We construct sub-grid scale models of incompressible fluids by considering expectations of semi-martingale Lagrangian particle trajectories. Our construction is based on the Lagrangian decomposition of flow maps into mean and fluctuation parts, and it is separated into the following steps. First, through Magnus expansion, the fluid velocity field is expressed in terms of fluctuation vector fields whose dynamics are assumed to be stochastic. Second, we use Malliavin calculus to give a regularised interpretation of the product of white noise when inserting the stochastic velocity field into the Lagrangian for Euler's fluid. Lastly, we consider closures of the mean velocity by making stochastic analogues of Talyor's frozen-in turbulence hypothesis to derive a version of the anisotropic Lagrangian averaged Euler equation.
Introduction
The modelling of multi-scale fluid flows is a well-known problem in fluid dynamics. In previously established work, fluid flows exhibiting multiple spatial scales have typically been modelled by considering averaging techniques and closures to capture the effects of the unresolvable scales on large-scale dynamics. These models are typically referred to as sub-grid scale models. One such model is the Generalised Lagrangian Mean (GLM) model introduced by Andrews & McIntyre [1] that considered the application of an arbitrary averaging operator to the Lagrangian expansion of ideal incompressible fluid equations and it is later closed in [10]. To preserve the geometric structure of ideal fluid equations during averaging, one option is to apply averaging to the variational formulation of ideal fluid flows, instead of applying to the equations directly. This approach resulted in several models such as the Lagrangian averaged Navier Stokes model [9]. More recently, fluid flows that exhibit multiple temporal scales have been modelled by stochastic partial differential equations where the evolution of the fast dynamics is represented by integration against driving stochastic processes such as Brownian motion and geometric rough paths. One class of structure-preserving stochastic perturbation, known as transport noise, is constructed by assuming a semi-martingale decomposition of the Lagrangian fluid flow map gt∈Diff(M) (for fluid domain M) in a stochastic variational principle, see, e.g., [13, 8].
Stochastic fluid flows naturally induce a type of ensemble averaging operator. This is the expectation operator E(⋅) over the probability space (Ω,F,P) that supports the driving stochastic processes. E is linear, a projection and commutes with spatial derivative operations. Consequently, the expectation satisfies the required axioms of an average seen in GLM theory [1]. Additionally, one has the analogue of the zero mean fluctuations under expectation operator as martingales. More precisely, for a Rd-martingale Yt, the martingale property of Itô integral implies that
E[∫0tζsidYsi]=0,
for any Ft-adapted vector fields ζt=(ζti)i=1d∈X(M).
In this work, we consider the averaging of stochastic fluid models using the expectation operator to construct probabilistic variants of sub-grid scale models for ideal fluids. Focusing on incompressible flows, the starting assumption is the semi-martingale Lagrangian fluid flow map gt∈SDiff(M) similar to the derivation of stochastic transport fluid models in [13, 16]. The main contributions of this work are as follows. In section 2, we define the mean velocity by assuming a factorisation of the full flow map into the fluctuation and mean maps. Then, we use Magnus expansion of the velocity field of the composite flow such that it is expressed by vector fields and their actions only, in a similar fashion to [15, 18]. This way, the expectation operator acts on vector space quantities only and can be easily defined. In section 3, using tools from Malliavin calculus, we consider a normalised KE Lagrangian functional such that it is well-defined when the input velocity contains white noise terms. To obtain a dynamical equation of the mean flow, we first make dynamical assumptions on the fluctuations vector fields in assumptions 1 and 2. Then, we explicitly computed the averaged Lagrangian and derived a stochastically averaged version of the anisotropic Lagrangian averaged Euler’s equation [15] through an Euler-Poincaré variational principle.
Expansion of Kinetic Energy Lagrangian
Fix a smooth, compact, boundary-less Riemannian manifold (M,g) of dimension d. Throughout this paper we equip (M,g) with the Levi-Civita connection ∇. Our aim is to derive deterministic dynamical equations for the mean flow velocity of Euler’s fluid flow from an averaged variational principle of the form,
where ℓ is the kinetic energy Lagrangian of divergence free vector fields. Here, we have denoted μ=volg the Riemannian volume. The covariant divergence div∇ appearing in (1) and used throughout is defined for any tensor T∈T(M) by div∇(T):=tr(∇T) by contracting the covariant ∇i index with, by convention, the first (possibly after raising) contravariant index of T. In the case of vector fields1, X∈Xdiv∇(M) implies that
We assume the full Lagrangian flow map gt∈SDiff(M) is a semi-martingale satisfying the stochastic differential equation,
dgt=ut(gt)dt+σt(gt)∘dWt,(2)
such that vt is formally associated with the vector field dtdgtgt−1∈Xdiv∇(M). Throughout this paper, we denote by Wt=(Wt1,…,Wtn) as a n-dimensional Brownian motion with identity covariance and abbreviate stochastic differentials such as σt(gt)∘dWt:=∑i=1nσti(gt)∘dWti. To define the mean flow velocity, we assume the factorisation of the stochastic flow map gt∈SDiff(M) into
gt=gtϵ=Ξtϵ∘gˉt,(3)
where gˉ:[0,T]→SDiff(M) is the mean flow map and Ξˉtϵ is a ϵ-parametrised semi-martingale flow map Ξϵ:Ω×[0,T]×[0,1)→SDiff(M), with following properties
Ξˉt0=e∀t∈[0,T],Ξ0ϵ=e∀ϵ∈[0,1),
such that Ξtϵ is a flow of near identity diffeomorphism. From the mean flow decomposition (2), the canonical definition of the mean flow velocity field is the velocity assumed with the mean flow map uˉt:=dtdgˉtgˉt−1. By assumption, uˉt is a deterministic object and it is natural to seek the dynamics of uˉt through the variational principle (1) where the stochastic fluctuation induced by the fluctuation map Ξtϵ is averaged. Via direct computation, one finds the vector field generated by gtϵ can be written in terms of the mean velocity uˉt
where the notation ∘dgtϵ is there to denote the Stratonovich interpretation of the stochastic differential dgtϵ. We observe that SDiff(M) does not necessarily possess a vector space structure and the term AdΞtϵuˉt is non-linear in Ξtϵ. In order to express terms of the form E[AdΞtϵuˉt] in the averaged Lagrangian in a workable form, we rewrite the fluctuations of uˉt defined in powers of ϵ via Taylor expansion as,
such that expectation is applied to ∘dgtϵgtϵ;−1∈X(M). The (formal) Lie algebra of divergence free vector fields is a linear space where these operations can be performed unambiguously. In the expansion (5), we will express the first and second order fluctuation fields ∘dv′ and ∘dv′′ from the deformations of Ξtϵ
Here, the deformation vector fields wt and χt are assumed to be semi-martingales which possess a unique decomposition into finite variation and (local) martingale parts.
Lemma 1. In terms of the deformation vector fields wt and χt, the fluctuation stochastic vector fields dv0, dv′ and dv′′ can be expressed as
To calculate higher order terms, we differentiate the decomposition, where the first term is a direct application of the definition of the adjoint representation (Lie chain rule),
To take the derivative of ∘dΞtϵΞtϵ;−1, we express Ξtϵ=exp(Ωtϵ) and differentiate Ωtϵ through the chain rule on the exponential map. This procedure is known as a Magnus expansion [14]. Although we work formally, the application of this type of theory to the infinite dimensional group of volume preserving diffeomorphisms can be found in [2, 4, 5]. We state the following formulae that can be obtained through known derivatives of the exponential map at the identity [11],
The time derivative ∘dΩtϵ is expressible as the series [3]
∘dΩtϵ=k=0∑∞k!BkadΩtϵk(∘dΞtϵΞtϵ;−1),(11)
where Bk denotes the Bernoulli numbers with B1=−21. From the initial assumption Ξt0≡e it must be the case that Ωt0≡∘dΩt0≡∘dΞt0≡0 from differentiating a constant.
Using the Magnus generator Ωtϵ, we can now compute first ϵ-derivatives of the velocity dΞtϵΞtϵ;−1 by termwise differentiation of the power series (10),
Note that only the k=0 term contributes due to the fact Ωt0=∘dΩt0=0. We may relate δΩt to the first order fluctuation wt using the k=0 term of the Magnus series (11),
In order to compute the second order fluctuation of dvϵ we require the connection to take second derivatives. We first differentiate AdΞtϵu through twice application of the Lie chain rule, where the shorthand (⋅)′ denotes ∂ϵ∂(⋅),
@page 3/9 starts_mid=0 ends_mid=1 where ξtϵ;′′=Ξtϵ;′′Ξtϵ;−1−∇Ξtϵ;′′Ξtϵ;−1Ξtϵ;′′. In particular, the term ∇wtwt arises when one attempts to compute a subsequent variation of Ξtϵ;′′Ξtϵ;−1. With the Magnus expansion, one can write this term as δ2Ωt without use of a connection, however a connection is required to relate this quantity to the fluctuations χt,wt, see the Remark 1 below.
For the second derivative of dΞtϵΞtϵ;−1 it suffices to check the k=0,1 terms in the series (11),
Remark 1 (Connection dependence and matrix Lie groups). The expressions (12), (14) define equations on any Lie algebra in the absence of any additional structure. As a consequence of ∘dΞt0=0, (13) can always be used to relate the variations of Ω and Ξ. The second order term δ2Ωt differs, as we must define an action of Lie algebra terms using external structures such as a connection. In order to proceed, one formally writes,
When g is finite dimensional, one may appeal to Ado’s Theorem to embed g⊂GL(N) for some N∈N and give meaning to δΞtδΞt−1=−δΞtΞt0;−1δΞtΞt0;−1=−δΞtδΞt=−w2 as a product of matrices. One can then show the second order expansion cancels significantly as,
under the property that ad reduces to the commutator of GL(N) matrices within this embedding. In the case of the diffeomorphism group, a concrete representation of (14) in terms of δΞt requires identifying δΞtδΞt−1:=−∇wtwt as a covariant derivative. In Euclidean space, we recover consistency with the ordinary chain rule where the connection is canonical,
In the above, we have denoted by ∇E the Euclidean connection and D the total (Frechêt) derivative for maps Rn↦Rn, with the second equality was derived from the inverse function theorem. In curved geometries, the choice of ∇ is non-canonical, although the Levi-Civita connection is a natural choice we take throughout.
Lemma 2.The second order fluctuation vector fieldsdv′′may be written equivalently as
which follows directly from the definition of being a torsion-free connection and the Leibniz property. Here, we have the covariant Hessian operator ∇∇ defined on tensor T∈T(M) of type (p,q) by ∇∇T=∇i∇jTdxi⊗dxj. Inserting (17) into the equation for dv′′ thus gives,
The curvature term R(u,w)w vanishes in flat geometries, but otherwise may be expressed as R(u,w)w=R(u,⋅):(w⊗w) as a contraction with a (2,1) tensor field R(u,⋅)=Rijαβuidxj⊗dxα⊗∂xβ∂ with a (2,0) tensor field w⊗w.
With the alternative formulation of Lemma 2, the full expansion of ∘dgtϵgtϵ;−1 to O(ϵ2) is given by,
Remark 2 (Expectation of advected quantities). Advected quantities at∈T(M) of the mean flow ut can also be perturbed in similar fashion to the perturbed stochastic velocity field ∘dgtϵgtϵ;−1 and expanded in (18) by considering the quantity atϵ:=Ξt∗ϵat=gt∗ϵa0. Applying the standard Lie chain rule in the ϵ-parameter,
dϵd(Ξt∗ϵatϵ)=Ξt∗ϵdϵdatϵ−LΞtϵgtϵ;−1atϵ,
where L:X(M)×T(M)→T(M) is the Lie derivative operator, we have the expansion
The direct substitution of the expansion (18) into the Lagrangian (1), dropping all cubic and higher order terms, results in the following formal action,
Here, I1 is the kinetic energy of the mean velocity ut; I2 and I3 are mixed products with stochastic fluctuations wt and χt, of the mean velocity ut, understood as stochastic integrals. The kinetic energy of the fluctuations I4 is ill-defined due to terms representing the formal product of white noise and has no classical interpretation.
In existing literature of stochastic variational principles where the action contains explicit integration against semi-martingales, e.g., [4, 7, 13, 12], terms such as I4 are typically dropped. Our approach is to instead retain these terms under an average, which allow for a renormalisation of I4 into a meaningful quantity. The method we employ is to interpret I4 using the Wick product, which defines products of white noise through Malliavin calculus.
We very briefly discuss this construction. Since our usage is only elementary, we do not dwell substantially on definitions. For precise details, we refer the reader to [?] and [17]. One defines a white noise probability space(Ω,B,μ) of random distributions inside a sample space of Rn valued tempered distributions Ω:=S∗(R+;Rn). A probability measure and Borel sigma algebra exists on this vector space via the Bochner-Milnos theorem.
One may then define L2(Ω)-integrable random variables taking values in the Schwartz class which may be characterised by convergent Itô-Weiner chaos expansions. The elements of L2(Ω) with a rapidly decreasing series expansion are known as the Hida test functions S(Ω), and elements its dual space S∗(Ω) are known as Hida distributions. Within this space, a distribution valued stochastic process W˙t may be defined. Crucial for us is that S∗(Ω) contains a closed, associative, commutative product that is defined by the formal Cauchy product of the series defined in an L2(Ω) orthogonal basis,
The Wick product satisfies several desirable properties, firstly, E[S⋄T]=E[S]⋅E[T], where S,T∈S∗(Ω) need not be independent random variables. Secondly, if either S,T is deterministic (i.e. S(t)=E[S](t):=c0(t)), then S⋄T=S⋅T reduces to the ordinary product (well defined as a chaos expansion times a constant). Finally, is the correspondence of an Itô diffusion with its S∗(Ω) valued “Wick calculus" representation
dXt=u(Xt)dt+σ(Xt)dWt⟺X˙t=u(Xt)+σ(Xt)⋄W˙t.
which can be proven rigorously using the Skorohod integral from Malliavin calculus. Interpreting white noise as a Hida distribution value quantity, W˙∈S∗(Ω) with E[W˙t]=0, we choose to interpret the white noise products of I4 in the Wick sense. With this choice the expectation distributes to cancel I4. In the other integrals I1,I2,I3 we have at least one deterministic factor, making the distinction between Wick and ordinary products trivial. Thus, we will proceed by making assumptions to the dynamics of wt and χt and take averages of all quantities.
Assumption 1 (Additive stochastic Taylor hypothesis)
We assume that the dynamics of the first order fluctuation vector field wt is driven by Brownian motion additively such that it can be written as
∘dwt+adwtutdt=ξ∘dWt,(21)
for some ξ∈Xdivn(M) that are prescribed, divergence free and temporally constant.
In this case, the Itô and Stratonovich interpretations coincide,
ξ∘dWt:=i=1∑nξi∘dWti=i=1∑nξidWti=ξdWt.
In particular, taking the expectation of (21) implies that E[wt]=:wt is advected by the mean flow ut
∂twt+adwtut=0.
which is equivalent for wt to satisfy Taylor’s frozen turbulence hypothesis in expectation. From the additive stochastic Taylor hypothesis (21), via direct calculation we find the dynamics of the fluctuation correlation tensor Ft:=E[wt⊗wt]∈X(M)⊗X(M) as
∂tFt+LutFt=(ξ⊗ξ),(22)
in geometric notation. In terms of index notations, it can be written as
Here, the Stratonovich to Itô correction for the additive stochastic Taylor hypothesis results in higher-order terms that are dropped in the expansion by ϵ. Noting that there is no canonical choice when applying the expectation, one may apply it to the velocity decomposition expressed on ∂tgtϵ∈TgtϵSDiff(M) or on vtϵ=∂tgtϵ,−1=TeSDiff(M)≃X(M), which corresponds to the Lagrangian and Eulerian expressions of the fluid velocity respectively. In taking the expectation of the Euler-Poincaré variational principle, we make the modelling choice to apply the expectation directly to the vector field vtϵ at fixed Eulerian coordinates. Thus, we obtain the expected mean flow expansion in terms of ϵ,
Assumption 2 (Expectation orthogonality hypothesis) We assume that the material derivative of the expectation of the second order fluctuation vector field χt is orthogonal to the mean flow uˉt under metric pairing. That is
∫Mg(uˉt,∂tχˉt+∇uˉtχˉt)=0,(25)
where χˉt:=E[χt]. Noting that uˉt is assumed to be deterministic and E[⋅] commutes with spatial derivatives and integrals, the condition (25) can be rewritten as
∫Mg(uˉt,∘dχt+∇uˉtχtdt)=dMt,
where Mt∈R is a martingale.
Under the assumptions [1] and [2], substituting the white noise velocity expansion (23) into the white noise Lagrangian (20) we have
Here, we have used the property that E[b(Xt)⋄W˙t]=0 which is analogous to the martingale property to Itô integrals in the second equality and the orthogonality condition (25) in the third equality. Noting that χˉt+div∇Ft=E[χt+div(wt⊗wt)]=E[χt+∇wtwt]∈Xdiv(M), through integration by parts and appropriate boundary conditions we have
Using the Euler-Poincaré constrained variations for uˉt, δuˉt=∂tuˉt−adηtuˉt for arbitrary ηt∈Xdiv∇(M) vanishing at the boundaries, Hamilton’s principle δS=0 yields the following Euler-Poincaré equation
Here, d:Λk(M)→Λk+1(M) is the exterior derivative, div∇(∇uˉt♭⋅Ft):=∇j(∇kuˉt♭Ftjk), and the pressure contribution arises from the equivalence class of the dual space to the space divergence free vector fields. The Lie derivative Luˉ acting on m∈Λ1(M) can be expressed using covariant derivatives as
Luˉm=∇uˉm+(∇uˉ)Tm.
Together with the evolution of F derived from the stochastic Taylor hypothesis [1], we have the closed system of equations
which are the forced anisotropic Lagrangian averaged Euler equations. Several simplification can be made by specialising to flat geometries such that curvature R vanishes. In this case, we have the following proposition.
Proposition 1.Assuming the correlation tensor Ft has the dynamics
∂tFt+LuˉtFt=Gt,(29)
for arbitrary, time dependent Gt∈X(M)⊗X(M) which can have functional dependence on uˉt, wt and Ft. Denoting by C(q)=div∇(∇q⋅Ft) for all arbitrary vector qt∈X(M), we have the commutation relation
[∂t+∇uˉt,C]qt=div∇(∇qtGt).(30)
Proof. In the case ∂tFt+LuˉtFt=0, it was shown in e.g. [18, 15], that [∂t+uˉt⋅∇,C]qt=0 for Cu=div∇E(∇u⋅Ft) for Euclidean connection ∇E. Since the contribution from Gt enters only in term ∂t(C(qt))=∂tdiv∇(∇qt⋅Ft), we have the required result through a similar direct computation and switching of covariant derivatives, which always holds under the assumption that R≡0.
Corollary 1.When the curvature R vanishes, the Euler-Poincaré equation can (27) be expressed equivalently as
where p=p+21∇(∣uˉt∣2) is the modified pressure.
Conclusion
In this work, we considered an averaged Euler-Poincaré variational principle to construct sub-grid scale ideal fluid models. In this construction, the total fluid flow was assumed stochastic and averaged through the expectation in analogy with the GLM theory. The first step of this construction, made in section 2, is the formulation of the mean velocity, defined by the factorisation of the total flow map into the fluctuation and mean maps. Then, the Magnus expansion of the stochastic flow map is considered so that the expectation operator is applied to vector fields and their actions. Our averages then occurred only at the tangent space of the identity, bypassing the difficulty in defining the expectation of diffeomorphism group-valued quantities. In section 3, we generalised the KE Lagrangian functional such that it is well-defined when the input velocity contains white noise terms using tools from Malliavin calculus. After making dynamical assumptions on the fluctuations vector fields in assumptions 1 and 2, we explicitly computed the averaged Lagrangian and derived a stochastically averaged version of the anisotropic Lagrangian averaged Euler’s equation through an Euler-Poincaré variational principle. Following this work, several open problems present themselves to be addressed in the future. First, the closure dynamics of wt and χt can be modified to include additional physics. For example, one can assume transport type stochastic perturbation for the dynamics of wt, dwt+adwtut+adwtξ∘dWt=0 as to introduce dissipation into the dynamics of the correlation tensor Ft. Second, the driving stochastic processes can be generalised from Brownian motions to geometric rough paths. For example, fractional Brownian motion or lifts of Brownian motion that incorporate a noise-induced drift. Such drifts are natural in the context of multiscale limits [6]. Lastly, the same modelling procedure can be generalised from incompressible Euler’s fluid using SDiff(M) to semi-direct products such that sub-grid scale models for a large class of fluid flows can be derived.
Acknowledgments. We wish to thank D. Holm, O. Street and J. Woodfield for several thoughtful suggestions during the course of this work. TD is supported during the present work by European Research Council (ERC) Synergy grant Stochastic Transport in Upper Ocean Dynamics (STUOD) – DLV-856408. RH is grateful for the support by Office of Naval Research (ONR) grant award N00014-22-1-2082, Stochastic Parameterization of Ocean Turbulence for Observational Networks.
Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article.
References
[1]Andrews, D.G., Mcintyre, M.E.: An exact theory of nonlinear waves on a Lagrangian-mean flow. Journal of Fluid Mechanics 89(4), 609–646 (1978). https://doi.org/10.1017/S0022112078002773
[2]Arnaudon, M., Chen, X., Cruzeiro, A.B.: Stochastic Euler-Poincaré reduction. Journal of Mathematical Physics 55 (2014). https://doi.org/10.1063/1.4893357
[3]Blanes, S., Casas, F., Oteo, J., Ros, J.: The magnus expansion and some of its applications. Physics Reports 470(5), 151–238 (2009). https://doi.org/10.1016/j.physrep.2008.11.001
[4]Chen, X., Cruzeiro, A.B., Ratiu, T.S.: Stochastic Variational Principles for Dissipative Equations with Advected Quantities. Journal of Nonlinear Science 33, 5 (2 2023). https://doi.org/10.1007/s00332-022-09846-1
[5]Cipriano, F., Cruzeiro, A.B.: Navier-Stokes equations and diffusions on the group of homeomorphisms of the torus. Communications in Mathematical Physics 275, 255–269 (10 2007). https://doi.org/10.1007/s00220-007-0306-3
[6]Cotter, C.J., Gottwald, G.A., Holm, D.D.: Stochastic partial differential fluid equations as a diffusive limit of deterministic Lagrangian multi-time dynamics. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 473(2205), 20170388 (2017). https://doi.org/10.1098/rspa.2017.0388
[7]Diamantakis, T., Holm, D.D., Pavliotis, G.A.: Variational Principles on Geometric Rough Paths and the Lévy Area Correction. SIAM Journal on Applied Dynamical Systems 22, 1182–1218 (6 2023). https://doi.org/10.1137/22M1522164
[12]Holm, D.D., Hu, R.: Stochastic effects of waves on currents in the ocean mixed layer. Journal of Mathematical Physics 62 (2021). https://doi.org/10.1063/5.0045010
[13]Holm, D.D.: Variational principles for stochastic fluid dynamics. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences 471(2176), 20140963 (2015)pubmed.ncbi.nlm.nih.gov/27547083
[14]Magnus, W.: On the exponential solution of differential equations for a linear operator. Communications on Pure and Applied Mathematics 7(4), 649–673 (1954). https://doi.org/10.1002/cpa.3160070404
[15]Marsden, J.E., Shkoller, S.: The anisotropic lagrangian averaged euler and navier-stokes equations. Archive for Rational Mechanics and Analysis 166, 27–46 (1 2003). https://doi.org/10.1007/s00205-002-0207-8
[17]Nunno, G.D., Øksendal, B., Proske, F.: Malliavin Calculus for Lévy Processes with Applications to Finance. Springer Berlin Heidelberg, Berlin, Heidelberg (2009). https://doi.org/10.1007/978-3-540-78572-9
[18]Shkoller, S.: The Lagrangian Averaged Euler (LAE-α) Equations with Free-Slip or Mixed Boundary Conditions, pp. 113–165. Springer-Verlag (2002). https://doi.org/10.1007/0-387-21791-6_5
Anna C. Green, Antonella Bianchi, Daniel D. Brown, Felice Feldmann, Jeremie Gobeil, Miron van der Kolk, Riccardo Maggiore, Jonathan W. Perry, Emma Prins, Mischa Salle et al. ·