.1 . A function Y\in C^{\alpha}([0,T];\mathbb{R}^{d}) is a solution to the RDE ( 1 ) if for any 0\leqslant s\leqslant t\leqslant T , \displaystyle Y_{s,t}-f(Y_{s})X_{s,t}-Z(Y_{s})_{s,t}=o(\left|t-s\right|). We delineate algebraic and analytic conditions on Z under which Y^{h} indeed converges and carry out the convergence argument in Section 2 . The application of the splitting method to stochastic equations \mathrm{d}u=F\,\mathrm{d}t+G\,\mathrm{d}W has also been widely studied (see, e.g., [ 4 , 8 , 10 ] ). An intuitive way to split up the problem is to consider \mathrm{d}u=F\,\mathrm{d}t and \mathrm{d}u=G\,\mathrm{d}W on separate time-steps. 1.3. Numerical analysis of RDEs Computational methods, especially explicit methods, of solving ( 1 ) are very natural in light of the sewing lemma, and have been studied by various authors, including recently, [ 2 , 12 ] . The associated literature on computational methods for SDEs is vast. In the standard reference [ 5 , Chapter 10] , a thorough study was made of local higher order Euler methods. Our construction is reminescent of the classical work [ 1 ] on which [ 5 , Chapter 10] is also based. All of our calculations can be extended without change to the case where Y takes values in \mathbb{R}^{n} , X takes values in \mathbb{R}^{d} , and f takes values in \mathbb{R}^{n\times d} . In [ 1 ] , Davie constructed solutions to ( 1 ) using a Milstein scheme, or a second-order Euler scheme, using the area \mathbb{X} . Following [ 11 ] , driving paths X of p -variation were considered in [ 1 ] , with 2\leqslant p<3 . Davie showed the more refined result that there is uniqueness only holds generally for f\in C^{p} . Two steps of a splitting scheme for ODEs added together is a slightly shifted Milstein scheme, interpolated between grid-points. We offer a different argument of convergence via the splitting scheme, which also provides a rate of convergence. As earlier alluded to, our splitting scheme, like Davie’s Milstein scheme, is explicit. As such it cannot witness non-uniqueness when \gamma<\frac{1}{\alpha}+1 . It may of interest in subsequent work to analyse the convergence of a splitting scheme that is implicit in its first equation in subsequent work. 2. Convergence and existence 2.1. Assumptions on Z As indicated, the primary object of this note is to use assumptions on Z to close estimates and give us convergence of approximating solutions to the operator splitting scheme in C^{\alpha}([0,T];\mathbb{R}^{d}) as h\downarrow 0 . We house our assumptions here and shall show their sufficiency in forthcoming calculations. (1) For every 0\leqslant s\leqslant t\leqslant T , uniformly in x\in\mathbb{R}^{d} , \displaystyle\left|Z(x)_{s,t}\right|\lesssim\left|t-s\right|^{2\alpha}. (4) (2) For every 0\leqslant s\leqslant t\leqslant T , x,y\in\mathbb{R}^{d} , \displaystyle\left|Z(x)_{s,t}-Z(y)_{s,t}\right|\lesssim\left|x-y\right|^{\gamma-2}\left|t-s\right|^{2\alpha}. (5) (3) For every 0\leqslant s\leqslant u\leqslant t\leqslant T , uniformly in x\in\mathbb{R}^{d} , define \delta Z(x)_{s,u,t}:=Z(x)_{s,t}-Z(x)_{s,u}-Z(x)_{u,t} as usual; we require \displaystyle\Big|\delta Z(x)_{s,u,t}-\big(\nabla f(x)f(x)X_{s,u}\otimes X_{u,t}+\nabla f(x)Z(x)_{s,u}X_{u,t}\big)\Big|\lesssim\left|t-s\right|^{3\alpha}. (6)