Skip to content
Publicly Available Published by De Gruyter February 7, 2018

Discrete Functional Analysis Tools for Some Evolution Equations

  • Thierry Gallouët EMAIL logo

Abstract

We present some discrete functional analysis tools for the proof of convergence of numerical schemes, mainly for equations including diffusion terms such as the Stefan problem or the Navier–Stokes equations in the incompressible and compressible cases. Some of the results covered here have been proved in previous works, coauthored with several coworkers.

MSC 2010: 65M08; 35Q30; 65M12

1 Introduction

The main purpose of this paper is to describe some tools, which have recently been developed, for the proof of convergence of numerical schemes for some evolution problems. We have two main examples in mind.

The first example is the case of the Navier–Stokes equations, with the three classical particular cases, namely the incompressible case, the case of variable density and the compressible case. We briefly describe the most difficult case, which is the compressible case. The main unknowns are the density ρ, the pressure p, and the velocity 𝒖. They are function of the time t which belongs to [0,T] for a given T>0, and of the space variable x which belongs to Ω, where Ω is a given bounded open connected set of d, d=2 or 3. With a given source term, f, in the momentum equation, and a perfect gas equation of state (EOS for short), the equations read

(1.1)tρ+div(ρ𝒖)=0,
(1.2)t(ρ𝒖)+div(ρ𝒖𝒖)-Δ𝒖+p=f,
(1.3)p=ργ.

The constant γ is larger than 1 for d=2 and 32 for d=3. To these equations we have to add some boundary condition, 𝒖=0 for instance, and an initial condition on ρ and 𝒖 (or on ρ𝒖).

Considering the most interesting case d=3, with a convenient discretization of these equations, for instance with the so-called Marker-And-Cell scheme, MAC scheme for short, see [13] for the seminal paper, it is possible to obtain an L(]0,T[,Lγ(Ω))-estimate on ρn and an L2(]0,T[,L6(Ω)3)-estimate on 𝒖n, where (ρn,pn,𝒖n)n is a sequence of approximate solutions (with a mesh size and a time step vanishing as n+). Then, up to a subsequence, it is possible to assume that, as n+, ρn weakly converges to some ρ in L2(]0,T[,Lγ(Ω)) and 𝒖n weakly converges to some u in L2(]0,T[,L6(Ω)3). A first difficulty is that this weak convergence is not sufficient to prove the convergence of the product ρn𝒖n to ρ𝒖 (and of ρn𝒖n𝒖n to ρ𝒖𝒖, note that, since γ>32, one has ρn𝒖n𝒖n bounded in L1(]0,T[,Lr(Ω)3×3) for some r>1). We tackle this difficulty in this paper (at least if γ2 for ρn𝒖n and γ3 for ρn𝒖n𝒖n). Another difficulty, not solved by the present paper, is to pass to the limit in the EOS (see, for instance, [11] for the stationary case).

The second example is the Stefan problem. The set Ω is a given bounded open set of d, d1. The function φ is a Lipschitz continuous function from to , nondecreasing and such that φ=0 on ]a,b[, for some real numbers a,b with a<b. We also assume that lim inf|s|+|φ(s)||s|>0. The unknowns are the functions ρ and u, from Ω×]0,T[ to , and the equation is

(1.4)tρ-Δu=0,quadu=φ(ρ).

Here also, we have to add a boundary condition, u=0, and an initial condition on ρ.

Numerous discretizations of this problem, for instance with Finite Volumes Methods (see [5]) or Gradient Discretization Methods (see [4]), lead to an L2(]0,T[,L2(Ω))-estimate on a sequence of approximate solutions (un,ρn)n. Then one has, up to a subsequence, that ρnρ weakly in L2(]0,T[,L2(Ω)) and unu weakly in L2(]0,T[,L2(Ω)), as n+.

With a mesh size and a time step vanishing as n+, it is quite easy to prove that tρ-Δu=0, in the distributional sense, but the difficulty is to prove that u=φ(ρ). The main step for proving u=φ(ρ) is to prove that

limn+0TΩρnun𝑑x𝑑t=0TΩρu𝑑x𝑑t.

Then, by applying, e.g., Minty’s trick, one obtains the desired identity. This will be detailed in Sections 3.1.2 (continuous setting) and 3.2.2 (discrete setting).

The common feature of these two examples is that we have two sequences, namely (ρn)n and (un)n, weakly converging in L2(]0,T[,Lq(Ω)) for a sequence and in L2(]0,T[,Lp(Ω)) for the other sequence, with p,q>1, 1p+1q=1. We denote by ρ and u the weak limits of these two sequences. Then the objective is to present some convenient additional hypothesis in order to prove that, with Q=Ω×]0,T[,

limn+Qρn(y)un(y)𝑑y=Qρ(y)u(y)𝑑y,

and to show that this additional hypothesis is satisfied in the case of the two examples described above.

For the sake of clarity, we will first consider the simple case corresponding essentially to the case of stationary equations (such as the stationary compressible Stokes equations). Then we will consider the more difficult case of evolution equations, including the Stefan problem or the Navier–Stokes equations, where the time derivative plays a different role than the space derivatives.

As usual, the discrete analysis follows closely the continuous analysis and for this reason we begin by the tools in the continuous setting.

This paper uses ideas which were developed in some previous papers written in collaboration with several authors, among whom Jérôme Droniou, Robert Eymard, Raphaèle Herbin and Jean-Claude Latché. Part of the results given in this paper were presented in [7].

2 Stationary Case

2.1 Continuous Setting

Let Ω be a bounded open set of d (d1), p,q>1, 1p+1q=1, and let (ρn)n and (un)n be sequences such that ρnρ weakly in Lq(Ω), unu weakly in Lp(Ω) as n+.

In general, one does not have, as n+,

(2.1)Ωρnun𝑑xΩρu𝑑x.

But, as it is well known, (2.1) is true if the sequence (un)n is bounded in H01(Ω) and p<2, with 2=+ if d=1 or 2, and 2=2dd-2 if d3 (note that 2=6 if d=3). This can be proved in two slightly different ways.

The first way is to use the compactness of the sequence (un)n (note that, in this paper, “compactness” always means “strong compactness”). Indeed, since H01(Ω) is compactly embedded in Lp(Ω), one has unu in Lp(Ω) and then (2.1) holds.

The second way is to use the compactness of the sequence (ρn)n in H-1(Ω). Since H01(Ω) is compactly embedded in Lp(Ω), by duality Lq(Ω) is compactly embedded in the dual space of H01(Ω), namely H-1(Ω). Here, one has identified, as usual, an element f of Lq(Ω) with the element of Lp(Ω) defined by gΩfg𝑑x. Then one has

(2.2)ρnρin H-1(Ω)as n+.

Since (un)n is bounded in H01(Ω), one has also (without extraction of a subsequence)

(2.3)unuweakly in H01(Ω)as n+.

Using (2.2) and (2.3), we obtain

Ωρnun𝑑x=ρn,unH-1,H01ρ,uH-1,H01=Ωρu𝑑x.

Remark 1.

Another way to use the compactness of the sequence (ρn)n in H-1(Ω) is to use, for all n, the solution of the Dirichlet problem with ρn as datum, namely the function wn such that

(2.4)wnH01(Ω),Ωwnvdx=Ωρnv𝑑xfor all vH01(Ω).

Using the weak convergence of ρn in Lq(Ω), one has wnw in H01(Ω), where w is the solution of

(2.5)wH01(Ω),Ωwvdx=Ωρv𝑑xfor all vH01(Ω).

Indeed, it is quite easy to prove that wn weakly converges to w in H01(Ω). Then, taking v=wn in (2.4) and using the convergence of ρn to ρ in H-1(Ω), one has

limn+Ωwnwndx=limn+Ωρnwn𝑑x=Ωρw𝑑x=Ωwwdx.

This proves the convergence of the H01(Ω)-norm of wn to the H01(Ω)-norm of w. This gives

limn+Ω(wn-w)(wn-w)𝑑x=limn+Ωwnwndx-2limn+Ωwnwdx+Ωwwdx=0,

which proves that wn converges to w in H01(Ω) (as n+). In order to conclude, we now use (2.3) and obtain

Ωρnun𝑑x=ΩwnundxΩwudx=Ωρu𝑑x,

since wn converges in L2(Ω)d to w and un weakly converges in L2(Ω)d to u.

2.2 Discrete Setting

The set Ω is always a bounded open set of d (d1) but adapted to a space discretization. For all n, one has ρnLn and unHn, where Ln and Hn are finite-dimensional spaces included in L(Ω). Let p,q>1 with 1p+1q=1. We assume that the sequence (un)n weakly converges to u in Lp(Ω) and the sequence (ρn)n weakly converges to ρ in Lq(Ω). We also assume, as in Section 2.1, that p<2 and we want (as in Section 2.1) a convenient additional condition giving (2.1).

In the case of conforming discretizations (such as Finite Element Methods), namely when HnH01(Ω), an easy condition is, as in Section 2.1, to assume that the sequence (un)n is bounded in H01(Ω). We are interested here by the more involved case when HnH01(Ω) but Hn is equipped with a norm, depending on n, “close” to the H01-norm. This is the case for Finite Volume Methods or Gradient Discretization Methods, see, for instance, [5], [6] and [4] for elliptic and parabolic equations and see [13] for the seminal paper on the Marker-And-Cell scheme for the Navier–Stokes equations. As in Section 2.1, we will present two different methods, using compactness of the sequence (un)n or compactness of the sequence (ρn)n.

2.2.1 Compactness of the Sequence (un)n

We first consider the case of classical Finite Volumes with admissible meshes, as in [5, Definition 9.1], see Figure 1. Roughly speaking for d=2 or 3 (the case d=1 is simpler), a classical Finite Volumes mesh consists in a family of disjoints “control volumes” which are open polygonal (if d=2) or polyhedral (if d=3) convex subsets of Ω whose closures cover Ω. The interface between two control volumes is contained in an hyperplane of d. For each control volume K, a point xK is given (it is not necessarily the center of gravity of K). Such a mesh is “admissible” if the segment joining the points xK and xL is orthogonal to the interface σ between the control volumes K and L (see Figure 1).

Figure 1 Here is an example of admissible mesh in the sense of [5].
Figure 1

Here is an example of admissible mesh in the sense of [5].

In this case, the space Hn is the space of functions which are constant on each control volume of the mesh n. We denote by hn the maximum of the diameter of the control volumes of the mesh n and we assume that limn+hn=0.

The space Hn is equipped with a discrete norm which mimics the H01-norm. We denote by int the interfaces which are in Ω and by ext the interfaces which are on the boundary of Ω. If σ is the interface between K and L, we denote by dσ the distance between xK and xL. If σ is an interface of K lying on the boundary of Ω, then dσ is the distance between xK and the boundary of Ω. Finally, mσ denotes the (d-1)-Lebesgue measure of σ. With these notation, the norm on Hn, which mimics the H01-norm, reads, if uK is the value of u in the control volume K,

u1,2,n2=σint,σ=K|Lmσdσ|uK-uLdσ|2+σext,σKmσdσ|uKdσ|2.

We assume that the sequence ((un)1,2,n)n is bounded. Then it is proven in [5, Theorem 9.3] that unu in L2(Ω). Now, we assume furthermore that dK,σdσ is bounded from below by a positive number (independently of K and n), where dK,σ is the distance between xK and xσ, the point belonging to σ and the segment joining xK and xL (σ=K|L). Then it is also proven in [5, Lemma 3.5] that the sequence (un)n is bounded in Lr(Ω) for any r<+ if d=1 or 2, and for r=6 if d=3. It is the so-called discrete Sobolev embedding. Then unu in Lp(Ω) since p<2. Finally, since ρnρ weakly in Lq(Ω) (and q=pp-1), we obtain (2.1) as desired.

The fact that unu in L2(Ω) (and therefore in Lp(Ω)) is a consequence of the Kolmogorov Compactness Theorem and of the following inequality, which is proven in [5, Lemma 9.3] with some C depending only on Ω and taking u=0 outside Ω,

(2.6)u(+η)-uL2(d)C|η|u1,2,nif uHn and ηd.

In order to prove (2.6) the admissibility of the mesh, namely the orthogonality between the segment joining xK and xL and K|L (the interface between K and L), is used in the proof of [5, Lemma 9.3]. Without this admissibility condition, we are not able to prove Inequality (2.6).

We consider now the case of non-admissible meshes so that we do not assume the orthogonality condition between the segment joining the points xK and xL and the interface σ between the control volumes K and L (see Fig. 1). But we assume (as before) that dK,σdσ is bounded from below by a positive number (independently of K and n). Then we can also conclude by using an inequality on the translates on u in the L1(d)-norm instead of the L2(d)-norm. This is done, for instance, in [6]. Indeed, the 1,2,n-norm is the same as above (see [6, norm defined by (43) or by (74) in Lemma 5.2]) and it is proven in [6, Lemma 5.5] that

(2.7)u(+η)-uL1(d)|η|du1,2,nif uHn.

Using again the Kolmogorov Compactness Theorem, we obtain the compactness of (un)n in L1(Ω) and therefore unu in L1(Ω). But, the estimate on un1,2,n also gives an estimate on un in Lr(Ω) for any r<+ if d=1 or 2, and for r=6 if d=3. This is again the discrete Sobolev embedding, given for instance in [6, Lemma 5.3]. Using this estimate, together with the L1(Ω)-convergence, we can deduce the convergence of un in Lp(Ω) (since p<2) and conclude as before that (2.1) holds.

2.2.2 Compactness of the Sequence (ρn)n

Although it is, of course, not necessary in this stationary case, we will try now to prove (2.1) by using compactness on ρn instead of compactness of un. The interest of this second method will appear in the evolution case where the time variable plays a different role than that of the space variables. However, we will consider here only the case p=q=2, the general case (p<2, q=pp-1) needs more work).

The main difficulty is that the bound on un is on a norm which depends on n (even if this norm is “close” to the H01-norm). The trick we propose here (presented in [7]) is to use the Sobolev space Hs(d) with some s]0,1[. We first recall the definition of the space Hs(d).

Definition 1.

For s0, Hs(d)={uL2(d):(1+||2)s2u^L2(d)}, where u^ is the Fourier transform of u. The norm for the space Hs(d) is given by

(2.8)us=(1+||2)s2u^L2(d).

With this norm, Hs(d) is a Hilbert space.

Then we set, for s>0,

Hs={uHs(d):u=0 a.e. on dΩ},

equipped with the Hs(d)-norm. The space Hs is a Hilbert space as a closed subspace of the Hilbert space Hs(d).

We begin with the case of admissible meshes. By using (2.6), it is possible to prove that the sequence (un)n is bounded in Hs for 0<s<12. We give the proof of this result in Lemma 4. Then, since Hs is a Hilbert space (with its natural norm), we have unu weakly in Hs. But, since s>0, we also have compactness of Hs in L2(Ω) since the Hs-norm of u allows a control on the translates of u, see Lemma 5. Then, by duality, identifying the space L2(Ω) with its dual space, one has compactness of L2(Ω) in (Hs). This gives ρnρ in (Hs) and we can conclude as in the continuous case, but with Hs instead of H01,

(2.9)Ωρnun𝑑x=ρn,un(Hs),Hsρ,u(Hs),Hs=Ωρu𝑑x.

In the case of non-admissible meshes, we can also conclude but we need a little more work. We have to work with (2.7) instead of (2.6). We recall that, for uHn, we also have an Lr(d)-estimate (in term of u1,2,n) on u for any r<+ if d=1 or 2, and r=6 if d=3. Then, taking r>2, we use the inequality, for all a>0 and ε>0,

a2εar+ε-1r-2a.

(Actually, this inequality follows from the fact that a2>εar implies ar-2<ε-1 which implies a<ε-1r-2. Therefore, for all ε>0 and all a>0, one has a2max{εar,ε-1r-2a}.) Taking ηd, a=|u(x+η)-u(x)| and integrating on d (recall that all functions are taken equal to 0 outside Ω), we obtain, for uHn,

u(+η)-uL2(d)2εu(+η)-uLr(d)r+ε-1r-2u(+η)-uL1(d).

It remains to choose ε=|η|r-2r-1. The Lr(d)-estimate on u (in terms of u1,2,n) and (2.7) give the existence of C depending only on Ω, r and the regularity of the mesh (to be more precise, we also use an homogeneity argument) such that

(2.10)u(+η)-uL2(d)C|η|r-22(r-1)u1,2,nif uHn and ηd.

In the case d=3, one takes r=6 and one has r-22(r-1)=25. It is now possible to conclude as in the case of admissible meshes. By using (2.10), the sequence (un)n is bounded in Hs for 0<s<r-22(r-1) (see Lemma 4). One has unu weakly in Hs. Here also, since s>0, we have compactness of Hs in L2(Ω) (see Lemma 5) and, by duality and identifying the space L2(Ω) with its dual space, compactness of L2(Ω) in (Hs). This gives ρnρ in (Hs) and we conclude with (2.9).

Remark 2.

To conclude this section, we can also remark that it is possible, similarly to the continuous case, to use the compactness of (ρn)n under the form nwnw in L2(Ω)d (and nunu weakly in L2(Ω)d) where n is a discretization of , wn is the solution of a discrete equivalent of (2.4) and w is the solution of (2.5). This method is used, for instance, in [4].

3 Evolution Case

3.1 Continuous Setting

Let Ω be a bounded open set of d (d1), T>0 and let (ρn)n and (un)n be sequences such that ρnρ weakly in L2(]0,T[,Lq(Ω)), unu weakly in L2(]0,T[,Lp(Ω)), with 1<p<2 and q=pp-1 (recall that 2=+ if d=1 or 2, and 2=2dd-2 if d3). As in Section 2.1, in general one does not have

(3.1)0TΩρnun𝑑x𝑑t0TΩρu𝑑x𝑑tas n+,

even if (un)n is bounded in L2(]0,T[,H01(Ω)) because there is no compactness of L2(]0,T[,H01(Ω)) in L2(]0,T[,L2(Ω)) (or L2(]0,T[,L2(Ω)) in L2(]0,T[,H-1(Ω))). (We take here p=q=2 for simplicity.) Of course, (3.1) holds if (un)n is bounded in H1(]0,T[,H01(Ω)) since there is compactness of H1(]0,T[,H01(Ω)) in L2(]0,T[,L2(Ω)). Similarly, (3.1) holds if (ρn)n is bounded in H1(]0,T[,L2(Ω)) (and (un)n bounded in L2(]0,T[,H01(Ω))) since there is compactness of H1(]0,T[,L2(Ω)) in L2(]0,T[,H-1(Ω)). (As usual, L2(Ω) is identified to its dual space.) But such hypotheses are quite strong and the objective is to obtain (3.1) using weaker hypotheses on (tun)n or (tρn)n. We consider, for instance, the two examples given in Section 1.

3.1.1 Compressible Navier–Stokes Equations

Let T>0, let Ω be a bounded open connected set of d, let d=2 or 3 and let γ such that γ>1 if d=2, γ>32 if d=3. For n, let (ρn,pn,𝒖n) be a (weak) solution of (1.1)–(1.3) with fn as datum instead of f and an homogeneous Dirichlet boundary condition on 𝒖n. We assume that fnf in L2(]0,T[,L2(Ω)). Under a convenient hypothesis on the initial condition (on density and velocity), it is possible to obtain an L(]0,T[,Lγ(Ω))-estimate on ρn and an L2(]0,T[,H01(Ω)d)-estimate on 𝒖n. Then, up to subsequences, it is possible to assume, as n+, that ρn weakly converges to some ρ in L2(]0,T[,Lγ(Ω)) and 𝒖n weakly converges to some 𝒖 in L2(]0,T[,H01(Ω)d).

In order to pass to the limit on the equation tρn+div(ρn𝒖n)=0 for proving that tρ+div(ρ𝒖)=0 (in the distributional sense), the main difficulty is to prove that, for any φCc(d×]0,T[),

(3.2)limn+0TΩρn𝒖nφdxdt=0TΩρ𝒖φdxdt.

This is not easy since we only have weak convergence of ρn and 𝒖n in Lebesgue spaces.

Let un be a component of the vector-valued function 𝒖n and consider that d=3 (the case d=2 is simpler). We do not have any space-time compactness of (un)n since we have no condition on tun. But, using the fact that (ρn)n is bounded in L2(]0,T[,Lγ(Ω)) and assuming that (𝒖n)n is bounded in L2(]0,T[,L6(Ω))d, we deduce, using γ>65, that (tρn)n is bounded in L1(]0,T[,W-1,1(Ω)) and this gives a compactness result for the sequence (ρn)n in L2(]0,T[,H-1(Ω)) by means of an adaptation of the well-known compactness results for evolution equation due to Lions, Aubin and Simon (see, for instance, Lions [15], Aubin [2] and Simon [16]). Indeed, one uses Theorem 1 with B=H-1(Ω), X=Lγ(Ω) and Y=W-1,1(Ω) (note that X is compactly embedded in B since γ>65, this is due, by duality, to the fact that H01(Ω) is compactly embedded in Lr(Ω) for r<6). Then one has ρnρ in L2(]0,T[,H-1(Ω)). Since 𝒖n𝒖 weakly converges in L2(]0,T[,H01(Ω)3), we then obtain (3.1) and, similarly, since φ is a regular function, (3.2). More precisely, for all ψCc(3×]0,T[)3, one has

0TΩρn𝒖nψ𝑑x𝑑t=ρn,𝒖nψL2(H-1),L2(H01)ρ,𝒖ψL2(H-1),L2(H01)=0TΩρ𝒖ψ𝑑x𝑑t,

which gives tρ+div(ρ𝒖)=0 (in the distributional sense).

The same difficulty appears in the momentum equation in order to pass to the limit on div(ρn𝒖n𝒖n). This will be possible if one proves that for all ψCc(3×]0,T[), one has

(3.3)limn+0TΩρnvnunψ𝑑x𝑑t=0TΩρvuψ𝑑x𝑑t,

where un and vn are two components of 𝒖n (and u, v the corresponding components of 𝒖).

We consider here also the case d=3 and we use the fact that γ>32. It gives that ρnun is bounded in L2(]0,T[,Lr(Ω)) with r=6γ6+γ>65. Since we already know that ρnunρu in the distributional sense, this gives that ρnunρu weakly in L2(]0,T[,Lr(Ω)). Using now the momentum equation, we prove that (t(ρnun))n is bounded in L1(]0,T[,W-1,1(Ω)) and this gives a compactness result for the sequence (ρnun)n in L2(]0,T[,H-1(Ω)) using Theorem 1 with B=H-1(Ω), X=Lr(Ω) and Y=W-1,1(Ω) (X is compactly embedded in B since r>65). Since vnv weakly converges in L2(]0,T[,H01(Ω)), we then obtain (3.3) and this allows us to pass to the limit in div(ρn𝒖n𝒖n).

3.1.2 Stefan Problem

Let T>0, let Ω be a bounded open set of d, d1. For n, let (ρn,un) be a (weak) solution of (1.4) with an homogeneous Dirichlet boundary condition on un and an initial condition on ρn (bounded in L2(Ω)). We recall that φ is a Lipschitz continuous function from to , nondecreasing, such that φ=0 on ]a,b[, for some real numbers a,b with a<b and lim inf|s|+|φ(s)||s|>0.

The natural estimates for this problem are an L2(]0,T[,L2(Ω))-estimate on the functions ρn and an L2(]0,T[,H01(Ω))-estimate on the functions un. Up to a subsequence, ρnρ weakly in L2(]0,T[,L2(Ω)) and unu weakly in L2(]0,T[,H01(Ω)). It is quite easy to pass to the limit in the equation tρn-Δun=0 and one obtains tρ-Δu=0, in the distributional sense, but it is less easy to prove that u=φ(ρ).

The way to prove that u=φ(ρ) consists in proving (3.1) and then to use the Minty trick given (in this simple case) in Lemma 1. In order to prove (3.1), one can use, as in Section 2.1, compactness on (un)n or compactness on (ρn)n (this was not the case in the case of the compressible Navier–Stokes equations described above).

The proof of compactness of (un)n in L2(]0,T[,L2(Ω)) (which leads to (3.1)) is not easy since there is no direct estimate on tun, but a trick due to Alt and Luckhaus [1] allows to obtain directly an estimate on the time-translates of un (without estimate on tun) in L2(]0,T[,L2(Ω)) and therefore gives the compactness of (un)n in L2(]0,T[,L2(Ω)). Then (3.1) holds and this (with the Minty trick) concludes the proof of u=φ(ρ).

Instead of proving compactness of (un)n in L2(]0,T[,L2(Ω)), it is perhaps simpler to prove compactness of the sequence (ρn)n in L2(]0,T[,H-1(Ω)). Indeed, since tρn-Δun=0 and since (un)n is bounded in L2(]0,T[,H01(Ω)), the sequence (tρn)n is bounded in L2(]0,T[,H-1(Ω)). This gives compactness of (ρn)n in L2(]0,T[,H-1(Ω)) using Theorem 1 with B=Y=H-1(Ω) and X=L2(Ω). Then one has

unuweakly in L2(]0,T[,H01(Ω)),
ρnρin L2(]0,T[,H-1(Ω)),

and this gives (3.1) since

0TΩρnun𝑑x𝑑t=ρn,unL2(H-1),L2(H01)ρ,uL2(H-1),L2(H01)=0TΩρu𝑑x𝑑t.

Here also, it remains to use the Minty trick in order to conclude that u=φ(ρ).

Remark 3.

As for the stationary case, for this Stefan problem, it is also possible to use the compactness of (ρn)n under the form wnw in L2(]0,T[,L2(Ω)d) (and unu weakly in L2(]0,T[,L2(Ω)d)), where wn is the solution of (2.4) and w is the solution of (2.5) (see, for instance, [4]).

Lemma 1 (Minty Trick).

Let Q be a bounded open set of RN, N1, and let φ be a continuous nondecreasing function from R to R. We assume that there exists CR+ such that

(3.4)|φ(s)|C|s|+Cfor all s.

(Note that the existence of C is true if φ is Lipschitz continuous.) Let (ρn)nN and (un)nN be sequences such that ρnρ weakly in L2(Q), unu weakly in L2(Q). We assume that un=φ(ρn) a.e., for all n, and that

limn+Qρn(y)un(y)𝑑y=Qρ(y)u(y)𝑑y.

Then u=φ(ρ) a.e.

Proof.

Since φ is nondecreasing, one has, for all ρ¯L2(Q),

0Q(ρn-ρ¯)(φ(ρn)-φ(ρ¯))𝑑y=Q(ρn-ρ¯)(un-φ(ρ¯))𝑑y.

Inequality (3.4) gives that φ(ρ¯)L2(Q). Then

0Qρnun𝑑y-Qρnφ(ρ¯)𝑑y-Qρ¯un𝑑y+Qρ¯φ(ρ¯)𝑑y.

Passing to the limit as n+ in this inequality yields

0Q(ρ-ρ¯)(u-φ(ρ¯))𝑑y.

Let ψCc(Q) and m>0. Taking ρ¯=ρ-1mψ in the previous inequality gives

0Q(u-φ(ρ-1mψ))ψ𝑑y.

The function (u-φ(ρ-1mψ))ψ converges a.e. to (u-φ(ρ)ψ, as m+, and is dominated by the function (|u|+C|ρ|+C|ψ|+C)ψ which belongs to L1(Q). Then, using the Dominated Convergence Theorem, we obtain, as m+,

0Q(u-φ(ρ))ψ𝑑y.

This inequality with ψ and -ψ gives 0=Q(u-φ(ρ))ψ𝑑y and then, since ψ is arbitrary, u=φ(ρ) a.e. in Q. ∎

In the previous examples, we use a compactness result for evolution equations which is essentially due to [15, 2, 16], a proof is given in [10]. We now give this theorem.

Theorem 1.

Let X, B, Y be three Banach spaces, XB, XY, such that the following hold:

  1. X is compactly embedded in B.

  2. For any bounded sequence (wn)n of X, if wn-wB0 and wnY0, then w=0.

Let T>0, 1p<+ and let (un)nN be a sequence such that

  1. (un)n is bounded in Lp(]0,T[,X),

  2. (tun)n is bounded in L1(]0,T[,Y).

Then there exists uLp(]0,T[,B) such that, up to a subsequence, unu in Lp(]0,T[,B).

Hypothesis (ii) of Theorem 1 on the spaces X, B, Y is perhaps a little bit curious. Indeed, we can distinguish two cases. A simple situation appears when BCY for some C>0 (this is the case, in particular, when Y=X and hypothesis (i) holds). Another situation is discussed in [16], when B is continuously embedded in Y, that is, YCB (for some C>0). Hypothesis (ii) in Theorem 1 covers a more general framework.

An example of Banach spaces X, B, Y satisfying the two hypotheses (i)(ii) is X=W01,1(Ω), B=L1(Ω), Y=W-1,1(Ω)=(W01,(Ω)), where, as usual, we identify an element of L1(Ω) with the corresponding linear form on W01,(Ω).

A main tool, crucial for the case where BCY is the following lemma essentially due to Lions [15], the proof of which is quite easy by contradiction.

Lemma 2.

Let X, B, Y are three Banach spaces, XB, XY, satisfying hypotheses (i)(ii) of Theorem 1. Then, for any ε>0, there exists Cε such that, for wX,

wBεwX+CεwY.

A particular case for which the conclusion of Lemma 2 is true, simpler than Lemma 2, is when B is a Hilbert space, X is a Banach space continuously embedded in B and Y=X but the norm in Y, denoted by Y, is the dual norm of X with respect to the scalar product of B, namely

uY=sup{(u|v)B:vX,vX1}.

Then, for any ε>0 and wX, one has

wBεwX+1εwY.

The proof is simple since

uB=(u|u)B12(uYuX)12εwX+1εwY.

In this simple case, the compactness of X in B is not needed for the conclusion of Lemma 2, but nevertheless this compactness is needed for Theorem 1 even in this case. We can also remark that in this simple case, the space Y is the space X with the norm Y and it is not necessarily a Banach space. Indeed, the fact that Y is a Banach space is never needed in Theorem 1, we only need Y to be a normed space since we can always complete this normed space in order to have a Banach space.

We recall the three spaces X, B, Y which were used in this section for the examples describes above (Navier–Stokes equations and Stefan problem).

For the compressible Navier–Stokes equations, in order to prove the compactness of the sequence (ρn)n in L2(]0,T[,H-1(Ω)) we choose B=H-1(Ω), X=Lγ(Ω) and Y=W-1,1(Ω). In order to prove the compactness of (ρnun)n in L2(]0,T[,H-1(Ω)), we choose B=H-1(Ω), X=Lr(Ω) (with r=6γ6+γ) and Y=W-1,1(Ω).

For the Stefan problem, we choose X=L2(Ω), B=Y=H-1(Ω).

Remark 4.

Another interesting example is the case of the incompressible Navier–Stokes equations for d=2 or 3 (see [3]), where we have to pass to the limit on div(𝒖n𝒖n). We set H={𝒖H01(Ω)d:div𝒖=0} (which is the natural space for the velocity) and we can use, for this case, Theorem 1 with p=2, B=L2(Ω), X=H, Y=H (with the identification of L2(Ω) with its dual space). It gives the compactness of the sequence of approximate velocities (𝒖n)n in L2(]0,T[,L2(Ω)d) which allows to pass to the limit in div(𝒖n𝒖n). In this example one has X continuously embedded in B (and, even, compactly embedded in B) but X is not dense in B so that B (which is identified with B) is not included in Y. However, one has HH (using the identification of B with B), which is the hypothesis needed in Theorem 1. This is a situation quite general. If E is a Banach space, continuously embedded in the Hilbert F, and if F is identified with F, then one has EE but one does not have FE except if E is dense in F (see, for instance, [10, Chapter 4]).

3.2 Discrete Setting

The objective of this subsection is to adapt the methods of Section 3.1 (in particular Theorem 1) to a discrete setting, in order to prove the convergence of numerical schemes. The set Ω is now a bounded open set of d adapted to a space discretization. The time interval is [0,T], T>0.

Let n; one has a time step kn such that T=knNn with some positive integer Nn and one has a space discretization which gives two finite-dimensional spaces Ln and Hn. As in the stationary case, Section 2.2, we consider the case where Ln and Hn are spaces of functions constant on control volumes defined by some meshes (which can be different for Ln and Hn as in interesting case of the MAC-scheme)

Assume that ρn and un are functions constant in time on each interval ](l-1)kn,lkn[, l=1,,Nn (but we could also assume that these functions are continuous in time and affine on each interval ](l-1)kn,lkn[, this will not change the results given hereafter). For all t](l-1)kn,lkn[, l=1,,Nn, one has

ρn(,t)=ρn(l)Lnandun(,t)=un(l)Hn.

Since we consider functions which are constant in time on each interval ](l-1)kn,lkn[, we have also to define discrete derivatives, namely for l{2,,Nn},

t,nun(,t)=t,kn(l)u=1kn(un(l)-un(l-1))for t](l-1)kn,lkn[,
t,nρn(,t)=t,kn(l)ρn=1kn(ρn(l)-ρn(l-1))for t](l-1)kn,lkn[,

and, for l=1,

t,nun(,t)=t,nρn(,t)=0for t]0,kn[.

We assume that limn+kn=0 and limn+hn=0, where hn is the maximum of the diameter of the control volumes of the meshes defining Ln and Hn. The sequences (ρn)n and (un)n weakly converge to ρ and u in L2(]0,T[,L2(Ω)) (we do not consider in this subsection the more general case of weak convergence in L2(]0,T[,Lq(Ω)) and L2(]0,T[,Lp(Ω)), with 1p+1q=1). We want, as in Section 3.1, a convenient additional condition giving (3.1). We will mimic the method of Section 3.1.

We begin with a discrete version of Lemma 2.

Lemma 3.

Let B be a Banach space and let (Bn)nN be a sequence of finite-dimensional subspaces of B. Let Xn and Yn be two norms on Bn such that, if (wnXn)nN is bounded, the following hold:

  1. Up to a subsequence, there exists wB such that wnw in B.

  2. If wn-wB0 and wnYn0, then w=0.

Then, for any ε>0, there exists C¯ε such that, for all nN and all wBn,

(3.5)wBεwXn+C¯εwYn.

Proof.

We first remark that Bn is a finite-dimensional subspace of B. Then, for all n and for all ε>0, there exists C¯ε satisfying (3.5) for all uBn. The problem is to find C¯ε independently of n.

We argue by contradiction. Assume that there exists ε>0 such that C¯ε does not exists. Then, for a subsequence of (Bn)n, still denoted (Bn)n, there exists (wn)n such that wnBn and

wnB>εwnXn+CnwnYn,

with limn+Cn=+.

By homogeneity, it is possible to assume that wnB=1. Then (wnXn)n is bounded and, by (i), up to a subsequence, wnw in B (so that wB=1). But wnYn0, so that w=0 (by ii), in contradiction with wB=1. ∎

We now give a discrete version of Theorem 1, the proof of which uses Lemma 3.

Theorem 2.

Let B be a Banach space, let 1q<+ and let (Bn)nN a family of finite-dimensional subspaces of B. Let Xn and Yn be two norms on Bn such that, if (wnXn)nN is bounded, the following hold:

  1. Up to a subsequence, there exists wB such that wnw in B.

  2. If wn-wB0 and wnYn0, then w=0.

The space Xn is the space Bn with norm Xn, the space Yn is Bn with norm Yn. Let T>0, kn>0, T=knNn, and let (un)nN be a sequence such that

  1. for all n, un(,t)=un(l)Bn for t((l-1)kn,lkn), l=1,,Nn,

  2. the sequence (un)n is bounded in Lq(]0,T[,Xn), that is to say that there exists C1>0 such that for all n, l=1Nnknun(l)XnqC1,

  3. the sequence (t,knun)n is bounded in L1(]0,T[,Yn), that is to say that there exists C2>0 such that for all n, l=2Nnknt,knun(l)YnC2.

Then there exists uLq(]0,T[,B) such that, up to a subsequence, unu in Lq(]0,T[,B).

See, for instance, [10, Theorem 4.51] for a proof of Theorem 2. Similar theorems are also in [3, 12].

Of course, the main example for the present paper is Bn=Ln or Bn=Hn. But it remains to choose B, Xn and Yn. We present these choices for the discretization of the two examples of Section 3.1, namely the compressible Navier–Stokes equations and the Stefan problem.

3.2.1 Navier–Stokes Equations

We begin with the case of the compressible Navier–Stokes equations as in Section 3.1.1 with a discretization using the MAC scheme, see [13, 3]. In this example, un is one component of the vector-valued function 𝒖n, which is the discrete velocity field, and each component of 𝒖n is a constant function on each control volume of its own mesh and for all time interval ](l-1)kn,lkn[ (it is the so-called staggered discretization). The discrete density ρn is a constant function on the control volumes of another grid, generally called the primal grid, denoted Ln, and for all time interval ](l-1)kn,lkn[.

We denote by H¯n the spatial space for the discrete velocity field. This space is equipped with a norm, denoted 1,2,n, which mimics the (H01)d-norm. Indeed, this norm contains, for each component of 𝒖n the norm 1,2,n defined in Section 2.2 in the case of admissible meshes.

The estimates that can be obtained on the approximate solutions mimics the ones of the continuous setting. One obtain an estimate in L(]0,T[,Lγ(Ω)) for ρn and an estimate on L2(]0,T[,H¯n) for 𝒖n, where H¯n is equipped with the norm 1,2,n (the way to obtain this estimate on 𝒖n is roughly explained below). In particular, this gives that the sequence (𝒖n)n is bounded in L2(]0,T[,L2(Ω)d). We also assume that γ2 so that the sequence (ρn)n is bounded in L2(]0,T[,L2(Ω)). Then we can assume, up to a subsequence, that ρnρ weakly in L2(]0,T[,L2(Ω)) and 𝒖n𝒖 weakly in L2(]0,T[,L2(Ω)d). We recall that we want to obtain (3.1) or, more generally, to prove the convergence of ρnun to ρu in a convenient Lebesgue space. In order to obtain (3.1), we will use, as in Section 3.1.1, a compactness result on ρn.

Taking, for instance, an implicit discretization of the mass balance (namely tρ+div(ρ𝒖)=0), we have with some convenient upwind discretization of div(ρ𝒖), for all l{2,,Nn},

(3.6)t,kn(l)ρn+divn(ρn(l)𝒖n(l))=0.

A crucial idea in the discretization of Navier–Stokes Equations with staggered grids is to deduce from (3.6) a discrete mass balance on the mesh, or on the meshes (in the case of the MAC scheme), associated to the velocity field. In the case of the MAC scheme, this discrete mass balance reads for each i=1,,d and for all l{2,,Nn},

t,kn(l)ρn,i+divn,i(ρn(l)𝒖n(l))=0,

where the function ρn,i is a reconstruction of ρn on the mesh associated to the corresponding component of 𝒖𝒏 and divn,i a discretization of div deduced from divn. We refer to [8, Section 3.3] for the first application of this idea, but not with the MAC scheme, and, for instance, [14] for the application of this idea with the MAC scheme. A main interest of this idea is that it gives, together with the discretization of the momentum equation, a kinetic energy balance and therefore an estimate on 𝒖n in L2(]0,T[,H¯n) where H¯n is equipped with the norm 1,2,n.

In order to apply Theorem 2, one takes Bn=Ln and Xn=L2(Ω) so that the sequence (ρn)n is bounded in L2(]0,T[,Xn). For the choice of Yn, we remark that a discrete integration by part leads to, for all l{2,,Nn},

Ωvt,kn(l)ρndx=-Ωvdivn(ρn(l)𝒖n(l))𝑑x=Ω(ρ𝒖)n(l)nvdxfor all vLn,

where n is a convenient discretization of . Then a natural choice of Yn is, for all wLn,

wYn=max{Ωwφ𝑑x:φLn,nφL(Ω)d+φL(Ω)=1}.

With this norm, one has t,kn(l)ρnYn(ρ𝒖)n(l)L1(Ω)d. Since (ρn)n is bounded in L2(]0,T[,L2(Ω)) and (𝒖n)n is bounded in L2(]0,T[,L2(Ω)d), one has a bound for (ρ𝒖)n in L1(]0,T[,L1(Ω)d) which gives a bound for t,knρn in L1(]0,T[,Yn).

We choose B=(Hs) for s such that 0<s<12 (see Section 2.2 for the definition of Hs) and we now prove that hypotheses (i)(ii) of Theorem 2 are satisfied. We already know that L2(Ω) is compactly embedded in (Hs) (see Section 2.2 and Lemma 5 in Section 4). This gives hypothesis (i). In order to prove hypothesis (ii), let wnLn such that (wn)n is bounded in L2(Ω), wnw in (Hs) and wnYn0. We want to prove that w=0. Let φCc(Ω). We define φn in Ln taking, for instance, the values of φ at the centers of the control volumes defining the space Ln. It is quite easy to prove that the L-norm of nφn is bounded by the L-norm of φ. Then one has

|Ωwnφndx|wnYn(nφnL(Ω)d+φnL(Ω))wnYn(φL(Ω)d+φL(Ω)).

This gives

limn+Ωwnφn𝑑x=0.

But, since (wn)n is bounded in L2(Ω) and wnw in (Hs), one has also wnw weakly in L2(Ω) (we use here the uniqueness of the limit in (Hs)). Then, since φnφ uniformly in Ω, one has

limn+Ωwnφn𝑑x=Ωwφ𝑑x.

We then conclude that Ωwφ𝑑x=0 for all φCc(Ω) and therefore that w=0 a.e. on Ω.

We can now apply Theorem 2, it gives ρnρ in L2(]0,T[,(Hs)). Indeed, the previous proof shows that it is also possible to prove that the sequence (t,knρn)n is bounded in L1(]0,T[,W-1,1(Ω)) and then to apply Theorem 2 with Yn=W-1,1(Ω), where W-1,1(Ω)=W01,(Ω).

We now conclude. Taking 0<s<12, we recall, see Section 2.2 and Lemma 4, that an estimate on v1,2,n gives an estimate on v in Hs. Then the sequence (𝒖𝒏)n is bounded in L2(]0,T[,(Hs)d) and therefore weakly convergent in this space (up to a subsequence). By the uniqueness of the weak limit in L2(]0,T[,L2(Ω)d), its limit is necessarily 𝒖 (and the convergence holds without extracting a subsequence) so that we finally obtain (3.1) for any component un of 𝒖n. Furthermore, it is possible to prove that for all φC(d×]0,T[,) one has

0TΩρn𝒖nnφdxdt=ρn,𝒖nφL2((Hs)),L2(Hs)+Rnρ,𝒖φL2((Hs)),L2(Hs)=0TΩρ𝒖φdxdt.

This gives, in particular, tρ+div(ρ𝒖)=0 in the distributional sense.

The previous proof gives also, for all ψC(d×]0,T[,d),

0Tρn𝒖nψ𝑑x𝑑t0Tρ𝒖ψ𝑑x𝑑t.

In the case d=3 (the most interesting case), one has an estimate on 𝒖n in L2(]0,T[,L6(Ω)d) (this is due to the discrete Sobolev inequality, see Section 2.1, proven in [5, Lemma 3.5]) and then an estimate on ρn𝒖n in L2(]0,T[,Lr(Ω)d). with r=6γ6+γ. This allows us to conclude that ρn𝒖nρ𝒖 weakly in L2(]0,T[,Lr(Ω)). If γ3, one has r2. Then it is also possible, if γ3, to pass to the limit (in the momentum equation) in the term div(ρn𝒖n𝒖n) as we did in Section 3.1.1. We use Theorem 2 with each component of 𝒖n and the corresponding reconstruction of ρn on the associated mesh (and this gives the space Bn). We choose B=(Hs) (with 0<s<12), the L2-norm for Xn and, for instance, the W-1,1(Ω)-norm for Yn. This gives the compactness in L2(]0,T[,(Hs)) for each component of the sequence (ρn𝒖n)n and we conclude using the boundedness in L2(]0,T[,Hs) of each component of the sequence (𝒖n)n.

3.2.2 Stefan Problem

For the case of the Stefan problem, the spatial discretization is the same for ρn and for un (with the notation of the beginning of Section 3.2, one has Ln=Hn). We recall that we are interested by the case where the space Hn is not included in H01(Ω) but Hn is equipped with a norm, depending on n, “close” to the H01-norm and denoted 1,2,n. We refer, for instance, to [5, 6, 4].

The discretization of the Stefan problem described in Section 1 gives that the couple (ρn,un) satisfy

t,knρn-Δnun=0,un=φ(ρn).

The discrete operator Δn from Hn to Hn is a convenient discretization of Δ as it is done, for instance, in [5, 6, 4]. We recall that φ is a Lipschitz continuous function from to , nondecreasing such that φ=0 on ]a,b[, for some real numbers a,b with a<b and lim inf|s|+|φ(s)||s|>0.

A natural estimate for this problem gives that the sequence (un)n is bounded in L2(]0,T[,Zn), where Zn is the space Hn with the norm 1,2,n. Then, using a discrete Poincaré estimate (see [5] for the case of admissible meshes or [6] for a more general case of meshes), one obtains that the sequence (un)n is bounded in L2(]0,T[,L2(Ω)). From this estimate on un, one deduces, with the hypothesis lim inf|s|+|φ(s)||s|>0, that (ρn)n is also bounded in L2(]0,T[,L2(Ω)). A consequence of these estimates is that we can assume, up to a subsequence, that ρnρ weakly in L2(]0,T[,L2(Ω)) and unu weakly in L2(]0,T[,L2(Ω)).

The weak convergence of ρn and un and some consistency property of the discretization of Δ lead to

tρ-Δu=0.

By using the estimate of un is L2(]0,T[,Zn), it is also classical to prove that uL2(]0,T[,H01(Ω)) (see [5] for the case of admissible meshes or [6] for a more general case of meshes).

Our purpose here is to give two ways to prove that u=φ(ρ). As in the continuous case, the first step is to prove (3.1) and then to conclude with the Minty trick (Lemma 1).

For proving (3.1), as in the continuous case (Section 3.1.2), we can use compactness of (un)n or compactness of (ρn)n.

The compactness of (un)n in not given by an application of Theorem 2 because we do not have any estimate on t,knun. However, it is sometimes possible to adapt the method of Alt–Luckhaus [1] to this discrete setting in order to obtain some estimates on the time-translates of un and then compactness of the sequence (un)n in L2(]0,T[,L2(Ω)). This gives (3.1) and then, using the Minty trick, u=φ(ρ). This way is, for instance, used in [5].

The second way is to prove some compactness on the sequence (ρn)n. We recall that the norm 1,2,n control the Hs-norm for some convenient s>0 (see Section 2.2 and Lemma 4 in Section 4). In the case of admissible meshes as in [5], we can take any s<12. in the case of more general meshes as in [6], we can take, for d=3, any s<25. Then one has unu weakly in L2(]0,T[,Hs) (since L2(]0,T[,Hs) is a Hilbert space). In order to prove (3.1), it suffices to prove that (ρn)n converges in L2(]0,T[,(Hs)) (as usual, L2(Ω) is identified with its dual space). We will prove this compactness on (ρn)n with Theorem 2 applied with

B=(Hs),Bn=Hn,Xn=L2(Ω),Yn=-1,2,n,

where -1,2,n is the dual norm of the norm 1,2,n, that is, for vHn,

v-1,2,n=max{Ωvw𝑑x:wHn,w1,2,n=1}.

By multiplying, for wHn, the equation t,knρn-Δnun=0 by w and using a discrete integration by part, the estimate on un in L2(]0,T[,Zn) gives an estimate on t,knρn in L2(]0,T[,Yn). In order to apply Theorem 2, it remains to verify hypotheses (i)(ii) of Theorem 2. Hypothesis (i) is due to compactness of L2(Ω) in (Hs) (which is a consequence, by duality, of the compact embedding of Hs in L2(Ω), Lemma 5 in Section 4). For proving hypothesis (ii), let (wn)n be a bounded sequence of L2(Ω) converging to w in (Hs) and to 0 for the Yn-norm. We first remark that wnw weakly in L2(Ω). Let ψCc(Ω). We can define ψn in Ln taking, for instance, the mean values of ψ on the control volumes defining the space Ln. It is possible to prove that ψn1,2,nCψH1(Ω), where C depends only the regularity of the mesh (see, for instance, [5, Lemma 9.4] for the case of admissible meshes). Then, under a regularity hypothesis on the sequence of meshes, one has

|Ωwnψndx|wnYnψn1,2,nCwnYnψH1(Ω).

This gives

limn+Ωwnψn𝑑x=0.

But, since wnw weakly in L2(Ω) and ψnψ uniformly in Ω, one has

limn+Ωwnψn𝑑x=Ωwψ𝑑x.

We then conclude that Ωwψ𝑑x=0 for all ψCc(Ω) and therefore that w=0 a.e. on Ω.

All the hypotheses of Theorem 2 are satisfied and we obtain the convergence of the sequence (ρn)n in L2(]0,T[,(Hs)). It gives (3.1) and we conclude with the Minty trick (Lemma 1) that u=φ(ρ).

Remark 5.

Here also, as in Section 2.2, it is possible to use the compactness of (ρn)n under the form nwnw in L2(]0,T[,L2(Ω)d) (and nunu weakly in L2(]0,T[,L2(Ω)d)), where n is a discretization of , wn is the solution of a discrete equivalent of (2.4) and w is the solution of (2.5). This method is used, for instance, in [4].

4 Appendix

We prove here that the space Nα,2(d) (which is a Nikolsky-space) is continuously embedded in the space Hs(d) for 0s<α1 (Lemma 4) and that the space Hs (with s>0 and Ω bounded) is compactly embedded in L2(d) (Lemma 5).

Definition 2.

Let 0<α1 and d1. The space Nα,2(d) is the set of elements u in L2(Ω) such that there exists C satisfying

u(+η)-uL2(d)C|η|αfor all ηd.

The norm in the space Nα,2(d) is defined by

uNα,2=uL2(d)+maxηd,η0u(+η)-uL2(d)|η|α.

With this norm, Nα,2 is a Banach space.

Lemma 4.

Let d1, 0s<α1. Then the space Nα,2(Rd) is continuously embedded in the space Hs(Rd) (see Definition 2.8).

Proof.

Let uNα,2(d) and C=uNα,2. For all ηd, one has

u(+η)^-u^L2(d)=u(+η)-uL2(d)C|η|α.

We recall that for all ηd, u(+η)^(ξ)=eiηξu^(ξ) a.e. Then

(4.1)d|eiηξ-1|2|u^(ξ)|2𝑑ξC2|η|2α.

Let e1,,ed be the canonical basis of d. Let j{1,,d} and t>0. With η=tej in (4.1) one obtains

d|eitξj-1|2t2α|u^(ξ)|2𝑑ξC2,

and then, for ε>0 and for all t>0,

d|eitξj-1|2t2α|u^(ξ)|2𝑑ξ1t1-εC21t1-ε.

Integrating this inequality between 0 and 1 and using the Fubini–Tonelli Theorem lead to

d|u^(ξ)|2(01|eitξj-1|2t2α+1-ε𝑑t)𝑑ξC2ε.

We use now the change of variable t|ξj|=τ so that (since |eiτ-1|=|e-iτ-1|)

(4.2)d|u^(ξ)|2|ξj|2α-ε(0|ξj||eiτ-1|2τ2α+1-ε𝑑τ)𝑑ξC2ε.

We set

aε=01|eiτ-1|2τ2α+1-ε𝑑τ.

Note that 0<aε<+ since |eiτ-1|2=(cos(τ)-1)2+(sin(τ))22τ2 and α1. With this definition of aε, one has

d|u^(ξ)|2|ξj|2α-ε𝑑ξ|ξj|<1|u^(ξ)|2𝑑ξ+1aε|ξj|1|u^(ξ)|2|ξj|2α-ε(0|ξj||eiτ-1|2τ2α+1-ε𝑑τ)𝑑ξ,

and then we deduce from (4.2) that

d|u^(ξ)|2|ξj|2α-ε𝑑ξuL2(d)2+C2εaεbεuNα,22,

with bε=1+1εaε. This gives

uα-ε22=d(1+|ξ|2)α-ε2|u^(ξ)|2𝑑ξdbεuNα,22

and concludes the proof. ∎

Lemma 5.

Let Ω be a bounded open set of Rd, d1, let s>0 and let

Hs={uHs(d):u=0 a.e. on dΩ}

be equipped with the norm s given in (2.8). Then Hs is compactly embedded in L2(Rd).

Proof.

Since Hs¯ is continuously embedded in Hs for 0s<s¯, we have only to prove the lemma for 0<s1.

Let 0<s1, we are going to prove that for all uHs and ηd,

(4.3)u(+η)-uL2(d)2us|η|s.

Inequality (4.3) gives that Hs is compactly embedded in L2(d). It is a consequence of the Kolmogorov Theorem in Lp-Spaces (see, for instance, [9, Theorem 8.16]). To this end, let ηd, η0. One has, as in Lemma 4,

u(+η)-uL2(d)2=u(+η)^-u^L2(d)2=d|eiηξ-1|2|u^(ξ)|2dξ.

We now remark that for all η, ξd, one has

(4.4)|eiηξ-1|24(1+|ξ|2)s|η|2s.

Indeed, since (for all x) |eix-1|2, estimate (4.4) is true if |η|1 and if |η|<1 but |ξ||η|1. If |η|<1 and |ξ||η|<1, estimate (4.4) is also true since |eix-1|22x2 and then |eiηξ-1|22|ξ|2|η|22|ξ|2s|η|2s (since s1), which gives (4.4). Using (4.4), we then obtain that

u(+η)-uL2(d)2d4(1+|ξ|2)s|η|2s|u^(ξ)|2dξ

and, finally, u(+η)-uL2(d)2us|η|s. This concludes the proof. ∎

5 Conclusion

This paper presents some tools useful for proving the convergence of numerical schemes. The main result is Theorem 2. Two applications are presented. The first one appears for the discretization of the Navier–Stokes equations and the second one for the Stefan problem. However, one may find other problems where the tools discussed here can be applied for proving the convergence of the numerical schemes. The ideas that are developed in this paper can probably also been extended to the analysis of high-order nonconforming methods (e.g., for the convergence analysis of discretization schemes for the Navier–Stokes equations, Leray–Lions problems, nonlinear elasticity models, etc.).

In the case of the discrete setting of the compressible Navier–Stokes equations when the equation of state does not allow an L2(]0,T[,L2(Ω))-estimate on ρ, it will be interesting to work with a sequence (ρn)n weakly convergent in L2(]0,T[,Lq(Ω)) with some 1<q<+ and a sequence (un)n weakly convergent in L2(]0,T[,Lp(Ω)) with 1p+1q=1 (we only consider in this paper the case p=q=2).

References

[1] H. W. Alt and S. Luckhaus, Quasilinear elliptic-parabolic differential equations, Math. Z. 183 (1983), no. 3, 311–341. 10.1007/BF01176474Search in Google Scholar

[2] J.-P. Aubin, Un théorème de compacité, C. R. Acad. Sci. Paris 256 (1963), 5042–5044. Search in Google Scholar

[3] E. Chénier, R. Eymard, T. Gallouët and R. Herbin, An extension of the MAC scheme to locally refined meshes: Convergence analysis for the full tensor time-dependent Navier–Stokes equations, Calcolo 52 (2015), 69–107. 10.1007/s10092-014-0108-xSearch in Google Scholar

[4] J. Droniou, R. Eymard, T. Gallouët, C. Guichard and R. Herbin, The gradient discretisation method: A framework for the discretisation and numerical analysis of linear and nonlinear elliptic and parabolic problems, preprint (2017), https://hal.archives-ouvertes.fr/hal-01382358/file/gdm.pdf. Search in Google Scholar

[5] R. Eymard, T. Gallouët and R. Herbin, Finite volume methods, Handbook of Numerical Analysis, North-Holland, Amsterdam (2000), 713–1020. 10.1016/S1570-8659(00)07005-8Search in Google Scholar

[6] R. Eymard, T. Gallouët and R. Herbin, Discretization of heterogeneous and anisotropic diffusion problems on general nonconforming meshes SUSHI: A scheme using stabilization and hybrid interfaces, IMA J. Numer. Anal. 30 (2010), no. 4, 1009–1043. 10.1093/imanum/drn084Search in Google Scholar

[7] T. Gallouët, Some discrete functional analysis tools, Finite Volumes for Complex Applications. VIII, Springer Proc. Math. Stat. 199, Springer, Cham (2017), 29–41. 10.1007/978-3-319-57397-7_3Search in Google Scholar

[8] T. Gallouët, L. Gastaldo, R. Herbin and J.-C. Latché, An unconditionally stable pressure correction scheme for the compressible barotropic Navier–Stokes equations, M2AN Math. Model. Numer. Anal. 42 (2008), no. 2, 303–331. 10.1051/m2an:2008005Search in Google Scholar

[9] T. Gallouët and R. Herbin, Mesure, intégration, probabilités, Ellipse, Paris, 2013. Search in Google Scholar

[10] T. Gallouët and R. Herbin, Equations aux dérivées partielles, lecture notes (2015), https://hal.archives-ouvertes.fr/cel-01196782. Search in Google Scholar

[11] T. Gallouët, R. Herbin, J.-C. Latché and D. Maltese, Convergence of the MAC scheme for the compressible stationary Navier–Stokes equations, preprint (2017), https://arxiv.org/abs/1607.01968v2; to appear in Math. Comp. 10.1090/mcom/3260Search in Google Scholar

[12] T. Gallouët and J.-C. Latché, Compactness of discrete approximate solutions to parabolic PDEs—application to a turbulence model, Commun. Pure Appl. Anal. 11 (2012), no. 6, 2371–2391. 10.3934/cpaa.2012.11.2371Search in Google Scholar

[13] F. Harlow and J. Welch, Numerical calculation of time dependent viscous incompressible flow of fluids with free surface, Phys. Fluids 8 (1965), 2182–2189. 10.1063/1.1761178Search in Google Scholar

[14] R. Herbin and J.-C. Latché, Kinetic energy control in the MAC discretization of the compressible Navier–Stokes equations, Int. J. Finite Vol. 7 (2010), no. 2, 6. Search in Google Scholar

[15] J.-L. Lions, Quelques méthodes de résolution des problèmes aux limites non linéaires, Dunod, Paris, 1969. Search in Google Scholar

[16] J. Simon, Compact sets in the space Lp(0,T;B), Ann. Mat. Pura Appl. (4) 146 (1987), 65–96. 10.1007/BF01762360Search in Google Scholar

Received: 2017-05-16
Revised: 2017-11-10
Accepted: 2017-11-21
Published Online: 2018-02-07
Published in Print: 2018-07-01

© 2018 Walter de Gruyter GmbH, Berlin/Boston

Downloaded on 26.4.2024 from https://www.degruyter.com/document/doi/10.1515/cmam-2017-0059/html
Scroll to top button