Randomisation of rough stochastic differential equations

Peter K. Friz, Khoa Lê, Huilin Zhang · 2025

Read Randomisation of rough stochastic differential equations by Peter K. Friz, Khoa Lê, Huilin Zhang (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-2503-06622.txt.

Introduction

Peter Friz, Khoa Lê, and Huilin Zhang study rough stochastic equations in a 2025 math paper. Such equations blend two traditions: the noise driven calculus of Itô and Lyons's pathwise rough paths. They now serve fields from filtering and control to finance and systems with shared noise. The trio take a unified view of rough Itô processes and ask a sharp question.

When does such a process, once randomized, turn into a doubly stochastic Itô process? And what can be said about its conditional laws in that case? Answers would bridge the pathwise and probabilistic camps that grew up apart. The work aims to give users of these equations in applied fields a firmer shared base.

Catalog record

CREATORPeter K. Friz, Khoa Lê, Huilin Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-06622

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Randomisation of rough stochastic differential equations Peter K. Friz, Khoa Lê, Huilin Zhang Abstract Rough stochastic differential equations (RSDEs) are common generalisations of Itô SDEs and Lyons RDEs and have emerged as new tool in several areas of applied probability, including non-linear stochastic filtering, pathwise stochastic optimal control, volatility modelling in finance and mean-fields analysis of common noise system. We here take a unified perspective on rough Itô processes and discuss in particular when and how they become, upon randomisation, “doubly stochastic” Itô processes, and what can be said about their conditional laws. 1 Introduction Inspired by the recent theory of rough stochastic differential equations (RSDEs) of [ 21 ] , rough Itô processes are continuous adapted stochastic processes, on some stochastic basis (\Omega,\mathfrak{F},(\mathfrak{F}_{t})_{t\geq 0},\mathbb{P}) , of the form X=X(\omega,\mathbf{Y})=X^{\mathbf{Y}}(\omega) with dynamics 1 1 1 All processes are multidimensional; extension to infinite dimensions are possible but not the purpose of this note. dX_{t}=A_{t}(\omega,\mathbf{Y})dt+\Sigma_{t}(\omega,\mathbf{Y})dB_{t}+(F_{t},F^{\prime}_{t})(\omega,\mathbf{Y})d\mathbf{Y}_{t},\quad 0\leqslant t\leqslant T, (1.1) where B=B(\omega) is Brownian motion and \mathbf{Y}=(Y,\mathbb{Y})\in\mathscr{C}{\equiv\mathscr{C}_{T}\subset\mathscr{C}^{\alpha}([0,T],\mathbb{R}^{d_{Y}})} , a class of deterministic rough paths, on a fixed Hölder scale with exponent \alpha\in(1/3,1/2] . All coefficients are assumed progressive in (t,\omega) and (F_{t},F^{\prime}_{t})(\omega,\mathbf{Y}) is a stochastic controlled rough path in the sense of [ 21 ] ; this allows to give bona fide integral meaning to ( 1.1 ) where the reader can also find an Itô formula for such processes. ( 1.1 ) generalizes classical Itô-diffusions X=X(\omega) , with dynamics dX_{t}=a_{t}(\omega)dt+\sigma_{t}(\omega)dB_{t}, which of course accommodates solutions to Itô stochastic differential equations, dX_{t}=b_{t}(X_{t},\omega)dt+\sigma_{t}(X_{t},\omega)dB_{t}. From a rough path perspective, ( 1.1 ) generalizes rough integration, with dynamics dX_{t}=(F_{t},F^{\prime}_{t})d\mathbf{Y}_{t}\Leftrightarrow(\delta X)_{s,t}\approx_{3\alpha}F_{s}(\delta Y)_{s,t}+F^{\prime}_{s}\mathbb{Y}_{s,t},(\delta F)_{(s,t)}\approx_{2\alpha}F_{s}^{\prime}(\delta Y)_{s,t}, (1.2) where (F,F^{\prime})=(F,F^{\prime})(\mathbf{Y}) is a given Y -controlled rough path. 2 2 2 By \Psi\approx_{\gamma}\Phi we mean |\Psi_{s,t}-\Phi_{s,t}|\leq C|t-s|^{\gamma} , for some constant C , uniformly on [0,T] . This accommodates solutions to “contextualized” (by which we mean explicit \mathbf{Y} -dependence in the coefficient fields) rough differential equation (RDE) dX_{t}=f(X_{t};\mathbf{Y})d\mathbf{Y}_{t} (1.3) via the usual Davie expansion F_{t}:=f(X_{t};\mathbf{Y}),\ F^{\prime}_{t}:=((Df)f)(X_{t};\mathbf{Y}) . This also accommodates the case of regular time-dependence f(t,X_{t};\mathbf{Y}) in ( 1.3 ); and more generally RDEs of the form dX_{t}=(f_{t},f_{t}^{\prime})(X_{t};\mathbf{Y})d{\mathbf{Y}_{t}}, (1.4) in which case F_{t}=f_{t}(X_{t};\mathbf{Y}),\ F^{\prime}_{t}=((Df_{t})f_{t}+f_{t}^{\prime})(X_{t};\mathbf{Y}) . Our main motivation for ( 1.1 ) comes from the recent theory of rough SDEs [ 21 ] dX_{t}=b_{t}(X_{t},\omega;\mathbf{Y})dt+\sigma_{t}(X_{t},\omega;\mathbf{Y})dB_{t}+(f_{t},f^{\prime}_{t})(X_{t},\omega;\mathbf{Y})d\mathbf{Y}_{t}, (1.5) solutions X=X^{\mathbf{Y}} of which provide natural examples of rough Itô processes, with (F_{t},F^{\prime}_{t})(\omega,\mathbf{Y})=(f_{t}(X_{t},\omega;\mathbf{Y}),\ ((Df_{t})f_{t}+f_{t}^{\prime})(X_{t},\omega;\mathbf{Y})). Extensions to mean-field a.k.a. McKean-Vlasov rough SDEs have been developed in [ 22 ] . The purpose of this note is to systematically relate the “hybrid” rough stochastic processes seen above to their “doubly” stochastic counterparts, when \mathbf{Y} above is replaced by some stochastic driving noise, often (but not always) an independent Brownian noise (seen as rough path). Conversely, one can start with doubly stochastic situations and ask what happens under partial conditioning of the noise. To the best of our knowledge, such questions were first raised in the context of filtering [ 16 ] . In Section 2 , we give motivation from non-linear stochastic filtering, volatility modelling in finance, pathwise stochastic optimal control and mean-fields SDE with common noise system; topics also closely related to [ 11 , 12 ] . We do not strive for precise statements, but try to explain how rough stochastic processes naturally emerge, together with motivation for what is to come: Section 3 deals with the mathematics of randomisation of rough Itô processes, one obstacle being the potentially uncountable number of rough paths that figure as parameter in ( 1.1 ). The remaining section discuss applications to RSDE and then also mean-field RSDEs. The body of work surrounding RSDEs, control and common noise also resonates with early remarks made by the authors of [ 10 ] , conjecturing the use of rough paths for conditioned forward stochastic dynamics (in their context of mean field games with common noise). Acknowledgement : PKF and HZ acknowledge support from DFG CRC/TRR 388 “Rough Analysis, Stochastic Dynamics and Related Fields”, Projects A07, B04 and B05. Part of this work was carried out during a visit of the first author to Shandong University. KL acknowledges supports from EPSRC [grant number EP/Y016955/1]. HZ is partially supported by the Fundamental Research Funds for the Central Universities, NSF of China and Shandong (Grant Numbers 12031009, ZR2023MA026), Young Research Project of Tai-Shan (No.tsqn202306054). 2 Motivating examples 2.1 Non-linear stochastic filtering We follow [ 16 , 8 ] . One has an observation process Y_{t}=\int_{0}^{t}h(X_{s},Y_{s})ds+B^{\perp}_{t} where the signal process X is assumed to have “correlated” dynamics, \displaystyle dX_{t} \displaystyle=\tilde{b}(t,X_{t},Y_{t})dt+\sigma(t,X_{t},Y_{t})dB_{t}(\omega)+f(t,X_{t},Y_{t})dB^{\perp}_{t}(\omega) \displaystyle=b(t,X_{t},Y_{t})dt+\sigma(t,X_{t},Y_{t})dB_{t}(\omega)+f(t,X_{t},Y_{t})dY_{t}(\omega). (2.1) Here, (B,B^{\perp}) and (B,Y) are independent Brownians under some original measure \mathbb{P}^{o} and \mathbb{P} , respectively, related by the Girsanov formula \displaystyle\left.\frac{d\mathbb{P}^{o}}{d\mathbb{P}}\right|_{\mathcal{F}_{t}}=\exp\left(\int_{0}^{t}h(s,X_{s},Y_{s})dY_{s}-\frac{1}{2}\int_{0}^{t}|h(s,X_{s},Y_{s})|^{2}ds\right)=:\exp(I_{t}). Formally at least, the conditional dynamics of X_{t} given the observation \{Y_{t}:0\leqslant t\leqslant T\} should be captured by the solution X^{\mathbf{Y}} to a rough SDE of the form ( 1.5 ) 3 3 3 Assuming regular t dependence in f . dX^{\mathbf{Y}}_{t}=b(t,X^{\mathbf{Y}}_{t},Y_{t})dt+\sigma(t,X^{\mathbf{Y}}_{t},Y_{t})dB_{t}(\omega)+(f,D_{y}f)(t,X^{\mathbf{Y}}_{t},Y_{t})d\mathbf{Y}_{t}, (2.2) together with the “rough stochastic” Girsanov exponent I_{t}^{\mathbf{Y}}=\int_{0}^{t}(h,(D_{x}hf+D_{y}h))(s,X^{\mathbf{Y}}_{s},Y_{s})d\mathbf{Y}_{s}-\frac{1}{2}\int_{0}^{t}|h(s,X^{\mathbf{Y}}_{s},Y_{s})|^{2}ds. Bearing exponential integrability, one defines a flow of measures \langle\mu^{\mathbf{Y}}_{t},\varphi\rangle:=\mathbb{E}[\varphi(X^{\mathbf{Y}}_{t})\exp(I^{\mathbf{Y}}_{t})], whose randomisation \bar{\mu}_{t}(\omega):=\mu^{\mathbf{Y}}_{t}|_{\mathbf{Y}=\mathbf{Y}^{\operatorname{Ito}}(\omega)} should satisfy \langle\bar{\mu}_{t}(\omega),\varphi\rangle:=\mathbb{E}[\varphi(X_{t})\exp(I_{t})|\mathcal{F}^{Y}_{t}]. The interest in this construction is a rough path factorisation of the Kallianpur-Striebel formula, {\mathbb{E}^{o}}[\varphi(X_{t})|\mathcal{F}^{Y}_{t}]=\left.\frac{\langle\mu^{\mathbf{Y}}_{t},\varphi\rangle}{\langle\mu^{\mathbf{Y}}_{t},1\rangle}\right|_{\mathbf{Y}=\mathbf{Y}^{\operatorname{Ito}}(\omega)}. A suitable class of rough paths, for 0\leqslant t\leqslant T , is the (Polish) rough path space \mathscr{C}^{0,2\alpha}([0,T],\mathbb{R}^{d_{Y}}) . In fact, one can also work with (cf. Section 3.1 for notation) \mathscr{C}_{T}=\{\mathbf{Y}\in\mathscr{C}^{0,2\alpha}([0,T]):[\mathbf{Y}]_{s,t}=I_{d_{Y}\times d_{Y}}\times(t-s),0\leqslant s\leqslant t\leqslant T\} to hard-wire the bracket of the observation (rough) path to be consistent with the quadratic variation of the observation process: \langle Y\rangle_{s,t}=\langle B^{\perp}\rangle_{s,t}=I_{d_{Y}\times d_{Y}}\times(t-s) . Equivalently, one can recast (X,I) above in Stratonovich form w.r.t. Y ; as was done in [ 16 ] to study Clark’s robustness problem. The appropriate randomisation is then \mathbf{Y}=\mathbf{Y}^{\operatorname{Strato}}(\omega) , so that one can take \mathscr{C}_{T}=\mathscr{C}^{0,2\alpha}_{g}([0,T],\mathbb{R}^{d_{Y}}) , the space of geometric rough paths. We mention that under natural assumptions, (\mu^{\mathbf{Y}}_{t}) is the unique solution to an intrinsically defined rough Zakai equation [ 5 , 8 ] . (Specialists may appreciate the fact that uniqueness results can be obtained by forward-backward arguments, without forward-backward filtration issues that arise in presence of Y -dependent coefficients in ( 2.1 ); cf. [ 17 ] for a recent discussion.) 2.2 Pathwise stochastic optimal control We follow many authors (Davis-Burstein, Lions-Souganidis, Buckdahn-Ma, Rogers, … ), leaving detailed references to [ 23 ] , in investigating the random field V(t,x;\omega):=\operatorname{essinf}_{\eta(.)}\mathbb{E}^{t,x}\left(g\left({X_{T}^{\eta}}\right)|\mathcal{F}_{T}^{W}\right), (2.3) where X^{\eta} has controlled stochastic dynamics 4 4 4 As noted in [ 19 ] , one must not control the final coefficient field f for otherwise the problem degenerates. dX^{\eta}_{t}=b(t,X^{\eta}_{t};\eta_{t})dt+\sigma(t,X^{\eta}_{t};\eta_{t})dB_{t}+f(t,X^{\eta}_{t})dW_{t}. (2.4) In some formal sense at least, the conditional dynamics of X^{\eta} , given some fixed environmental noise \{W_{t}:0\leqslant t\leqslant T\} , should be captured by the solution X^{\eta,\mathbf{Y}} to the controlled rough SDE 5 5 5 As explained in ( 1.3 ),( 1.4 ), having f^{\prime}\equiv 0 , amounts to a regular t dependence in f which is assumed here. dX^{\eta,\mathbf{Y}}_{t}=b(t,X^{\eta,\mathbf{Y}}_{t};\eta_{t})dt+\sigma(t,X^{\eta,\mathbf{Y}}_{t};\eta_{t})dB_{t}+(f,0)(t,X^{\eta,\mathbf{Y}}_{t})d\mathbf{Y}_{t}. (2.5) With \eta=\eta_{t}(\omega;\mathbf{Y}) these are RSDEs of the form ( 1.5 ). One reverts to a deterministic value function V(s,x;\mathbf{Y}):=\inf_{\eta(.)}\mathbb{E}^{s,x}\left(g\left({X_{T}^{\eta,\mathbf{Y}}}\right)\right), (2.6) with immediate regularity results (derived from estimates in [ 21 ] for solutions of RSDEs), validity of a dynamic programming principle, existence of \varepsilon -optimal controls \eta=\eta_{t}(\omega;\mathbf{Y}) etc. Rough HJB equations for ( 2.6 ) have been considered several

Read the complete text