Space-time variational saddle point formulations of Stokes and Navier-Stokes equations

DOI:10.1051/m2an/2013124 www.esaim-m2an.org

SPACE-TIME VARIATIONAL SADDLE POINT FORMULATIONS OF STOKES AND NAVIER–STOKES EQUATIONS

Rafaela Guberovic

, Christoph Schwab

and Rob Stevenson

Abstract. The instationary Stokes and Navier−Stokes equations are considered in a simultaneously space-time variational saddle point formulation, so involving both velocitiesuand pressurep. For the instationary Stokes problem, it is shown that the corresponding operator is aboundedly invertiblelinear mapping betweenH1andH₂, both Hilbert spacesH1andH2being Cartesian products of (intersections of) Bochner spaces, or duals of those. Based on these results, the operator that corresponds to the Navier−Stokes equations is shown to mapH1intoH2, with a Fr´echet derivative that, at any (u, p)∈H1, is boundedly invertible. These results are essential for the numerical solution of the combined pair of velocities and pressure as function of simultaneously space and time. Such a numerical approach allows for the application of (adaptive) approximation from tensor products of spatial and temporal trial spaces, with which the instationary problem can be solved at a computational complexity that is of the order as for a corresponding stationary problem.

Mathematics Subject Classification. 35A15, 35Q30, 76D05, 76D07, 76M30.

Received October 5, 2012. Revised August 1, 2013.

Published online April 24, 2014.

1. Introduction

1.1. Background and motivation

The classical approach to the existence of weak solutions of the instationary, incompressible Navier−Stokes equations views these equations as an infinite-dimensional dynamical system (see,e.g., [23], Chap. III and the references there). In line with this view, most methods for thenumerical solutionof the instationary (Navier–) Stokes equations aretime marchingmethods: assuming that some approximate solution on time tis available, for a sufficiently small time incrementΔt >0, an approximate solution on timet+Δtis computed by solving a corresponding stationary problem.

Keywords and phrases. Instationary Stokes and Navier−Stokes equations, space-time variational saddle point formulation, well-posed operator equation.

∗ This work was supported in part by the European Research Council (ERC) under the FP7 programme ERC AdG 247277 (to CS).

1 Seminar for Applied Mathematics, ETH Z¨urich, CH 8092 Z¨urich, Switzerland. [email protected];

[email protected]

2 Korteweg-de Vries Institute for Mathematics, University of Amsterdam, P.O. Box 94248, 1090 GE Amsterdam, The Netherlands.[email protected]

Article published by EDP Sciences c EDP Sciences, SMAI 2014

R. GUBEROVICET AL.

Because of the generally lacking global smoothness of the solution, efficient numerical schemes have to be adaptive. With suitable time-marching schemes, it is possible to adapt both the spatial “mesh”, and the time stepΔtdepending ont. We refer to [2] for ana posteriori error analysis of such an approach.

Combined space-time adaptivity, where Δtis adapted also depending on the spatial location, are not easily accommodated by classical time stepping schemes, although some studies on local time stepping have appeared, seee.g.[8,9,17].

In any case, due to the character of time marching, it seems hard to guarantee a kind of quasi-optimal distribution of the ‘grid-points’ over space and time, and no mathematical results in this direction are presently known to us.

To develop an alternative for time marching schemes, in [4,18] we studied simultaneously space-time varia- tional formulationsof linearparabolicevolution equations. The operators defined by such variational formulations were shown to be boundedly invertible between a Hilbert space H₁ and the dual of another Hilbert space H₂, bothH₁and H₂ being Cartesian products of Bochner spaces or intersections of those.

By equipping these Bochner spaces with Riesz bases, beingtensor productsof temporal and spatial wavelet collections, the space-time variational problem was written as anequivalent well-posed, bi-infinite, symmetric positive definite matrix-vector system by forming normal equations. By running on this system an adaptive wavelet scheme, in its original form being proposed in [3], a sequence of approximations is produced in linear computational complexity that converges with the best possible nonlinear approximation rate from the basis, i.e., the rate of the so-called best N-term approximations.

Because of the application of tensorized wavelet collections in space and time, under mild (Besov) smoothness conditions the latter rate isequal(in some situations up to log-factors) to the best possible approximation rate for the solution of a corresponding stationary problem from the spatial wavelet basis,i.e., there is (hardly) any increase in the order of computational complexity as a consequence of the additional time dimension. Numerical results illustrating this fact are given in [4].

Besides the computational realization of the best possible nonlinear approximation rate, the latter property is a major advantage when the approximate solution is needed as function of simultaneously space and time, as it is the case for example in time-dependent optimal control problems, see [10]. Indeed, with time marching schemes this would require the availability of the approximate solutions simultaneously at all discrete times, requiring a huge amount of memory.

The results concerning the adaptive wavelet solution method generalize to simultaneously space-time vari- ational formulations of nonlinear parabolic evolution equations when they define a (two times continuously differentiable) mapping from H₁ into H₂, and the Fr´echet derivative at the solution is boundedly invertible between H₁ andH₂ (see [20]). The latter condition is satisfied for example for a semi-linear equation with a time-independent spatial operator.

Aiming at the application of space-time variational formulations to the incompressible instationary (Navier–) Stokesequations, there aretwopossibilities.

Thefirstone is to reduce these equations to problems for the divergence-free velocities only. Then the Stokes or linearized Navier−Stokes equations read as a linear parabolic evolution problems, and the aforementioned results concerning well-posed space-time variational formulations apply. The reduction to divergence-free velocities is also the standard approach followed in the literature for demonstrating existence and uniqueness of solutions (seee.g.[23], Chap. III).

In [21], we theoretically investigated the application of the adaptive wavelet scheme to the space-time vari- ational divergence-free velocities formulation of the instationary Stokes problem. Wavelets suitable for this formulation were constructed forrectangulardomains in [21,22].

1.2. This paper

The approach to tackle the instationary (Navier–) Stokes equations by a reduction to equations for the divergence-free velocities has the obvious disadvantage that no results for the pressure are obtained. Moreover,

the numerical solution of these equations by an adaptive wavelet scheme requires a divergence-free wavelet basis, that seems to be realizable in restricted settings only.

Therefore, in this paper as thesecondpossibility, we studysimultaneously space-time variational saddle point formulationsof the (Navier–) Stokes equations for the combined pair of velocities and pressure. For both free- slip and no-slip boundary conditions, we prove that the Stokes operator defined by this variational formulation isboundedly invertiblebetween a Hilbert spaceH₁and the dualH₂ of another Hilbert spaceH₂. In order to be able to arrive at this result, we have to assumeH²-regularity of the Poisson or of the stationary Stokes operator, which imposes certain smoothness or convexity conditions on the spatial domain.

In the space-times variational formulation of the present paper, both trial- and test-spaces H₁ and H₂ are Cartesian products of (intersections of) of Bochner spaces for velocities and pressure. Based on the results [18]

for the space-time variational formulations of parabolic evolution equations, the velocity components of the test- and trial-spaces are as expected, and so are the corresponding pressure components of either test- or trial- space. The space for the remaining pressure component, now being fully determined by the instationary Stokes operator, is less standard being the dual of the intersection of two Bochner spaces. In any case for polytopal spatial domains, countable tensor product wavelet bases can be constructed for these spaces, which are separable Hilbert spaces (although separability is not useda fortiori in the present paper).

To the best of our knowledge,well-posedness, i.e., bounded invertibility of the instationary Stokes operator for the combined pair of velocities and pressure has not been established before. Compare the discussion at the end of ([23], Chap. III, Sect. 1.5), where regularity of the pair of velocities and pressure is established only under additionalsmoothness conditions on the right-hand side.

With the spaces H₁ and H₂ as above, additionally it will be shown that the instationary Navier–Stokes operator maps H₁ into H₂ (for no-slip conditions on two- and three-dimensional domains, and for free-slip conditions on two-dimensional domains). A generalization of the results for the instationary Stokes operator to the linearized instationary Navier–Stokes operator – the difference being a lower order spatial differential operator– shows that the latter, at any (u, p)∈H₁, is a boundedly invertible operator betweenH₁ andH₂. A first consequence is that any solution (u, p)∈H₁of the instationary Navier−Stokes equations is locally unique.

Secondly, assuming that a sufficiently accurate initial approximation is available, it shows that the adaptive wavelet solver can be used to approximate the solution with thebest possible nonlinear approximation rate in space-time tensorized bases.

Finally, since also Lipschitz continuity of the instationary Navier–Stokes operator will be shown, using a fixed-point argument we conclude existence of a space-time variational Navier–Stokes solution in H₁, albeit under a small data hypothesis.

This paper is organized as follows: in Section 2, necessary and sufficient conditions are recalled for bounded invertibility of generalized linear saddle point problems. In Sections 3 and 4, these conditions are verified for space-time variational formulations of the instationary Stokes problem with free- and no-slip boundary conditions, respectively. In Section5, the aforementioned mapping properties of the instationary Navier–Stokes operator with homogeneous initial datum are verified.

Throughout, for positive constantsc₁, c₂,c₁ c₂we will mean that c₁ can be bounded by a multiple of c₂, independently of parameters on which c₁ and c₂ may depend. Obviously, c₁ c₂ is defined as c₂ c₁, and c₁c₂as c₁c₂ andc₁c₂.

2. Generalized saddle point problems

For reflexive Banach spacesU,V,P, andQ, and for bounded bilinear formsa:U×V →R,b:P×V →R, andc:U×Q→R, we consider the problem of finding (u, p)∈U×P that, for givenf ∈V, g∈Q, satisfy

a(u, v) +b(p, v) +c(u, q) =f(v) +g(q) (v∈V, q∈Q). (2.1) In this section, we collect sufficient and necessary conditions for the corresponding L : (u, p) → (f, g) ∈ L(U×P, V×Q) to be boundedly invertible. These conditions can already be found in [1], and a Hilbert space

R. GUBEROVICET AL.

setting, in [15]. Since some intermediate results will be used in the following sections, we have chosen to include the short arguments.

(Bv)(p) =b(p, v) = (Bp)(v) and (Cu)(q) =c(u, q) = (Cq)(u).

We recall that for a closed subspace Z of a Banach space X the polar set Z⁰ ⊂ X is defined by {f ∈ X : f(Z) = 0}.

Theorem 2.1. For given, bounded bilinear formsa,bandcas in (2.1), the variational problem (2.1)defines a boundedly invertible linear mappingU×P →V×Q if and only if the following three conditions are satisfied:

(i) for allf ∈(kerB), there exists a uniqueu∈kerC such that a(u, v) =f(v) (v∈kerB), (ii) for allg∈Q, there exists someu∈U such that c(u, q) =g(q)(q∈Q),

(iii) for allf ∈(kerB)⁰, there exists a uniquep∈P such that b(p, v) =f(v) (v∈V).

Proof. Suppose (i)–(iii) are satisfied, and let f ∈ V, g ∈ Q. By condition (ii), there exists a ¯u ∈ U with c(¯u, q) = g(q) (q ∈ Q). Condition (i) shows that there exists a u₀ ∈ kerC with a(u₀, v) = f(v)−a(¯u, v) (v∈kerB). Sou:=u₀+ ¯usolves

a(u, v) +c(u, q) =f(v) +g(q) (v ∈kerB, q∈Q).

Thisuis unique. Indeed, letu₁, u₂∈U be two solutions, then

a(u₁−u₂, v) =c(u₂−u₂, q) (v∈kerB, q∈Q).

By taking v = 0, we find u₂−u₁ ∈ kerC. Now by taking q = 0, we infer that u₁ = u₂ by (i). Finally, condition (iii) shows that there exists a unique p∈P that solves

b(p, v) =f(v)−a(u, v) (v∈V, q ∈Q), (2.2) so that (2.1) has a unique solution (u, p)∈U×P. An application of theopen mapping theoremshows that (2.1) defines boundedly invertible linear mappingU×P →V×Q.

Conversely, let (2.1) define a boundedly invertible linear mapping. Then condition (ii) follows easily. From A B

C 0

:U×P →V×Qbeing boundedly invertible, we have that ranB×{0}= A B

C 0

{0}×P

is closed, and thus that ranB is closed. By an application of theclosed range theorem, we conclude that (kerB)⁰ = ranB, which is (iii). Now let f ∈ (kerB). By an application of Hahn−Banach’s theorem extend it to f ∈ V, and takeg = 0. Then for the solution (u, p) of (2.1), it holds thatu∈kerC, anda(u, v) =f(v) (v ∈kerB). Now suppose that the last problem has two solutionsu₁, u₂ ∈kerC. Fori = 1,2, define p_i ∈ P as the solution of b(p, v) =f(v)−a(u_i, v) (v ∈V). Then both (u₁, p₁) and (u₂, p₂) solve (2.1), and we concludeu₁=u₂, which

completes the proof of (i).

Proposition 2.2. Having bounded b andc, conditions equivalent to (ii) and (iii) are, respectively, (ii) inf_0=q∈Qsup_{0=u∈U q}^c(u,q)

QuU >0, (iii) inf_0=p∈Psup_{0=v∈V p}^b(p,v)

PvV >0.

Proof. The equivalence of (ii) and (ii) follows from the equivalence of (a) and (e) in Lemma2.3stated below.

Another application of Lemma2.3shows that (iii)implies thatB ∈ L(P, V) is a homeomorphism onto (kerB)⁰, which implies (iii). Conversely, since ranB ⊂(kerB)⁰ by definition, (iii) implies that (kerB)⁰ = ranB and

thatB is injective, and so, by Lemma2.3, that (iii) is valid.

Lemma 2.3. For reflexive Banach spaces X and Y, and for T ∈ L(X, Y), the following statements are equivalent:

(a) inf_0=y∈Y sup_0=x∈Xx^{(T x)(y)}

XyY >0,

(b) T∈ L(Y, X)is a homeomorphism onto its range, (c) T injective andranT is closed,

(d) T injective andranT= (kerT)⁰, (e) T is surjective.

Proof. (a)⇔(b) and (b)⇒(c) follow easily.

(c)⇒(b) is a consequence of theopen mapping theorem.

(c)⇔(d) follows from theclosed range theorem.

(e)⇒(c): Since ranT is closed, the closed range theorem shows that ranT is closed, and that (kerT)⁰ = ranT =Y, so that, by an application of theHahn−Banach theorem, kerT=∅.

(c)⇒(e): Since ranT is closed, the closed range theorem shows that ranT = (kerT)⁰ =Y because T is

injective.

In the following, we shall the above existence results to verify the well-posedness of space-time variational saddle-point formulations of the Navier−Stokes equations. All the ensuing developments will require the pre- ceding results in the particular setting of Hilbert spaces.

3. The instationary Stokes problem with free-slip boundary conditions, as a well-posed operator equation

LetΩ⊂Rⁿ be a bounded Lipschitz domain. Given vector fields˜f on (0, T)×Ωandu₀onΩ, and functionsg on (0, T)×Ω, andg_i(1≤i≤n−1) on (0, T)×∂Ω, we consider the instationary inhomogeneous Stokes problem withfree-slipboundary conditions of finding for someν >0 a velocity fielduand corresponding pressurepthat

satisfy ⎧

⎪⎪

⎪⎪

⎪⎨

⎪⎪

⎪⎪

⎪⎩

∂u

∂t −νΔ_xu+∇_xp=˜f on (0, T)×Ω, div_xu=g on (0, T)×Ω, u·n= 0 on (0, T)×∂Ω,

∂u∂n·τ_i =g_i on (0, T)×∂Ω,1≤i≤n−1, u(0,·) =u₀ onΩ,

(3.1)

whereτ₁, . . . ,τ_n−1is an orthonormal set of tangent vectors.

Integrating the first equation against smooth vector fieldsv, that as function ofxhave vanishing normals at

∂Ω, and that as function oft vanish att=T, and by applying integration by parts in space and time, and by integrating the second equation against smooth functionsq, we arrive at a variational problem of the form (2.1),

where ⎧

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎨

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎩

a(u,v) =− ^T

0 Ω

u·^∂v_∂t dxdt+

T 0 Ω

ν∇xu:∇xvdxdt, b(p,v) =− ^T

0 Ω

pdivvdxdt, c(u, q) = ^T

0 Ω

qdivudxdt, f(v) =

T 0 Ω

˜f·vdxdt+

T 0 ∂Ω

n−1

i=1

(v·τi)g_idxdt+

Ω

u₀·v(0,·) dx, g(q) = ^T

0 Ω

g qdxdt.

(3.2)

R. GUBEROVICET AL.

We will need the following assumption on the domain Ω concerning H²-regularity of the Poisson problem with homogeneous Neumann boundary conditions.

Assumption 3.1. The bounded Lipschitz domainΩ⊂Rⁿ is such that for any h∈L_2,0(Ω) :=L₂(Ω)/R, the solutionu∈H¹(Ω)/Rof

Ω∇u· ∇vdx=

Ω

hvdx (v∈H¹(Ω)/R), is inH²(Ω), withu_H²_(Ω)h_L2(Ω).

This assumption is known to be satisfied whenΩis convex, or when it has aC²-boundary.

Theorem 3.2. With the spaces H˜²(Ω) :={p∈H²(Ω) : _∂n^∂p = 0on ∂Ω}/R,H¹(Ω) :={w∈H¹(Ω)ⁿ :w·n= 0 on∂Ω} and

U:=L₂(0, T;H¹(Ω)), P :=

L₂(0, T;L_2,0(Ω))∩H_0,{T¹ _}

0, T; ˜H²(Ω) , V:=L₂(0, T;H¹(Ω))∩H_0,{T}¹ (0, T;H¹(Ω)), Q:=L₂(0, T;L_2,0(Ω)),

and under Assumption3.1, the mappingL: (u, p)→(f, g)as in (2.1)with bilinear forms from (3.2)defines a boundedly invertible linear mapping U×P →V×Q.

(Here and in what follows, we denote for a Banach spaceB and a summability index 1≤p <∞byL_p(0, T;B) the space of strongly measurable functionsu: (0, T)→B such that (0, T)t→ u(t)B ∈L_p(0, T). As usual, dual spaces should be interpreted with respect to the identificationsL₂(Ω) L₂(Ω), L_2,0(Ω) L_2,0(Ω), or L₂(0, T;L_2,0(Ω)) L₂(0, T;L_2,0(Ω)), respectively. ForΓ ⊂ {0, T},H_0,Γ¹ (0, T) denotes the closure inH¹(0, T) of the set ofw∈C^∞(0, T)∩H¹(0, T) with suppw∩Γ =∅.)

To prove this theorem, in the following, we will verify the conditions of the abstract existence and uniqueness result, Theorem2.1.

The bilinear forms a: U×V → R,b : P ×V → R, and c : U×Q→R are bounded. For b, this follows from div ∈ L(H¹(Ω), L_2,0(Ω)) and div ∈ L(H¹(Ω),H˜²(Ω)), the latter, because of the density of D(Ω) in H¹(Ω), being equivalent to∇ ∈ L( ˜H²(Ω),H¹(Ω)). We conclude thatI⊗div_x∈ L(V, P), being equivalent to b:V×P →Ris bounded.

Knowing the boundedness of a, b, and c, next we verify the conditions (i)–(iii) of Theorem2.1. We start with condition (ii). Foru∈U,q∈Q, one hasc(u, q) =−_T

0

Ω∇_xq·udxdt. Since Ωis a bounded Lipschitz domain,

∇ ∈ L(L2,0(Ω),(H₀¹(Ω)ⁿ)) is a homeomorphism onto its range (3.3) ([14],cf.[23], Chap. 1, Rem. 1.4(ii)). By an application of Lemma2.3, this means that

inf_{0=q∈L2,0(Ω)}sup_0=u∈H1 0(Ω)ⁿ

Ωqdivudx

q_L2,0(Ω)u_H¹_(Ω)ⁿ >0, and so also that

inf_{0=q∈L2,0(Ω)}sup_0=u∈H1(Ω)

Ωqdivudx

q_L2,0(Ω)u_H¹_(Ω)ⁿ >0.

Since, additionally, (u, q)→

Ωqdivudxis bounded onH¹(Ω)×L_2,0(Ω), one has that∇ ∈ L(L_2,0(Ω),H¹(Ω)), and soI⊗ ∇x ∈ L(Q,U) are homeomorphisms onto their ranges by Lemma2.3. Knowing the boundedness of c:U×Q→R, the latter is equivalent to condition (ii) of Theorem2.1.

To show condition (i) of Theorem2.1, we give a characterization of the kernels of B:=I⊗div∈ L(V, P), C:=I⊗div∈ L(U, Q).

We set

H¹(Ω) :={u∈H¹(Ω) : divu= 0},H⁰(Ω) := clos_L2(Ω)nH¹(Ω),H⁻¹(Ω) :=H¹(Ω) (to be interpreted with respect to the identificationH⁰(Ω)H⁰(Ω)).

Since divH¹(Ω)⊂L_2,0(Ω), it holds that

kerC=L₂(0, T;H¹(Ω)). (3.4)

Lemma 3.3. With H˜⁻¹(Ω) := clos_H1(Ω)H¹(Ω), it holds that

kerB=L₂(0, T;H¹(Ω))∩H_0,{T¹ _}(0, T; ˜H⁻¹(Ω)).

Proof. Since both L₂(0, T;L_2,0(Ω)) and H_0,{T}¹

0, T; ˜H²(Ω)) are continuously embedded in, e.g., L₂(0, T; ˜H²(Ω)), these two spaces form a so-calledBanach couple, also known as a compatible couple of Banach spaces. Since moreover their intersection is dense in both spaces, the dual of their intersection is isomorphic to the sum of their duals, seee.g.([12], Chap. 1, Thm. 3.1),i.e.,

P =

L₂(0, T;L_2,0(Ω))∩H_0,{T¹ _}

0, T; ˜H²(Ω)

L₂(0, T;L_2,0(Ω)) +H_0,{T¹ _}

0, T; ˜H²(Ω) .

So v ∈ kerB if and only if v ∈ L₂(0, T;H¹(Ω))∩H_0,{T¹ _}(0, T;H¹(Ω)) and (Bv)(p) = 0 for all p ∈ L₂(0, T;L_2,0(Ω)) +H_0,{T¹ _}

0, T; ˜H²(Ω)

. This is equivalent to v ∈L₂(0, T;H¹(Ω)) and ((I⊗div)v)(p) = 0 for all p∈ L₂(0, T;L_2,0(Ω)), i.e., v ∈L₂(0, T;H¹(Ω)) by (3.4), together with v ∈ H_0,{T¹ _}(0, T;H¹(Ω)) and ((I⊗div)v)(p) = 0 for allp∈H_0,{T¹ _}

0, T; ˜H²(Ω)

. The second condition means that withN := ker(div ∈ L(H¹(Ω),H˜²(Ω))), it holds thatv∈H_0,{T¹ _}(0, T;N), so what is left to show is thatN = ˜H⁻¹(Ω).

By div ∈ L(H¹(Ω),H˜²(Ω)), N contains ˜H⁻¹(Ω). To prove that N ⊂ H˜⁻¹(Ω), it suffices to show the reversed inclusion for their polar sets

{u∈H¹(Ω) :u,w_L2(Ω)= 0,w∈N} ⊃ {u∈H¹(Ω) :u,w_L2(Ω)= 0,w∈H˜⁻¹(Ω)}. (3.5) The set on the right is contained in{u∈H¹(Ω) :u,w_L2(Ω)= 0,w∈ D(Ω),divw= 0}. As shown by De Rham ([7],cf.[23], Chap. 1, Prop. 1.1), a distributionuthat vanishes on all divergence-free test functions is a gradient of another distribution. If, additionallyu∈H¹(Ω), then necessarilyu∈ ∇H˜²(Ω).

The adjoint of div∈ L(H¹(Ω),H˜²(Ω)) is−∇ ∈ L( ˜H²(Ω),H¹(Ω)). The latter operator is bounded, and so closed, and it is an homeomorphism with its image, so which in particular is closed. Theclosed range theorem now implies that the space on the left in (3.5) is equal to ∇H˜²(Ω). This completes the proof.

Next, we show that, under conditions, the space ˜H⁻¹(Ω) in the characterization of kerBcan be replaced by H⁻¹(Ω).

Lemma 3.4. If theL₂(Ω)ⁿ-orthogonal projector ontoH⁰(Ω)is bounded onH¹(Ω), thenH˜⁻¹(Ω) =H⁻¹(Ω).

R. GUBEROVICET AL.

Proof. As shown in,e.g., ([23], Chapt. 1, Thm. 1.4), the closure of the set of divergence-free test functions in L₂(Ω)ⁿ is {u∈L₂(Ω)ⁿ : divu= 0,u·n= 0 on∂Ω}, and so this space is contained in H⁰(Ω). On the other hand, if for (u_k)_k ⊂H¹(Ω),u_k→uinL₂(Ω)ⁿ, and so inD(Ω), then divu= 0, and sou_k→uinH(div;Ω), in particular meaning thatu·n= lim_k→∞u_k·n= 0 on∂Ω. We conclude that

H⁰(Ω) ={u∈L₂(Ω)ⁿ : divu= 0,u·n= 0 on∂Ω}.

Let Π denote the L₂(Ω)ⁿ-orthogonal projector onto H⁰(Ω). From H¹(Ω) ⊂ H⁰(Ω)∩H¹(Ω), we have H¹(Ω)⊂imΠ|_H¹_(Ω). On the other hand, if Π is bounded on H¹(Ω), then imΠ|_H¹_(Ω) ⊂H⁰(Ω)∩H¹(Ω) = {u∈H¹(Ω) : divu= 0}=H¹(Ω),i.e.,

H¹(Ω) = imΠ|_H¹_(Ω). (3.6)

If, for some (f_n)_n⊂H¹(Ω),f_n →f inH¹(Ω), then, viewed as functionals onH¹(Ω),f_n →f inH⁻¹(Ω),i.e., H˜⁻¹(Ω)⊂H⁻¹(Ω).

Conversely, letf ∈H⁻¹(Ω). Then there exists a (f_n)_n⊂H¹(Ω) withf_n→finH⁻¹(Ω). For anyu∈H¹(Ω), f_n((I−Π)u) =f_n,(I−Π)u_L2(Ω)ⁿ =(I−Π)f_n,u_L2(Ω)ⁿ= 0. So, after trivially extendingf to a functional onH¹(Ω) by means off(im(I−Π)) = 0, by the boundedness ofΠonH¹(Ω) and (3.6) we havef−f_nH¹_(Ω) = sup_0=u∈H1(Ω)|(f−fn)(Πu)|

u_H1(Ω)n sup_0=u∈H1(Ω)|(f−fn)(Πu)|

Πu_H1(Ω)n =f−f_nH⁻¹_(Ω), orH⁻¹(Ω)⊂H˜⁻¹(Ω).

Theorem 3.5. Under Assumption3.1, we haveH˜⁻¹(Ω) =H⁻¹(Ω).

Proof. As shown in,e.g., ([23], Chap. 1, Thm. 1.4), one has the followingHelmholtz decomposition

L₂(Ω)ⁿ=H⁰(Ω)⊕^⊥∇(H¹(Ω)/R). (3.7)

The L₂(Ω)ⁿ-orthogonal projectorΠ ontoH⁰(Ω) is known as the Leray projector. Given u∈L₂(Ω)ⁿ, ∇z = (I−Π)uis the solution ofu− ∇z,∇w_L2(Ω)ⁿ= 0 (w∈H¹(Ω)/R).

Whenu∈H¹(Ω), thiszsolves the Poisson problemwithNeumann boundary conditions

⎧⎪

⎪⎨

⎪⎪

⎩

−Δz= divuonΩ,

∂z

∂n= 0 on∂Ω,

Ωzdx= 0.

Under Assumption3.1, this Poisson problem is H²(Ω)-regular, and so

∇z_H¹_(Ω)ⁿz_H²_(Ω)divu_L2(Ω)ⁿ u_H¹_(Ω)ⁿ,

i.e., I−Π and thusΠ is bounded onH¹(Ω). Now the result follows from Lemma3.4.

Using that onH¹(Ω)×H¹(Ω), (w,v)→

Ων∇w:∇vdxisboundedand satisfies aG˚arding inequality, we have the following result about the well-posedness of the variational formulation of the parabolic problem that results from the reduction of the instationary Stokes problem, with the homogeneous constraint div_xu= 0, to a system of equations for the divergence-free velocities only:

Theorem 3.6. With

X :=L₂(0, T;H¹(Ω)), Y :=L₂(0, T;H¹(Ω))∩H_0,{T¹ _}(0, T;H⁻¹(Ω)), A:=u→(v→a(u,v)) is a boundedly invertible linear mapping fromX toY.

Proof. The statement is equivalent to A being boundedly invertible fromY to X, which in turn, by making the change of variableT−t tot, is equivalent to the statement that

u→

v→ ^T

0 Ω

∂u

∂t ·vdxdt+

T 0 Ω

ν∇xu:∇xvdxdt

,

from L₂(0, T;H¹(Ω))∩H_0,{0}¹ (0, T;H⁻¹(Ω)) to L₂(0, T;H⁻¹(Ω)) is boundedly invertible. The boundedness of this mapping follows easily. The mapping corresponds to a variational formulation of a parabolic problem with homogeneous initial datum in the space of divergence-free velocities. The boundedness of the inverse is a consequence of (u,v) →

Ων∇xu : ∇xvdx being bounded and coercive on H¹(Ω)×H¹(Ω), and it is shown,e.g., as a special case of ([18], Thm. 4.1), where a possible inhomogeneous initial condition is imposed

weakly.

The characterizations of the kernels given by (3.4), Lemma3.3, and Theorem3.5, together with Theorem3.6 imply condition (i) of Theorem2.1.

Condition (iii) of Theorem2.1is equivalent to (iii) which, by Lemma2.3, is equivalent to B=I⊗div_x:L

L₂(0, T;H¹(Ω))∩H_0,{T}¹ (0, T;H¹(Ω)), L₂(0, T;L_2,0(Ω))∩H_0,{T¹ _}(0, T; ˜H²(Ω))

is surjective. (3.8)

Note that since I ⊗ div_x is not injective, to prove (3.8) it is generally not sufficient to show that I ⊗ div_x is surjective both as mapping in L(L2(0, T;H¹(Ω)), L₂(0, T;L_2,0(Ω))) and as mapping in L(H_0,{T¹ _}(0, T;H¹(Ω)), H_0,{T¹ _}(0, T; ˜H²(Ω))).

Below, we will construct a mapping div⁺ with div◦div⁺=I, such that

div⁺∈ L(L_2,0(Ω),H¹(Ω)), div⁺∈ L( ˜H²(Ω),H¹(Ω)). (3.9) Since, consequently,I⊗div⁺_x is a right-inverse for the mapping from (3.8), this will imply the surjectivity of the latter mapping.

We define div⁺:g→uby ⎧

⎨

⎩

u+∇p=f onΩ, divu=g onΩ, u·n= 0 on∂Ω,

wheref = 0, or, more precisely, by its variational formulation to find (u, p)∈L₂(Ω)ⁿ×H¹(Ω)/Rsuch that

Ω

u·v+

Ω∇p·v+

Ω∇q·u=f(v) +g(q) ((v, q)∈L₂(Ω)ⁿ×H¹(Ω)/R). (3.10) From the fact that∇ ∈ L(H¹(Ω)/R, L₂(Ω)ⁿ) is a homeomorphism onto its range, applications of Lemma2.3, Proposition 2.2, and Theorem 2.1 confirm the well-known fact that this variational problem, for general f ∈ L₂(Ω)ⁿ, defines a boundedly invertible operator fromL₂(Ω)ⁿ×H¹(Ω)/Rto its dual.

Under Assumption3.1, forf ∈H¹(Ω) andg∈L_2,0(Ω), the solutionpof −Δp=g−divf onΩ,

∂p

∂n = 0 on∂Ω,

is in ˜H²(Ω), and u := f − ∇p ∈ H¹(Ω). We infer that the mapping (f, g) → (u, p) defined by (3.10) is in L(H¹(Ω)×L_2,0(Ω),H¹(Ω)×H˜²(Ω)), and so, by considering the adjoint and using the symmetry of the left- hand side of (3.10) in (u, p) and (v, q), it is inL(H¹(Ω)×H˜²(Ω),H¹(Ω)×L_2,0(Ω)). We conclude that (3.9) and thus condition (iii) of Theorem2.1are valid.

R. GUBEROVICET AL.

Having verified all conditions of Theorem 2.1, the proof of Theo- rem3.2is now completed.

Finally in this section, we derive well-posedness of an alternative variational formulation. The variational formulation (3.2) of our Stokes problem (3.1) was derived by applying integration by parts over time. This has the advantage that the initial conditionu(0,·) = u₀ enters the variational formulation as a natural boundary condition, i.e., in the right-hand side. In any case for a homogeneous initial condition, i.e., u(0,·) = 0, an alternative variational formulation is obtained by not applying integration by parts over time. It reads as a variational formulation of the form (2.1), where

⎧⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎨

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎩

a(u,v) =

T 0 Ω

∂u

∂t ·vdxdt+

T

0 Ων∇xu:∇xvdxdt, b(p,v) =− ^T

0 Ω

pdivvdxdt, c(u, q) = ^T

0 Ω

qdivudxdt, f(v) =

T 0 Ω

˜f·vdxdt+

T 0 ∂Ω

n−1

i=1

(v·τi)g_idxdt g(q) = ^T

0 Ω

g qdxdt.

(3.11)

Theorem 3.7. With

U:=L₂(0, T;H¹(Ω))∩H_0,{0}¹ (0, T;H¹(Ω)), P:=L₂(0, T;L_2,0(Ω)),

V:=L₂(0, T;H¹(Ω)), Q:=

L₂(0, T;L_2,0(Ω))∩H_0,{0}¹

0, T; ˜H²(Ω) ,

and under Assumption 3.1, the mapping L: (u, p)→(f, g)as in (2.1) with bilinear forms from (3.11) defines a boundedly invertible linear mapping U×P →V×Q.

Proof. Denoting the spacesU,V, P, and Q, and operator L from Theorem 3.2 here asU,¯ V, ¯¯ P, ¯Q, and ¯L, and defining (Rw)(t, x) =w(T−t, x), we have

(L(u, p))(v, q) = (L(Rv,¯ −Rq))(Ru,−Rp) = (L¯(Ru,−Rp))(Rv,−Rq).

From L¯ ∈ L(V¯ ×Q,¯ U¯×P¯) being a boundedly invertible, andRU=V,¯ RP = ¯Q, RV=U, and¯ RQ= ¯P,

the proof is completed.

4. The instationary Stokes problem, with no-slip boundary conditions, as a well-posed operator equation

LetΩ⊂Rⁿ be a bounded Lipschitz domain. Given vector fields˜f on (0, T)×Ωandu₀ onΩ, and a function g on (0, T)×Ω, we consider the instationary inhomogeneous Stokes problem withno-slip boundary conditions to find the velocitiesuand pressurepthat satisfy

⎧⎪

⎪⎪

⎨

⎪⎪

⎪⎩

∂u

∂t −νΔxu+∇xp=˜f on (0, T)×Ω, div_xu=g on (0, T)×Ω, u= 0 on (0, T)×∂Ω, u(0,·) =u₀ onΩ.

By integrating the first equation against smooth vector fields v, that as function ofxvanish at∂Ω, and that as function oftvanish att=T, and by applying integration by parts in space and time, and by integrating the second equation against smooth functionsq, and by applying integration by parts, we arrive at a variational problem of the form (2.1), where

⎧⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎨

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎪

⎪⎩

a(u,v) =− ^T

0 Ω

u·∂v

∂t dxdt+

T 0 Ω

ν∇_xu:∇_xvdxdt, b(p,v) =

T 0 Ω

v· ∇pdxdt, c(u, q) =− ^T

0 Ω

u· ∇qdxdt, f(v) =

T 0 Ω

˜f·vdxdt+

Ω

u₀·v(0,·) dx, g(q) = ^T

0 Ω

g qdxdt.

(4.1)

Remark 4.1. With ˆH²(Ω) := {p ∈ H²(Ω) : ∇p = 0 on∂Ω}/R, following the exposition in Section 3, an obvious choice for the spacesU, P andV, Qfor the variablesu, p andv, q, would be

L₂(0, T;H₀¹(Ω)ⁿ),

L₂(0, T;L_2,0(Ω))∩H_0,{T¹ _}

0, T; ˆH²(Ω) , L₂(0, T;H₀¹(Ω)ⁿ)∩H_0,{T¹ _}(0, T;H⁻¹(Ω)ⁿ), L₂(0, T;L_2,0(Ω)),

where H⁻¹(Ω) = H₀¹(Ω) with respect to the identification L₂(Ω) L₂(Ω). With this choice, the resulting spaceH¹(Ω) of divergence free spatial functions would read as{u∈H₀¹(Ω)ⁿ: divu= 0}, with, as in Section3, its closureH⁰(Ω) inL₂(Ω)ⁿbeing{u∈L₂(Ω)ⁿ: divu= 0,u·n= 0 on∂Ω}. Now when following the analysis from Section 3, the problem is that the L₂(Ω)ⁿ-orthogonal projector onto H⁰(Ω), i.e., the Leray projector, does not preserve no-slip boundary conditions, and therefore is not bounded onH₀¹(Ω)ⁿ.

In view of Remark 4.1, later, in Theorem 4.3, we will select trial- and test-spaces by making a shift in smoothness indices for the spatial variables.

Before that, first we study the stationary Stokes problem with homogeneous Dirichlet boundary conditions of findingu∈H₀¹(Ω)ⁿ,p∈L_2,0(Ω) such that, for givenf ∈H⁻¹(Ω)ⁿ, g∈L_2,0(Ω),

Ω

ν∇u:∇vdx−

Ω

pdivvdx+

Ω

qdivudx=f(v) +g(q) ((v, q)∈H₀¹(Ω)ⁿ×L_2,0(Ω)). (4.2) Since, using that Ω is a bounded Lipschitz domain, ∇ ∈ L(L_2,0(Ω), H⁻¹(Ω)ⁿ) is homeomorphism onto its range, see (3.3), applications of Lemma2.3, Proposition2.2, and Theorem2.1confirm the well-known fact that this variational problem defines a boundedly invertible mapping betweenH₀¹(Ω)ⁿ×L_2,0(Ω) and its dual.

We will need the following assumption on the domainΩaboutH²(Ω)ⁿ×H¹(Ω)-regularity of this stationary Stokes problem.

Assumption 4.2. The bounded Lipschitz domainΩ⊂Rⁿis such that for anyf ∈L₂(Ω)ⁿ,g∈H¹(Ω)/R, the solution (u, p) of (4.2) belongs to H²(Ω)ⁿ×H¹(Ω) andu_H²_(Ω)ⁿ+p_H¹_(Ω)f_L2(Ω)ⁿ+g_H¹_(Ω).

This assumption is known to be satisfied for domainsΩinR² orR³that either have aC²-boundary, or that are convex with a piecewise smooth boundary. See [5,11] for the two- or three-dimensional case, respectively.

R. GUBEROVICET AL.

Theorem 4.3. With

U:=L₂(0, T;L₂(Ω)ⁿ), P :=

L₂(0, T;H¹(Ω)/R)∩H_0,{T}¹

0, T; (H¹(Ω)/R) , V:=L₂(0, T; (H₀¹(Ω)∩H²(Ω))ⁿ)∩H_0,{T¹ _}(0, T;L₂(Ω)ⁿ),

Q:=L₂(0, T;H¹(Ω)/R),

and under Assumption4.2, the mappingL: (u, p)→(f, g)as in (2.1)with bilinear forms from (4.1)defines a boundedly invertible linear mapping U×P →V×Q.

(Here, dual spaces should be interpreted with respect to the identifications L_2,0(Ω) L_2,0(Ω), or L₂(0, T;L_2,0(Ω)) L₂(0, T;L_2,0(Ω)), respectively.)

Before proving this theorem, we give some more auxiliary results dealing with the stationary problem.

Lemma 4.4. It holds that

ker(∇∈ L(L2(Ω)ⁿ,(H¹(Ω)/R))) ={u∈L₂(Ω)ⁿ: divu= 0,u·n= 0on ∂Ω}:=H˘⁰(Ω), ker(∇ ∈ L(H₀¹(Ω), L_2,0(Ω))) ={u∈H₀¹(Ω)ⁿ: divu= 0}:=H˘¹(Ω), ker(∇ ∈ L((H²(Ω)∩H₀¹(Ω)¹)ⁿ, H¹(Ω)/R)) ={u∈(H²(Ω)∩H₀¹(Ω))ⁿ: divu= 0}:=H˘²(Ω).

Proof. Since in the last two cases∇=−div by definition, we only have to verify the first statement,i.e., that N := ker(∇∈ L(L₂(Ω)ⁿ,(H¹(Ω)/R))) =H˘⁰(Ω).

Foru∈H˘⁰(Ω),p∈H¹(Ω)/R, one has

Ω∇p·udx= 0,i.e.,H˘⁰(Ω)⊂N. To prove the reversed inclusion, we have to show that

{u∈L₂(Ω)ⁿ :u,w_L2(Ω)= 0,w∈N} ⊃

u∈L₂(Ω)ⁿ:u,w_L2(Ω)= 0,w∈H˘⁰(Ω)

. (4.3)

The set on the right is part of{u∈L₂(Ω)ⁿ:u,w_L2(Ω)= 0,w∈ D(Ω),divw= 0}. As shown by De Rham ([7], cf.[23], Chap. 1, Prop. 1.1), a distributionuthat vanishes on all divergence-free test functions is a gradient of another distribution. If, additionallyu∈L₂(Ω)ⁿ, then necessarilyu∈ ∇(H¹(Ω)/R).

Since∇:H¹(Ω)/R→L₂(Ω)ⁿ) is bounded, and so is closed, and moreover since this mapping is an homeo- morphism with its image, which is therefore closed, the closed range theoremtells us that the space on the left

in (4.3) is equal to∇(H¹(Ω)/R), which completes the proof.

It holds thatH˘²(Ω)→H˘¹(Ω)→H˘⁰(Ω) with dense embeddings. Fori∈ {1,2}, we setH˘⁻ⁱ(Ω) := (H˘ⁱ(Ω)), where this dual space should be interpreted with respect to the identification (H˘⁰(Ω)) H˘⁰(Ω).

The stationary Stokes problem (4.2) with g = 0 can be reduced to a problem involving divergence-free velocities only. It reads as findingu∈H˘¹(Ω) that solves

Ω

ν∇u:∇vdx=f(v)

v∈H˘¹(Ω)

. (4.4)

Under Assumption4.2, forf ∈L₂(Ω)ⁿ we haveu∈H˘²(Ω) with

u_H²_(Ω)ⁿf_L2(Ω)ⁿ. (4.5)