Arxiv:2007.13692V2 [Physics.Flu-Dyn] 1 Oct 2020 1.A

Home , Actuarial science, Wiener process

arXiv:2007.13692v2 [physics.flu-dyn] 1 Oct 2020 1.A. nan in n aifisdiv satisfies and density, eedlnal ntedniyw ee o[ to refer we density the on linearly depend h torus the ocsare forces (1.1) where (1.2) ftedniy htis that density, the of nrywe h udi neulbim Further, equilibrium. in is fluid the when energy (1.3) where Let (1.4) (1.5) iazakafih FA,Ot-anrPaz5 -00Vi A-1090 address 5, E-mail Platz Otto-Wagner (FMA), Finanzmarktaufsicht o akrudrgrigtebrtoi airSoe equatio Navier-Stokes barotropic the regarding background For h opesbeNve-tkseutoswt est dependent density with equations Navier-Stokes compressible The κ h aorpcNve-tksequations. Navier-Stokes barotropic The n ≥ ∇ ν ailr ocscnas etetdb hsapoc.Det h Ha stochas the along to theorem Due circulation approach. Kelvin a this satisfies by system treated field equ be mean dens also McKean-Vlasov with can the equation Navier-Stokes forces that barotropic Capillary shown the yields is limit It field sto mean (HIPS). a as system interpreted particle is ing components (stochastic) fast and ministic) Abstract. dmninldmi ihproi onaycniin.Tevelocity, The conditions. boundary periodic with domain -dimensional u AITNA NEATN ATCESSE FOR SYSTEM PARTICLE INTERACTING HAMILTONIAN A ρ ≥ eacntn.Tecplaytensor, capillary The constant. a be 0 u M = stevsoiycecet h orsodn oc is force corresponding The coefficient. viscosity the is 0 = = ρ h div ( u, : ,x t, R [email protected] C n ∇i S h eopsto fteeeg facmrsil udpre nos into parcel fluid compressible a of energy the of decomposition The / ,aeatm-eedn etrfil n ucin epciey d respectively, function, and field vector time-dependent a are ), = Z u = n κρ = . p X ∇ = P C u ∆ ρ ˙ ˙ ∂ ρ u i ρ = = 2 = S j . U ∂ ij −∇ − κ j ′ e ( u OPESBEFLOW COMPRESSIBLE div( j ρ h yrsai pressure, hydrostatic The . o nw function known a for ) = u ρ u S ∆ ρu IO HOCHGERNER SIMON ν − ij ( 1. ρ ∇ ) ρ = + ⊤ Introduction − u 1 νρ 2 1 ∇ ) . h∇ 43 ∇ p ∂ n eeecstherein. references and ] ρ ρ, + 1 i C u + ∇ ρ j sdfie as defined is , S − νρ + ρ 1 i stesrs esr endby defined tensor, stress the is ∇ ∂ osdracmrsil aorpcfluid barotropic compressible a Consider div I j div( u ∇ − i S U u div + ρ hc oesteseicinternal specific the models which + ) p ∇ ⊗ sasmdt egvni terms in given be to assumed is , hsi aitna interact- Hamiltonian chastic ν ∇ C enna ρ ∇ itna tutr the structure miltonian i arninpaths. Lagrangian tic t eedn viscosity. dependent ity to soitdt the to associated ation ρ swt icste which viscosities with ns u + icst n capillary and viscosity νρ ∆ u. o (deter- low u = u ( fie on efined ,x t, ,and ), 2 HIPS FOR COMPRESSIBLE FLOW

The capillary tensor (1.5) appears in this form also in [3, Equ. (4)] and goes back to Kor- teweg [37]. Analytic aspects of the barotropic system with capillary forces (1.1)-(1.2) are treated in [10]. Navier-Stokes equations with more general third-order spatial derivative terms are discussed in [36].

1.B. Description of results. This paper is concerned with a mean field representation of solutions to the compressible Navier-Stokes system (1.1)-(1.2), and this mean field is derived from a stochastic Hamiltonian interacting particle system (Hamiltomian IPS or HIPS). The HIPS picture follows from a combination of a many particle approach to fluid dynamics and a decomposition of the energy into slow (deterministic) and fast (stochastic) components. The interpretation as a slow-fast decomposition is consistent with the multi-time formulation of [15], where it is shown that advection along stochastic transport fields in the Eulerian representation of stochastic fluid dynamics can be obtained by homogenization. To describe the many particle approach, consider a tiny Eulerian volume (fixed in space and time) ∆V which is divided into a very large number N of equal subvolumes. It is assumed that the continuum hypothesis holds in each of the infinitesimal subvolumes ∆V α. Thus there is a mass density ρα for each grid index α. Since the subvolumes are equal, so are the initial α α conditions for ρ , and ρ |0 = ρ|0 which is the initial condition for the overall mass density in ∆V . Now, the fluid parcels in all of the subvolumes interact because the deterministic component of the energy of the blob of fluid in ∆V depends on the total momentum and the overall mass density. (‘Momentum’ shall always refer to momentum per unit volume, i.e. its dimension is density times velocity.) On the other hand, molecular diffusion is modeled as a set of N independent (multidimensional) Brownian motions such that all N individual parcels undergo their own stochastic process. Since molecules are incompressible these processes are set up as stochastic perturbations along divergence free vector fields. Hence the fluid parcel in ∆V consists of N identical subparcels, and each subparcel follows the flow of the ensemble of subparcels (is dragged along or advected) but also undergoes its own diffusion. Heuristically, this means that the interaction in the IPS is due to the deterministic part of the motion where each infinitesimal subparcel follows the mean flow of the ensemble. The corresponding total Hamiltonian (3.22) for the fluid in M is then obtained by integrating the energies over the infinitesimal subvolumes ∆V α and summing over all indices α. Since this Hamiltonian describes a system consisting of N subparcels, it is a function N HN : T ∗(Diff(M)sF(M) → R which is the N-fold product of the phase space of compressible fluid mechanics. (The semi- direct product notation is explained in Section 2.C.) The canonical symplectic structure on the phase space then yields a Hamiltonian system of Stratonovich SDEs by adapting the construction of [39] to the infinite dimensional setting. This system of N interacting SDEs will be called the ‘HIPS equations of motion’. In Section 3.A it is explained how the Hamiltonian HN is a sum of terms involving: translational kinetic energy of the particle ensemble, equilibrium internal energy associated to the hydrostatic pressure, equilibrium internal energy associated to capillary forces, non-equilibrium internal energy due to expansion/compression along the flow, and stochastic energy associated to molecular bombardment. The HIPS equations of HIPS FOR COMPRESSIBLE FLOW 3 motion are derived in Section 3.B, see (3.25)-(3.27). Under the assumption that the mean field limit exists, as N →∞, the stochastic mean field equations are obtained in Section 3.C. The mean field SDEs (3.32)-(3.33) represent the Eulerian description of the motion of a fluid parcel associated to a subdivision ∆V α for very large N. The expected fluid flow is then obtained by averaging over momenta and mass densities of all the smaller fluid parcels. But these are just the mean fields. It therefore remains to calculate the equations of motions for the latter. These equations are deterministic and given in (3.34)-(3.35). In Section 3.D it is shown that, if the stochastic perturbation corresponds to a Brownian motion for each Fourier mode (i.e., is a cylindrical Wiener process in the space of solenoidal vector fields), then the momentum and mass density mean fields solve the compressible Navier- Stokes system (1.1)-(1.2). This is the content of Theorem 3.4. Moreover, since the HIPS is invariant under the particle relabeling symmetry (invariance with respect to the group of diffeomorphisms) Noether’s Theorem implies that a Kelvin Circulation Theorem holds along stochastic Lagrangian paths (Proposition 4.1). The density dependence of the viscosity, as manifest by the ρ-factor in (1.3), is a consequence of the form of the Hamiltonian (3.22). The compressible Navier-Stokes equations are often expressed with respect to a density-independent viscosity, but it is not clear how to realize this independence in the HIPS framework. Mean field representations of solutions to the incompressible Navier-Stokes equation have been previously obtained by [13, 34] by considering a Weber functional along stochastic La- grangian paths. In [29, 30] the above described Eulerian (H)IPS formulation has been used to also obtain a stochastic mean field representation for solutions to the incompressible Navier- Stokes equation. To the best of my knowledge, Theorem 3.4 is the first stochastic representation of solutions to the compressible Navier-Stokes equations. Related approaches to fluid dynamics from the perspective of stochastic variational principles include [54, 14, 18], which characterize solutions to the incompressible Navier-Stokes equations by variational principles for stochastic Lagrangian paths, and generalizations in [4, 11].

1.C. Applications: NatCat models and Solvency II capital requirements. Uses of the Navier-Stokes equation range from semi-conductor engineering to astrophysics and it is not the goal of this section to attempt a review of these topics. Rather, I want to briefly describe an application where the interaction between academia and industry is perhaps not very well established. This concerns models of natural perils (NatCat models) that are used in the insurance industry to calculate risk capital requirements. These risk capital requirements are a determining factor for the solvency of a given company. In the EU the relevant regulatory framework is called Solvency II ([55]). There exist different NatCat models, which are in general proprietary, for natural disasters such as earthquakes, flooding, tropical cyclones, and extratropical cyclones (European winter storms). Storm models are often based on numerical weather prediction (NWP) systems, and therefore inherit all their advantages and flaws. However, it is not claimed that the stochastic HIPS formulation of this paper is appropriate to generate stochastic NatCat storm scenarios. Thus the material in this section is only intended as additional background information concerning possible applications of stochastic fluid mechanics, and it is logically independent from the main result, Theorem 3.4. 4 HIPS FOR COMPRESSIBLE FLOW

Numerical weather prediction (NWP) and climate modeling. NWP systems and climate models are important for a number of apparent reasons such as daily weather forecasts or climate change quantification. Geophysical flows are modeled by the compressible Navier-Stokes equations (without capillary term, i.e. C = 0 in (1.1)). These equations are deterministic. However, any implementation of these equations needs to introduce temporal and spatial discretizations. Physical processes which occur below these chosen grid scales (‘subgrid phenomena’) cannot be ac- counted for by any numerical model of the deterministic flow equations. Since the subgrid processes are inherently unknown and uncertain it seems reasonable to model these by a stochastic dynamics point of view. Indeed, such a position has been adopted quite early by Kraichnan [38]. See also [46] for a modern version and fluid equations with stochastic force terms. Reviews regarding NWP systems and climate models are contained in [8, 9, 47]. These also address the need for stochastic parameterizations of unknown subgrid processes. Recent advances in the stochastic modeling of geophysical flows include the ‘location uncertainty’ approach of Mémin, Resseguier and collaborators [44, 49, 50, 51] as-well as the ‘stochastic advection by Lie transport (SALT)’ theory of Holm and collaborators [32, 15, 16, 22, 2]. The SALT approach is based on the observation that subgrid phenomena represent unknown physical processes and should therefore be derived from a stochastic variational principle. As a consequence, these models preserve circulation along stochastic Lagrangian paths.

NatCat models and solvency capital requirement (SCR). With the implementation of the Sol- vency II regulatory regime ([55]) per 1. January 2016, applied insurance mathematics has become a surprisingly diverse and multi-disciplinary subject. The relevant tools extend beyond classical actuarial science to, e.g., modeling of local general accounting principles ([31, 26]), no-arbitrage principles and stochastic interest rate models ([53, 24]), and stochastic fluid dynamics. The significance of the last point in this (certainly incomplete) list is briefly explained below. One of the basic quantitative principles of Solvency II can be summarized as follows: In- surance and reinsurance undertakings are required to quantify all relevant risk factors over a one-year horizon and derive a corresponding loss distribution. Now, the own funds (i.e. excess of assets over liabilities) have to cover the 99.5 percentile of this distribution (‘survival of the 200 year event’). This 99.5 percentile corresponds to the so-called solvency capital requirement (SCR). The SCR can be calculated from a prescribed standard formula or a company specific internal model. Medium and large sized companies generally use internal models. If a company has chosen the internal model approach and covers windstorm risks, then the corresponding loss distribution for a one-year period has to be derived. To do so the following three-step procedure, or a variant of it, is used ([27]):

(1) Hazard module: physical model. A large set of windstorm scenarios is generated. These are the so-called stochastic scenarios of the NatCat model. (2) Vulnerability module. This step quantiﬁes the vulnerability (i.e., damage done) of the insured structures for each of the stochastic scenarios. HIPS FOR COMPRESSIBLE FLOW 5

(3) Financial loss module. Finally, structural damage is converted to loss by taking into account contract specifics (e.g., sum insured) for each damaged structure and, possibly, reinsurance. Since the relevant time period for NatCat models is one year they operate on a new temporal scale when compared to short term weather forecast models and medium or long term climate models. The most advanced NatCat models are based on a coupling of a global circulation model (GCM) and a NWP system. Since these models are proprietary it is not possible to cite a suitable model documentation. However, [12] contains a review of the basic method, which is still quite up to date. In particular, it is explained how (deterministic) NWP systems are used to generate a set of scenario events by statistically sampling the initial conditions. “Using NWP technology, a large set of potential future storms is generated by taking data sets comprising the initial pressure fields of historical storms, per- turbing them both temporally and spatially, and moving them forward in time through the application of a set of partial differential equations governing fluid flow. The resulting event set is rigorously tested to ensure that it provides an appropriate representation of the entire spectrum of potential storm experience – not just events of average probability, but also the extreme events that make up the tail of the loss distribution.” ([12]) This point will be taken up again in Section 5. A recent reanalysis of European winterstorm events, which also highlights the potential financial risks for the insurance sector, has been carried out by [33].

2. Notation and preliminaries 2.A. Diffeomorphism groups. Let M = T n = Rn/Zn. We fix s> 1+ n/2 and let Diff(M)s ∞ s s denote the infinite dimensional C -manifold of H -diffeomorphisms on M. Further, Diff(M)0 denotes the submanifold of volume preserving diffeomorphisms of Sobolev class Hs. Both, s s Diff(M) and Diff(M)0, are topological groups but not Lie groups since left composition is only continuous but not smooth. Right composition is smooth. The tangent space of Diff(M)s s s s s (resp. Diff(M)0) at the identity e shall be denoted by g (resp. g0). Let X (M) denote the s s vector fields on M of class H and X 0(M) denote the subspace of divergence free vector fields s s s s s of class H . We have g0 = X 0(M) and g = X (M). The superscript s will be dropped from now on. We use right multiplication Rg : Diff(M) → Diff(M), k 7→ k ◦ g = kg to trivialize the ∼ g −1 tangent bundle T Diff(M) = Diff(M) × g, ξg 7→ (g, (T R ) ξg), and similarly for Diff(M)0. The Riemannian metric on Diff(M) × g ∼= T Diff(M) is defined by

hhξg, ηgii = hξ(g(x)), η(g(x))i dx ZM for ξ, η ∈ g, where dx is the standard volume element in M, and h., .i is the Euclidean inner product. See [7, 25, 40, 45] for further background. 6 HIPS FOR COMPRESSIBLE FLOW

2.B. Derivatives. The adjoint with respect to hh., .ii to the Lie derivative L, given by LX Y = ∇X Y −∇Y X, is ⊤ ⊤ (2.6) LX Y = −∇X Y − div(X)Y − (∇ X)Y i j ⊤ j j with ∇X Y = hX, ∇iY = X ∂iY ej and (∇ X)Y = (∂iX )Y ei with respect to the ⊥ ⊤ standard basis ei, i =1,...,nP. The notation ad(X)Y = [X,YP ]= −LX Y and ad(X) = −LX will be used. The variational derivative of a functional F : g → R will be denoted by δF/δX, that is δF ∂ (2.7) hh ,Y ii = F (X + tY ) δX ∂t 0 for X,Y ∈ g.

2.C. Semi-direct product structure. The configuration space of compressible fluid mechanics on M is the semi-direct product Diff(M)sF(M) where F(M) denotes functions (also of Sobolev class s) on M. The semi-direct product structure is defined by the right action (2.8) R(φ,g)(ψ, f)=(ψ ◦ φ, f ◦ φ + g). The corresponding phase space is trivialized with respect to this right multiplication as T ∗(Diff(M)sF(M)) ∼= Diff(M)sF(M) × g∗ ×F(M)∗. The variables µ ∈ g∗ and ρ ∈ F(M)∗ represent momentum and mass density, respectively. Making use of the Euclidean volume form dx, the duals can be identified as g∗ = Ω1(M) and F(M)∗ = F(M). Details on Hamiltonian mechanics on semi-direct products are given in [41], where also the case of compressible ideal fluids is treated.

2.D. Stochastic dynamics. Let (Ω, F, (Ft)t∈[0,T ],P ) be a filtered probability space satisfying the usual assumptions as specified in [48]. In the following, all stochastic processes shall be understood to be adapted to this filtration. The symbol

(2.9) δt will be used to denote the Stratonovich differential to distinguish it from the variational derivative δ. The Ito differential does not appear in this paper. The exterior differential is d.

2.E. Brownian motion in g0. Let Z+ Z n := {k ∈ n : k1 > 0 or, for i =2,...,n,k1 = ... = ki−1 =0,ki > 0}. Z+ ⊥ ⊥ Rn For k ∈ n let k1 ,...,kn−1 denote a choice of pairwise orthogonal vectors in such that ⊥ ⊥ |ki | = |k| and hki ,ki = 0 for all i =1,...,n − 1. Consider the following system of vectors in g0 ([17, 19]): 1 1 A = coshk, xik⊥, B = sinhk, xik⊥, A = e (k,i) |k|s+1 i (k,i) |k|s+1 i (0,j) j n where ej ∈ R is the standard basis and s is the Sobolev index from Section 2.A. By slight abuse of notation we identify these vectors with their corresponding right invariant vector ﬁelds on Diﬀ(M)0. HIPS FOR COMPRESSIBLE FLOW 7

Further, in the context of the ζr vectors we shall make use of the multi-index notation Z+ r =(k, i, a) where k ∈ n and a =0, 1, 2 such that

ζr = A(0,i) with i =1,...,n if a =0

ζr = A(k,i) with i =1,...,n − 1 if a =1

ζr = B(k,i) with i =1,...,n − 1 if a =2

Thus by a sum over ζr we shall mean a sum over these multi-indices, and this notation will be used throughout the rest of the paper.

It can be shown (see [19, Appendix] for details) that the ζr form an orthogonal system of basis vectors in g0, such that

(2.10) ∇ζr ζr =0 and, for X ∈ X (M),

s (2.11) ∇ζr ∇ζr X = c ∆X X s n−1 1 + 2 where c =1+ n k∈Zn |k| s is a constant and ∆ is the vector Laplacian. P p r Proposition 2.1 ([14, 17, 18, 23]). Let Wt = ζrWt , where Wt are independent copies of Brownian motion in R. Then W deﬁnes (a versionP of) Brownian motion (i.e., cylindrical Wiener process) in g0.

3. HIPS for compressible flow This section is concerned with the HIPS equations of motion and the mean ﬁeld limit. Background on mean ﬁeld theory can be found in [52, 1, 20, 42, 35].

3.A. The Hamiltonian. Consider a barotropic fluid in M = Rn/Zn. At a (macroscopic) position x ∈ M consider a tiny volume element ∆Vx. Suppose ∆Vx is further divided into N α infinitesimal volume elements, ∆Vx , labeled by α = 1,...,N, which are all assumed to have identical dimensions. In each volume element there is a mass density ρα = ρα(x). Thus it is assumed that the continuum description of the fluid also holds at the level of the subdivsion. α α Let the fluid element in volume ∆Vx have velocity v (x). Thus the energy of the particle ensemble in the total infinitesimal volume is determined by the velocities, vα, and densities, ρα, in the subdivisions. To arrive at the total energy (3.22) we consider five contributions: (1) Translational kinetic energy; (2) Equilibrium internal energy U which gives rise to the hydrostatic pressure p; (3) Equilibrium capillary energy; (4) Non-equilibrium expansion/compression energy; (5) Stochastic energy due to molecular bombardment; 8 HIPS FOR COMPRESSIBLE FLOW

Translational kinetic energy. The total velocity at x, v(x), is the weighted average N ρα(x)vα(x) (3.12) v(x)= α=1 . P N α α=1 ρ (x) P The total momentum, µ(x), is therefore

(3.13) µ(x)= ρα(x)vα(x)= µα Xα Xα and the translational kinetic energy KN (x) of the particle system is

1 1 µβ(x) (3.14) KN (x)= hµ(x), v(x)i = µα(x), β 2 2 P ρν(x) D Xα ν E P

Equilibrium internal energy. Notice that the overall mass density in ∆Vx is expressed, in α this subdivision picture, as ρ(x) = ρ (x)/N. Consequently, the mass contained in ∆Vx is α α ρ(x)·∆Vx = ρ dx where ∆Vx ≈P∆Vx/N is identified with the infinitesimal volume element dx. P The barotropicity assumption implies that the specific equilibrium internal energy in ∆V depends on the overall mass density ρα/N. Then the equilibrium internal energy in ∆V is P (3.15) ραU ρα/N dx. X X Capillary energy. Let κ ≥ 0 be a constant. The capillary energy is defined as

1 α α (3.16) κ 2 ∇ ρ , ∇ ρ /N dx D X X E which depends, again, on the overall mass density. This form coincides with the capillary contribution to the Helmholtz free energy [3, Equ. (5)].

Non-equilibrium expansion energy. Let ν ≥ 0 be a constant. Assume temporarily that ∆Vx is a box which is aligned along Cartesian coordinates e1, e2, e3 and that the ﬂow has only velocity components pointing in the direction of the e1-axis. Let L denote the set of labels α α such that the corresponding subvolumes ∆V constitute the left wall of ∆Vx while R denotes those which correspond to the right wall (viewed along e1) of the volume. Now, if ραvα ραvα (3.17) α∈L − α∈R P α P α α∈L ρ α∈R ρ P P is greater than 0, then particles moving into ∆V are faster than those moving out, and the corresponding energy diﬀerence should contribute to the (non-equilibrium) internal energy of the system. Since this expression is proportional to minus the divergence of the barycentric velocity, we propose a non-equilibrium internal energy expression ραvα (3.18) − ν ραdiv α dx. P α X α ρ P HIPS FOR COMPRESSIBLE FLOW 9

α Stochastic energy. The energy due to molecular bombardment along a vector ﬁeld ξr in ∆Vx ≈ dx is α α r (3.19) hρ v ,ξri δtW dx. See [39, 28, 32, 29, 30]. This stochastic perturbation corresponds to individual molecules im- α parting their velocities, namely ξr, on the macroscopic ﬂuid element in ∆V . Since individual molecules are incompressible the ξr are assumed to be divergence free.

Total energy. The above subdivision formulation implies that the total configuration space N α α α is of the form Πα=1(Diff(∆V )sF(∆V ))). But now the ∆V , which are all identical by assumption, are identified with the infinitesimal element dx in M. Thus each index α corresponds to a copy of M, and this can be done since the for the state variables, velocity and density, it does not make a difference whether these are regarded at the ∆V or at the ∆V α level. Therefore, letting the position x ∈ M range over the full domain, the total configuration N N space is Πα=1(Diff(M)sF(M)) = (Diff(M)sF(M)) . Let us switch from velocity and density to momentum (density), µα = ραvα, and density, ρα, as state variables. The phase space is thus T ∗(Diff(M)sF(M))N =(T ∗(Diff(M)sF(M)))N . We use the Euclidean metric to identify each copy in the phase space, which is the regular dual, as T ∗(Diff(M)sF(M)) = T (Diff(M)sF(M)) = (Diff(M)sF(M)) × (X (M)sF(M)) where the last identification follows from right-multiplication in the semi-direct product group, see Section 2.C. The resulting Hamiltonian of the IPS will therefore be a function N (3.20) HN : T (Diff(M)sF(M)) → R. For N N (3.21) Γ = Φα, f α; µα, ρα ∈ T (Diff(M)sF(M)) α=1 the total Hamiltonian is the sum of (3.14), (3.15), (3.16), (3.18) and (3.19), and given in semi-martingale notation as 1 µβ (3.22) HN (Γ) = µα, β dx δ t 2 Z P ργ t M D Xα γ E P ρβ + ραU β dx δ t Z PN t M Xα + κ ∇ρα, ∇ρα/N dx δ t Z t M D Xα Xα E µβ − ν ραdiv β dx δ t Z P ργ t M Xα γ P + ε hµα,ξ i dx δ W r,α Z r t M Xj,α 10 HIPS FOR COMPRESSIBLE FLOW

r,α r s where W are pairwise independent Brownian motions such that [W , W ]t = δr,st, the solenoidal vector fields ξr ∈ X 0(M) are fixed and ε ≥ 0 is a constant. The Hamiltonian is right-invariant by construction as it does not depend on (Φα, f α) ∈ Diff(M)sF(M).

3.B. HIPS equations of motion. The phase space (3.20) is an N-fold direct product of a tangent bundle identiﬁed with its dual. It carries therefore the corresponding direct prod- N α α N uct canonical symplectic form. Since the Hamiltonian H does not depend on (Φ , f )α=1 ∈ (Diﬀ(M)sF(M))N we can pass via Lie-Poisson reduction to the phase space (X (M)sF(M))N . The Hamiltonian IPS equations of motion follow therefore from the variational derivatives (again, in the semi-martingale notation)

δHN µβ (3.23) = β δ t + ν∇ log ρβ δ t + ε ξ δ W r,α δµα P ργ t t r t γ Xβ Xr P N N r,α = u δtt + ν∇ log ρ δtt + ε ξr δtW Xr δHN 1 µβ µβ ρβ ρβ ρβ (3.24) = − β , β + U β + β U ′ β α P γ P γ P P P δρ 2D γ ρ γ ρ E N N N P β P ρ 1 − κ∆ β − ν div µβ δ t PN ρβ t β Xβ P 1 N N N N ′ N N 1 N = − hu ,u i + U(ρ )+ ρ U (ρ ) − κ∆ρ − ν div(µ ) δtt 2 ρN by using the N-fold product of the semi-direct product structure ([41]). Here the abbreviations β β N N Pβ µ N Pβ ρ N µ µ := N , ρ := N and u := ρN are used.

Remark 3.1. The variational derivatives, δHN /δµα and δHN /δρα, also depend on µβ and ρβ with β =6 α. Hence the right hand sides in the equations (3.25)-(3.27) below cannot be viewed as vector ﬁelds on X (M)sF(M), but only as Cartesian projections of vector ﬁelds on the full space (X (M)sF(M))N .

Remark 3.2. Equations (3.23) and (3.24) depend only on the empirical averages µN and ρN .

Remark 3.3. The quantity uN is the speciﬁc momentum (viewed as a vector ﬁeld) of the ensemble average. But it is (for non-constant density) not equal to the empirical average of N P µα/ρα velocities, that is u =6 N . HIPS FOR COMPRESSIBLE FLOW 11

The stochastic Hamilton equation associated to (3.22) for (Φα,µα, ρα) are:

N α δH α (3.25) δtΦt = ◦ Φt δµα N N α α r,α = u + ν∇ log ρ ◦ Φt δtt + ε ξr ◦ Φt δtW Xr ⊤ N α α α −1 α δH α (3.26) δtµt = −ad (δtΦt ) ◦ (Φt ) µt − ⋄ ρt δρα N α N α α ⊤ N α α α δH = −∇ N α µ − div(δH /δµ )µ − (∇ δH /δµ )µ − ρ ∇ δH /δµ δρα α N N α ⊤ N ⊤ N α = −∇uN +ν∇ log ρN µ − div(u + ν∇ log ρ )µ − (∇ u + ν∇ ∇ log ρ )µ − ρα − (∇⊤uN )uN +(ρN )−1∇ (ρN )2U ′(ρN ) − κ∇∆ρN N −2 N N N −1 N + ν(ρ ) div(µ )∇ρ − ν(ρ ) ∇div(µ ) δtt ⊤ α r,α − ε ad ξr µ δtW Xr α α α −1 (3.27) δtρt = −L(δtΦt )◦(Φt ) ρt α N α N α r,α = −div ρ u + νρ ∇ log ρ δtt − ε div ρ ξr δtW Xr Here ρα is viewed as a density whence L is the Lie derivative of a density, not of a function. The momentum variable is identiﬁed, via the Euclidean metric, as an element

(3.28) µα ∈ X (M) whence the transpose Lie derivative L⊤ is used instead of L∗. The diamond notation in (3.26) is deﬁned by f ⋄ ρ = ρ∇f and this term arises because of the semi-direct product structure.

3.C. Mean field limit. Assume the mean field limit of (3.25)-(3.27) exists, for N → ∞. Since all subvolumes ∆V α and their enclosed fluid elements are identical it suffices to consider α = 1,

1 1 1 1 (3.29) (Φt , ft ,µt , ρt ) −→ (Φt, ft,µt, ρt) as N →∞. Hence

(3.30) µN −→ E[µ]=:µ ¯ and ρN −→ E[ρ]=:ρ ¯ and µN (3.31) uN = −→ µ/¯ ρ¯ =: u ρN as N →∞. 12 HIPS FOR COMPRESSIBLE FLOW

Note that µ and ρ are stochastic processes while u is deterministic. The mean ﬁeld limit equations of motion for µ and ρ are

⊤ ⊤ (3.32) δtµ = −∇u+ν∇ logρ ¯µ − div(u + ν∇ logρ ¯)µ − (∇ u)µ − ν(∇ ∇ logρ ¯)µ + ρ(∇⊤u)u − ρρ¯−1∇ ρ¯2U ′(¯ρ) + κρ∇∆¯ρ −2 −1 − νρρ¯ div(¯µ)∇ρ¯ + νρρ¯ ∇div(¯µ) δtt ⊤ r − ε ad(ξr) µ δtW Xr r (3.33) δtρt = −div(ρu + νρ∇ logρ ¯) δtt − ε div(ρξr) δtW Xr

These equations are linear in µ and ρ, and depend otherwise on the mean ﬁelds E[µ]=¯µ and E[ρ]=¯ρ. Let p :=ρ ¯2U ′(¯ρ). Using thatµ ¯ =ρu ¯ , the equations for the expectationsµ ¯,ρ ¯ are therefore

⊤ (3.34) µ¯˙ = −∇u+ν∇ logρ ¯µ¯ − div(u + ν∇ logρ ¯)¯µ − ν(∇ ∇ logρ ¯)¯µ ε2 −∇p + κρ¯∇∆¯ρ − νρ¯−1div(¯µ)∇ρ¯ + ν∇div(¯µ)+ L⊤ L⊤ µ¯ 2 ξr ξr Xr ε2 (3.35) ρ¯˙ = −div ρu¯ + ν∇ρ¯ + div div(¯ρξ )ξ 2 r r Xr

3.D. Barotropic Navier-Stokes equation. Assume that the perturbation vector ﬁelds ξr 2 s are given by ζr, deﬁned in Section 2.E and that ε c /2= ν. Then (2.11) implies

ε2 ε2 (3.36) L⊤ L⊤ µ¯ = ν∆¯µ and div div(¯ρζ )ζ = ν∆¯ρ. 2 ζr ζr 2 r r Xr Xr

(The explicit calculation is carried out in [30, Lemma 4.3].) Hence equations (3.34) and (3.35) become

⊤ (3.37) µ¯˙ = −∇u+ν∇ logρ ¯µ¯ − div(u + ν∇ logρ ¯)¯µ − ν(∇ ∇ logρ ¯)¯µ −∇p + κρ¯∇∆¯ρ − νρ¯−1div(¯µ)∇ρ¯ + ν∇div(¯µ)+ ν∆¯µ (3.38) ρ¯˙ = −div(¯ρu) HIPS FOR COMPRESSIBLE FLOW 13

Therefore, ˙ ∂ µ¯ = ∂t (¯ρu)= −div(¯ρu)u +ρ ¯u˙ = −ρ¯∇u+ν∇ logρ ¯u − div(¯ρu)u − νdiv(¯ρ∇ logρ ¯)u − ν(∇⊤∇ logρ ¯)¯ρu −∇p + κρ¯∇∆¯ρ − νρ¯−1h∇ρ,¯ ui∇ρ¯ − νdiv(u)∇ρ¯ + ν∇ h∇ρ,u¯ i +ρu ¯ + ν∆(¯ρu) = −ρ¯∇u+ν∇ logρ ¯u − div(¯ρu)u − ν(∆¯ρ)u −1 − ν∇u∇ρ¯ + νρ¯ hu, ∇ρ¯i∇ρ¯ −∇p + κρ¯∇∆¯ρ − νρ¯−1h∇ρ,¯ ui∇ρ¯ − νdiv(u)∇ρ¯ + ν(∇⊤∇ρ¯)u + ν(∇⊤u)∇ρ¯ + νdiv(u)∇ρ¯ + νρ¯∇div(u)+ ν(∆¯ρ)u

+2ν∇∇ρ¯u + νρ¯∆u ⊤ = −ρ¯∇uu − div(¯ρu)u −∇p + κρ¯∇∆¯ρ + ν(∇ u)∇ρ¯ + νρ¯∇div(u)+ ν∇∇ρ¯u + νρ¯∆u Deﬁne the stress tensor, S, by

j i (3.39) Sij = νρ¯ ∂iu + ∂ju and the corresponding force

⊤ (3.40) div S = ∂iSijej = ν(∇ u)∇ρ¯ + νρ¯∇div(u)+ ν∇∇ρ¯u + νρ¯∆u. X The capillary tensor, C, is deﬁned (see [3, Equ. (4)]) by

(3.41) C = κ ρ¯∆¯ρ + 1 h∇ρ,¯ ∇ρ¯i I −∇ρ¯ ⊗∇ρ¯ 2 and satisﬁes div C = κρ¯∇∆¯ρ. Note that ν ≥ 0 and κ ≥ 0 are constants, and thatµ ¯, resp.ρ ¯ are a time dependent vector ﬁeld, resp. function by construction. It follows that:

s 2 Theorem 3.4. The mean ﬁeld equations (3.32) and (3.33) imply, if ξr = ζr and c ε /2= ν, that the expectations µ¯ = E[µ] and ρ¯ = E[ρ] satisfy the compressible Navier-Stokes equations

−1 −1 (3.42) u˙ = −∇uu − ρ¯ ∇p +ρ ¯ div S + div C (3.43) ρ¯˙ = −div(¯ρu) where u =µ/ ¯ ρ¯.

4. Stochastic Kelvin Circulation Theorem In [21] it is shown that stochastic Euler-Poincaréfluid equations are characterized by preserving circulation along Lagrangian paths. Since (3.32)-(3.33) are obtained as a mean field limit of a Hamiltonian IPS, and can be viewed of as mean field generalization of the stochastic fluid system in [21], there should be a Kelvin Circulation Theorem: 14 HIPS FOR COMPRESSIBLE FLOW

Proposition 4.1. Let C be a smooth closed loop which is transported by the Lagrangian flow Φt, defined through the mean field limit (3.29) and characterized by

−1 r (4.44) δtΦt ◦ Φt = ut + ν∇ logρ ¯ δtt + ε ξr δtWt . X Let µt and ρt be solutions of (3.32) and (3.33). Then

−1 ♭ (4.45) δt ρt µt =0 Z(Φt)∗C where ♭ is the Euclidean isomorphism to one-forms (since µ is treated as a vector ﬁeld).

Proof. Equations (3.25) and (3.29) yield

r (4.46) δtΦ= u + ν∇ logρ ¯ ◦ Φ δtt + ε ξr ◦ Φ δtW X −1 Now, (3.32) and (3.33) imply that Xt := ρt µt satisﬁes −2 −1 (4.47) δtX = −ρ ( δtρ)µ + ρ δtµ

⊤ ⊤ = −∇u+ν∇ logρ ¯X − (∇ u)X − ν(∇ ∇ logρ ¯)X −∇p˜ δtt ⊤ r − ε −∇ξr X − (∇ ξr)X δtW X with 1 (4.48)p ˜ := − hu,ui + U(¯ρ)+¯ρU ′(¯ρ) − κ∆¯ρ − νρ¯−1div(¯µ). 2 Hence, with parameterization C = c([0, 1]):

♭ ∗ ♭ ♭ δt Xt = δt Φt Xt = δt Xt ◦ Φt .T Φt Z(Φt)∗C ZC ZC 1 ′ = δthXt ◦ Φt, T Φt.c (s)i ds Z0 1 ′ = h( δtXt) ◦ Φt + TXt.δΦt, T Φt.c (s)i Z0 r ′ + hXt ◦ Φt, T (ut δtt + ν∇ logρ ¯ δtt + ε ξr δtWt ).T Φt.c (s)i ds X 1 r = ( δtXt) ◦ Φt +(∇ut δtt+ν∇ logρ ¯ δtt+ε P ξr δtWt Xt) ◦ Φt Z0 D ⊤ ⊤ ⊤ r ′ + ((∇ ut δtt + ν∇ ∇ logρ ¯ δtt + ε ∇ ξr δtWt )Xt) ◦ Φt, T Φt.c (s) ds X E

r = δtXt + ∇ut δtt+ν∇ logρ ¯ δtt+ε P ξr δtWt Xt Z(Φt)∗C ♭ ⊤ ⊤ r +(∇ ut δtt + ν∇ logρ ¯ δtt + ε ∇ ξr δtWt )Xt X =0

♭ since, by (4.47), the integrand equals −(∇p˜) δtt = −dp˜ δtt. HIPS FOR COMPRESSIBLE FLOW 15

5. Conclusions 5.A. HIPS approach to the compressible Navier-Stokes equation. The mean field system (3.32)-(3.33) is derived from the interacting particle point of view under the basic assumption that the equations of motion follow from stochastic Hamiltonian mechanics. Therefore, circulation is preserved along stochastic Lagrangian paths. If the perturbation fields ξr run over the orthogonal system ζr, defined in Section 2.E, such that the stochastic perturbation is given by a cylindrical Wiener process, then the mean fields E[µ] and E[ρ] solve the compressible Navier-Stokes equations. (Theorem 3.4 and Proposition 4.1.) While the HIPS formulation relies on a system of N interacting SDEs, the mean field equations (3.32)-(3.33) is a single SDE system for momentum and mass density. In contrast to statistical mechanics, this mean field formulation is obtained without any closure assumptions. (However, in this paper the existence of the mean field limit is not proved but assumed.) In the mean field limit (3.29), the HIPS evolution equation (3.25) becomes

−1 r (5.49) ( δtΦt) ◦ Φt = E[µt]/E[ρt]+ ν∇ log E[ρt] δtt + ε ξr δtWt X which is of similar form as the LA SALT advection field L r (5.50) E[ut ] δtt + ξr δtWt X L of [22, 2], where ut is a stochastic velocity field. However, there are a few crucial differences: while (5.50) is the starting point for LA SALT theory, the HIPS formulation is based on the ensemble Hamiltonian (3.22) and the Lie transport along (5.49) in the mean field equations (3.32)-(3.33) is a consequence of the Hamiltonian structure of the IPS (and the ensuing passage to the mean field limit). Moreover, unless the density is constant, it is not clear how to identify the drift in (5.49) with the expectation of a velocity. Thus, both, the starting points and the advection fields are different. However, the perturbation fields ξr can be interpreted in the same manner.

5.B. NatCat modeling of windstorm events. NatCat models used to calculate the solvency capital requirement (SCR as defined in [55]) for storm risks rely on NWP systems. These NWP systems are deterministic and, to arrive at a set of ‘stochastic’ NatCat scenarios, the initial conditions are statistically sampled. As discussed in the Introduction, all such numerical schemes suffer from subgrid phenomena, and for geophysical flow models a well-established means for treating these deficiencies is by stochastic fluid mechanics ([2, 9, 32, 44, 49]). Since SCR calculation is concerned with predicting extreme events in the 99.5 percentile, and not only average storm patters, it seems reasonable to expect that also NatCat models would benefit from a stochastic dynamics approach. However, it is not claimed that the stochastic HIPS formulation of this paper is appropriate to generate stochastic NatCat storm scenarios.

References [1] N.U. Ahmed, X. Ding, A semilinear Mckean-Vlasov stochastic evolution equation in Hilbert space, Stochastic Processes and their Applications Volume 60, Issue 1, November 1995, Pages 65-85. https://doi.org/10.1016/0304-4149(95)00050-X 16 HIPS FOR COMPRESSIBLE FLOW

[2] D. Alonso-Oran, A. Bethencourt de León, D. Holm, S. Takao, Modeling the Climate and Weather of a 2D Lagrangian-Averaged Euler–Boussinesq Equation with Transport Noise, https://doi.org/10.1007/s10955-019-02443-9 [3] D.M. Anderson, G.B. McFadden, A.A. Wheeler, Diffuse-Interface Methods in Fluid Mechanics, Annual Review of Fluid Mechanics 30 Nr. 1 (1997). https://doi.org/10.1146/annurev.fluid.30.1.139 [4] M. Arnaudon, X. Chen, A.B. Cruzeiro, Stochastic Euler-Poincaréreduction, J. Math. Phys. 55(8):081507 (2014). [5] A. Arnaudon, A.L. De Castro, D.D. Holm, Noise and dissipation on coadjoint orbits J. Nonlinear Sci. 28, No. 1, pp. 91-145 (2018). [6] V.I. Arnold, Sur la géométrie différentielle de groupes de Lie de dimension infinie et ses applications à l’hydrodynamique des fluides parfaits, Annales de l’Institut Fourier 16 No. 1 (1966), p. 319-361. [7] V. Arnold, B. Khesin, Topological Methods in Hydrodynamics, Springer 1998. [8] P. Bauer, A. Thorpe, G. Brunet, The quiet revolution of numerical weather prediction, Nature 525, 47–55 (2015). https://doi.org/10.1038/nature14956 [9] J. Berner, et al., Stochastic Parameterization: Toward a New View of Weather and Climate Models, Bull. Amer. Meteor. Soc. 98 (2017), pp. 565–588. https://doi.org/10.1175/BAMS-D-15-00268.1 [10] D. Bresch, B. Desjardins, C.-K. Lin, On some compressible fluid models: Korteweg, lubrication, and shallow water systems Comm. Partial Differential Equations, 28(3-4):843–868, (2003). [11] X. Chen, A.B. Cruzeiro, T.S. Ratiu, Constrained and stochastic variational principles for dissipative equations with advected quantities, arXiv:1506.05024, 2015. [12] K.M. Clark, The Use of Computer Modeling in Estimating and Managing Future Catastrophe Losses, The Geneva Papers on Risk and Insurance. Issues and Practice Vol. 27, No. 2, Special issue on risk management (2002), pp. 181-195. [13] P. Constantin, G. Iyer, A stochastic Lagrangian representation of the three-dimensional incompressible Navier-Stokes equations, Comm. Pure Appl. Math. 61 (2008), no. 3, 330-345. [14] F. Cipriano, A.B. Cruzeiro, Navier-Stokes equation and diffusions on the group of homeomorphisms of the torus, Comm. Math. Phys., 275 Issue 1 (2007), pp 255-269 [15] C.J. Cotter, G.A. Gottwald, D.D. Holm, Stochastic partial differential fluid equations as a diffusive limit of deterministic Lagrangian multi-time dynamics Proc. R. Soc. A 473 (2017) 20170388. [16] D. Crisan, F. Flandoli, D.D. Holm, Solution properties of a 3D stochastic Euler fluid equation, J Nonlinear Sci 29, 813–870 (2019). https://doi.org/10.1007/s00332-018-9506-6 [17] A.B. Cruzeiro, P. Malliavin, Nonergodicity of Euler fluid dynamics on tori versus positivity of the Arnold–Ricci tensor, J. Functional Analysis 254 (2008) pp. 1903–1925. [18] A.B. Cruzeiro, Hydrodynamics, probability and the geometry of the diffeomorphism group, Seminar on Stochastic Analysis, Random Fields and Applications VI pp. 83-93 (2011). [19] A.B. Cruzeiro, E. Shamarova, Navier-Stokes equations and forward–backward SDEs on the group of diffeomorphisms of a torus, 119 Issue 12 (2009), pp. 4034-4060. [20] D. Dawson, J. Vaillancourt, Stochastic McKean-Vlasov equations, NoDEA 2, 199–229 (1995). https://doi.org/10.1007/BF01295311 [21] T.D. Drivas and D.D. Holm, Circulation and energy theorem preserving stochastic fluids, Proc. Royal Soc. Edinburgh Section A: Mathematics, pages 1–39 (2019). [22] T.D. Drivas, D.D. Holm, J.-M. Leahy, Lagrangian averaged stochastic advection by Lie transport for fluids, J Stat Phys (2020). https://doi.org/10.1007/s10955-020-02493-4 [23] G. Da Prato, J. Zabczyk, Stochastic equations in infinite dimensions, CUP, 2nd Ed. 2014. [24] S. Desmettre, S. Hochgerner, S. Omerovic, W. Stockinger, S. Thonhauser, Mean field Libor market models and valuation of long term guarantees, in preparation. [25] D. Ebin, J.E. Marsden, Groups of diffeomorphisms and the motion of an incompressible fluid, Ann. of Math. 92(1) (1970), pp. 1037-1041. [26] F. Gach, S. Hochgerner, K. Laimer, Estimation of future discretionary benefits for traditional life insurance liabilities under Solvency II, in preparation. [27] P. Grossi, H. Kunreuther (Eds.), Catastrophe Modeling: A New Approach to Managing Risk, Huebner International Series on Risk, Insurance and Economic Security, Springer 2005. HIPS FOR COMPRESSIBLE FLOW 17

[28] S. Hochgerner, T. Ratiu, Geometry of non-holonomic diffusion, J. Euro. Math. Soc. Volume 17, Issue 2 (2015), pp. 273–319. [29] S. Hochgerner, Stochastic mean field approach to fluid dynamics, J. Nonlinear Science 28 No. 2 (2018) pp. 725-737. [30] S. Hochgerner, A Hamiltonian mean field system for the Navier–Stokes equation Proc. R. Soc. A 474 (2018). http://doi.org/10.1098/rspa.2018.0178 [31] S. Hochgerner, F. Gach, Analytical validation formulas for best estimate calculation in traditional life insurance, Eur. Actuar. J. 9, pp. 423–443 (2019). [32] D.D. Holm, Variational principles for stochastic fluid dynamics Proc. R. Soc. A 471 (2015) https://doi.org/10.1098/rspa.2014.0963 [33] A. Hunter, D.B. Stephenson, T. Economou, M. Holland, I. Cook, New perspectives on the collective risk of extratropical cyclones, Q. J. R. Meteorol. Soc. 142:243–256 (2015). https://doi.org/10.1002/qj.2649 [34] G. Iyer, A stochastic perturbation of inviscid flows, Commun. Math. Phys. 266, No. 3, 631-645 (2006). [35] P.E. Jabin, Z. Wang, Mean field limit for stochastic particle systems, In: Bellomo N., Degond P., Tadmor E. (eds) Active Particles, Volume 1. Modeling and Simulation in Science, Engineering and Technology. Birkhäuser. https://doi.org/10.1007/978-3-319-49996-3 10 [36] A. Jüngel, Effective velocity in compressible Navier–Stokes equations with third-order derivatives, Nonlin- ear Analysis 74 (2011) 2813–2818. [37] D.J. Korteweg, Sur la forme que prennent les équations du mouvements des fluides si l’on tient compte des forces capillaires causées par des variations de densitéconsidèrables mais continues et sur la théorie de la capillaritédans l’hypothèse d’une variation continue de la densité, Arch. Neerl. Sci. Exactes Nat. Ser. II 6:1–24 (1901). [38] R.H. Kraichnan, Small-scale structure of a scalar field convected by turbulence, Phys. Fluids, 11 (1968), pp. 945-963. [39] J.-A. Lazaro-Cami, J.-P. Ortega, Stochastic Hamiltonian dynamical systems, Rep. Math. Phys. 61, Issue 1 (2008), Pages 65-122. [40] J. Marsden, D. Ebin, A. Fischer, Diffeomorphism groups, hydrodynamics and relativity, Proc. of the 13th Biennial Seminar of Canadian Mathematical Congress, (J. Vanstone, ed.), (1972), 135-279 [41] J. Marsden, T. Ratiu, A. Weinstein, Semidirect Products and Reduction in Mechanics, Trans. Amer. Math. Soc. 281(1), 147-177 (1984). doi:10.2307/1999527 [42] S. Méléard, Asymptotic behaviour of some interacting particle systems; McKean-Vlasov and Boltzmann models In: Talay D., Tubaro L. (eds) Probabilistic Models for Nonlinear Partial Differential Equations. Lecture Notes in Mathematics, vol 1627. Springer (1996). https://doi.org/10.1007/BFb0093177 [43] A. Mellet, A. Vasseur, On the Barotropic Compressible Navier–Stokes Equations, Communications in Partial Differential Equations 32(3) (2007) DOI:10.1080/03605300600857079. [44] E. Mémin, Fluid flow dynamics under location uncertainty, Geophysical & Astrophysical Fluid Dynamics, 108(2): pp. 119-146, (2014). [45] P. Michor, Some Geometric Evolution Equations Arising as Geodesic Equations on Groups of Diffeomor- phism, Including the Hamiltonian Approach. IN: Phase space analysis of Partial Differential Equations. Series: Progress in Non Linear Differential Equations and Their Applications, Vol. 69. Bove, Antonio; Colombini, Ferruccio; Santo, Daniele Del (Eds.). Birkhauser Verlag 2006. Pages 133-215. [46] R. Mikulevicius and B.L. Rozovskii, Stochastic Navier–Stokes Equations for Turbulent Flows, SIAM J. Math. Anal., 35(5), 1250–1310 (2006). https://doi.org/10.1137/S0036141002409167 [47] T.N. Palmer, Stochastic weather and climate models, Nat Rev Phys 1, 463–471 (2019). https://doi.org/10.1038/s42254-019-0062-2 [48] P.E. Protter, Stochastic integration and differential equations, Stochastic Modelling and Applied Proba- bility, 2nd Ed., Springer 2005. [49] V. Resseguier, E. Mémin, B. Chapron, Geophysical flows under location uncertainty, Part I: Random transport and general models, Geophysical & Astrophysical Fluid Dynamics 111(3): pp. 149-176 (2017). [50] V. Resseguier, E. Mémin, B. Chapron, Geophysical flows under location uncertainty, Part II: Quasi- geostrophic models and efficient ensemble spreading, Geophysical & Astrophysical Fluid Dynamics 111(3): pp. 177-208 (2017). 18 HIPS FOR COMPRESSIBLE FLOW

[51] V. Resseguier, E. Mémin, B. Chapron, Geophysical flows under location uncertainty, Part III: SQG and frontal dynamics under strong turbulence, Geophysical & Astrophysical Fluid Dynamics 111(3): pp. 209- 227 (2017). [52] A. Sznitman, Topics in propagation of chaos, In: Hennequin (ed.) Ecole d’Etéde Probabilités de Saint- Flour XIX — 1989. LN in Math., vol 1464. Springer (1991). https://doi.org/10.1007/BFb0085169 [53] J. Teichmann, M.V. Wüthrich, Consistent yield curve prediction, Astin Bulletin 462 (2016), pp. 191-224. https://doi.org/10.1017/asb.2015.30. [54] K. Yasue, A variational principle for the navier-stokes equation, J. Functional Analysis 51(2):133–141, (1983). [55] Directive 138/2009/EC (Solvency II) https://www.eiopa.europa.eu/rulebook-categories/directive-1382009ec-solvency-ii-directive en, downloaded on 21. June 2020.