ESAIM: M2AN 44 (2010) 921945 DOI: 10.

1051/m2an/2010047

ESAIM: Mathematical Modelling and Numerical Analysis www.esaim-m2an.org

NONLINEAR FILTERING FOR OBSERVATIONS ON A RANDOM VECTOR FIELD ALONG A RANDOM PATH. APPLICATION TO ATMOSPHERIC TURBULENT VELOCITIES

Christophe Baehr 1, 2
Abstract. To lter perturbed local measurements on a random medium, a dynamic model jointly with an observation transfer equation are needed. Some media given by PDE could have a local probabilistic representation by a Lagrangian stochastic process with mean-eld interactions. In this case, we dene the acquisition process of locally homogeneous medium along a random path by a Lagrangian Markov process conditioned to be in a domain following the path and conditioned to the observations. The nonlinear ltering for the mobile signal is therefore those of an acquisition process contaminated by random errors. This will provide a Feynman-Kac distribution ow for the conditional laws and an N particle approximation with a O( 1 ) asymptotic convergence. An application to N nonlinear ltering for 3D atmospheric turbulent uids will be described. Mathematics Subject Classication. 82B31, 65C35, 65C05, 62M20, 60G57, 60J85.
Received May 20, 2009. Revised December 23, 2009. Published online August 26, 2010.

Introduction
In Applied Sciences the knowledge of a (physical, biological, etc.) state system needs measurements. Often it is not possible to recover the measured values everywhere in the state space but only on discrete points or along a path or barely on a grid. Our interest is focused on the case of measurements along a random path in a random medium. The other situations can be considered as a reduction or combinations of this case. This paper is devoted to the denition of a new process called Acquisition Process and using this process to nonlinear ltering of corrupted measurements of a stochastic vector elds made along a random path. The physical application concerns measurements of atmospheric turbulent uids with a mobile sensor. Stochastic nonlinear ltering requires a model for the signal to be denoised. If the studied system can be represented as a set of equations, global or local, for mobile measurements into this system there are in general no model. We propose here a method using a local model of the random medium to derive a dynamic model for mobile measurements signal.
Keywords and phrases. Nonlinear ltering, Feynman-Kac, stochastic model, turbulence.
1 Mto-France-CNRS, CNRM-GAME URA1357, 42 avenue Coriolis, 31057 Toulouse Cedex 1, France. e e [email protected] 2 Associated member of the Laboratory of Statistics and Probability of the Toulouse Mathematics Institute (UMR 5219), 118 route de Narbonne, 31062 Toulouse Cedex 9, France.

Article published by EDP Sciences

c EDP Sciences, SMAI 2010

922

C. BAEHR

The representation of the solution of partial dierential equations by the expectation of a stochastic process is well known for linear parabolic equation (see [12]). For nonlinear partial dierential equations it is not so clear. Without a general theory sometimes it needs particle systems to have an approximated solution [13]. In the particular case of the Navier-Stokes equation, there is nonlinear McKean-Vlasov equations to represent the small scale dynamics as random walks. Some are exact but need implicit calculations ([7] or [6]), some have uniqueness and existence proved only for the dimension 2 ([11] or [14]), some are phenomenological but able to approach 3D inhomogeneous ows [15] or stratied ows [8]. The measurements along a path are local processes, living on a line. This kind of measurements do not allow the reconstruction of the entire observed eld. In Section 1, using a local stochastic representation of the medium and the known sensor trajectory, we will dene a process and will give (Sect. 2) a procedure to estimate locally the parameters of the medium. Then to denoise the measurements corrupted by instruments, we suggest a ltering algorithm to compute the conditional distribution of states knowing partial observations. By the physical context under consideration, this lter have to deal with mean-eld processes, indeed the local estimations are connected with large scale structures, and these connections will be made through the average of large scale elds. Particle lter are well adapted to compute at the same time the ltering conditional laws and the mean-eld terms. We propose in Section 3 some algorithmic solutions to solve this probabilist estimation problems. Application to measurements of atmospheric wind will be presented in Section 4. This application will use all the material presented in the previous sections and a particular representation of the Navier-Stokes equations adapted to the atmospheric physics.

1. Acquisition process
In the problem of the examination of a multidimensional eld along a path, there is a clear separation between the trajectories and the other components of the state space. Then we consider the phase space as the Cartesian product E E where E is the conguration space where lie the trajectories and E the vector space. Now we need some denitions. They could have a very wide range of uses and we will detail few of them. Denition 1.1 (Acquisition System). Let E Rd , d N , be a metric locally compact space of point called conguration space. E is endowed with the -algebra E. Let E Rd , d N , be a vector space called phase space endowed with the -algebra E . Let (, F , (Ft )t0 , P) be a complete ltered probability space. Let T < be a real number and x E a point of the conguration space. Let Xt be the (E, E)-valued random variable family on (, F , Ft ) indexed by the time t [0, T ] and Xt,x be the (E , E )-valued random variable family on (, F , Ft ) indexed by the time t [0, T ] and the point x E. Then, the pair of applications Ft -measurable, (Xt , Xt,x ) is called the Acquisition System of the random vector eld. The process Xt is called the Acquisition Path and the family Xt,x is called the Acquisition Field. This is the choice of the path process Xt , the Acquisition Field Xt,x and their possible coupling which makes sense to the Acquisition System and denes the Acquisition Process which is the statement of the eld Xt,x along the path Xt . Denition 1.2 (Acquisition Process). Let (Xt , Xt,x ) be a (E, E) (E , E ) -valued Acquisition System on the probability space , F , Ft , P . For all t [0, T ], the Acquisition Process is dened by the (E , E )-valued process on (, F , Ft ) At with At = Xt,Xt Example 1.1. Let x be a xed point in E. For all t > 0 we consider the stationary acquisition path Xt = x. Then the acquisition system is (x, Xt,x ) and the acquisition process is At = Xt,x which can be called, by analogy with physical applications, the Eulerian Acquisition of the eld.
def

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

923

Example 1.2. Let Xt,x be a (E , E )-valued random bounded vector eld C . Let x0 be a point in E and for all 0 t T , let Wt be a Ft -Brownian Motion on (, F , Ft , P). We dene the stochastic ow (see [4]):
x Xt 0 = x0 + t 0
x Xs,Xs 0 dWs

where signs the Stratonovich integral. Then the Acquisition Process is At = Xt,X x0 . By analogy with Physics, we call this process a Lagrangian t Acquisition. Example 1.3. In the last example which illustrates the previous one, we consider an Eulerian velocity vector eld Ut,x for a uid medium. Let D be a domain of the conguration space E, we dene for all t [0, T ], the Eulerian velocity eld Ut,x as a solution of the Navier-Stokes Equation:
t + Ut,x x Ut,x = f (t, x)

Ut,x = x pt,x + Ut,x

xD x on D

(1.1)

U0,x = U0 (x)

where f is a well-dened function for each time t. With the point of departure x0 D, the location of a uid element is given by a solution of the Stochastic Partial Dierential Equation:
x dXt 0 x X0 0

= =

x x Ut,Xt 0 dt + t,Xt 0 dWt x0

(1.2)

where (Wt )t0 is a cylindrical Brownian Motion and t,x a known nonzero function. With the regularity hypothesis given by [14], Mikulevicius and Rozovskii claim the existence of mild solutions for the two systems (in small time). The pair (Xt , Ut,x ) is an ((E, E)(E , E ))-valued Acquisition System. We dene the Acquisition x Process by At = Ut,Xt 0 . Then At is the Lagrangian velocity usually denoted Vt of a uid element started at x0 x and carried out by the ow Xt 0 . With these rst three examples, we can notice that the stochastic process At located Xt do not require the knowledge of the vector eld everywhere. More, the nature of At is utterly linked to the process Xt . Thus, the probabilistic properties of the eld (Markovian, ergodicity, etc.) will not be given only relatively to the random variable At , but more precisely relatively to the pair (Xt , At ). For all t [0, T ], we suppose the absolute continuity of the law of the pair (Xt , At ) w.r.t. Lebesgue measure with the density pXt ,At . We assume that for any x E, E pXt ,At (x, a)da > 0 and the existence of the conditional density pAt |Xt . By the denition of the conditional expectation knowing Xt for all bounded measurable functions f from (E, E ) to R: E(f (Xt , At )|Xt ) =
E

f (Xt , a)pAt |Xt (a|Xt ) da

the conditional law is given by the joint pdf: pAt |Xt (a|x) = pXt ,At (x, a) pXt ,At (x, a)da E

In physical experiments or in engineering applications, it could be necessary to evaluate only the expectation for At instead of for the pair (Xt , At ). It is the case of measurements on a random medium where no information of the geometry of the medium is available (uid ow, small conguration of a nuclear reactor, etc.).

924

C. BAEHR

Then the classic method requires to use a weak regularizing C kernel K of the Dirac measure, being a cut-o parameter. Then for all t [0, T ], for any measurable bounded function f , the conditional expectation of At given Xt is approximated by: E (f (At )|Xt ) =
EE

f (a) K (Xt , z) pAt |Xt (a|z)dadz.

More exactly, using the joint law pXt ,At , we dene the regularized integral: F (f (At ))(Xt ) =
EE

f (a)K (Xt , z)pXt ,At (z, a)dzda.

We assume here that EE K (Xt , z)pXt ,At (z, a)dzda is not null. Then the regularized expectation can be expressed in terms of F : F (f (At ))(Xt ) E (f (At )|Xt ) = F (1)(Xt ) with some necessary regularity properties on the density given by a supplementary hypothesis, it holds
0

lim E (f (At )|Xt ) = E(f (At )|Xt ).

Now, here come two applications directly connected to our previous examples. Application 1.1. This is the trivial case where for a xed point x E, for all time t > 0, we have the stationary Acquisition Path Xt = x. We dene the Acquisition Field as the Eulerian velocities vector eld Ut,x given by the NSE (1.1). Then Ut = Ut,x is an Acquisition Process and for all bounded measurable function f , we have p.s. E(f (Ut,Xt )|Xt = x) = E(f (Ut )). This is the usual Eulerian velocities mean usually written in uid mechanics literature (with a chaotic point of view) by the Ensemble Average f (U ) t,x . In the sequel we will use this remark and replace for a uid, the ensemble average f (U ) t,x by the conditional expectation E(f (Ut,Xt )|Xt = x). Application 1.2. Another direct application, to stay in the uid mechanics domain, let Ut,x be the Acquisition x Field given by the Eulerian velocities vector eld. Let Xt 0 be the Acquisition Path dened by the trajectory associated to the eld with the point of departure x0 E. This trajectory is given by the solution of the equation (1.2).
x The Lagrangian Acquisition is dened by Vt = Ut,Xt 0 and using the regularised expectation we have

E(f (Vt )|Xt ) = lim

F (f (Vt ))(Xt ) 0 F (1)(Xt )

with F (f (Vt ))(Xt ) = EE f (v) K (Xt , x) pXt ,Vt (x, v)dxdv. For example, in the case of homogeneous turbulent ows using the Kraichnan model of turbulence, the regularizing kernel can be given by the Gaussian function: K (Xt , x) = G (Xt x) = e
Xt x 2 2

(1.3)

Pope in [15] proposes other kernels built using spline functions and phenomenological considerations. Remark 1.1. A ltering problem can be considered as the symmetrical problem of the Acquisition estimation. Indeed an observation Yt E of a state Xt E are connected together by the observational equation Yt = H(t, Xt , Wt ) where Wt is a canonical random process, and H is a bounded transfer function. Clearly the transfer function H(t, Xt , ) = Ht,Xt () plays the role of a random eld and Yt is the Acquisition Process
def

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

925

of the acquisition eld H along the path Xt in E. The acquisition problem consists in the computation of E(f (Yt )|Xt ) while the ltering problem is the calculation of E(f (Xt )|Yt ). This remark is fruitful to suggest an algorithm estimating the Acquisition expectations.

2. Estimation for discrete acquisition in locally homogeneous medium

As said previously the estimation of the medium parameter along a path is a pointwise process and we have to derive the dynamics of the mobile measurements from a local model. In this section, we propose a sequential algorithm for the computation of the Acquisition Process expectation given the Acquisition Path. We consider an Acquisition System (X, X ) with a discrete time dynamics. Let be n N, t > 0 a time step, and for all i {0, . . . , n} we denote ti = it and (Xi , Xi,x ) = (Xti , Xti ,x ). Denition 2.1 (Locally homogeneous Acquisition Process). For all n N, let (Xn , Xn,x ) be an (E E )valued Acquisition System and An = Xn,Xn an Acquisition Process. The Acquisition System is said to be locally homogeneous if: E is a metrisable locally convex space with a convex set covering A = iI Ai , where I is an index set. n N and x E, there exists n > 0 and i I such that: B(x, n ) Ai with B(x, n ) = z E s.t. |x z| n and y B(x, n ) = Bn (x), and for all a E we have P(An da | Xn = x) = P(An da | Xn = y). This is the law of An given Xn belonging to the ball Bn (x), denoted Law(An | Xn Bn (x)).
In the locally homogeneous case, it will be possible to evaluate the expectation E(f (Xn , An ) | Xn Bn ). First of all, we highlight that: def

E(f (Xn , An ) | Xn Bn )

=
EE

f (x, a) pAn |Xn Bn (a|x)dxda =

E(f (Xn , An ) Bn (Xn )) E(Bn (Xn ))

f (x, a)
EE

Bn (x) pXn ,An (x, a) dxda (x)pXn (x)dx E Bn

(2.1)

As described in Figure 1, this denition is directly extendable to the trajectorial processes. For n > 0, we denote X[0,n] = {Xi , i {0, . . . , n}} the discrete path of the process Xn . Let B[0,n] be the hull of the balls set (Bi )in : B[0,n] =
i{0,...,n} We denote X[0,n] B[0,n] the event {X0 B0 (x0 ), . . . , Xn Bn (xn )}. Bi (xi ).

In this trajectorial situation we attempt to the computation of the probability measures n such that for any bounded measurable function f we get the acquisition expectation along the path X[0,n] : n (f ) = E(f (Xn , An )|X[0,n] B[0,n] ). The structure of the conditional expectation n suggests with its product of (indicatrice) functions a possible Feynman-Kac shape (see [9] for a complete description of Feynman-Kac measures).

2.1. Mean-eld interpretation for the discrete acquisition process

Up to now, we have introduced the Acquisition Systems in their general form, with examples coming from uid mechanics. To go further, we have to specialize the description mainly taking into account a local dynamical structure. General Acquisition Processes indeed have structures close the ones occur in a ltering problem.

926

C. BAEHR

Figure 1. Set of homogeneity balls for a Acquisition Path.

There is an a-priori evolution of the medium (i.e. without interaction with the Acquisition Path) and then a conditioning to the Acquisition Field statement along the Acquisition Path. This procedure can be decomposed in two steps, one for the mediums prediction stage and the second for a selection along the Acquisition Path. Since the medium is only locally known by the Acquisition, the model can be local and we choose to use a probabilistic representation with a Lagrangian dynamics. We have also seen that a Lagrangian dynamics is a specic Acquisition Process and we realize the coupling between this Lagrangian Acquisition and the Acquisition Process associated with the random path of measurements. For a path initiated at x0 E, we denote Xi = Xi,X x0 the discrete Lagrangian Acquisition. Let Z0 , . . . , Zn i be a discrete path in the conguration space E independent of the Lagrangian Acquisition dynamics. In the situation of a locally homogeneous medium, the knowledge of the Acquisition Field can be restricted to a ball Bn excluding the values of the eld outside this ball. Then to achieve the estimation we have to couple two Acquisition Systems: The rst one is given by the Lagrangian model (1 point model) and put out the Lagrangian Acquisition x with the point of departure x0 E: (Xn 0 , Xn,X x0 ) = (Xn , Xn ). n The second Acquisition system is the statement of the Lagrangian Acquisition along the independent Acquisition Path Zn : (Zn , An ). To express what An could be, we have to consider a set of Lagrangian Acquisition starting from dierent initial points and arriving to the dierent points of the Acquisition Path. Thus, we have the explanation of Figure 2: At the step n 1, the Acquisition Path is at Zn1 , and acquires the value Xn1 in the phase space
0 0 given by the Lagrangian Acquisition Xn1 = Zn1 started from xn1 and then An1 = Xn1,Zn1 . 0 n1 is to emphasize the departure point of the Lagrangian Acquisition. The superscript x0 At time n we change of Lagrangian Acquisition to acquire those started from the initial point xn and 0 xn arrived at Zn . Then we got An0 = Xn,Zn .

xn1

xn1

But we do not have any informations of the point of departure for each Lagrangian trajectory. It would be suitable to compute the expectation An = Ax PX0 (dx). This expectation is hard to approach, it is an n integration along the ancestral lines. Therefore, using the locally homogeneous medium hypothesis, we propose to give in each point of the Acquisition Path Zn a parameter n > 0 and a ball Bn (Zn ) = {x E : |x Zn | n }. Then we estimate

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

927

Figure 2. Coupling between a Lagrangian Acquisition and an Acquisition along an independent random path. the two probability measures n and n such that for any test function f : n (f ) = n (f ) =
E(f (Xn , An ) | X0 B0 (Z0 ), . . . , Xn Bn (Zn )) E(f (Xn , An ) | X0 B0 (Z0 ), . . . , Xn1 Bn1 (Zn1 )).

Complementary with the Lagrangian hypothesis, the pair (Xn , Xn ) is supposed to be Markovian and the evolution is described by the mean-eld transition kernel: Mn+1,n ((x, x ), d(z, z )) = P((Xn+1 , Xn+1 ) d(z, z )|(Xn , Xn ) = (x, x )) where n is the probability law of (Xn , Xn ). Later we will call the evolution using the transition kernel Mn+1,n a mutation. The mean-eld hypothesis is made to take into account the macroscopic structures of the random eld locally represented by (Xn , Xn ). Being inspired by (2.1), we can write: n (f ) = E(f (Xn , Xn ) E(
n p=0
Bp (Zp ) (Xp ))

n p=0

Bp (Zp ) (Xp ))

n (f ) =

E(f (Xn , Xn ) E(
n1 p=0

Bp (Zp ) (Xp ))

n1 p=0

Bp (Zp ) (Xp ))

which confer on the probability measures n and n their Feynman-Kac distribution structure. Without any change we do not only consider the conditioning given the ball Bp (Zp ) but given the cylinder Bp (Zp )E and we dene the potential Gp for (Xp , Xp ):
Gp (Xp , Xp ) = Bp (Zp )E (Xp , Xp ) .

Associated to this potential function, we dene the selection kernel:

Z Sn+1,n+1 (x, x ), d(y, y ) = Bn+1 (Zn+1 )E (x, x )(x,x ) (d(y, y ))
+ 1 Bn+1 (Zn+1 )E (x, x )

Bn+1(Zn+1 )E (y, y )n+1 (d(y, y ))

n+1 (Bn+1 (Zn+1 ) E )

where n+1 (d(y, y )) = P((Xn+1 , Xn+1 ) d(y, y ) | X0 B0 (Z0 ), . . . , Xn Bn (Zn )).

928

C. BAEHR

Figure 3. Discrete evolution of the Acquisition Process for a Lagrangian dynamics. Starting from the Acquisition initial distribution 0 (d(x, x )) = P((X0 , X0 ) d(x, x )) we have the sequential scheme, for n 0:
n n n+1 n+1 n+1

Mn+1,

Z Sn+1,

Z which is a stochastic dynamical system: n+1 = n Sn,n Mn+1,n .

In terms of the state vector evolution, Figure 3 illustrates the procedure which corresponds to:
Selection Mutation (Xn , Xn ) (Xn+1 , Xn+1 ) (Xn+1 , Xn+1 ).

After the selection step, the state (Xn+1 , Xn+1 ) is the estimation of the Acquisition System, where Xn+1 is the location of the system and Xn+1 is the estimation of the required Acquisition Process An+1 = Xn+1,Zn+1 . The coupling between the two Acquisition Systems is a localization after a Lagrangian mutation. Written like this, the evolution is not Markovian and the potential can nullify. The Figure 3 is quite abusive, we must have a covering of the trajectory Z0 , . . . , Zn by balls (Bj (Zj ))0jn with non-empty intersections as seen in Figure 2. For each time step n and each state point (x, x ), we suppose that 0 < Mn+1,n (Bn+1 (Zn+1 )E )(x, x ). This hypothesis means that the probability to arrive in a ball Bn+1 (Zn+1 ) coming from x is not null. This hypothesis is important, without we lost the Markovian nature of the process. It has a non-trivial implications. For example in physical applications, the time-discretization should be tight enough. Now we dene the modied Markov kernel and the new potential function to restore Markovian property: Mn+1,n (x, x ), d(y, y ) = and Gn+1 (x, x ) = Mn+1,n (Bn+1 (Zn+1 )E )(x, x ). Mn+1,n (x, x ), d(y, y )
Bn+1 (Zn+1 )E (y, y )

Mn+1,n (Bn+1 (Zn+1 )E )(x, x )

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

929

The Markovian kernel Mn+1,n ((x, x ), d(y, y )) is a local transition restricted to the domain Bn+1 (Zn+1 ) n+1 (x, x ) gives the chance to evolve in B (Zn+1 ) E . This implies that the while the potential function G n+1 Lagrangian Acquisition follows the path Z0 , . . . , Zn+1 and the potential can not be degenerate. We start at initial step from the cylinder B0 (Z0 ) E using the probability B and the law 0 of the pair 0 (X0 , X0 ). We use for all the time step the Feynman-Kac potential and kernel (Gn , Mn+1,n ) to produce the B B Feynman-Kac distribution at each time step n. We denote n and n the stochastic ows of the process restricted to the cylinders Bn+1 (Zn+1 ) E . This corresponds to an interacting mean-eld process, solution of B the nonlinear equation B = B Kn+1,B ,n . n+1 n n B We know that the dynamical system B = B Kn+1,B ,n have a non-unique McKean representation and n+1 n n we choose to write: B Kn+1,B ,n = Sn,B Mn+1,n . n n

The non-unicity is brought by the selection kernel Sn,B . Using the potential function Gn+1 we choose a n particular form for Sn,B , called genetic selection: n Gn+1 (y, y )B (d(y, y )) n Sn+1,B ((x, x ), d(y, y )) = Gn+1 (x, x )(x,x ) (d(y, y )) + [1 Gn+1 (x, x )] n B (Gn+1 ) n The choose of the genetic selection is the justication of the name mutation given to the prediction step. This dynamical system produces a sequential scheme for the evolution laws restricted to the Acquisition Path:
n B B B n n+1 n+1 n+1

Mn+1,

Sn+1,B

and in terms of the state space points we have the algorithm:

n+1 Selection B B (Xn , XnB ) (Xn , XnB ) (Bn (Zn )E ) (Bn (Zn )E )

Mutation restricted to B

B B (Xn+1 , Xn+1 ) .
(Bn+1 (Zn+1 )E )

The Acquisition ow restricted to the balls centered to the Acquisition Path is therefore a solution of the dy namical system B = B Sn,B Mn+1,n with initial B and describes the coupling between the two Acquisition n+1 n 0 n Systems along the random path Z0 , . . . , Zn and it is a mean-eld process. Now we have to develop a particle approximation of this Acquisition Process.

2.2. Particle approximation of acquisition processes on locally homogeneous medium

The Acquisition Process is given by the solution of stochastic integrals. These integrals have only approximated solutions. We describe briey the interacting particle algorithm. The algorithm will be more detailed in the next section where the Acquisition Process will be used to lter pointwise measurements on a random medium. We have by hypotheses a E-valued Acquisition Path (Z0 , . . . , Zn ) and because of the locally homogeneous medium a set of real numbers 0 < (i )0in with the existence of homogeneity balls Bi (Zi ) for i {0, . . . , n}. To establish the asymptotic approximations of the particle system, the Markovian transition have to be suciently regular, at least for a class of function. So we assume that the Markovian transition Mn+1,n follows a regularity hypothesis: For any measurable bounded function f and any measure , n N, there exists some constants cn () and a nite set of bounded functions h with h 1 such that for any measure Mn+1, (f ) Mn+1, (f ) cn () f
h

|(h) (h)|.

930

C. BAEHR

i For the initial step, we consider a ball B0 (Z0 ) E and a particle system with cardinal N > 0, (0 , 0i )1iN , i.i.d. following B and we denote: 0

B,N = 0

1 N

i , i .
i=1
0 0

The empirical law B,N is an approximation of the exact law B and the technical Lemma 3.1 gives the 0 0 existence for all p > 0 of a constant C0 (p) such that: E( B,N B 0 0
1 p p H)

C0 (p) I(H). N

i At the beginning of the nth time step, we have a particle system (n , ni )1iN located in the ball Bn (Zn )E and following the sequential scheme Selection i i i i Mutation i (n , ni ) (n+1 , n+1 ) (n+1 , n+1 ). B B The exact ow B = B Kn+1,B ,n is approached by B,N obtained by the empirical transition Kn+1,B ,n n+1 n n+1 1 of the discrete law B,N = N i=1 (n ,ni ) . i n As we will see in the next section, to estimate n the a-priori law of the random medium it is necessary to have an auxiliary system with d particles evolving separately with the transition kernel Mn+1,n . Then the following theorem holds: N
n n

Theorem 2.1. For all n 0 and for all p 1, there exists some constants Cn and Cn such that E( B,N B n n
1 p p H)

Cn C + n I(H) N d

The proof of this theorem is similar to the proof of Theorem 3.2. The denitions of the subset H, the integral entropy I(H) and the norm . H are given in Section 3. One of the diculties of this particle estimation, where the Acquisition Path is recovered with a discrete set of balls, is the possibility to nullify the potential function, Gn (x, x ) = Bn (Zn )E (x, x ) or the possibility to fail the positivity hypothesis for Mn+1,n (Bn+1 (Zn+1 )E )(x, x ). In these cases, the algorithm is stopped and the estimated state is put to a cemetery point . The estimation of this stopping time is always an open problem. Nevertheless denoting N the rst time of extinction, for some supplementary hypothesis, it is possible to estimate: Theorem 2.2 ([9], Thm. 7.4.1). Suppose we have, for any n 0, E( N 1 it yields P( N n) a(n)eN/b(n) .
n p=0
Bp (Zp ) (Xp )) > 0. Then for all

For a complete proof, see [9]. This inequality is not optimal, but could indicate that with a large enough number of particles, the extinction becomes weakly probable. This could be seen as a constraint in numerical experiments and suggests that the recovering set of ball have to be wide enough without failure of the medium homogeneity hypothesis.

3. Non-linear particle filter for acquisition processes

We present in this section the specic lter necessary to denoise corrupted Acquisition Process on homogeneous medium. To perform this ltering, we have to develop an original particle algorithm and prove its convergence. This ltering is not only suitable to Acquisition Process but is available for all mean-eld processes.

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

931

So we use generic notations which can be applied to all situations where the dynamical processes move according to a probability law. Considering Xn a discrete process taking values in a Polish space (E, F , P), we denote n the law of the process Xn . For the initial step, we suppose that X0 is a random variable distributed with the probability 0 . For any n 0, we assume that the pair (Xn , n ) is Markovian with the dynamical equation: Xn+1
X X = Xn + b(Xn , n )t + n Wn

X where Wn is a sequence of i.i.d. centered Gaussian random variables with variance t, b is a locally bounded X function and Lipschitz in law, t is the time step and n is a positive bounded constant. This equation is a discrete version of a McKean SDE widely studied by Sznitman [16] and Mlard [13], and we ee know, with the previous hypotheses, the pathwise existence and uniqueness of the solution. For the interacting particle systems, Sznitman [16] proved the propagation of chaos property and gave some estimations for the particle approximation errors. Bossy and Talay [5] complete these works with the convergence estimations of the discretization Euler scheme. We choose a particular shape for the function b and suppose that for any x E and for any probability measure on E:

b(x, ) = D(x) +

B(x, y)(dy)

where D is a bounded function and B is a bounded function for the rst variable, which means that there exists C B such that for all x E, B(., x ) C B . Moreover we suppose that the function b : (x , ) E M1 (E) b(x , ) is Lipschitz with the parameter LB for its second variable and for all x E and (, ) M1 (E) M1 (E), b(x, ) b(x, ) LB H where H is the seminorm for H the sub-class of functions dened as follows: We consider H a subset of bounded measurable functions h such that h 1 completed with the unity function (see [9] for details). For two measures and in M1 (E), we denote the seminorm in H by:
H

= sup{ |(h) (h)| s.t. h H }.

To measure a class of H, we count the minimal number of covering Lp -balls with radius > 0. This number is 1 denoted N (, H), and we dened the associated integral entropy I(H) = 0 log(1 + N (, H))d. Now we given a very useful technical lemma for convergence estimation of empirical processes. Lemma 3.1 (technical lemma). For any p 1, any measure in M1 (E) and its N -empirical approximation N , N N , we have 1 E( N H ) p c [p/2]! I(H). (3.1) This technical lemma and its proof are given by Del Moral in [9]. Our interest is put on the nonlinear ltering to estimate Xn , a hidden Markov mean-eld process observed by the random process Yn . For any time step n 0, we have the system: X X Xn+1 = Xn + b(Xn , n )t + n Wn Y Y Yn = h(Xn ) + n Wn (3.2) X0 0 Y0 = y0
Y where h : E F is a bounded function and we assume that F is a Polish space. Wn is a sequence of i.i.d. X Y centered reduce Gaussian random variables independent of Wn , and n a positive bounded constant.

932

C. BAEHR

We denote the two Feynman-Kac ows (n , n ). For all bounded measurable test function f , the predictor is dened by n (f ) = E[f (X0 . . . Xn ) | Y0 . . . Yn1 ], and the lter ow by n (f ) = E[f (X0 . . . Xn ) | Y0 . . . Yn ]. The stochastic ltering problem consists in the calculation of the Law(X0 . . . Xn |Y0 . . . Yn ). When the mean-eld law is unknown (unfortunately, it is the general case), we have to approach this law with a (auxiliary) particle system. Thus, for any time n 0, we need two particle systems, the rst one (X i,N )1iN , i,d N > 0 is dedicated to the ltering of the perturbed observations Yn and the second system (Zn )iid , d > 0 is devoted to learn the a-priori law of the medium. i,d For this second system, the point of departure is given by the particles (Z0 )1id distributed according to i,d 0 = Law(X0 ) and we suppose that for all n > 0, each particle Zn moves with the transition kernel Mn+1,n i,d i,d i,d X Z,i where n is the law corresponding to the dynamical equation Zn+1 = Zn + b(Zn , n )t + n Wn , where Z,i X Wn is a sequence of i.i.d. centered Gaussian random variables with variance t independent of Wn Y and Wn . d 1 d In fact n is unknown and each particle uses the empirical law n1 = d i=1 Z i,d . So for i {1, . . . , d},
i,d i,d Zn evolves according to Mn,n1 (Zn1 , ). d The error to the exact law can be evaluated by:
n1

d d d d n n = n1 Mn,n1 n1 Mn,n1 [n n1 Mn,n1 ]. d d

In an usual manner,

d to a Gaussian random variable. We retain only the rst part of the expression and consider that (n n ) is d d approached with (n1 Mn,n1 n1 Mn,n1 ) for which the technical Lemma 3.1 gives an usable error control. Back to the ltering problem, we assume that for any time n 0, the dynamics/observation pair (Xn , Yn ) is Markovian and follows the system (3.2). For any measurable bounded test function f , the ltering process moves forward according to the stochastic dynamical system:

d d d[n n1 Mn,n1 ] converges weakly to 0, function by function, or almost surely, d

n+1 (f ) = n Kn+1,n ,n (f ) = n Sn,n Mn+1,n (f ). Now we use a set of N particles to estimate the ltering laws n and n , and a set of d particles to sample the mean-eld law n . i,d The d particles (Zn )1id are in interaction through their mean-eld, and follow the empirical dynamics:
i,d Zn+1 i,d i,d d X Z,i = Zn + b(Zn , n )t + n Wn .

i,d These particles Zn ignore the observations and evolve without selection step. N N In the particle approximation of the nonlinear lter n+1 (f ) n Sn,n Mn,n (f ), the empirical lterN d ing process, the selection kernel and the mutation transition are computed using the interaction particle set i,N (Xn )1iN and the dynamical model:

i,N Xn+1

i,N i,N d X X,i = Xn + b(Xn , n )t + n Wn

X,i where Wn is a sequence of i.i.d. centered Gaussian random variables with variance t.

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

933

N We have to control the error estimation of the law n+1 (f ) by the empirical law n+1 (f ) and we prove the theorem:

Theorem 3.2. With the hypotheses mentioned above, for any n 0 and p 1, there exists nite constants Cn (p) > 0 and Cn (p) > 0 such that
N E( n n p 1 p H)

Cn (p) Cn (p) + N d

I(H).

Proof. The complete proof is in the paper of Baehr and Le Gland [2] (or in [1] for this particular case) where this case and others can be found. But we can give a sketch of the proof which use triangular inequalities and argue by induction. At the initial step, the technical Lemma 3.1 gives the inequality
N E( 0 0
1 p p H)

cp (0) I(H) N

where C0 (p) = c0 (p) and C0 (p) = 0. We show that if the statement occurs at step n, then it also holds at step n + 1. We use the following decomposition:
N N N N d n+1 (f ) n+1 (f ) = n+1 (f ) n Kn+1,n ,n (f )

() () ( )

+ +

N N n Kn+1,n ,n (f ) n Kn+1,n ,n (f ) N d N n Kn+1,n ,n (f ) n Kn+1,n ,n (f ).

C N The quantity () is a uctuation term where we obtain the control E(()2 | n ) N f 2 for some constant C independent of N and d. The dierence ( ) uses the induction hypothesis, and with the notation Qn+1,n (z, dy) = Gn (z).Mn+1,n (z, dy), the equalities N n Kn+1,n ,n (f ) N

n Kn+1,n ,n (f )
N N N = n Kn+1,n ,n (f ) n Kn+1,n ,n (f ) + n Kn+1,n ,n (f ) n Kn+1,n ,n (f ) N N = [1 n (Gn )] N n n N n (Gn ) n (Gn ) N Qn+1,n (f ) + (n n )Kn+1,n ,n (f )

N [1 n (Gn )] N N (n n )Qn+1,n (f ) + (n n ) N n (Gn ) N + (n n )Kn+1,n ,n (f )

Gn n (Gn )

n Qn+1,n (f )

which gives the straightforward Lp estimation. For the last term (), the work is put on the mean-eld law estimation and we have to use the Lipschitz hypothesis for the function b. For any bounded measurable function f , the dierence between the exact and empirical McKean kernel is expressed by:
N N n Kn+1,n ,n (f ) n Kn+1,n ,n (f ) = N d N N f (y)n (dx)Sn,n (x, dz)[Mn+1,n Mn+1,n ](z, dy). N d

Recalling that for two M(E)-valued measures and and any x E, Mn+1, f (x) Mn+1, f (x) = dyf (y) e
A(x,y,)2 X2 2n t

A(x,y,)2 X2 2n t

934

C. BAEHR

with A(x, y, ) = y x b(x, )t and using the Lipschitz hypothesis plus the Gaussian properties, it yields for any p > 0 1 1 N N N E n Kn+1,n ,n (f ) n Kn+1,n ,n (f ) p p Cn (p) I(H). N d d Using the induction, the technical Lemma 3.1 and the Marcinkiewicz inequality, the sketch of the proof is therefore completed. The ltering of an a-priori mean-eld process requires two particle systems, one with cardinality N , the 1 other with cardinality d, to produce bounding errors O( 1 + d ). N Now we can modify the problem and consider a dynamical system conditioned to the observation Yn , or more exactly to the law of the lter n and following the SDE: X X Xn+1 = Xn + b(Xn , n )t + n Wn Y Y Yn = h(Xn ) + n Wn X0 0 Y0 = y0 where n is the Law(Xn |Y0 . . . Yn ). The ltering of this kind of process needs only one particle systems with N interacting particles, and we prove the theorem: Theorem 3.3. With the hypotheses of this section, for all n 0 and for all p > 0 there exists constants 0 Cn (p) < such that
N E( n n
1 p p H)

Cn (p) I(H). N

Proof. We argue by induction with analogous arguments to the previous proof. For n = 0 invoking the technical lemma, the statement holds. We assume that it holds for n > 0 and for all p > 0. For any measurable bounded function f , the dierence N n+1 (f ) n+1 (f ) can be decomposed in:
N n+1 (f ) n+1 (f ) = N N n+1 (f ) n Kn+1,n ,n (f ) N N

()

N N + n Kn+1,n ,n (f ) n Kn+1,n ,n (f ) () N N N N + n Kn+1,n ,n (f ) n Kn+1,n ,n (f ). ( ) N

Terms () and ( ) are already known. To be controlled the term () uses the Lipschitz hypothesis of the function b and a supplementary inequality gives the induction. Indeed by construction, we have n = n Sn,n , then the bounding veries the inequality
N N |n Sn,n (f ) n Sn,n (f )| |n Sn,n (f ) n Sn,n (f )| + |n Sn,n (f ) n Sn,n (f )|. N N N N

With some calculations on the selection kernel we have the last bound:
N sup E|n Sn,n (f ) n Sn,n (f )| 2

f 1

1 n (Gn ) Gn n (Gn )

N sup E|n (f ) n (f )|. f 1

We use the recurrence hypothesis and obtain the required Lp inequality for n + 1. In this second case, with a modied dynamical equation, with only N particles we have Feynman-Kac distributions errors bounded by O( 1 ). This remark could be a clue to reduce the numerical cost of a lter N

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

935

for physical systems with mean-eld dynamics. Accepting an error in the model we have a faster algorithm. But in counterpart, the ltering estimation is very dependent on a good representation of the observations by the used observation equation, otherwise the lter could experience serious disturbances.

4. Application to atmospheric turbulence measurements

In the previous two sections, we have developed a tool to inspect locally a random eld, and a ltering algorithm to denoise the sort of processes implied by these local representations. Now we propose to apply our methods to a physical problem. We consider the measurements of turbulent uids with a special focus on observed atmospheric wind. First, we have to choose local stochastic Lagrangian models for turbulent uids. It will be necessary to adapt them to the measurement problem. For the adapted models, we will be able to couple the Acquisition Process dynamics and the nonlinear ltering for mean-eld processes to write a performing algorithm. We suggest here a method of closure for the dynamical system by conditioning the evolution (kernel) to the observation. With our particle approximation we will see that we have the opportunity to learn the missing parameters with the large scale observations dynamics. For our applications, we intend to simulate data using the chosen model and the acquisition process or to use recorded real data chosen for their quality with the minimum of instruments noises, for not to say without noise at all. These data will be called reference signals. Then, we add numerical noise with various possible type of probability laws to obtain the perturbed signal. Using this noisy signal our algorithms have to estimate the parameters of the random medium and estimate a ltered signal. As a result we can compare our estimation to the reference data which is one realization of the random medium along the path.

4.1. Adapted stochastic Lagrangian models for nonlinear lter

This section is devoted to the adaptation of two models for turbulent uids. The rst one is the Simplied Langevin Model (SLM) for incompressible homogeneous turbulence proposed by Pope (see [15] for example). The second model is proposed by Das and Durbin (see [8]) for relative dispersion in stratied turbulent ows. In our ltering problem, there are no other informations on the uid ows except the observations themselves. Thus we have to use these observations and their dynamics to learn the missing parameters. The ltering of such turbulent ows needs to constrain the model to the observations in order to give relevant estimation of the Eulerian quantities of large eddies. This conditioning to the observations of a Markovian kernel is a new method of closure for a stochastic model. Now, we detail the dierent necessary adaptations. Consider for all t [0, T ] with T , the (E E , E E )-valued Acquisition System (Xt , Ut,x ) on (, F , P). Then vector eld Ut,x is the (E, E)-valued Eulerian uid velocities. The Eulerian average is given for the stationary Acquisition Path Xt = x E and for all bounded measurable function f by the equality f (U ) t,x = E[f (Ut,x )] P ps (see Sect. 1, Application 1.1).
x x Let x0 E be an initial xed point. We consider the Lagrangian Acquisition System (Xt 0 , Ut,Xt 0 = Vt ) x0 where Xt is the ow coming from x0 . For any measurable function f we have x Ex0 (f (Vt )|Xt 0 )

= 0

F 0 (f (Vt )) x0 x (Xt ) F 0 (1) x

(4.1)

using the weakly regularizing kernel G with parameter 0 mentioned in section as (1.3):
x F 0 (f (Vt ))(Xt 0 ) = x x f (v)G (Xt 0 , z)pXt
x0

,Vt

(z, v)dzdv
x0

with a convergence of the approximation (4.1) as 0 only for smooth densities pXt

,Vt

(z, v).

936

C. BAEHR

This expectation is dependent on the initial point x0 . If we assume that the transportation by the Lagrangian ow is without density variation (no loss or gain of mass), we can take as point of departure a random variable X0 with the distribution 0 and all the Acquisition Processes arrived at time t in x E, and we evaluate the expectation E[f (Vt ) | Xt = x, X0 ] and with the same regularization technique for any measurable bounded function f (f (Vt ))(x) t
def

E [f (Vt )|Xt = x, X0 ]
x0

and without mass density variations, we have P ps, for smooth densities pXt E [f (Vt )|Xt = x] = lim (f (Vt ))(x). t
0

,Vt

This approximation (f (Vt ))(x) will be used to replace the (unreachable) Eulerian average f (U ) t,x . t In its continuous version for 1D or 2D isotropic ows, the simplied SLM given by Pope for incompressible isotropic homogeneous turbulence is the It process o dXt dVt = = Vt dt x p dt ( 1 + 3 C0 ) kt (Vt v ) dt + C0 t dBt 2 4 t (4.2)

where Vt is the Lagrangian velocity, x p is the gradient of mean pressure, t is the turbulent dissipation rate, kt the Eulerian average of the turbulent kinetic energy, are Eulerian means, C0 is the Kolmogorov constant and Bt is a Wiener process. The incompressibility of the uid requires an hypothesis: the divergence of the velocity eld have to be null. This hypothesis is guaranteed by the pressure term x p dt (see [15]). As usually for It processes, the time discretization uses a classical explicit Euler scheme, here with a time o step t > 0. In the Pope model, the quantities x p and t act as exogenous parameters. These two commands of the model realize a micro-macro coupling for the Lagrangian system with the large Eulerian scale. First we express a remark due to the particular shape of equation (4.2): the asymptotic behaviour of the expectation of the velocities increments as the parameter tends to 0 is E[Vn (Vn )(Xn )] 0. n Since x p
n 0

is independent of the pair (Xn , Vn ), it holds: E[Vn | Xn ] = x p n t. (4.3)

Then we propose to model the right handside of (4.3) by a random variable E -valued such that E[Vn ] = E[n ]. Here we consider also that the density (more exactly 1 ) is included in the term n . Our lter will have to learn this random variable n using the observation process. Back to the continuous model we remark that E[dVt dVt ] = C0 t dt, and we propose another model for n taking the expectation of a random variable (yet denoted) n : E[n ] =
def def

E[Vn Vn ] C0 t

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

937

By our models for the two commands, there will be some errors on the velocity representation, then we have X to add a random error Wn on the location equation. Thus the discrete adapted Pope model is the system: Xn+1 Vn+1 = =
X X Xn + Vn t + n Wn E(n ) Vn + E(n ) + C1 kn [Vn (Vn )(Xn )]t + n V C0 E(n )Wn .

(4.4)

If we use this model without any care, the incompressible hypothesis is disrupted, but using the ltering method and observation of incompressible uids, we hope that this quality will be restored. Now, the system (4.4) has to be modied to express the coupling between the Lagrangian Acquisition system and the Observation Acquisition by a conditioning of the probability laws. This coupling is a new closure method and we think with this technique to improve the accuracy and the computational performance of the ltering. Indeed, as seen in Section 3, when we condition a Markov kernel by observations, we reduce the number of particles necessary to lter the mean-eld process. In return, the accuracy will be very linked to a good representation of the observation process. Let us formalize the physical situation to be closer to the ltering problem. The system (4.4) can be rewritten in the dynamical system: Xn+1 = Xn + b(Xn , n )t + (Xn )Wn where Xn = (Xn , Vn1 , Vn ) is the conditional mean-eld uid-signal and n is probability law of the lter, n = Law(X[0,n] |Y[0,n] ), b is a function with the hypotheses of Section 3 and a Lipschitz constant depending on . In this model is a function of the state vector Xn . We have to assume that this is a positive, non-zero and bounded function. This model is more general than the system (3.2). But with the hypotheses, the control is easy to obtain (see [1], Thm. 6.5.1). In the physical applications, this hypothesis relative to the nite energy of the physical system will be insure. This is our stochastic model and we have to use it to lter the measurements of a turbulent uid.
Y Y We consider the perturbed observations of the velocities Yn = H(Xn ) + n Wn where n is still a bounded positive constant, and the ltering problem is to compute the Law(X[0,n] | Y[0,n] ). Xn is then a conditional Markov process with the transition law Mn+1,n where

Mn+1,n ((x, u), d(z, v)) = PXn+1 |Xn ,Y0 (d(z, v)|(x, u)).
n

Section 3 claims that the particle approximation requires a system of N particles and provides an error in O 1 . N In its spread out form the uid-system is: X X Xn+1 = Xn + Vn t + n Wn Vn+1 = Vn + E(n |Y0 . . . Yn )t vG (X x)P(Xn ,Vn |Y0 ...Yn ) (d(x,v)) 0 C1 E(n |Yn...Yn ) [Vn G (X nx)P(Xn ,Vn |Y0 ...Yn ) (d(x,v)) ] t k n V + C0 E(n |Y0 . . . Yn ) Wn Y Yn = H(Vn ) + Y Wn with kn = 1 [w G (zx)P(Xn ,Vn |Y0 ...Yn ) (d(x,v)) ]2 P(Xn ,Vn |Y0 ...Yn ) (d(z, w)). 2 Now we have to add the localization step to couple the system with the Acquisition Path of the sensor. This second coupling means that the dierent conditional expectation and the Eulerian average . are in fact approximated by E(. | X[0,t] Bt (Z[0,t] ), Yt ). This third part of the algorithm is an acceptation/rejection step with a redistribution in the set of balls Bt (Z[0,t] ). For 3D atmospheric ows, the SLM model made for homogeneous ows is inappropriate and we suggest to build a model following the proposition of Das and Durbin (see [8]) for dispersion in a stratied turbulent ow.
vG (zx)P(Xn ,Vn |Y0 ...Yn ) (d(x,v))

938

C. BAEHR

Our stochastic model can be seen as the dispersion of a uid particle with respect to a virtual element following the mean ow: dXt = Vt dt dVh,t = h p dt C1 t (Vh,t V h,t )dt 2 kt 1 d V V + (C2 1) (Wt W t ) dzh,t dt + (C0 t ) 2 dBt h dWt = d W t C1 kt (Wt W t )dt (4.5) 2 t 1 W + (1 C5 ) t g(t t )dt + (C0 t ) 2 dBt dt = d t (C1 C1 t (t t )dt 2 kt 1 (Wt W t ) d dz t dt + (C ) 2 dBt where Vh,t is the horizontal speed, h is horizontal gradient operator (the partial derivatives of the horizontal V W components), W is the vertical speed, is the temperature and Bt h , Bt , Bt are independent Wiener processes. The system (4.5), using the dispersion of a uid particle with respect to the mean ow, is our rst adjustment of the Das and Durbin model to be usable in our ltering problem. We add some simplications particularly on the constants: C0 becomes the universal Kolmogorov constant with C0 = 2.1, and using the paper of Das and Durbin [8] we set the other constants to C1 = 1.8, C1 = 2.5, C2 = 0.6 and C5 = 1/3, C will be learned as t can be. The buoyancy coecient t is approximated by the reciprocal of the local mean temperature (t )(Xt ), t Vh t = (t1)(Xt ) . It remains hidden parameters, the vertical gradients h = d dz and = ddz , where z is the t t t vertical dimension. The particle lter will learn these gradients using the selected and well-adapted particles. Applying the same methods used for the SLM model, we write:
h h Ah = E[Vn Vn1 ] n h h E[Vn Vn1 ]2 C0 t E[Wn Wn1 ]2 W En = C0 t E[n n1 ]2 C = C0 t h En =

AW = E[Wn Wn1 ] n A = E[n n1 ] n

For the kinetic energy, we use dierent forms for the horizontal and the vertical components to restore the h h h W stratied medium: kn = 1 ([Vn (Vn )]2 )(Xn ) and kn = 1 ([Wn (Wn )]2 )(Xn ). For the temperature n n 2 n 2 n 2 h 2 W 2 2 h 2 W equation, we choose kn = (kn ) + (kn ) and En = (En ) + (En )2 . These models are made to learn from the observations the hidden parameters and close the system by conditioning the transition law. The dynamical system is therefore: X X Xn+1 = Xn + Vn t + h Bn Vh h h h n n+1 = Vn + Ah C1 Eh [Vn (Vn )(Xn )] t n n 2 kn h 1 h + (C2 1) [Wn (Wn )(Xn )] h t + (C0 En ) 2 Bn n n W W C 1 En n+1 = Wn + AW 2 kW [Wn (Wn )(Xn )] t n n n W 1 W + (1 C5 ) n g [n (n )(Xn )] t + (C0 En ) 2 Bn (4.6) n C 1 En n+1 = n + An C1 2 kn [n n (n )(Xn )] t 1 [Wn (Wn )(Xn )] t + (C ) 2 Bn n n X0 = x0 V0h = v0 , W0 = w0 0 = 0 .

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

939

We have seen previously that the Markov kernels have to be conditioned to the path, this has a consequence for A the dynamic model. Indeed denoting Vn the Acquisition velocity (velocity of the uid of along the Acquisition Z L Path), Vn the course velocity along the Acquisition Path, and Vn the velocity of the Lagrangian uid particle following (4.6) using the implicit derivation:
A L Z X Vn+1 = Vn+1 + GA Vn t + X GA Bn n,Z n,Z

(4.7)

where GA is the gradient of uid velocity along the Acquisition Path. In the case of a particle resolution of n,Z the system, GA is approached by GA,N the empirical gradient in the direction [Zn Zn1 ] computed using n,Z n,Z the selected particles. B The Markov kernel associated to (4.7) is denoted Mn+1,n . In 1D ow (resp. 3D ow) we consider h Xn = (Xn , Vn ) (resp. Xn = (Xn , Vn , Wn , n )). The superscript Bn means conditioned to be in the ball Bn (Zn ). The discrete ltering algorithm is then described by the following sequence:
n n n+1 B Xn n Xn n Xn+1 Xn+1 B n+1

w.r.t. Sn,

B w.r.t. Mn+1,

Z,B w.r.t Sn+1,

where n+1 = Law(Xn+1 | X[0,n] Bn (Z[0,n] ), Y[0,n] ). These two examples of adaptation of turbulent uid models are necessary to lter simulated or real data in 1D, 2D or 3D atmospheric wind. Others models or parameterizations or learning techniques are possible. In the next section we will use the 3D stratied model (4.6) to lter real data.

4.2. Application to 3D real wind measurements

We have described all the algorithms for the various parts of the problem put down by the ltering of the mobile measurements of a uid. It is necessary to test our methods on simulated or real data. To check the ability of our lter to perform good estimations, we propose to use a reference signal, real or not. We perturb this signal with a random noise using simple or complex probability laws. Then, our algorithms are used to deliver a ltered signal. By this procedure, we are able to compare the ltered signal with the initial reference signal. The use of real data requires some precautions especially on the high quality of the measurements to be considered as a reference. We have to interpret the results with care. The reference signal is one realization of the random medium taking into account its parameters, the ltered signal is built with the estimated parameters of the observed medium. So, there is a natural dierence between the two series, but they have to give the same turbulence characteristics particularly on large scale and in their energetic cascades, seen with their Power Density Spectra (PSD). The quality of the estimations have to be evaluated with several manner. The rst one is the ltering of uid velocities simulated data. This examination enables to test the sequences of the algorithms. We do not present these results on 1D or 2D simulated uids and prefer to examine here directly the capacity of the method with real 3D ows. Real or simulated, the rst comparison will be visual with the examination of their ltered and reference series and the PSD. We remind that atmospheric turbulent ows, in the domain of our measurements, follow the empirical Kolmogorov law (K41) (see for example Frischs book [10]). In log-log diagram, the PSD have to follow a slope in 5/3 and we expect this kind of energetic cascade for the reference signal and for the ltered one. Another possible control of the lter quality can be the calculation of some large scale classical turbulence indicators. We will present these calculations in a further publication. Here, we present the results of ltering using atmospheric data recorded at 5 Hz with an ultrasonic anemometer in the experimental eld of the French Weather Service Research Center in Toulouse, France, at two moments: rst on the 12th of May 2006 between 12h03 and 12h07 UTC and second on the 14th of June 2006 between 15h33 and 15h37 UTC.

940
2 0 2 4 6 8 10 12 900 1100 1300 1500 1700

C. BAEHR

20 10 0 10 20 30
1900
3 2 1 0 1

10 20 10 0 10 20

10

10

10

10

2 1 0 1 2 3 4 5 6 900 1100 1300 1500 1700 1900

10

10

10

10

10

Figure 4. Left, series of horizontal wind (velocity (ms1 ) vs. time step number) and right, PSD (with a log-log scale, power (dB) vs. frequency (Hz)). The series on the top is the U component and the bottom the V component. Time series are taken the 12th of May 2006 between 12h03 and 12h07 UTC. In light blue the perturbed signal to be denoised, in black the reference signal and in red the ltered with 800 particles. (Figure in color available online at www.esaim-m2an.org/.)

The choice of this two dates and the hours is only determined by the measurement quality. The ultrasonic anemometer gives a measure of the sonic temperature, which is converted in a real temperature with thermodynamics laws. We choose this method of measurement for the temperature to insure a colocalization of all the measurements (wind and temperature). But it is well known that ultrasonic anemometer are far from being the best thermometer. So the temperature signal is taken as reference but it is a stopgap. For further experimental campaigns we will add to the anemometer a thin wire thermometer to measure a fast temperature. In the two cases we choose to add a random noise built using the local empirical variance of the reference signals and a Gaussian law. This noise is an upper limit for turbulence, and there is a memory eect which can give dramatic errors if the lter does not return the right parameter of the uid. Sometimes our noise is quite big, barely too strong, but it provides a good test on the robustness of our lter. For the rst experiment on 12th of May, Figures 4 (horizontal wind) and 5 (vertical wind and temperature) show in light blue dash line the noisy signal of the three components of wind and temperature, the black line is the signal of reference to retrieve. The lter uses the algorithm described in the previous sections with 800 particles for the approximation. The red signal is the output of our lter using a SCILAB code. Even if the perturbations are strong, the ltered signal compares well with the reference component. On the diagram of PSD, the correction is very sharp, with a decreasing of power following the reference spectrum,

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

941

3.0 2.5 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 900

10 0 10 20 30 1100 1300 1500 1700 1900

3 2 1 0 1

10

10

10

10

10

293.0 292.5 292.0 291.5 291.0 290.5 290.0 289.5 289.0 288.5 288.0 900

10 0 10 20 30 1100 1300 1500 1700 1900

3 2 1 0 1

10

10

10

10

10

Figure 5. Series (x-axis is time step number). On the top, the vertical velocity (ms1 ), on the bottom, the temperature (K) and PSD (power (dB) vs. frequency (Hz)). Time series are taken the 12th of May 2006 between 12h03 and 12h07 UTC. In light blue the perturbed signal to be denoise, in black the reference signal and in red the ltered with 800 particles. (Figure in color available online at www.esaim-m2an.org/.)

far from the perturbed signal. In detail, there are slight dierences between the ltered and the reference signals, it comes from our method itself. Indeed, the algorithm estimates the characteristic parameters of the uid through the model used by the lter, and returns a ltered signal, while the output is one possible realization of the medium. This is the reason why the spectral correction is very good, while the series show dierences. In conclusion, the noise is entirely subtracted out of the perturbed signal. Our method, which estimates some characteristic terms, retrieves also turbulence parameters at high frequency which are accessible by the prediction model. It could be possible to evaluate the turbulent dissipation rate or buoyancy coecient or 3-d gradients of Eulerian averages, etc., and of course the Acquisition Path of the measurement sensors. Figure 6 gives an example of the series of turbulence dissipation rates and the vertical gradient of temperature. These estimations with high frequency of turbulent quantities are something very new. Usually in meteorological sciences these parameters are estimated with long time series, empirical means and doubtful ergodic hypothesis. Our method provide estimations of these quantities at the same frequency as the measurements and a real sense for the Eulerian averages. This result is very promising but we do not have to forget that the ltered signal is very dependent on the chosen model. If the model is not pertinent relatively of the dynamics of the random medium, it could have some disturbances, sometimes annoying. To illustrate this fact, we can present an experiment where the ltered temperature signal has a poor quality. Figures 7 (horizontal wind) and 8 (vertical wind and temperature) show,

942

C. BAEHR

Figure 6. Left, estimation of n parameter (m2 s3 ) and right gradient of mean temperature (Km1 ) with 800 particles. Time series are taken the 12th of May 2006 between 12h03 and 12h07 UTC. X-axis is time step number.

Figure 7. Left, series of horizontal wind (velocity (ms1 ) vs. time step number) and right, PSD (with a log-log scale, power (dB) vs. frequency (Hz)). The series on the top is the U component and the bottom the V component. Time series are taken the 14th of June 2006 between 15h33 and 15h37 UTC. In light blue the perturbed signal to be denoise, in black the reference signal and in red the ltered with 800 particles. (Figure in color available online at www.esaim-m2an.org/.)

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

943

8 6 4 2 0 2 4 6 1000 30 1200 1400 1600 1800 2000

3 2 1 0 1

10 0 10 20

10

10

10

10

10

308 306 304 302 300 298 296 294 1000 1200 1400 1600 1800 2000 10 20 30
3 2 1 0 1

10 0

10

10

10

10

10

Figure 8. Series (x-axis is time step number). On the top, the vertical velocity (ms1 ), on the bottom, the temperature (K) and PSD (power (dB) vs. frequency (Hz)). Time series are taken the 14th of June 2006 between 15h33 and 15h37 UTC. In light blue the perturbed signal to be denoise, in black the reference signal and in red the ltered with 800 particles. (Figure in color available online at www.esaim-m2an.org/.) with the same color code, the three signals (noised, ltered and of reference) for the second date on the 14th of June 2006 at 15h30 UTC. The (horizontal or vertical) wind is always properly ltered, but the temperature response is not correct as indicates by the PSD. There is two leads. The rst one is the relevance of the model for this specic real meteorological situation. The second is the quality on the measurement used as reference. In the stratied Lagrangian model, the four components are coupled and the ltering of the wind is dependent on a good estimation of the temperature. But the three wind components behave correctly, while the temperature signal seems very dierent from the reference. At this point there is no unique answer, the problem could be on the measurement of temperature, or in the choice of the model or its adaptation or may be there is a specic behaviour of the atmosphere not taken into account by the model (water vapor, water content, etc.). To insure that the problem does not come from too big numerical perturbations, we have proceeded with various level of noises without any improvement of the ltered temperature structure. The conclusion could be that the dynamics of the reference temperature and the dynamics of the temperature deduced by the model are dierent, while the 3D wind is correctly represented, but we have no argument yet to say if the problem lies in the chosen reference for temperature. For all the distinct experiments we have made, with various regime of wind, period of the year, hour of the day, etc., in the majority of the cases the ltering is good with an accurate correction. Nevertheless, when the atmospheric turbulence is too weak (early morning, no wind situation, etc.), the ltered signal remains

944

C. BAEHR

too turbulent just because the laminar regime is not well tted by our Lagrangian models. Out the case of laminar ows, sometimes there are some problems with the temperature ltering while the wind is accurately estimated. The situation shown previously, the 14th of June 2006, is an example of this fact. For the moment, there exists no satisfying explanations. It illustrates the diculties of this exercise in real conditions and gives some clues for the further experiments and improvements.

5. Outcomes and further developments

In this paper, we have described how to lter perturbed observations of a random medium taken along a random path and applied the result to real meteorological data. First, we dene precisely the Acquisition Process as the statement along a path of a random medium. We provide some examples of use and describe an algorithm of estimation in the case of homogeneous medium represented by its Lagrangian ows. Subsequently we have developed new algorithms for the ltering of mean-eld processes and suggest particle approximations. We give their asymptotic behaviour when the number of particles goes to innity. Then to apply to uid velocities measurements we have adapted some existing physical models to be able to learn turbulence parameters. These adaptations correspond to the coupling of the model with the Acquisition Path and the Observational Acquisition Process. We have seen that these couplings with a conditioning of the model to the observations are a new manner to close a dynamical system. For the physical applications, we have chosen two real meteorological situations. We have corrupted the reference signal with colored Gaussian noises, occasionally with very strong perturbations. Our lter provides quite good estimations with pertinent series of ltered signals and a sharp correction of the energetic cascades. High frequency estimations for turbulent parameters with a sense of the Eulerian means are an important innovation given by the Acquisition Process and stochastic ltering for mean-eld processes. But nonlinear ltering never lives in a perfect world and we have seen that problems may occur especially with laminar ows or in the treatments of the temperature. About the temperature, the question of the pertinence of ultrasonic anemometer for the measurements or the pertinence of our stochastic model for temperature component had not be decided. But this work is a rst stage. For atmospheric data assimilation we have successfully used the Acquisition Process to enhance the performance of an Ensemble Kalman Filter with a coupled Local Particle Filter (see the short report of Baehr and Pannekoucke [3]). We hope to use in future developments these dierent techniques to estimate turbulent parameters for high resolution atmospheric forecasting model assimilation. Using the Acquisition Process and the data fusion particle techniques, it is also our intention to contribute to the design of new integrated systems for airborne turbulence measurements and with a more physical modeling to lter numerically more complex atmospheric set of parameters including water content, chemical or aerosol concentration, droplets counting, etc. Theoretical works are also in progress to modify the algorithm presented here and not break the incompressibility hypothesis like the modication of the model proposed here does. Back to the mathematical part of this work, there is also some study to complete the understanding of the lter based on Acquisition Process and particularly to characterize the closure of a model by observations. The article of Baehr and Pannekoucke [3] explain that in high dimension genetic ltering algorithms require a critical number of particles to insure the non-divergence of the lter. We have to deal studies about this aspect and propose new lter coupling various nonlinear lters and separating the treatment of linear and nonlinear components of a system and nd strategies of piloting/tuning of the selection parameters to be adapted to models with large degrees of freedom. Local Particle Filter with Acquisition Process presented here should be a partial answer.
Acknowledgements. The author thanks the referee for his very careful review and comments that include many points and improve signicantly the clarity of this paper.

FILTERING FOR RANDOM VECTOR FIELD ALONG A RANDOM PATH

945

References
[1] C. Baehr, Modlisation probabiliste des coulements atmosphriques turbulents an den ltrer la mesure par approche e e e particulaire. Ph.D. Thesis University of Toulouse III Paul Sabatier, Toulouse Mathematics Institute, France (2008). [2] C. Baehr and F. Legland, Some Mean-Field Processes Filtering using Particles System Approximations. In preparation. [3] C. Baehr and O. Pannekoucke, Some issues and results on the EnKF and particle lters for meteorological models, in Chaotic Systems: Theory and Applications, C.H. Skiadas and I. Dimotikalis Eds., World Scientic (2010). [4] G. BenArous, Flots et sries de Taylor stochastiques. Probab. Theor. Relat. Fields 81 (1989) 2977. e [5] M. Bossy and D. Talay, Convergence rate for the approximation of the limit law of weakly interacting particles. 2: Application to the Burgers equation. Ann. Appl. Prob. 6 (1996) 818861. [6] B. Busnello and F. Flandoli and M. Romito, A probabilistic representation for the vorticity of a 3d viscous uid and for general systems of parabolic equations. Proc. Edinb. Math. Soc. 48 (2005) 295336. [7] P. Constantin and G. Iyer, A stochastic Lagrangian representation of 3-dimensional incompressible Navier-Stokes equations. Commun. Pure Appl. Math. 61 (2008) 330345. [8] S. Das and P. Durbin, A Lagrangian stochastic model for dispersion in stratied turbulence. Phys. Fluids 17 (2005) 025109. [9] P. Del Moral, Feynman-Kac Formulae, Genealogical and Interacting Particle Systems with Applications. Springer-Verlag (2004). [10] U. Frisch, Turbulence. Cambridge University Press, Cambridge (1995). [11] G. Iyer and J. Mattingly, A stochastic-Lagrangian particle system for the Navier-Stokes equations. Nonlinearity 21 (2008) 25372553. [12] I. Karatzas and S.E. Shreve, Brownian Motion and Stochastic Calculus. Springer-Verlag (1988). [13] S. Mlard, Asymptotic behaviour of some particle systems: McKean Vlasov and Boltzmann models, in Probabilistic Models ee for Nonlinear Partial Dierential Equations, Lecture Notes in Math. 1627, Springer-Verlag (1996). [14] R. Mikulevicius and B. Rozovskii, Stochastic Navier-Stokes Equations for turbulent ows. SIAM J. Math. Anal. 35 (2004) 12501310. [15] S.B. Pope, Turbulent Flows. Cambridge University Press, Cambridge (2000). [16] A.S. Sznitman, Topics in propagation of chaos, in Ecole dEt de Probabilits de Saint-Flour XIX-1989, Lecture Notes in e e Math. 1464, Springer-Verlag (1991).