betaFeedback

WorksPhysics and Astronomy

Stochastic geometric mechanics with diffeomorphisms

Abstract

Noether's celebrated theorem associating symmetry and conservation laws in classical field theory is adapted to allow for broken symmetry in geometric mechanics and is shown to play a central role in deriving and understanding the generation of fluid circulation via the Kelvin-Noether theorem for ideal fluids with stochastic advection by Lie transport (SALT).

Noether’s theorem in geometric mechanics

Euler-Poincaré reduction

Geometric mechanics deals with group-invariant variational principles. In this setting, Noether’s theorem Noether [ [23, 24] plays a key role. Given the tangent lift action G×TM→TMG \times TM \to TM of a Lie group GG on the tangent bundle TMTM of a manifold MM1 on which GG acts transitively, Noether’s theorem states that each Lie symmetry of a Lagrangian L:TM→RL : TM \to\mathbb{R} defined in the action integral S=∫L(q,v)dtS = \int L(q,v)\mathrm{d}t for Hamilton’s variational principle δS=0\delta S = 0 with (q,v)∈TM(q,v) \in TM implies a conserved quantity for the corresponding Euler-Lagrange equations defined on the cotangent bundle T∗MT^*M. The conserved quantities arising from Noether’s theorem in the case where the configuration manifold MM is a Lie group GG were studied by Smale, in Smale [1970a,b], where it was shown that the reduction procedure TG→TG\G≃gTG \to TG \backslash G \simeq\mathfrak{g} leads to dynamics which take place on the dual g∗\mathfrak{g}^* of the Lie algebra g\mathfrak{g}. The dynamical variable m∈g∗m \in\mathfrak{g}^* in the dual Lie algebra is now called the momentum map (Smale called it angular momentum). In general, the configuration manifold MM is not a Lie group. However, when a Lie group GG acts transitively on a configuration manifold MM the proof of Noether’s theorem induces a cotangent-lift momentum map J:T∗M→g∗J : T^*M \to\mathfrak{g}^*. The momentum map induced this way is an infinitesimally equivariant Poisson map taking functions on the cotangent bundle T∗MT^*M of MM to the dual Lie algebra g∗\mathfrak{g}^* of the Lie group GG. The momentum map J:T∗M→g∗J : T^*M \to\mathfrak{g}^* is equivariant and Poisson, even if GG is not a Lie symmetry of the Lagrangian in Hamilton’s principle. Momentum maps naturally lead from the Lagrangian side to the Hamiltonian side. The Hamiltonian dynamics on T∗MT^*M involves symplectic transformations. However, as we shall discuss below, for the class of Hamiltonians which can be defined as H∘J:g∗→RH \circ J : \mathfrak{g}^* \to\mathbb{R}, the momentum map induces Euler-Poincaré motion on the Lagrangian side and Lie-Poisson motion on the Hamiltonian side. To illustrate these remarks, we return to the situation in which the configuration manifold, MM, is a Lie group, GG.

For hyperregular Lagrangians, the Legendre transform to the Hamiltonian side is invertible and one may reconstruct the solution on GG from its representation on T∗G\G≃g∗T^*G \backslash G \simeq\mathfrak{g}^*. In that case, solving the equations describing the evolution of the momentum map on the dual Lie algebra g∗\mathfrak{g}^* is equivalent to solving the equations on the cotangent bundle T∗GT^*G when the configuration manifold is GG. When the Lie group GG acts transitively, freely and properly on the configuration manifold MM, then one may reconstruct the solution on MM from its representation on T∗G\G≃g∗T^*G \backslash G \simeq\mathfrak{g}^*. The last statement is proven for finite-dimensional Lie groups GG in, e.g., Abraham and Marsden [ [1].

The Lie-group reduced equations defined on the dual Lie algebra g∗\mathfrak{g}^* via Smale’s procedure of reduction by symmetry T∗G\G≃g∗T^*G \backslash G \simeq\mathfrak{g}^* are called Euler-Poincaré equations after Poincaré [ [25]. Provided the Lagrangian is hyperregular, the Euler-Poincaré reduction procedure can be expressed in terms of the cube of linked commutative diagrams shown in Figure 1.

Cube of commutative diagrams: $L:T G\to\mathbb{R}$ and $H:T^*G\to\mathbb{R}$ connected by Legendre transform; Euler-Lagrange eqns and Hamilton’s eqns; reduced functionals $\ell:\mathfrak{g}\to\mathbb{R}$ and $\hbar:\mathfrak{g}^*\to\mathbb{R}$ connected by Legendre transform; Euler-Poincaré eqns and Lie-Poisson eqns

Figure 1. The cube of commutative diagrams for geometric mechanics on Lie groups. Euler-Poincaré reduction (on the left side) and Lie-Poisson reduction (on the right side) are both indicated by the arrows pointing down. The diagrams are all commutative, provided the Legendre transformation and reduced Legendre transformation are both invertible.

To summarise the notation in Figure 1, GG denotes the configuration manifold which is assumed to be isomorphic to a Lie group, TGTG is the tangent bundle, T∗GT^*G is the cotangent bundle, TG∖G≃gTG\setminus G\simeq\mathfrak{g} is the Lie algebra and T∗G∖G≃g∗T^*G\setminus G\simeq\mathfrak{g}^* is the dual of the Lie algebra. The Lagrangian is a functional L:TG→RL:TG\to\mathbb{R} and the Hamiltonian is a functional H:T∗G→RH:T^*G\to\mathbb{R}. Euler-Poincaré reduction takes advantage of Lie group symmetries to transform the Lagrangian and Hamiltonian into group-invariant variables, which leads to a reduced Lagrangian ℓ:g→R\ell:\mathfrak{g}\to\mathbb{R} and a reduced Hamiltonian ℏ:g∗→R\hbar:\mathfrak{g}^*\to\mathbb{R}. The diagram comprising the face of the cube involving these functionals in Figure 1 commutes if the Legendre transform is a diffeomorphism. This is guaranteed if the Lagrangian or Hamiltonian is hyperregular. The Euler-Lagrange equations and Hamilton’s equations are related via a change of variables, which also holds for the Euler-Poincaré equations and the Lie-Poisson equations. Many finite dimensional mechanical systems may be described naturally in this framework. The classic example is the rotating rigid body, discussed from the viewpoint of symmetry reduction by Poincaré [ [25]. In his 1901 paper, Poincaré also raised the issue of symmetry breaking, by introducing the vertical acceleration of gravity, which breaks the SO(3)SO(3) symmetry for free rotation and restricts it to SO(2)SO(2) for rotations about the vertical axis.

Stochasticity may also be included in the framework of Euler-Poincaré reduction by symmetry. The first attempt to include noise consistently in finite-dimensional symplectic Hamiltonian mechanics was by Bismut [ [5] and reduction by symmetry of stochastic systems was studied by Lázaro-Camí and Ortega [ [21].

Plan of the paper. In the present work, we will review Euler-Poincaré reduction of stochastic infinite dimensional variational systems with symmetry breaking. The infinite dimensional case is interesting because it is the natural setting for fluid dynamics, quantum mechanics and elasticity. The foundations of the finite dimensional stochastic geometric mechanics are established in Cruzeiro et al. [ [12]. We will explore the infinite dimensional case in context of fluid dynamics, where symmetry under the smooth invertible maps of the flow domain is broken by the spatial dependence of the initial mass density.

Sobolev class diffeomorphisms

Consider an nn-dimensional compact and oriented smooth manifold MM, equipped with a Riemannian metric ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle. This will be the spatial domain of flow and X∈MX\in M will denote the initial position of any given fluid particle. The manifold MM is acted upon by a group of Sobolev class diffeomorphisms. In Ebin and Marsden [ [13] it is shown that the space of C∞C^\infty diffeomorphisms, defined by D={g∈C∞(M,M)∣g is bijective and g−1∈C∞(M,M)}\mathcal{D}=\{g\in C^\infty(M,M)\mid g\text{ is bijective and }g^{-1}\in C^\infty(M,M)\}, is not the convenient setting to study fluid dynamics, but that one should use Ds={g∈Hs(M,M)∣g is bijective and g−1∈Hs(M,M)}\mathcal{D}^s=\{g\in H^s(M,M)\mid g\text{ is bijective and }g^{-1}\in H^s(M,M)\}, the space of Sobolev class diffeomorphisms with weak derivatives. The reason for this choice is that the smooth diffeomorphisms constitute a Fréchet manifold for which there is no inverse or implicit function theorem and no general solution theorem for ordinary differential equations. Each of these latter features would prohibit the study of geodesics.

The space of Sobolev class diffeomorphisms is both a Hilbert manifold and a topological group if s>n/2+1s>n/2+1, as was shown by Ebin [ [14]. The Hilbert manifold structure implies the existence of function inverses and the implicit function theorem, as well as the existence of a general solution theorem for ordinary differential equations. This additional structure also implies that one can construct the tangent space of Ds\mathcal{D}^{s} in the usual way and study geodesics. The space Ds\mathcal{D}^{s} is the configuration space for continuum mechanics and each g∈Dsg \in\mathcal{D}^{s} is called a configuration. A fluid trajectory starting from X∈MX \in M at time t=0t = 0 is given by x(t)=gt(X)=g(X,t)x(t) = g_t(X) = g(X,t), with Ds∋g:M×R+→M\mathcal{D}^{s} \ni g : M \times\mathbb{R}^{+} \to M being a continuous one-parameter subgroup of Ds\mathcal{D}^{s}. In the deterministic case, computing the time derivative of this one-parameter subgroup gives rise to the reconstruction equation, given by

∂∂tgt(X)=u(gt(X),t),(1)\frac{\partial}{\partial t}g_t(X) = u(g_t(X),t), \tag*{(1)}

where ut(⋅)=u(⋅,t)∈Xsu_t(\cdot) = u(\cdot,t) \in\mathfrak{X}^{s} is a time dependent vector field with flow gt(⋅)=g(⋅,t)g_t(\cdot) = g(\cdot,t). The initial data is given by g(X,0)=Xg(X,0) = X. Here Xs=Hs(TM)\mathfrak{X}^{s} = H^{s}(TM) denotes the space of Sobolev class vector fields on MM, which is also the Lie algebra associated to the Sobolev class diffeomorphisms.

Stochastic advection by Lie transport (SALT)

In the setting of stochastic advection by Lie transport (SALT), which was introduced by Holm [ [16], the deterministic reconstruction equation in (1) is replaced by the semimartingale

dg(X,t)=u(gt(X),t)dt+∑i=1Mξi(gt(X))∘dWti,(2)\mathrm{d}g(X,t) = u(g_t(X),t)\mathrm{d}t + \sum_{i=1}^{M}\xi_i(g_t(X)) \circ\mathrm{d}W_t^i, \tag*{(2)}

where the symbol ∘\circ means that the stochastic integral is taken in the Stratonovich sense. The initial data is given by g(X,0)=Xg(X,0) = X. The WtiW_t^i are independent, identically distributed Brownian motions, defined with respect to the standard stochastic basis (Ω,F,(Ft)t≥0,P)(\Omega,\mathcal{F},(\mathcal{F}_t)_{t\geq0},\mathbb{P}). Such a noise was shown to arise from a multi-time homogenisation argument in Cotter et al. [ [10]. The ξi(⋅)∈Xs\xi_i(\cdot) \in\mathfrak{X}^{s} are called data vector fields and are prescribed. These data vector fields represent the effects of unresolved degrees of freedom on the resolved scales of fluid motion and account for unrepresented processes. They are determined by applying empirical orthogonal function analysis to appropriate numerical and/or observational data. For instance, for an application to the two dimensional Euler equations for an ideal fluid, see Cotter et al. [ [8] and for an application to a two-layer quasi-geostrophic model, see Cotter et al. [ [7]. Stochastic models enable the use of a variety of methods in data assimilation, which are discussed in Cotter et al. [ [9]. It is not difficult to make sense of (1), but understanding (2) is more complicated. In Bethencourt de Leon et al. [ [4], a stochastic chain rule is shown to exist. This stochastic chain rule is called the Kunita-Itô-Wentzell (KIW) formula and helps interpret the semimartingale in (2). The KIW formula will also be used later to prove the stochastic Kelvin circulation theorem. First, however, the space Ds\mathcal{D}^{s} needs to be given more structure.

The space Ds\mathcal{D}^{s} inherits a weak Riemannian structure from the underlying manifold MM in a natural way. For g∈Dsg \in\mathcal{D}^{s} and V,W∈TgDsV,W \in T_g\mathcal{D}^{s}, one can define the following bilinear form

(V,W)=∫M⟨V(X),W(X)⟩g(X)μ(dX),(3)(V,W) = \int_M \langle V(X),W(X)\rangle_{g(X)}\mu(\mathrm{d}X), \tag*{(3)}

where μ\mu is the volume form on MM induced by the metric. The Riemannian structure induced by (3) is weak because the topology is of type L2L^2, which is strictly weaker than the HsH^s topology. This bilinear form is a linear functional on the Hilbert space TgDsT_g\mathcal{D}^{s} and can be used to define the dual space Tg∗DsT_g^*\mathcal{D}^{s}. The pairing between V∈TgDsV \in T_g\mathcal{D}^{s} and α∈Tg∗Ds\alpha\in T_g^*\mathcal{D}^{s} is given by

⟨α,V⟩=∫Mα(X)⋅V(X).(4)\langle\alpha,V\rangle= \int_M \alpha(X)\cdot V(X). \tag*{(4)}

Hence the metric on MM and the volume form μ(dX)\mu(\mathrm{d}X) can be used to construct the isomorphism between TDsT\mathcal{D}^{s} and T∗DsT^*\mathcal{D}^{s} as V(X)↦α(X)=V♭(X)μ(dX)V(X) \mapsto\alpha(X) = V^♭(X)\mu(\mathrm{d}X), where b:TM→T∗Mb : TM \to T^*M is one of the musical isomorphisms that are induced by the metric on MM. The group Ds\mathcal{D}^{s} is not a Lie group; since right multiplication is smooth, but left multiplication is only continuous. Hence Ds\mathcal{D}^{s} is a topological group with a weak Riemannian structure. In general, these properties are not sufficient to guarantee the existence of an exponential map. However, Ebin and Marsden [ [13] showed that an exponential map can exist in many important cases. In particular, they showed that the geodesic spray associated to (3) (with and without forcing) is smooth.2 The smoothness of the geodesic spray persists even though HsH^s diffeomorphisms are considered rather than smooth diffeomorphisms. Combined with the existence of an exponential map, the smoothness property implies a regular interpretation of the Euler–Poincaré equations on Ds\mathcal{D}^{s}, provided that one uses right translations and right representations of the group on itself and its Lie algebra, as shown in Holm et al. [ [17]. However, due to the presence of the volume form μ(dm)\mu(dm), the bilinear form (3) is not right-invariant under the action of the entire Hs\mathcal{H}^{s} diffeomorphism group, although there is right-invariance under the action of the isotropy subgroup Dμs={g∈Ds∣g∗μ=μ}\mathcal{D}^{s}_{\mu}=\{g\in\mathcal{D}^{s}\mid g_{*}\mu=\mu\}. Since this subgroup is a proper subgroup, as it is smaller than Ds\mathcal{D}^{s} itself. Thus, one speaks of symmetry breaking.

In deriving the equations of ideal deterministic fluid dynamics, one needs to keep track of the volume form as well. The appropriate mathematical setting for this is an outer semidirect product group. This means that one constructs a new group from two given groups with a particular type of group operation. For continuum mechanics, the ingredients are Ds\mathcal{D}^{s} and V∗V^{*}, where V∗V^{*} is a vector space of tensor fields. This vector space is the space of advected quantities and it will always contain at least the volume form μ\mu.

Definition 1.1 (Advected quantity). A fluid variable is said to be advected, if it keeps its value along Lagrangian particle trajectories. Advected quantities are sometimes called tracers, because the evolution histories of scalar advected quantities with different initial values (labels) trace out the Lagrangian particle trajectories of each label, or initial value, via the push-forward of the full diffeomorphism group, i.e., at=gt∗a0=a0gt−1a_t=g_t{}_{*}a_0=a_0g_t^{-1}, where gtg_t is a time-dependent curve on the manifold of diffeomorphisms that represents the fluid flow.

Remark 1 (Advected quantities as order parameters). When several advected quantities are involved, the space V∗V^{*} is the direct sum of several vector spaces, where each summand space hosts a different advected quantity. In general, each additional advected quantity decreases the dimension of the isotropy subgroup. For example, consider an ideal deterministic fluid with a buoyancy variable bb, then the Lagrangian corresponding to the model will depend on μ\mu and bb in a parametric manner. This Lagrangian will be right invariant under the action of the isotropy subgroup Dμ,bs={g∈Ds∣g∗μ=μ and g∗b=b}\mathcal{D}^{s}_{\mu,b}=\{g\in\mathcal{D}^{s}\mid g_{*}\mu=\mu\ \text{and}\ g_{*}b=b\}. Hence, advected quantities are order parameters and each additional order parameter breaks more symmetry. For the sake of notation, one usually writes D0s\mathcal{D}^{s}_{0} for the isotropy subgroup, no matter how many advected quantities there are. One then uses aa to represent all advected quantities and a0a_0 to denote the initial value of the advected quantities.

Semidirect product group adjoint & coadjoint actions

The semidirect product group action is constructed in the following way. The representation of Ds\mathcal{D}^{s} on a vector space VV by push-forward, which is a left representation, as shown by Marsden et al. [ [22]. The representation of the group on itself and on its Lie algebra is a right representation. In terms of analysis, this means that all representations are smooth and no derivatives need to be counted. The group action of the semidirect product group is given by

∙:(Ds×V)×(Ds×V)⟶(Ds×V)\bullet:(\mathcal{D}^{s}\times V)\times(\mathcal{D}^{s}\times V)\longrightarrow(\mathcal{D}^{s}\times V)
(g1,v1)∙(g2,v2):=(g1∘g2,v2+(g2)∗v1)(5)(g_1,v_1)\bullet(g_2,v_2):=(g_1\circ g_2,v_2+(g_2)_{*}v_1) \tag*{(5)}

with g1,g2∈Dsg_1,g_2\in\mathcal{D}^{s} and v1,v2∈Vv_1,v_2\in V. The semidirect product group is often denoted as DsⓈV=(Ds×V,∙)\mathcal{D}^{s}\circledS V=(\mathcal{D}^{s}\times V,\bullet). In the group action above, (g2)∗v1(g_2)_{*}v_1 denotes the push-forward of v1v_1 by g2g_2 and ∘\circ denotes composition. Note that the group affects both slots in (5), but the vector space only appears in the second slot. The identity element of the semidirect product group is (e,0)(e,0) where e∈Dse\in\mathcal{D}^{s} is the identity diffeomorphism and 0∈V0\in V is the zero vector. An inverse element is given by

(g,v)−1=(g−1,−(g−1)∗v)=(g−1,−g∗−1v),(6)(g,v)^{-1}=(g^{-1},-(g^{-1})_{*}v)=(g^{-1},-g^{-1}_{*}v), \tag*{(6)}

where g∗g^{*} denotes the pull-back of vv by gg. To understand how reduction works for semidirect products, it is helpful to know how the group acts on its Lie algebra and on the dual of its Lie algebra. Duality will be defined with respect to the sum of the pairing (4) and the dual linear transformation [⋅,⋅]∗[\cdot,\cdot]^{*} on VV. This pairing induces another pairing in a natural way on Xs×V\mathfrak{X}^{s}\times V. Consider two at least C1C^{1} one parameter subgroups (gt,vt),(g~t,v~t)∈Ds×V(g_t,v_t),(\widetilde{g}_t,\widetilde{v}_t)\in\mathcal{D}^{s}\times V. Using these one parameter subgroups, one can compute the inner automorphism, or adjoint action of the group on itself. This adjoint action is defined by conjugation

AD⁡:(Ds×V)×(Ds×V)⟶(Ds×V),\operatorname{AD}:(\mathcal{D}^{s}\times V)\times(\mathcal{D}^{s}\times V)\longrightarrow(\mathcal{D}^{s}\times V),
AD⁡(gt,vt)(g~t,v~t):=(gt,vt)∙(g~t,v~t)∙(gt,vt)−1=(gt∘g~t∘gt−1,gt∗(v~t−vt+g~t∗vt)).(7)\begin{aligned} \operatorname{AD}_{(g_t,v_t)}(\widetilde{g}_t,\widetilde{v}_t):=(g_t,v_t)\bullet(\widetilde{g}_t,\widetilde{v}_t)\bullet(g_t,v_t)^{-1} \\ &=(g_t\circ\widetilde{g}_t\circ g_t^{-1},g_t{}_{*}(\widetilde{v}_t-v_t+\widetilde{g}_t{}_{*}v_t)). \tag*{(7)} \end{aligned}

To see how the group acts on its Lie algebra, one can compute the derivative with respect to ϵ\epsilon and evaluate at ϵ=0\epsilon=0 in the adjoint action of the group on itself. Let Xs∋u~=ddϵ∣ϵ=0g~ϵ\mathfrak{X}^{s}\ni\widetilde{u}=\left.\frac{d}{d\epsilon}\right|_{\epsilon=0}\widetilde{g}_{\epsilon} and V∋b~=ddϵ∣ϵ=0v~ϵV\ni\widetilde{b}=\left.\frac{d}{d\epsilon}\right|_{\epsilon=0}\widetilde{v}_{\epsilon}. This choice for a vector field is guided by the deterministic reconstruction equation in (1). For any tensor Sϵ∈Tsr(M)S_\epsilon\in T^r_s(M) whose dependence on ϵ\epsilon is at least C1C^1 it holds that

ddϵg~ϵ∗Sϵ=g~ϵ∗(ddϵSϵ−Lu~Sϵ).(8)\frac{d}{d\epsilon}\tilde{g}_\epsilon^*S_\epsilon=\tilde{g}_\epsilon^*\left(\frac{d}{d\epsilon}S_\epsilon-\mathcal{L}_{\tilde{u}}S_\epsilon\right). \tag*{(8)}

Important here is that the Lie derivative does not commute with pull-backs and push-forwards that depend on parameters, see Abraham and Marsden [ [1]. The adjoint action of the group on its Lie algebra can be computed as

Ad⁡:(Ds×V)×(Xs×V)→(Xs×V),\operatorname{Ad}:(\mathcal{D}^s\times V)\times(\mathfrak{X}^s\times V)\to(\mathfrak{X}^s\times V),
Ad⁡(gt,vt)(u~,b~):=ddϵ∣ϵ=0AD⁡(gt,vt)(g~ϵ,v~ϵ)=(gt∗u~,gt∗b~−gt∗Lu~vt).(9)\begin{aligned} \operatorname{Ad}_{(g_t,v_t)}(\tilde{u},\tilde{b}):=\left.\frac{d}{d\epsilon}\right|_{\epsilon=0}\operatorname{AD}_{(g_t,v_t)}(\tilde{g}_\epsilon,\tilde{v}_\epsilon) \\ &=(g_{t*}\tilde{u},g_{t*}\tilde{b}-g_t^*\mathcal{L}_{\tilde{u}}v_t). \tag*{(9)} \end{aligned}

By means of the pairing on Xs×V\mathfrak{X}^s\times V, one can compute the dual action to the adjoint action (9). This is called the coadjoint action of the group on the dual of its Lie algebra. Let (m~,a~)∈(Xs×V)∗(\tilde{m},\tilde{a})\in(\mathfrak{X}^s\times V)^*, then the coadjoint action is given by

Ad⁡∗:(Ds×V)×(Xs×V)∗→(Xs×V)∗,\operatorname{Ad}^*:(\mathcal{D}^s\times V)\times(\mathfrak{X}^s\times V)^*\to(\mathfrak{X}^s\times V)^*,
⟨Ad⁡(gt−1,−gt−1vt)∗(m~,a~),(u~,b~)⟩:=⟨(m~,a~),Ad⁡(gt,vt)(u~,b~)⟩,\left\langle\operatorname{Ad}^*_{(g_t^{-1},-g_t^{-1}v_t)}(\tilde{m},\tilde{a}),(\tilde{u},\tilde{b})\right\rangle:=\left\langle(\tilde{m},\tilde{a}),\operatorname{Ad}_{(g_t,v_t)}(\tilde{u},\tilde{b})\right\rangle,
Ad⁡(gt−1,−gt−1vt)∗(m~,a~)=(gt∗m~+vt⋄gt∗a~,gt∗a~).(10)\operatorname{Ad}^*_{(g_t^{-1},-g_t^{-1}v_t)}(\tilde{m},\tilde{a})=(g_{t*}\tilde{m}+v_t\diamond g_{t*}\tilde{a},g_{t*}\tilde{a}). \tag*{(10)}

Definition 1.2 (The diamond operator). The coadjoint action (10) features the diamond operator, which is defined for a∈V∗a\in V^*, u∈Xsu\in\mathfrak{X}^s and fixed v∈Vv\in V as

⟨v⋄a,u⟩Xs∗×Xs:=−⟨a,Luv⟩V∗×V.(11)\langle v\diamond a,u\rangle_{\mathfrak{X}^s{}^*\times\mathfrak{X}^s}:=-\langle a,\mathcal{L}_u v\rangle_{V^*\times V}. \tag*{(11)}

Note that the diamond operator is the dual of the Lie derivative regarded as a map L(⋅)v:Xs→V\mathcal{L}_{(\cdot)}v:\mathfrak{X}^s\to V, hence v⋄(⋅):V∗→Xs∗v\diamond(\cdot):V^*\to\mathfrak{X}^s{}^*. The diamond operator shows how an element from the dual of the vector space acts on the dual of the Lie algebra.

When evaluated at t=0t=0, the tt-derivatives of Ad⁡\operatorname{Ad} in (9) and Ad⁡∗\operatorname{Ad}^* in (10) define, respectively, the adjoint and coadjoint actions of the Lie algebra on itself and on its dual. Denote by Xs∋u=ddt∣t=0gt\mathfrak{X}^s\ni u=\left.\frac{d}{dt}\right|_{t=0}g_t and V∋b=ddt∣t=0vtV\ni b=\left.\frac{d}{dt}\right|_{t=0}v_t. The adjoint action of the Lie algebra on itself is

ad⁡:(Xs×V)×(Xs×V)→(Xs×V),\operatorname{ad}:(\mathfrak{X}^s\times V)\times(\mathfrak{X}^s\times V)\to(\mathfrak{X}^s\times V),
ad⁡(u,b)(u~,b~):=ddt∣t=0Ad⁡(gt,vt)(u~,b~),\operatorname{ad}_{(u,b)}(\tilde{u},\tilde{b}):=\left.\frac{d}{dt}\right|_{t=0}\operatorname{Ad}_{(g_t,v_t)}(\tilde{u},\tilde{b}),
ad⁡(u,b)(u~,b~)=(−Luu~,Lub~−Lu~b)=(−[u,u~],Lub~−Lu~b).(12)\begin{aligned} \operatorname{ad}_{(u,b)}(\tilde{u},\tilde{b})=(-\mathcal{L}_u\tilde{u},\mathcal{L}_u\tilde{b}-\mathcal{L}_{\tilde{u}}b) \\ &=(-[u,\tilde{u}],\mathcal{L}_u\tilde{b}-\mathcal{L}_{\tilde{u}}b). \tag*{(12)} \end{aligned}

where the bracket [⋅,⋅][\cdot,\cdot] in (12) is the commutator of vector fields. The minus sign is due to fact that group acts on itself from the right. The coadjoint action of the Lie algebra on its dual can be obtained by computing the dual to (12) or by taking the derivative with respect to tt and evaluate at t=0t=0 in (10). Either way, one arrives at

ad⁡∗:(Xs×V)×(Xs×V)∗→(Xs×V)∗,\operatorname{ad}^*:(\mathfrak{X}^s\times V)\times(\mathfrak{X}^s\times V)^*\to(\mathfrak{X}^s\times V)^*,
⟨ad⁡(u,b)∗(m~,a~),(u~,b~)⟩:=⟨(m~,a~),ad⁡(u,b)(u~,b~)⟩,\left\langle\operatorname{ad}^*_{(u,b)}(\tilde{m},\tilde{a}),(\tilde{u},\tilde{b})\right\rangle:=\left\langle(\tilde{m},\tilde{a}),\operatorname{ad}_{(u,b)}(\tilde{u},\tilde{b})\right\rangle,
ad⁡(u,b)∗(m~,a~)=(Lum~+b⋄a~,−Lua~).(13)\operatorname{ad}^*_{(u,b)}(\tilde{m},\tilde{a})=(\mathcal{L}_u\tilde{m}+b\diamond\tilde{a},-\mathcal{L}_u\tilde{a}). \tag*{(13)}

in which (12) implies the last line in (13). Alternatively, one can also obtain (13) by taking the derivative with respecto to tt in (10) and evaluate at t=0t=0.

Remark 2 (Coadjoint action and the diamond operator). The coadjoint action is an important operator in geometric mechanics and representation theory. It was shown by Kirillov [ [18] and in further work by Kostant [1970] and Souriau [ [28] that the coadjoint orbits of a Lie group GG have the structure of symplectic manifolds and are connected with Hamiltonian mechanics. See Kirillov [1999] for a review. The computations of the adjoint and coadjoint actions for the semidirect product group is valuable for fluid mechanics, as they introduce the two fundamental operators that appear in the equations of motion. The Lie derivative is responsible for transport of tensors along vector fields and its dual action given by the diamond operator encodes the symmetry breaking. In particular, the diamond operator introduces the effect of symmetry breaking into the Euler-Poincaré equations of motion.

Deterministic geometric fluid dynamics

With the adjoint and coadjoint actions defined, one can derive continuum mechanics equations with advected quantities by using symmetry reduction. Euler-Poincaré reduction for a semidirect product group Ds⋉V\mathcal{D}^{s} \ltimes V as developed in Holm et al. [ [17] is sketched below in Figure 2.

The cube of continuum mechanics in the semidirect product group setting

Figure 2. The cube of continuum mechanics in the semidirect product group setting. Reduction is indicated by the arrows pointing down.

As shown by comparison of Figure 2 with Figure 1, several new features arise in semidirect product Lie group reduction which differ from Euler-Poincaré reduction by symmetry when the configuration space itself is a Lie group. These differences can be conveniently explained by introducing the physical concept of an order parameter. As discussed earlier, the order parameters in continuum mechanics are the elements of V∗V^{*} which are advected by the action of the diffeomorphism group Ds\mathcal{D}^{s}. The advection is defined simply as the semidirect product action on the elements of V∗V^{*}. The introduction of each additional advected state variable (or, order parameter) into the physical problem further breaks the original symmetry Ds\mathcal{D}^{s}. The remaining symmetry of the Lagrangian in Hamilton’s principle is the isotropy subgroup Da0s\mathcal{D}^{s}_{a_{0}} of the initial conditions, a0a_{0}, for the entire set of advected quantities, aa. The action of the diffeomorphism group Ds\mathcal{D}^{s} on these initial conditions then describes their advection as the action of Ds\mathcal{D}^{s} on its coset space Ds∖Da0s=V∗\mathcal{D}^{s}\setminus\mathcal{D}^{s}_{a_{0}}=V^{*}. Once the initial values of the order parameters, a0a_{0}, have been set, one must still define a Legendre transform from the Lagrangian formulation into the Hamiltonian formulation and vice versa. The Legendre transform in the setting of semidirect products is a partial Legendre transform, since it transforms between TDsT\mathcal{D}^{s} and T∗DsT^{*}\mathcal{D}^{s} or TDs∖Da0s≃XsT\mathcal{D}^{s}\setminus\mathcal{D}^{s}_{a_{0}}\simeq\mathfrak{X}^{s} and T∗Ds∖Da0s≃Xs∗T^{*}\mathcal{D}^{s}\setminus\mathcal{D}^{s}_{a_{0}}\simeq\mathfrak{X}^{s*} only after having fixed the value a0a_{0} of the order parameters, which live in V∗V^{*}. This coset reduction is what Figure 2 shows. The remaining invariance of a functional under the action of the isotropy subgroup is called its particle relabelling symmetry.

Our exploration continues on the Lagrangian side in Figure 2. Consider a fluid Lagrangian L:TDs×V∗→RL:T\mathcal{D}^{s}\times V^{*}\to\mathbb{R}. By fixing the value of a0∈V∗a_{0}\in V^{*}, one can construct La0:TDs→RL_{a_{0}}:T\mathcal{D}^{s}\to\mathbb{R}. If this Lagrangian is right invariant under the action of the isotropy subgroup Da0s\mathcal{D}^{s}_{a_{0}}, then one can construct

L(ddtg∘g−1,e,a0)=La0(ddtg∘g−1,e)=La0(ddtg∘g−1)=ℓ(ddtg∘g−1,g∗a0).(14)\begin{aligned} L\left(\frac{d}{dt}g\circ g^{-1},e,a_{0}\right)=L_{a_{0}}\left(\frac{d}{dt}g\circ g^{-1},e\right) \\ &=L_{a_{0}}\left(\frac{d}{dt}g\circ g^{-1}\right)=\ell\left(\frac{d}{dt}g\circ g^{-1},g_{*}a_{0}\right). \tag*{(14)} \end{aligned}

Here ∘\circ means composition of functions. The same procedure applies to the Hamiltonian. Since the coadjoint action is known, it is straightforward to formulate the Lie-Poisson equations. The details of Hamiltonian semidirect product reduction and also more information on the Lagrangian semidirect product reduction can be found in Holm et al. [ [17].

The coadjoint action of the Lie algebra on its dual is also required for the Lagrangian semidirect product reduction. One can use the deterministic reconstruction equation to see that the argument of the Lagrangians in (14) is

ddtg∘g−1=u.(15)\frac{d}{dt}g\circ g^{-1}=u. \tag*{(15)}

Using this information, one can integrate the Lagrangian in time to construct the action functional. By requiring the variational derivative of the action functional to vanish, one can compute the equations of motion. However, due to the removal of symmetries, the variations are no longer free.

Stochastic geometric fluid dynamics

In the situation where noise is present, that is, when the reconstruction equation is (2), the Euler–Poincaré variations become stochastic. Consider g:R2→Dsg:\mathbb{R}^{2}\to\mathcal{D}^{s} with gt,ϵ=g(t,ϵ)g_{t,\epsilon}=g(t,\epsilon) to be a two parameter subgroup with smooth dependence on ϵ\epsilon, but only continuous dependence on tt. Let us denote

dχt,ϵ(X)=(dgt,ϵ∘gt,ϵ)(X)=ut,ϵ(X)dt+∑i=1Nξi(X)∘dWti\mathrm{d}\chi_{t,\epsilon}(X)=(\mathrm{d}g_{t,\epsilon}\circ g_{t,\epsilon})(X)=u_{t,\epsilon}(X)\mathrm{d}t+\sum_{i=1}^{N}\xi_i(X)\circ\mathrm{d}W_t^i

and

vt,ϵ(X)=(∂∂ϵgt,ϵ∘gt,ϵ)(X).v_{t,\epsilon}(X)=\left(\frac{\partial}{\partial\epsilon}g_{t,\epsilon}\circ g_{t,\epsilon}\right)(X).

When a ∘\circ symbol is followed by dWt\mathrm{d}W_t it means Stratonovich integration and in every other context the ∘\circ symbol is used to denote composition. Note that the data vector fields ξi\xi_i are prescribed and hence will not have a dependence on ϵ\epsilon.

In order to compute with these stochastic subgroups and their associated vector fields, one needs a stochastic Lie chain rule. The Kunita-Itô-Wentzell (KIW) formula is the stochastic generalisation of the Lie chain rule (8). A proof of the KIW formula is given in Bethencourt de Leon et al. [ [4] for differential kk-forms and vector fields. That proof includes the technical details on regularity that will be omitted here. In the KIW formula, the kk-form is allowed to be a semimartingale itself. Let KK be a continuous adapted semimartingale that takes values in the kk-forms and satisfies

Kt=K0+∫0tGsds+∑i=1N∫0tHsi∘dBsi,(16)K_t=K_0+\int_0^t G_s\mathrm{d}s+\sum_{i=1}^{N}\int_0^t H_s^i\circ\mathrm{d}B_s^i, \tag*{(16)}

where the BtiB_t^i are independent, identically distributed Brownian motions. The drift of the semimartingale KK is determined by GG and the diffusion by HiH_i, both of which are kk-form valued continuous adapted semimartingales with suitable regularity. Let gtg_t satisfy (2), then Bethencourt de Leon et al. [ [4] shows that the following holds

d(gt∗Kt)=gt∗(dKt+LutKtdt+LξiKt∘dWti).(17)\mathrm{d}(g_t^*K_t)=g_t^*\left(\mathrm{d}K_t+\mathcal{L}_{u_t}K_t\mathrm{d}t+\mathcal{L}_{\xi_i}K_t\circ\mathrm{d}W_t^i\right). \tag*{(17)}

Equation (16) helps to interpret the dKt\mathrm{d}K_t term in the KIW formula (17). This formula will be particularly useful in computing the variations of the variables in the Lagrangian. To compute these variations, one needs the variational derivative.

The variational derivative. The variational derivative of a functional F:B→RF:\mathcal{B}\to\mathbb{R}, where B\mathcal{B} is a Banach space, is denoted δF/δρ\delta F/\delta\rho with ρ∈B\rho\in\mathcal{B}. The variational derivative can be defined by the first variation of the functional

δF[ρ]:=ddϵ∣ϵ=0F[ρ+ϵδρ]=∫δFδρ(x)δρ(x) dx=⟨δFδρ,δρ⟩.(18)\delta F[\rho]:=\left.\frac{\mathrm{d}}{\mathrm{d}\epsilon}\right|_{\epsilon=0}F[\rho+\epsilon\delta\rho]=\int\frac{\delta F}{\delta\rho}(x)\delta\rho(x)\,\mathrm{d}x=\left\langle\frac{\delta F}{\delta\rho},\delta\rho\right\rangle. \tag*{(18)}

In the definition above, ϵ∈R\epsilon\in\mathbb{R} is a parameter, δρ∈B\delta\rho\in\mathcal{B} is an arbitrary function and the first variation can be understood as a Fréchet derivative. A precise and rigorous definition can be found in Gelfand et al. [ [15]. With the definition of the functional derivative in place, the following lemma can be formulated.

Lemma 3. With the notation as above, the variations of uu and any advected quantity aa are given by

δu(t)=dv(t)+[dχt,v(t)],δa(t)=−Lv(t)a(t),(19)\delta u(t)=\mathrm{d}v(t)+[\mathrm{d}\chi_t,v(t)],\qquad\delta a(t)=-\mathcal{L}_{v(t)}a(t), \tag*{(19)}

where v(t)∈Xsv(t)\in\mathfrak{X}^{s} is arbitrary.

Proof. The proof of the variation of a(t)a(t) is a direct application of the Kunita-Itô-Wentzell formula to a(t,ϵ)=gt,ϵ∗a0a(t,\epsilon)=g_{t,\epsilon}*a_0. Note that the data vector fields ξi\xi_i are prescribed and do not depend on ϵ\epsilon. Denote by xt,ϵ=gt,ϵ(X)x_{t,\epsilon}=g_{t,\epsilon}(X). Then one has

dgt,ϵ(X)=dxt,ϵ=ut,ϵ(xt,ϵ)dt+∑i=1Nξi(xt,ϵ)∘dWti=:dχt,ϵ(xt,ϵ).(20)\mathrm{d}g_{t,\epsilon}(X)=\mathrm{d}x_{t,\epsilon}=u_{t,\epsilon}(x_{t,\epsilon})\mathrm{d}t+\sum_{i=1}^{N}\xi_i(x_{t,\epsilon})\circ\mathrm{d}W_t^i=:\mathrm{d}\chi_{t,\epsilon}(x_{t,\epsilon}). \tag*{(20)}

The vector field associated to the ε\varepsilon-dependence of the two parameter subgroup is given by

∂∂εgt,ε=∂∂εxt,ε=vt,ε(xt,ε).(21)\frac{\partial}{\partial\varepsilon}g_{t,\varepsilon}=\frac{\partial}{\partial\varepsilon}x_{t,\varepsilon}=v_{t,\varepsilon}(x_{t,\varepsilon}). \tag*{(21)}

Computing the derivative with respect to ε\varepsilon of (20) gives

∂∂εdxt,ε=∂∂ε(dχt,ε(xt,ε))=(∂∂εut,ε+vt,ε⋅∂∂xt,εdχt,ε)(xt,ε).(22)\frac{\partial}{\partial\varepsilon}\mathrm{d}x_{t,\varepsilon}=\frac{\partial}{\partial\varepsilon}\left(\mathrm{d}\chi_{t,\varepsilon}(x_{t,\varepsilon})\right) =\left(\frac{\partial}{\partial\varepsilon}u_{t,\varepsilon}+v_{t,\varepsilon}\cdot\frac{\partial}{\partial x_{t,\varepsilon}}\mathrm{d}\chi_{t,\varepsilon}\right)(x_{t,\varepsilon}). \tag*{(22)}

where the independence of the data vector fields ξi\xi_i on ε\varepsilon was used. Taking the differential with respect to time of (21) gives

d(∂∂εxt,ε)=d(vt,ε(xt,ε))=(dvt,ε(xt,ε)+dxt,ε⋅∂∂xt,εvt,ε)(xt,ε).(23)\mathrm{d}\left(\frac{\partial}{\partial\varepsilon}x_{t,\varepsilon}\right)=\mathrm{d}\left(v_{t,\varepsilon}(x_{t,\varepsilon})\right) =\left(\mathrm{d}v_{t,\varepsilon}(x_{t,\varepsilon})+\mathrm{d}x_{t,\varepsilon}\cdot\frac{\partial}{\partial x_{t,\varepsilon}}v_{t,\varepsilon}\right)(x_{t,\varepsilon}). \tag*{(23)}

One can then evaluate at ε=0\varepsilon=0 and call upon equality of cross derivative-differential to obtain the result by subtracting. Since gt,εg_{t,\varepsilon} depends on tt in a C0C^0 manner, the integral representation is required. The particle relabelling symmetry permits one to stop writing the explicit dependence on space,

δu(t) dt=dv(t)+[dχt,v(t)].(24)\delta u(t)\,\mathrm{d}t=\mathrm{d}v(t)+[\mathrm{d}\chi_t,v(t)]. \tag*{(24)}

This completes the proof of formula (19) for the variation of u(t)u(t).

The notation in (20) needs careful explanation, because it comprises both a stochastic differential equation and a definition. The symbol dχt,ε\mathrm{d}\chi_{t,\varepsilon} is used to define a vector field, whereas dxt,ε\mathrm{d}x_{t,\varepsilon} denotes a stochastic differential equation. This lemma makes the presentation of the stochastic Euler-Poincaré theorem particularly simple.

Theorem 4 (Stochastic Euler-Poincaré). With the notation as above, the following are equivalent.

i) The constrained variational principle

δ∫t1t2ℓ(u,a) dt=0(25)\delta\int_{t_1}^{t_2}\ell(u,a)\,\mathrm{d}t=0 \tag*{(25)}

holds on Xs×V∗\mathcal X^s\times V^*, using variations δu\delta u and δa\delta a of the form

δu=dv+[dχt,v],δa=−Lva,(26)\delta u=\mathrm{d}v+[\mathrm{d}\chi_t,v],\qquad\delta a=-\mathcal L_v a, \tag*{(26)}

where v(t)∈Xsv(t)\in\mathcal X^s is arbitrary and vanishes at the endpoints in time for arbitrary times t1,t2t_1,t_2.

ii) The stochastic Euler-Poincaré equations hold on Xs×V∗\mathcal X^s\times V^*

dδℓδu+Ldχtδℓδu=δℓδa⋄a dt,(27)\mathrm{d}\frac{\delta\ell}{\delta u}+\mathcal L_{\mathrm{d}\chi_t}\frac{\delta\ell}{\delta u}=\frac{\delta\ell}{\delta a}\diamond a\,\mathrm{d}t, \tag*{(27)}

and the advection equation

da+Ldχta=0.(28)\mathrm{d}a+\mathcal L_{\mathrm{d}\chi_t}a=0. \tag*{(28)}

Proof. Using integration by parts and the endpoint conditions v(t1)=0=v(t2)v(t_1)=0=v(t_2), the variation can be computed to be

δ∫t1t2ℓ(u,a) dt=∫t1t2⟨δℓδu,δu⟩+⟨δℓδa,δa⟩ dt=∫t1t2⟨δℓδu,dv+[dχt,v]⟩+⟨δℓδa dt,−Lva⟩=∫t1t2⟨−dδℓδu−Ldχtδℓδu+δℓδa⋄a dt,v⟩=0.(29)\begin{aligned} \delta\int_{t_1}^{t_2}\ell(u,a)\,\mathrm{d}t =\int_{t_1}^{t_2}\left\langle\frac{\delta\ell}{\delta u},\delta u\right\rangle+\left\langle\frac{\delta\ell}{\delta a},\delta a\right\rangle\,\mathrm{d}t \\ &=\int_{t_1}^{t_2}\left\langle\frac{\delta\ell}{\delta u},\mathrm{d}v+[\mathrm{d}\chi_t,v]\right\rangle+\left\langle\frac{\delta\ell}{\delta a}\,\mathrm{d}t,-\mathcal L_v a\right\rangle\\ &=\int_{t_1}^{t_2}\left\langle-\mathrm{d}\frac{\delta\ell}{\delta u}-\mathcal L_{\mathrm{d}\chi_t}\frac{\delta\ell}{\delta u}+\frac{\delta\ell}{\delta a}\diamond a\,\mathrm{d}t,v\right\rangle\\ &=0. \tag*{(29)} \end{aligned}

Since the vector field vv is arbitrary, one obtains the stochastic Euler-Poincaré equations. Finally, the advection equation (28) follows by applying the KIW formula to a(t)=gt∗a0a(t)=g_{t*}a_0.

Remark 5. The stochastic Euler-Poincaré theorem is equivalent to the version presented in Holm [ [16], which uses stochastic Clebsch constraints. In Holm [ [16] one can also find an investigation the Itô formulation of the stochastic Euler-Poincaré equation.

Stochastic Lie-Poisson formulation. The stochastic Euler-Poincaré equations have an equivalent stochastic Lie-Poisson formulation. To obtain the Lie-Poisson formulation, one must Legendre transform the reduced Lagrangian. The Legendre transformation in the presence of stochasticity becomes itself stochastic in the following way

m:=δℓδu,ℏ(m,a) dt+∑i=1N⟨m,ξi⟩∘dWti=⟨m,dχt⟩−ℓ(u,a) dt.(30)m := \frac{\delta\ell}{\delta u},\quad\hbar(m,a)\,dt+\sum_{i=1}^{N}\langle m,\xi_i\rangle\circ dW_t^i=\langle m,d\chi_t\rangle-\ell(u,a)\,dt. \tag*{(30)}

The stochasticity enters the Legendre transformation because the momentum map mm is coupled to the stochastic vector field dχtd\chi_t. The left hand side of the transformation determines the Hamiltonian, which is a semimartingale. The underlying semidirect product group structure has not changed, it is still the HsH^s diffeomorphisms with a vector space, but the Hamiltonian has become a semimartingale. This implies that in the stochastic case the energy is not conserved, because Hamiltonian depends on time explicitly. Note that (30) emphasises that the Lagrangian does not feature stochasticity in this framework. Instead, the Lagrangian represents the physics in the problem, which does not change. The stochasticity is supposed to account for the difference between observed data and deterministic modelling. The stochastic Lie-Poisson equations are given by

d(m,a)=−ad⁡(δhδm,δhδa)∗(m,a) dt−∑i=1Nad⁡(ξi,0)∗(m,a)∘dWti,(31)d(m,a)=-\operatorname{ad}^{*}_{\left(\frac{\delta h}{\delta m},\frac{\delta h}{\delta a}\right)}(m,a)\,dt-\sum_{i=1}^{N}\operatorname{ad}^{*}_{(\xi_i,0)}(m,a)\circ dW_t^i, \tag*{(31)}

where ad⁡∗\operatorname{ad}^{*} is given in (13). Since both the drift and the diffusion part use the same operator in (31), the stochastic Lie-Poisson equations preserve the same family of Casimirs (or integral conserved quantities) as the deterministic Lie-Poisson equations. The stochastic Euler-Poincaré theorem has a stochastic Kelvin-Noether circulation theorem as a corollary.

Let the manifold MM be a submanifold of Rn\mathbb{R}^n with coordinates XX. Then the volume form can be expressed with respect to a density. That is, μ(dnX)=ρ0(X)dnX\mu(d^nX)=\rho_0(X)d^nX. By pushing forward ρ0\rho_0 along the stochastic flow gtg_t, one obtains ρ\rho. Let Cs\mathcal{C}^s be the space of loops γ:S1→Ds\gamma:S^1\to\mathcal{D}^s, which is acted upon from the left by Ds\mathcal{D}^s. Given an element m∈Xsm\in\mathfrak{X}^s, one can obtain a 1-form by formally dividing mm by the density ρ\rho.

The circulation map K:Cs×V∗→Xs∗∗\mathcal{K}:\mathcal{C}^s\times V^*\to\mathfrak{X}^{s**} is defined by

⟨K(γ,a),m⟩=∮γmρ.(32)\langle\mathcal{K}(\gamma,a),m\rangle=\oint_\gamma\frac{m}{\rho}. \tag*{(32)}

Given a Lagrangian ℓ:Xs×V∗→R\ell:\mathfrak{X}^s\times V^*\to\mathbb{R}, the Kelvin-Noether quantity is defined by

I(γ,u,a):=∮γ1ρδℓδu.(33)I(\gamma,u,a):=\oint_\gamma\frac{1}{\rho}\frac{\delta\ell}{\delta u}. \tag*{(33)}

One can now formulate the following stochastic Kelvin-Noether circulation theorem.

Theorem 6 (Stochastic Kelvin-Noether). Let ut=u(t)u_t=u(t) satisfy the stochastic Euler-Poincaré equation (27) and at=a(t)a_t=a(t) the stochastic advection equation (28). Let gtg_t be the flow associated to the vector field dχtd\chi_t. That is, dχt=dgt∘gt−1=ut dt+∑i=1Nξi∘dWtid\chi_t=dg_t\circ g_t^{-1}=u_t\,dt+\sum_{i=1}^{N}\xi_i\circ dW_t^i. Let γ0∈Cs\gamma_0\in\mathcal{C}^s be a loop. Denote by γt=gt∘γ0\gamma_t=g_t\circ\gamma_0 and define the Kelvin-Noether quantity I(t):=I(γt,ut,at)I(t):=I(\gamma_t,u_t,a_t). Then

dI(t)=∮γt1ρδℓδa⋄a dt.(34)dI(t)=\oint_{\gamma_t}\frac{1}{\rho}\frac{\delta\ell}{\delta a}\diamond a\,dt. \tag*{(34)}

Proof. The statement of the stochastic Kelvin-Noether circulation theorem involves a loop that is moving with the stochastic flow. One can transform to stationary coordinates by pulling back the flow to the initial condition. This pull-back yields

I(t)=∮γt1ρδℓδu=∮γ0gt∗(1ρδℓδu)=∮γ01ρ0gt∗(δℓδu).(35)I(t)=\oint_{\gamma_t}\frac{1}{\rho}\frac{\delta\ell}{\delta u}=\oint_{\gamma_0}g_t^*\left(\frac{1}{\rho}\frac{\delta\ell}{\delta u}\right)=\oint_{\gamma_0}\frac{1}{\rho_0}g_t^*\left(\frac{\delta\ell}{\delta u}\right). \tag*{(35)}

An application of the Kunita-Itô-Wentzell formula (17) leads to

dI(t)=∮γ01ρ0gt∗(dδℓδu+Ldχtδℓδu)=∮γ01ρ0gt∗(δℓδa⋄a) dt.(36)dI(t)=\oint_{\gamma_0}\frac{1}{\rho_0}g_t^*\left(d\frac{\delta\ell}{\delta u}+\mathcal{L}_{d\chi_t}\frac{\delta\ell}{\delta u}\right)=\oint_{\gamma_0}\frac{1}{\rho_0}g_t^*\left(\frac{\delta\ell}{\delta a}\diamond a\right)\,dt. \tag*{(36)}

since uu satisfies the stochastic Euler-Poincaré theorem. Transforming back to the moving coordinates by pushing forward with gtg_t yields the final result.

Thus, Theorem 6 explains how particle relabelling symmetry gives rise to the Kelvin-Noether circulation theorem via Noether’s theorem. When the only advected quantity present is the mass density, the loop integral of the diamond terms vanishes. This means that circulation is conserved according to Noether’s theorem for an incompressible fluid, or for a barotropically compressible fluid. The presence of other advected quantities breaks the symmetry further and introduces the diamond terms which generate circulation, as one can see in the Kelvin-Noether circulation theorem in equation (34). Consequently, the symmetry breaking due to additional order parameters can provide additional mechanisms for the generation of Kelvin-Noether circulation in ideal fluid dynamics.

Outlook

Stochastic geometric mechanics is an active field of mathematics which has recently established its utility for a broad range of applications in science. Basically, everything that can be done with Hamilton’s principle for deterministic geometric mechanics can also be made stochastic in the sense of Stratonovich. This is possible because the variational calculus in Hamilton’s principle requires only the product rule and chain rule from ordinary calculus. The happy emergence of the new science of stochastic geometric mechanics was celebrated with the publication of the book Albeverio et al. [ [2]. This book showcases some of the recent developments in stochastic geometric mechanics. Another collection of recent developments can be found in Castrillón López et al. [ [6]. An ongoing development is in the direction of rough geometric mechanics, initiated with a rough version of the Euler-Poincaré theorem in Crisan et al. [ [11]. Remarkably, variational principles which are driven by geometric rough paths again only require the product rule and the chain rule. Other directions involve the inclusion of jump processes, fractional derivatives and non-Markovian processes in geometric mechanics. For example, recent work by Albeverio et al. [ [3] shows that SDEs driven by semimartingales with jumps have weak symmetries and a corresponding extension of the reduction and reconstruction technique is discussed.

Acknowledgments

We are enormously grateful for many encouraging discussions over the years with T.S. Ratiu, F. Gay-Balmaz, C. Tronci, S. Albeverio, A.B. Cruzeiro, F. Flandoli, and also with our friends in project STUOD (stochastic transport in upper ocean dynamics) and in the geometric mechanics research group at Imperial College London. The work of DDH was partially supported by European Research Council (ERC) Synergy grant STUOD - DLV-856408. EL was supported by EPSRC grant [grant number EP/L016613/1] and is grateful for the warm hospitality at the Imperial College London EPSRC Centre for Doctoral Training in the Mathematics of Planet Earth during the course of this work.

References

  1. [1]Ralph Abraham and Jerrold E Marsden. Foundations of mechanics, volume 36. Benjamin/Cummings Publishing Company Reading, Massachusetts, 1978.
  2. [2]Sergio Albeverio, Ana Bela Cruzeiro, and Darryl Holm. Stochastic Geometric Mechanics: CIB, Lausanne, Switzerland, January-June 2015, volume 202. Springer, 2017.
  3. [3]Sergio Albeverio, Francesco C De Vecchi, Paola Morando, Stefania Ugolini, et al. Weak symmetries of stochastic differential equations driven by semimartingales with jumps. Electronic Journal of Probability, 25, 2020.DOI
  4. [4]Ayhtami Bethencourt de Leon, Darryl D Holm, Erwin Luesink, and So Takao. Implications of Kunita–Itô–Wentzell formula for k-forms in stochastic fluid dynamics. Journal of Nonlinear Science, pages 1–34, 2020.
  5. [5]Jean-Michel Bismut. Mécanique aléatoire. In Ecole d’Eté de Probabilités de Saint-Flour X-1980, pages 1–100. Springer, 1982.
  6. [6]Marco Castrillón López, David Martín de Diego, Tudor S. Ratiu, and Cesare Tronci. Journal of Geometric Mechanics: Special issue dedicated to Darryl D. Holm on the occasion of his 70th birthday, volume 11. 2019.
  7. [7]Colin Cotter, Dan Crisan, Darryl D Holm, Wei Pan, and Igor Shevchenko. Modelling uncertainty using circulation-preserving stochastic transport noise in a 2-layer quasi-geostrophic model. arXiv preprint arXiv:1802.05711, 2018.DOI
  8. [8]Colin Cotter, Dan Crisan, Darryl D Holm, Wei Pan, and Igor Shevchenko. Numerically modeling stochastic Lie transport in fluid dynamics. Multiscale Modeling & Simulation, 17(1):192–232, 2019a.
  9. [9]Colin Cotter, Dan Crisan, Darryl D Holm, Wei Pan, and Igor Shevchenko. A particle filter for Stochastic Advection by Lie Transport (SALT): A case study for the damped and forced incompressible 2D Euler equation. arXiv preprint arXiv:1907.11884, 2019b.arxiv.org/abs/1907.11884
  10. [10]Colin J Cotter, Georg A Gottwald, and Darryl D Holm. 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.DOI
  11. [11]Dan Crisan, Darryl D Holm, James-Michael Leahy, and Torstein Nilssen. Variational principles for fluid dynamics on rough paths. arXiv preprint arXiv:2004.07829, 2020.
  12. [12]Ana Bela Cruzeiro, Darryl D Holm, and Tudor S Ratiu. Momentum maps and stochastic Clebsch action principles. Communications in Mathematical Physics, 357(2):873–912, 2018.DOI
  13. [13]David G Ebin and Jerrold Marsden. Groups of diffeomorphisms and the motion of an incompressible fluid. Ann. Math, 92(1):102–163, 1970.DOI
  14. [14]David Gregory Ebin. On the space of Riemannian metrics. PhD thesis, Massachusetts Institute of Technology, 1967.DOI
  15. [15]Izrail Moiseevitch Gelfand, Richard A Silverman, et al. Calculus of variations. Courier Corporation, 2000.
  16. [16]Darryl D Holm. 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
  17. [17]Darryl D Holm, Jerrold E Marsden, and Tudor S Ratiu. The Euler–Poincaré equations and semidirect products with applications to continuum theories. Advances in Mathematics, 137(1):1–81, 1998.
  18. [18]Alexander A Kirillov. Unitary representations of nilpotent Lie groups. RuMaS, 17(4):53–104, 1962.DOI
  19. [21]J Lázaro-Cami and J Ortega. Stochastic Hamiltonian dynamical systems. Reports on Mathematical Physics, (1), 2008.
  20. [22]Jerrold E Marsden, Tudor Raţiu, and Alan Weinstein. Semidirect products and reduction in mechanics. Transactions of the american mathematical society, 281(1):147–177, 1984.DOI
  21. [23]Emmy Noether. Invariante Variationsprobleme. Nachrichten von der Gesellschaft der Wissenschaften zu göttingen, 235-257. E. Noether: Gesammelte Abhandlungen, ed. N. Jacobson, pages 248–270, 1918.
  22. [24]Emmy Noether. Invariant variation problems. Transport Theory and Statistical Physics, 1(3):186–207, 1971.DOI
  23. [25]Henri Poincaré. Sur une forme nouvelle des équations de la mécanique. CR Acad. Sci., 132:369–371, 1901.
  24. [28]Jean-Marie Souriau. Structure des systèmes dynamiques. Dunod, Paris, 1970.

Paper details

vlnc:W_6k986phj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents