Kernel-based collocation methods for Zakai equations

Kernel-based collocation methods for Zakai equations

Yumiharu Nakano
Department of Mathematical and Computing Sciences, School of Computing
Tokyo Institute of Technology
W8-28, 2-12-1, Ookayama, Meguro-ku, Tokyo 152-8550, Japan
e-mail: nakano@c.titech.ac.jp
September 25, 2019
Abstract

We examine an application of the kernel-based interpolation to numerical solutions for Zakai equations in nonlinear filtering, and aim to prove its rigorous convergence. To this end, we find the class of kernels and the structure of collocation points explicitly under which the process of iterative interpolation is stable. This result together with standard argument in error estimation shows that the approximation error is bounded by the order of the square root of the time step and the error that comes from a single step interpolation. Our theorem is well consistent with the results of numerical experiments.

Key words: Zakai equations, kernel-based interpolation, stochastic partial differential equations, radial basis functions.

AMS MSC 2010: 60H15, 65M70, 93E11.

1 Introduction

We are concerned with numerical methods for Zakai equations, linear stochastic partial differential equations of the form

 (1.1) du(t,x)=L0u(t,x)dt+m∑k=1Lku(s,x)dWk(t),0≤t≤T,

with initial condition , where the process is an -dimensional standard Wiener processes on a complete probability space . Here, for each , the partial differential operator is given by

 L0f(x) =12d∑i,j=1∂2∂xi∂xj(aij(x)f(x))+d∑i=1∂∂xi(bi(x)f(x)), Lkf(x) =βk(x)f(x)+d∑i=1∂∂xi(γik(x)f(x)),k=1,…,m

where is -valued, is -valued, is -valued, is -valued, and is -valued, all of which are defined on . The conditions for these functions are described in Section 2 below.

It is well known that solving Zakai equations is amount to computing the optimal filter for diffusion processes. We refer to Rozovskii [17], Kunita [13], Liptser and Shiryaev [14], Bensoussan [4], Bain and Crisan [2], and the references therein for Zakai equations and its relation with nonlinear filtering. It is also well known that for linear diffusion processes the optimal filters allow for finite dimensional realizations, i.e., they can be represented by some stochastic and deterministic differential equations in finite dimensions. For nonlinear diffusion processes, it is difficult to obtain such realizations except some special cases (see Beneš [3] and [4]). Thus one may be led to numerical approach to Zakai equations for computing the optimal filter. Several efforts have been made to obtain approximation methods for the equations during the past several decades. For example, the finite difference method (see Yoo [20], Gyöngy [9] and the references therein), the particle method (see Crisan et al. [6]), a series expansion approach (Lototsky et al. [15]), Galerkin type approximation (Ahmed and Radaideh [1] and Frey et al. [7]) and the splitting up method (Bensoussan et al. [5]).

In this paper, we examine the approximation of by a collocation method with kernel-based interpolation. Given a points set and a positive definite function , the function

 I(f)(x):=N∑j=1(A−1f|Γ)jΦ(x−xj),x∈Rd,

interpolates on . Here, , is the column vector composed of , , and denotes the -th component of for . Thus, with time grid , the function recursively defined by

 uh(ti,x) =uh(ti−1,x)+L0I(uh(ti−1,⋅)(x)(ti−ti−1) +m∑k=1LkI(uh(ti−1,⋅)(x)(Wk(ti)−Wk(ti−1)),i=0,…,n,x∈Rd,

is a good candidate for an approximate solution of (1.1). The approximation above can be seen as a kernel-based (or meshfree) collocation method for stochastic partial differential equations. The meshfree collocation method is proposed by Kansa [11], where deterministic partial differential equations are concerned. Since then many studies on numerical experiments and practical applications for this method are generated. As for rigorous convergence, Schaback [18] and Nakano [16] study the case of deterministic linear operator equations and fully nonlinear parabolic equations, respectively. However, at least for parabolic equations, there is little known about explicit examples of the grid structure and kernel functions that ensure rigorous convergence. An exception is Hon et.al [10], where an error bound is obtained for a special heat equation in one dimension. A main difficulty lies in handling the process of the iterative kernel-based interpolation. A straightforward estimates for involves the condition number of the matrix , which in general rapidly diverges to infinity (see Wendland [19]). Thus we need to take a different route. Our main idea is to introduce a condition on the decay of when becomes large and to choose an appropriate approximation domain whose radius goes to infinity such that the interpolation is still effective. From this together with standard argument in error estimation we find that the approximation error is bounded by the order of the square root of the time step and the error that comes from a single step interpolation. See Lemma 3.7 and Theorem 3.4 below.

The structure of this paper is as follows: Section 2 introduces some notation, and describes the basic results for Zakai equations and the kernel-based interpolation, which are used in this paper. We derive an approximation method for the filter and prove its convergence in Section 3. Numerical experiments are performed in Section 4.

2 Preliminaries

2.1 Notation

Throughout this paper, we denote by the transpose of a vector or matrix . For we set . For a multiindex of nonnegative integers, the differential operator is defined as usual by

 Dα=∂|α|1∂xα11⋯∂xαdd

with . For an open set , we denote by the space of continuous real-valued functions on with continuous derivatives up to the order , with the norm

 ∥f∥Cκ(O)=max|α|1≤κsupx∈O|Dαf(x)|.

Further, we denote by the space of infinitely differentiable functions on with compact supports. For any and any open set , we denote by the space of all measurable functions such that

 ∥f∥Lp(O):={∫O|f(x)|pdx}1/p<∞.

For , we write for the space of all measurable functions on such that the generalized derivatives exist for all and that

 ∥f∥2Hκ(O):=∑|α|1≤κ∥Dαf∥2L2(O)<∞.

In addition, for , we write for the space of all measurable functions on such that the generalized derivatives exist for all and that

 ∥f∥2Hκ+r(O):=∥f∥2Hκ(O)+∑|α|1=κ∫O∫O|Dαf(x)−Dαf(y)|2|x−y|d+2rdxdy<∞.

For we use the notation . By we denote positive constants that may vary from line to line and that are independent of introduced below.

2.2 Zakai equations

We impose the following conditions for the equation (1.1):

Assumption 2.1.
1. All components of the functions , , , , and are infinitely differentiable with bounded continuous derivatives of any order.

2. For any ,

 ξT(a(x)−γ(x)γ(x)T)ξ≥0,ξ∈Rd.

It follows from Assumption 2.1(i) and Gerencsér et.al [8, Theorem 2.1] that there exists a unique predictable process such that the following are satisfied:

1. for any , where with and for any ;

2. for ,

 (2.1) (u(t),φ)=(u0,φ)+∫t0(u(s,⋅),L∗0φ)ds+m∑k=1∫t0(u(s,⋅),L∗kφ)dWk(s),0≤t≤T,a.s.

Here, denots the inner product in , and for each , the partial differential operator is the formal adjoint of . Moreover, satisfies

 E[sup0≤t≤T∥u(t,⋅)∥2Hν(Rd)]≤C∥π0∥2Hν(Rd),ν∈N.

Further, as in [17, Proposition 3, Section 1.3, Chapter 4], there exists a version , with respect to , of such that for and that for any and ,

 (2.2) Dα~u(t,x)=Dαπ0(x)+∫t0DαL0~u(s,x)dx+m∑k=1∫t0DαLk~u(s,x)dWk(s),a.s.,(t,x)∈[0,T]×Rd.

In particular, is a solution to the Zakai equation in the strong sense, i.e., satisfies

 ~u(t,x)=~u0(x)+∫t0L0~u(s,x)dx+m∑k=1∫t0Lk~u(s,x)dWk(s),a.s.,(t,x)∈[0,T]×Rd.

We remark that in (2.2) the stochastic integral is taken to be a continuous version with respect to . With this version, (2.2) holds with probability one uniformly on .

2.3 Kernel-based interpolation

In this subsection, we recall the basis of the interpolation theory with positive definite functions. We refer to [19] for a complete account. Let be a radial and positive definite function, i.e., for all and for every , for all pairwise distinct and for all , we have

 ℓ∑i,j=1αiαjΦ(yi−yj)>0.

Let be a finite subset of and put . Then is invertible and thus for any the function

interpolates on . If , then

 NΦ(Rd):={f∈C(Rd)∩L1(Rd):ˆf/√ˆΦ∈L2(Rd)},

called the native space, is a real Hilbert space with inner product

 (f,g)NΦ(Rd)=(2π)−d/2∫Rdˆf(ξ)¯¯¯¯¯¯¯¯¯¯ˆg(ξ)ˆΦ(ξ)dξ

and the norm . Here, for , the function is the Fourier transform of , defined as usual by

Moreover, is a reproducing kernel for . If satisfies

 (2.3) c1(1+|ξ|2)−κ≤ˆΦ(ξ)≤c2(1+|ξ|2)−κ,ξ∈Rd,

for some constants and , then we have from Corollary 10.13 in [19] that and

 (2.4) c1∥f∥Hκ(Rd)≤∥f∥NΦ(Rd)≤c2∥f∥Hκ(Rd),f∈Hκ(Rd).

Namely, the native space coincides with the Sobolev space with equivalent norm. Further, we mention that (2.4) and Corollary 10.25 in [19] implies

 (2.5) ∥I(g)∥Hκ(Rd)≤C∥g∥Hκ(Rd),∥g−I(g)∥Hκ(Rd)≤C∥g∥Hκ(Rd),g∈Hκ(Rd).

The so called Wendland kernel is a typical example of satisfying (2.3)–(2.5), which is defined as follows: for a given , set the function satisfying , , where

 ϕd,τ(r)=∫∞rrτ∫∞rτrτ−1∫∞rτ−1⋯r2∫∞r2r1max{1−r1,0}νdr1dr2⋯drτ,r≥0

with . For example,

 ϕ1,2(r) ≐max{1−r,0}5(8r2+5r+1), ϕ1,3(r) ≐max{1−r,0}7(315r3+285r2+105r+15), ϕ1,4(r) ≐max{1−r,0}9(5760r4+6795r3+3555r2+945r+105), ϕ2,4(r) ≐max{1−r,0}10(9009r4+9450r3+4410r2+1050r+105), ϕ2,5(r) ≐max{1−r,0}12(215040r5+283185r4+172620r3+59430r2+11340r+945),

where denotes equality up to a positive constant factor.

Then, and . Furthermore, satisfies (2.3)–(2.5) with .

3 Collocation method for Zakai equations

Let us describe the collocation methods for (2.1). In what follows, we always consider the version of , and thus by abuse of notation, we write for . Moreover, we restrict ourselves to the class of Wendland kernels described in Section 2.3. Suppose that the open rectangle for some is the set of points at which the approximate solution is to be computed. We take a set of grid points That is, we choose a set consisting of pairwise distinct points such that

 Γ={x1,…,xN}⊂(−R,R)d.

To construct an approximate solution of Zakai equation, we first take a set of time discretized points such that . The solution of the Zakai equation approximately satisfies

 u(ti,x)≃u(ti−1,x)+m∑k=0Lku(ti−1,x)ΔWk(ti),

where and . Since , we see

 u(ti,x)≃u(ti−1,x)+m∑k=0LkI(u(ti−1,⋅))(x)ΔWk(ti).

Thus, we define the function , a candidate of an approximate solution parametrized with a parameter , by

 (3.1) uh(t0,x) =u0(x),x∈Rd, uh(ti,x) =uh(ti−1,x)+m∑k=0LkI(uh(ti−1,⋅))(x)ΔWk(ti),x∈Rd,i=1,…,n.

With this definition, the -dimensional vector of the collocation points satisfies

 uh0 =(u0(x1),…,u0(xN))T, uhi =uhi−1+m∑k=0(AkA−1uhi−1)ΔWk(ti),i=1,…,n.

Here, we have set with . This follows from

 Lkuh(ti,xj)=N∑ℓ=1(A−1uhi)ℓLkΦ(x−xℓ)|x=xj=(AkA−1uhi)j.

To discuss the error of the approximation above, set and consider the Hausdorff distance between and , and the separation distance defined respectively by

 Δ1x=supx∈(−R,R)dminj=1,…,N|x−xj|,Δ2x=12mini≠j|xi−xj|.

Then suppose that , , , and are functions of .

Assumption 3.1.
1. The parameters , , , and satisfy , , , and as .

2. There exist and , positive constants independent of , such that

 |(A−1)ij|≤c1(Δ2x)dN

for with , and that

 c3(Δ2x)−(1+λ)d≤R1/2≤c4(Δ1x)−(τ−3/2)/d.
Remark 3.2.

Notice that . Thus the condition implies as .

Remark 3.3.

Suppose that is quasi-uniform in the sense that

 c5RN−1/d≤Δ2x≤c6RN−1/d,c′5RN−1/d≤Δ1x≤c′6RN−1/d

hold for some positive constants . In this case, a sufficient condition for which the latter part of Assumption 3.1 (ii) holds is

 c7N(1−1/(1+2d(1+λ))1d≤R≤c8N(1−d/(d+2τ−3))1d

with , for some positive constants and .

The approximation error for the Zakai equation is estimated as follows:

Theorem 3.4.

Suppose that Assumptions 2.1 and 3.1 hold. Suppose moreover that . Then, there exists such that

 maxi=1,…,nsupx∈(−R,R)dE[|u(ti,x)−uh(ti,x)|2]≤C(Δt+(Δ1x)2τ−3),h≤h0.

The rest of this section is devoted to the proof of Theorem 3.4. To this end, for every , put

 I(x)={i∈{1,…,N}:|x−xi|≤1}.

In what follows, denotes the cardinality of a finite set .

Lemma 3.5.

Suppose that . Then there exists such that

 #I(x)≤ν⌊q−d⌋,x∈(−R,R)d.
Proof.

Fix . Put for notational simplicity. It follows from the definition of that

 ⋃i∈I(x)Bq/2(xi)⊂B3/2(x).

Further, is disjoint. Indeed, otherwise, there exists , such that and . This implies , and so . Since we have assumed that ’s are pairwise distinct, we have . Denote by for the Lebesgue measure for . Then we have . Thus, for some that is independent of . ∎

Lemma 3.6.

Suppose that Assumption 3.1 (i) and hold. Then, there exists such that for any multi-index with and , we have

 ∥Dαf−DαI(f)∥L∞(−R,R)d≤C(Δ1x)τ+1/2−|α|1∥f∥Hτ+(d+1)/2(Rd),h≤h0.
Proof.

This result is reported in [19, Corollary 11.33] for more general domains. However, a simple application of that result leads to an ambiguity of the dependence of the constant on . Here we will confirm that we can take to be independent of .

Let with . Set and , . Then, and

 supy∈(−1,1)dminξ∈~Γ|ξ−y|=supy∈(−R,R)dminj=1,…,N∣∣∣xjR−yR∣∣∣=Δ1xR.

Since as and , we can apply [19, Theorem 11.32] to to obtain

 (3.2) |Dα~f(y)|≤C(Δ1x/R)τ+1/2−|α|1|~f|Hτ+(d+1)/2((−1,1)d),h≤h0

for some . Here, for an open set and ,

 |g|2H(d+1)/2+τ(O)=∑|α|=k∫O∫O|Dαg(x)−Dαg(y)|2|x−y|d+2sdxdy

with being the smallest integer that exceeds and . It is straightforward to see that

 Dα~f(y)=R|α|1(Dαf)(Ry),|~f|Hk+s((−1,1)d)=Rk+s−d/2|f|Hk+s((−R,R)d).

Substituting these relations into (3.2), we have

 (3.3) |Dαf(y)|≤C(Δ1x)τ+1/2−|α|1|f|Hτ+(d+1)/2((−R,R)d).

This and (2.5) yield

 ∥Dαf−DαI(f)∥L∞((−R,R)d) ≤C(Δ1x)τ+1/2−|α|1|f−I(f)|Hτ+(d+1)/2((−R,R)d) ≤C(Δ1x)τ+1/2−|α|1∥f∥Hτ+(d+1)/2(Rd).

Thus the lemma follows. ∎

Observe that for any ,

 I(f)(x)=N∑j=1(A−1f|Γ)jΦ(x−xj)=N∑j=1f(xj)Qj(x),x∈Rd,

where

 Qj(x)=N∑i=1(A−1)ijΦ(x−xi),j=1,…,N.

The following result tells us that the process of iterative kernel-based interpolation is stable, which is a key to our convergence analysis.

Lemma 3.7.

Suppose that Assumption 3.1 and hold. Then,

 suph>0supx∈(−R,R)dN∑j=1|DαQj(x)|<∞,|α|1≤2.
Proof.

Fix and . Set for simplicity. First consider the set

 J(~x):={j∈{1,…,N}:|i−j|>c2q−λd,i∈I(~x)}.

Then, by Assumption 3.1 (ii),

 |A−1ij|≤c1qdN,j∈J(~x),i∈I(~x).

This together with Lemma 3.5 leads to

 (3.4)

Now again by Assumption 3.1 (ii) and Lemma 3.5, there exists such that

 #{j:j∉J(~x)}≤~ν⌊R1/2⌋.

Then, by Kergin interpolation (see Kergin [12]) there exists a polynomial on with degree at most that interpolates at for all . This leads to

 (3.5) ∑j∉J(~x)|DαQj(~x)|=∑j∉J(~x)sgn(DαQj(~x))DαQj(~x)=∑j∉J(~x)p(xj)DαQj(~x).

By Corollary 3.9 in [19], this polynomial satisfies for . This together with Proposition 11.6 in [19] and Assumption 3.1 (ii) implies that

 max|α|1≤1supx∈(−R,R)d|Dαp(x)|≤C0

for some that is independent of and . In particular, is Lipschitz continuous on with Lipschitz coefficient . Hence the function

 ~p(x):=infy∈(−R,R)d{p(y)+C0|x−y|},x∈Rd,

is Lipschitz continuous on with the same Lipschitz coefficient and satisfies on . Further, for to be specified later, define the function on by

 ¯p(x)=ε−d∫Rd~p(y)ζ(x−yε)dy,

where is a -function such that for , for , and for for some . It is straightforward to verify that this function satisfies

 (3.6) |¯p(x)−~p(x)|≤C1ε,|Dα¯p(x)|≤C1ε−|α|1,|α|1≤ν1,x∈Rd

for some that is independent of and . Here . Then consider the function defined by , . With these modifications and in view of (3.4)–(3.6), we obtain

 N∑j=1|DαQj(~x)|