Variational Methods for

Computational Fluid Dynamics

François Alouges and Bertrand Maury

2

Contents

1 Models for incompressible fluids 5
1.1 Some basics on viscous fluid models . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 Mathematical framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Free in-/outlet conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.1 Prescribed normal stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Some typical problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.1 Flow around an obstacle . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.2 Some simple fluid-structure interaction problems . . . . . . . . . . . . . . 16

2 Navier-Stokes equations, numerics 19
2.1 The incompressibility constraint, Stokes equations . . . . . . . . . . . . . . . . . 19
2.2 Time discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Generalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.2 Going back to Navier-Stokes equations . . . . . . . . . . . . . . . . . . . . 23
2.2.3 The method of characteristics . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Projection methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3 Moving domains 27
3.1 Lagrangian standpoint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Arbitrary Lagrangian Eulerian (ALE) point of view . . . . . . . . . . . . . . . . 29
3.2.1 Actual computation of the domain velocity . . . . . . . . . . . . . . . . . 31
3.2.2 Navier–Stokes equations in the ALE context . . . . . . . . . . . . . . . . 32
3.3 Some applications of the ALE approach . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.1 Water waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.2 Falling water . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.4 Implementation issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4 Fluid-structure interactions 35
4.1 A non deformable solid in a fluid . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Rigid body motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Enforcing the constraint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.4 Yet other non deformable constraints . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

5 Duality methods 41
5.1 General presentation of the approach . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2 The divergence-free constraint for velocity fields . . . . . . . . . . . . . . . . . . . 44
5.3 Other types of constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3

. . . . . . . . . . . . . .3 oscillating cylinder . . . . . . .2. . .3 Some stable finite elements for Stokes equations . . . . 63 8. 63 8. . . . . . . . . . . . . . . . . .2 Rotating cylinder . . . . . . . 45 5. . . . . . . . . . . . . 50 6. . . . . . . . . . . .1 Which mesh for which Reynolds number ? . . . . . . . . . . . . . . . . .1 Glass forming process . . . . . . . . .3 Airfoil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 7 Complements 57 7. . . . 64 8. . . . . . . . . . . . . . 68 A. . . . . . . . . . . . . . . .3. . . .2 Elliptic problems and Finite Element Method . . .2 Energy balance . . . . . 65 8.1 Benchmark . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Prescribed flux . .3. . . . . . . . . . . . . . . . 64 8. . . . . . .1 Mixed finite element formulations . . . . . . . . . . . . .1 Hilbert spaces. . . . . . . . . . . .2. . .1 General . . . . . . 46 6 Stokes equations 49 6. . . . . . . . . . . .3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64 8. . . . . . . . . . . . . . . . . . . . . . . .4. . . . . . . .3 Moving objects . . . . . . . . . . . . . . . . . . . . . 65 A Appendix 67 A. . . . . . . . . . . . . . . . . . . . . 64 8. . . . . . . . . . .4 Free surface . . . . . . . . . . . . . . . . . . . . . . . . .1 Free fall . . . . . . . . . . . . . . . . . .1. . . . . . 63 8. . . 57 7. . . .4. . . . . . . . . . . . . . . . . . . .3. . . . . . . . . . . . . . . . . . . . . . . . . . .1. . . . . 49 6. . . . . .2 Fluid structure interaction . . . . . . . . . . 64 8. . . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Differential calculus. . . . . . . . . . . 64 8. . . . . . . . . . . . . . . . . . . . .2 Drag and lift of an obstacle . . . . .3 Inclined plane . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. . . 64 8. . . . . 64 8. . . . . .2 Density waves . . . .4. . . . . . . . . . .1 Obstacle problem . . . . . . . . . . . . . . . . .1 A general estimate without inf-sup condition . . . . . . . . Integration by parts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Visualization . . . . . . . . . . . . . . . . . . . 63 8. . .4 CONTENTS 5. . . . . . . . . . .3. . . . . . . . . . . . . . . . . . .1 Estimation of interaction forces . . .5 Free convection . 65 8. . . .4 Surface tension .4. . . . . . . . 60 7. . . . . . . . .2 Estimates with the inf-sup condition . .1. . . . . . . .5 Rayleigh-Taylor instability . . . . . . . . . . 64 8. . . . . . . . .1. . . . . . . . . . . 58 7. . . . . . . . . . . . . . . . . . . .3. . .4 Waves on the beach . . . . . . . . . . . . . . . . . . . . . . . . 60 8 Projects 63 8. . . . . . . . . . . . . . . . . .3 Prescribed mean velocity . . . . . . . . . . . . . . . 46 5. . . . . . . . . . . . minimization under constraint . . . . . . . . . . . . . . . . . . . 52 6. . . 64 8. . . 71 . . . 67 A. . . . . . . . . .

and if the mapping n 7−→ F(n) is linear. ∂ω ω The left-hand side of Eq. x a point in the domain occupied by the fluid. Since all integrals are now expressed as volume integrals. We denote by Fε (n) the force exerted on Dε (n) by the fluid located on the side of Dε (n) to which n points. 5 . If Fε (n)/ε goes to F(n) when ε goes to 0. t) the local density. and the net force exerted on ω by the fluid located outside ω. Z Z σ·n = ∇ · σ.1 Some basics on viscous fluid models Definition 1 (Stress tensor) We consider a fluid. Newton’s law write Z d ρu = sum of external forces. which writes. dt ω(t) ω(t) ∂t ∂ω(t) and the last term can be transformed onto a volume integral: Z Z ρu(u · n) = ∇ · (ρu ⊗ u) . The motion of a fluid that admits a stress tensor follows an evolution equation that can be formalized in a very general way. Consider a fluid element ω(t). we deduce that the integrand vanishes identically.1 now writes Z Z Z d ∂(ρu) ρu = + ρu(u · n). We denote by ρ = ρ(x.j . by u the velocity field. ∂ω(t) ω(t) where u ⊗ u stands for the symmetric matrix (ui uj )i. 1. we call stress tensor at x the tensor σ that represents this linear mapping: F(n) = σ · n. and perpendicular to n. n a unit vector.1) dt ω(t) R The right-hand side is the sum of the body force contribution ω f . (1.e a set of particle which we follow in their motion. i. and by f a body force acting on the fluid (like gravity). and Dε (n) a disc (or a segment is the space dimension is 2) centered at x. according to Definition 1. and since the fluid element is arbitrary. with area ε (length ε in dimension 2). possibly in motion.Chapter 1 Models for incompressible fluids 1.

∂t Model 2 (Force balance equation for an non-inertial fluid) Whenever inertia can be considered as negligible. t) and velocity field u(x. MODELS FOR INCOMPRESSIBLE FLUIDS Model 1 (General evolution equation for an inertial fluid) We consider a fluid with density ρ(x. where Id is the identity tensor. where (u · ∇) u is such that d X ∂ui ((u · ∇) u)i = uj . (1. According to the previous models and definitions. t). i. there exists a scalar field p. Newton’s law is replaced by instantaneous force balance for all fluid element. t). Then d d X ∂f f (X(t)) = uj . ∂t In the case where the fluid is homogeneous (ρ is uniform) and incompressible (the divergence of the velocity field is 0). which expresses −∇ · σ = f .e. an inertial perfect fluid is such that −∇ · σ = ∇ · (p Id) = ∇p. submitted to a body force f . dt ∂xj j=1 .2) ∂t Mass conservation writes ∂ρ + ∇ · (ρu) = 0. and we obtain Euler’s equation ∂ (ρu) + ∇ · (ρu ⊗ u) + ∇p = f . called pressure. ∂xj j=1 Remark 1 Consider the velocity field u as stationary. Denote by X(t) the path of a particle (advected by the velocity field). Definition 2 (Perfect fluid) A fluid is said to be perfect if it admits a stress tensor which is diagonal. and consider a scalar function f defined on the domain. one has ∇ · (ρu ⊗ u) = ρ (u · ∇) u. We assume that this fluid admits a stress tensor σ(x. such that σ(x) = −p Id. Newton’s law writes ∂ (ρu) + ∇ · (ρu ⊗ u) − ∇ · σ = f .6 CHAPTER 1.

t) ≈ u(x. t) +  +  · (y − x). and a last contribution that accounts for local deformations (symmetric part of the velocity gradient). | {z }  2 | {z } | {z } 2  T ranslation Rotation Def ormation The motion of a material segment xy can therefore be decomposed into 3 contributions: a translational motion at the local velocity.1: Local decomposition of a two-dimensional velocity field Newtonian fluids. Real fluids are characterized by a resistance to deformation.1. a rotational motion (skew symmetric part of the velocity gradient). t) + ∇u(x.1 for an illustration of this decomposition for a two dimensional velocity field. 1.1. See Fig. t) · (y − x)    ∇u − t ∇u ∇u + t ∇u  = u(x. The velocity field can be expanded in the neighborhood of x as u(y. We consider a fluid element moving in the neighborhood of t 7→ x(t). SOME BASICS ON VISCOUS FLUID MODELS 7 + = + Figure 1. Definition 3 (Strain tensor) .

respectively. as d X ∂ui ∇·u= = 0. MODELS FOR INCOMPRESSIBLE FLUIDS Considering a fluid moving according to the velocity field u. called viscosity. and homogeneous (the density is constant). ∂t ρU L . i. and a scalar field p = p(x. L the length scale (a characteristic dimension of the domain).j = ui . it can be taken out of the time derivative. Laplacian) operator with respect to the adimen- sional space variable. t) (pressure). Besides. U L T Denoting by ∇⋆ (resp. we obtain ∂u⋆ µ ⋆ + (u⋆ · ∇⋆ ) u⋆ − ∆ ⋆ u⋆ + ∇ ⋆ p ⋆ = f ⋆ . ∆⋆ ) the gradient (resp. such that  σ = 2µT − p Id = µ ∇u + t ∇u − p Id . up to a diagonal contribution: Definition 4 (Newtonian fluid) A fluid is said newtonian is there exists a positive parameter µ. and T = L/U the associated time scale. We introduce the dimensionless quantities u x t u⋆ = . Let U be the order of magnitude of the velocity. 2 2 ∂u1 ∂u1 ∂u2  + 2 ∂x2 ∂x2 ∂x2 The simplest real fluid model is built by considering that the stress tensor is proportional to this strain tensor. incompressible. ∂xi i=1 we have ! d X ∂uj ∇ · (u ⊗ u) = ∇ · (ui uj )i. Model 3 (Incompressible Navier-Stokes equations) An inertial fluid that is newtonian. such that the volume of a fluid element remains constant: ∇ · u = 0. the strain tensor is defined as   ∂u1 ∂u1 ∂u1 t 2 + ∇u + ∇u 1 ∂x1 ∂x2 ∂x2  T= =   (in 2d). Since ρ is constant. t⋆ = . in the non-inertial and inertial settings. and which is subject to a body force f . ∂xi i=1 1≤j≤d The latter quantity expresses the derivative of the velocity in its own direction. We furthermore assume that the fluid is homogeneous: ρ is supposed to be uniform and constant.e. x⋆ = . Adimensional form of the Navier-Stokes equations. We consider now incompressible fluids.8 CHAPTER 1. and it is denoted by (u · ∇) u. The incompressible Stokes and Navier-Stokes equations are straightforward expressions of the Newtonian character. follows the incompressible Navier-Stokes equations     ρ ∂u + (u · ∇) u − µ∆u + ∇p = f  ∂t   ∇ · u = 0.

so that in 3d.3) by v and 1 This strong assumption is sometimes ruled out. and f ⋆ = f L/(ρU 2 ) the adimensional forcing term.20).2. y) such that (u. have been introduced to limit the cost of numerical simulation by discretizing at a scale larger than η. . and involve extra unknown pertaining to those phenomena (like the kinetic energy k associated to smaller scales in the k − ε model).1.e. taking the scalar product of the first equation of (1.e. i.3) ∇·u = 0 If we consider the situation where the fluid is enclosed in a domain delimited by physical. with ∂u/∂t = 0. but allow a non zero tangential velocity. 1. Exercise 1 One considers a circular domain Ω centered at 0. it is usually considered1 that the fluid sticks to the wall. The quantity Re = ρU L/µ is called the Reynolds number. It is usually considered that dissipation occurs at some space scale η (smaller size of the eddies). it can be considered that inertial effects are negligible. This remark concerns Direct Navier-Stokes simulations (called DNS). Remark 2 The description of phenomena associated to flows at high Reynolds number goes far beyond the scope of this course. p) is a solution to (3. Show that the rigid velocity field   −y u=ω x is a steady solution to the incompressible Navier-Stokes equations. Those approaches rely on assumptions regarding what happens at scales smaller than the mesh size. Other methods. and the understanding of turbulence (which refers to situations where complex motions occur over a large large a space scales) still raises many open questions. which expresses as homogeneous Dirichlet boundary conditions u = 0 on the boundary ∂Ω. subject to a body force f in a regime where inertia can be neglected. Let us simply say here that high Reynolds flows can be pictured as containing eddies over a large range of sizes. impermeable walls. L/η is of the order Re3/4 . If one aims at discretizing the space in order to “capture” (i. so that Newton’s law for fluid elements simplifies to local force balance Model 4 (Incompressible Stokes equations) A fluid that is newtonian and incompressible. When this number is small compared to 1. According to Kolmogorov’s theory. or methods based on the so-called k − ε model. and ω a given angular velocity. This formula gives a precious indication in the context of numerical simulations. the total number of vertices is Re9/4 . follows the incompressible Stokes equations ( −µ∆u + ∇p = f (1. represent on the mesh) the smallest eddies. like Large Eddy Simulation (LES) method. It quantifies the relative impor- tance of inertia compared to viscous effects. that there exists a pressure field p(x. the number of mesh vertices in each direction scales at Re3/4 . In some situations it is in particular more relevant to use the so-called Navier conditions.2 Mathematical framework A variational formulation is obtained by considering a test function v which vanishes on the boundary of the domain Ω. which still preserve the impervious character of the wall. starting to the global size of the observed phenomenon. down to much smaller scales. MATHEMATICAL FRAMEWORK 9 where p⋆ = p/(ρU 2 ) is the adimensional pressure.

∇ · u = 0).4)   Z     q∇ · u = 0 ∀q ∈ X. and this non- linearity makes the problem much more difficult to solve. and which vanish on the boundary ∂Ω. X = L20 (Ω)   Z Z Z   µ ∇u : ∇v − p∇ · v = f ·v ∀v ∈ V Ω Ω Ω (1. page 68. and  Z  2 2 L0 (Ω) = q ∈ L (Ω) .2) (page 67). The problem may now be expressed in an appropriate mathematical sense as follows    V = H01 (Ω)d . Ω where H01 (Ω)d is the Sobolev space of vector fields the components of which are L2 function with square integrable gradient. so that this indeterminacy shall not be met. since only the gradient of the pressure appears in (1. which still presents unresolved issues. q=0 .3). Note that. . in the three-dimensional setting. any field v ∈ V is such that Z Z ∇·u= u · n = 0. in the most interesting case from the modeling standpoint. Ω Proposition 1 Let f be given in L2 (Ω)d . Proof : Problem (1.4) is the saddle-point formulation of a minimization problem of the type (A. as we shall deal with domain with inlet and outlets. this technical problem will actually disappear. the incompressibility constraint can be expresses similarly in a weak form. which comes from the fact that Newton’s law is written for a fluid element in motion. Ω Ω Ω Considering now a scalar test function q. which leads to Z Z Z µ ∇u : ∇v − p∇ · v = f · v. Because of the advective term (u · ∇) u.e. i. only uniqueness up to an additive constant may be expected2 . For this reason. where u minimizes Z Z µ 2 v 7−→ J(v) = |∇v| − f · v. Weak solutions (in the spirit of Definition 5 below) can be defined. which is the (opposite of the) divergence operator is surjective. Ω ∂Ω so that B maps V onto scalar fields with zero means. This enlightens the necessity to consider the set of pressure with zero mean. it is sufficient to prove that B. and there existence under very general assumptions is known since the 2 In the context of lung modeling. In particular. Problem (1.4) admits a unique solution (u. As V consists of velocity fields which vanish on the boundary.e. the Navier-Stokes equations are nonlinear. where K = ker B is the set of divergence free fields. MODELS FOR INCOMPRESSIBLE FLUIDS integrating by part. 2 Ω Ω among all those fields in H 1 (Ω) which are divergence-free (i. p) ∈ V ×X. we shall prescribe an extra constraint on this variable: the mean value over the domain must be zero. A huge literature is dedicated to this problem.10 CHAPTER 1. there is no general proof of exis- tence of smooth solutions defined globally in time. We refer to [10] for a proof of this property. By Proposition 12.

which ends up in the system as internal energy (the temperature increases). T ). time derivative of the kinetic energy (stored by the system). H) = v . Energy balance is obtained by multiplying the momentum equation by the velocity itself. (0. and by integrating by part. which expresses how the different types of energy interact. The two standpoints are related. dt Ω 2 Ω Ω Time integration over an interval (0. t)|2 dx < +∞ . let us start with the so-called energy balance. Considering now a divergence-free velocity field u(x. 0)| 2 − µ |∇u|2 + f · u.2. 2. 0 Ω The theoretical framework is based on a variational formulation of the problem. 3. T ) gives Z Z Z T Z Z TZ ρ ρ |u(x. In practice. which leads to the following formulation of the problem: . power dissipated by viscous effects (lost by the system). Those physical considerations are integrated in the mathematical framework through the space in which the problem is set. and in particular the role played by a priori estimates. We introduce n o L2 V = v ∈ H01 (Ω)d . We refer the reader to [19. 9] for details on the analysis of the Navier-Stokes equations. T . Assuming that the velocity is sufficiently regular to allow those operations. Assuming that a finite energy is supplied to the system. this balance will contain three contribution: 1. which make it possible to build weak solutions by means of compactness arguments. In the case of a fluid enclosed in a domain with no-slip condition on the wall. T ) writes ( Z ) u ∈ L∞ (0. 14. but uniqueness of those weak solutions is an open issue. power of external forces (supplied to the system). and so is the dissipated energy. T ). (1. T . dissipation transforms kinetic energy into heat. ess sup |v(·. the kinetic energy will remain bounded.1. MATHEMATICAL FRAMEWORK 11 celebrated paper by Leray [13]. as uniqueness of smooth solutions holds. (1. V ) = v . H = V . We shall focus in this chapter on the notion of weak solution. and integrating by part the viscous term. To emphasize the role played by a priori estimates. |∇v| dx dt < +∞ . ∇ · v = 0 . t) over Ω×(0.6) L2 where V is the complete closure of V for the L2 norm. Note that the system is not closed in terms of energy as temperature is not accounted for. minus the dissipated energy during (0. the fact that the kinetic energy is bounded over (0.5) Ω 2 Ω 2 0 Ω 0 Ω which expresses that the kinetic energy is the sum of its initial value and the work of external forces. T ) is equivalent to the requirement that u belongs to  Z TZ  2 2 L (0. T )| 2 = |u(x. which is obtained by multiplying the momentum equation by a test function v ∈ V . we obtain Z Z Z d ρ 2 2 |u| + µ |∇u| = f · u.T ) Ω Similarly. boundedness of the dissipated energy over the interval (0.

We consider a forcing term f ∈ L2 (0. . in the sense of Definition 5. . 0) = u0 . v) = a(u. um i| ≤ kϕk kum k ≤ kϕk2 + kum k2 . MODELS FOR INCOMPRESSIBLE FLUIDS Definition 5 (Weak solution of the Navier-Stokes equations) Let V and H be defined by (1. 2 0 2 0 We have 1 1 |hϕ . We define the operator A ∈ L(V. V ′ ). Proof : We give here a sketch of the proof. t) is a weak solution to the Navier- Stokes equations. v) = λ(w. for all v ∈ V . . We aim out pointint out the difficulties that we can expect in extending this approach to the situation that we are interested in. 19] for a detailed proof. for any t ∈ (0. λn ). for which the Cauchy Lipschitz Theorem ensures existence and uniqueness of a maximal solution. v) + c(u. ·) is symmetric. H). regular domain. V ′ ) by (Au. v) + a(u. . Denoting by Vm the linear space spanned by w1 . The eigenvalue problem therefore admits a sequence of solutions (wn . there exists a unique u such that a(u. T . Following [14]. and (·. v) = hϕ . v). T ). vi ∀v ∈ V. V ) ∩ L∞ (0. Theorem 1 Let Ω be a bounded.5) can be written in the abstract setting. by looking for a solution of the form m X um = ujm (t)wj . .7) dt where V and H are two Hilbert spaces. and an initial condition u0 ∈ H. t) in (0. with initial condition u0 .7) can be written in Vm . if u ∈ L2 (0. ·) denotes the scalar product in H × H.6). Z Z Z Z d ρ u · v + ρ (u · ∇) u · v + µ ∇u : ∇v = f ·v ∀v ∈ V dt Ω Ω Ω Ω u(·. T . 2 2 . and by |·| the norm in H. based on the Faedo-Galerkin approach. which consists in introducing a special basis of V . v) ∀v ∈ V. we shall restrict our attention here to the way a priori estimates allow to prove well-posedness of a finite dimensional version of the original problem. T . t) 7→ u(x. and elliptic over V × V . A−1 is compact. . um i. and (wn ) is an Hilbert basis of H. continuous.12 CHAPTER 1. we say that (x. We obtain a system of m ordinary differ- ential equations with unknown functions t 7→ ujm (t). T > 0. Thus. v). so that A−1 is in L(H). (1. We denote by k·k the norm in V .7) is verified for all test functions in Vm . namely the case of open domains (with fluid getting in or out of the domain). The incompressible Navier-Stokes equations admit at least a solution u(x. . We refer to [14. the sequence (λn ) goes to infinity. where 0 < λ0 ≤ λ1 ≤ . the variational problem (1. T ): Z t Z t 1 2 1 2 |um (t)| + a(um . um ) = |um (0)| + hϕ . For any f ∈ H. v. The bilinear form a(·. Since A−1 f ≤ C |f |. wm . v) = (f. V ⊂ H with compact and dense injection. j=1 such that (1. Consider the eigenvalue problem a(w. Let us first note that the problem can be written in an abstract form as d (u. so that we can use it as a scalar product in V . Now the energy balance (1.

so that finally Z t Z t 2 2 |um (t)| + a(um . i. T .1. we consider a pipe-like domain Ω whose boundary ∂Ω = Γ is decomposed into three components: The lateral component Γw corresponds to a rigid wall.2: Pipe-like domain with kum k2 = a(um . V ) ∩ L∞ (0. T .e. The sequence (um ) is bounded in L2 (0. on which homogeneous . |um (t)| can not blow up in finite time. H). defined up to the end T of the time interval.3. The proof is ended by compactness arguments: um weakly converges (up to a subsequence) to a limit u ∈ L2 (0. FREE IN-/OUTLET CONDITIONS 13 Γin Ω Γw Γw Γout Figure 1. and the solution given by the Cauchy Lipschitz theorem is global.3 Free in-/outlet conditions To fix the ideas. 0 0 As a consequence. V ) ∩ L∞ (0. T . H). We refer to [14. um ) = |um (0)| + kϕk2V ′ . 9] for details. um ). 1. and the rest of the proof mainly consists in proving that u is a weak solution to the Navier-Stokes equations. T . 19.

Computation have been performed with the software FreeFem++[8]. and Γout the outlet. The variational formulation corresponding to the free outlet conditions is Z Z Z Z Z µ ∇u : ∇v − p∇ · v = f ·v− Pin n · v − Pout n · v ∀v ∈ V. The velocity field on the left (parabolic profile) corresponds to free outlet conditions (1. We introduce n o V = v ∈ H 1 (Ω)d . the fact that some fluid is entering the domain will make it much more difficult to establish a priori estimates.3 illustrates the difference between the 2 types of conditions. (1.3. We aim at modeling the flow of some viscous fluid through the pipe. but which makes it possible to recover the exact solution in case of Poiseuille’s flow in a cylinder. 1.9) correspond to the situation where the actual fluid domain continues beyond the boundary. we shall call those conditions free outlet (or inlet) conditions. Neumann conditions (1. As Conditions (1.9) (free outlet) raise theoretical issues.8) This option may make sense from a modeling point of view. MODELS FOR INCOMPRESSIBLE FLUIDS Dirichlet boundary conditions are prescribed.9) which does not make clear sense from a physical point of view (the non symmetric tensor ∇u has no mechanical significance). if we are interested in a situation where the pipe continues beyond Γout . in a way which allows to keep the possibility to couple this pipe to other components of a multi-element model. But it might not be the case. Mathematical framework. Γin is the inlet. v|Γw = 0. As a first step. yet it corresponds to a situation where the viscous fluid is separated by Γout (or Γin ) from a fluid in which the stress tensor is diagonal (like a perfect gas). we start with the Stokes problem. and it would make clear sense if Γout were indeed a free surface separating the viscous fluid and such a medium.14 CHAPTER 1.1 Prescribed normal stress The simplest approach for setting boundary conditions consists in considering that the normal component of the stress tensor (see Definition 1. For the Navier-Stokes equations. (1.8). and the set of pressure X = L2 (Ω). and the plot on the right represents the same zone for free surface conditions (1. page 5) is known. Figure 1. It turns out that a better way to account for this continuing pipe consists in prescribing a boundary condition of the type µ∇u · n − pn = −Pext n (1. The stress tensor is σ = µ(∇u + t ∇u) − p Id so that free outlet conditions (expressing force balance in the normal direction) reads µ(∇u + t ∇u) · n − pn = −Pext n. Ω . as we shall see.9).10) Ω Ω Ω Γin Γout together with the weak expression of the divergence free constraint Z q∇ · u = 0 ∀q ∈ X.8) (free surface) or (1.

2. which is straightforward as v may have non zero values on Γin ∪ Γout .1. The problem admits a unique solution.3: Free outlet conditions based on the velocity gradient (left) and on the strain tensor (right). Ω Γin Γout The bilinear form involves the symmetrized velocity gradient. Ellipticity of this bilinear form is . Proof : In the present situation. Remark 3 If one considers free surface conditions (1. Proposition 2 We consider the Stokes problem on domain Ω represented in Figure 1. Considering R a pressure R field q ∈ L2 (Ω). the velocity may have non zero values on some part of the boundary. This makes the proof of the surjectivity of the divergence operator B slightly different. with homogeneous Dirichlet on the lateral boundary Γw and free in-/ out-let conditions (1. and the pressure is no longer defined up to a constant. FREE IN-/OUTLET CONDITIONS 15 Figure 1.3. the first step consists in building a velocity field v such that ∇ · v = q. by considering the pressure minus its mean value over Ω.9) on Γin ∪ Γout . The rest of the proof is then identical to the case of homogenous boundary conditions.8). the variational formulation is modi- fied: Z Z t  t  µ ∇u + ∇u : ∇v + ∇v − p∇ · v Ω Ω Z Z Z = f ·v− Pin n · v − Pout n · v ∀v ∈ V.

which ensures existence of a constant such that Z Z . MODELS FOR INCOMPRESSIBLE FLUIDS a consequence of Korn’s inequality.16 CHAPTER 1.

.

.

∇v + t ∇v.

4 Some typical problems We describe here some situations for which the exact solution cannot be given analytically. thereby conditioning the price to pay to make the plane flying. whereas drag is opposed to motion. this example calls for a special care of boundary . 1.4. The approach may lead to a large exterior domain. γ γ The lift force maintains the aircraft in the air (it balances the total weight of the flying plane). Fx = −ex · σ · n. The question of boundary conditions on the outer boundary raises some interesting issues.2 Some simple fluid-structure interaction problems In real life application. Spring-mass-fluid system.4): consider a flow of a viscous fluid around a fixed obstacle O.1 Flow around an obstacle A great amount of computational strategies to approximate solutions to the Navier-Stokes equa- tion have been developed during the last decades to address the following problem (represented in the two dimensional setting by Fig. it is reasonable to assume the velocity to be fixed on Γ = ∂Ω. The problem consists in solving incompressible Navier-Stokes equations in the fluid part Ω. which is the case here thanks to the no-slip condition on the lateral boundary. We present here here two situations where the number of degrees of freedom for the structure is finite. A first model consists in assuming that the outer boundary is far enough from ω. The domain Ω is bounded on the right side by a piston with mass m. one is interested in estimating the forces exerted by the fluid on the obstacle. We consider the situation represented in Fig. it is no longer block-diagonal for the form that is based on the symmetrized tensor.4. one is commonly interested in modeling the way a fluid interacts with another medium. 1. Ω Ω as soon as V is a subset of H 1 (Ω)d which does not contain any rigid motion other than 0. 1.2 ≥ C |∇v|2 ∀v ∈ V. In the case where ω represent an airfoil. by elaborating appropriate boundary conditions on the outer boundary.5 (left). Whereas the matrix corresponding to free outlet conditions is block-diagonal (scalar Laplace operator for each component of the velocity). Those conditions complicate the numerical resolution. like an elastic structure. equal to a “velocity at infinity” U∞ . This problem raises issues which are still the object of active research. so that the flow there is not affected by the presence of the obstacle. 1. quantities of interested are the lift force (portance in french) Fy . In this case. A great amount of investigations has been undertaken to propose alternative strategies to overcome this problem. and the drag (traı̂née) Fx Z Z Fy = −ey · σ · n . 1. attached to a spring with stiffness k > 0. with no-slip conditions on the boundary γ of the obstacle. both on the theoretical and numerical aspects. Despite its formal simplicity. and thereby high computational costs.

dt2 Γm Fluid particle flow. dt γ dt where J is the moment of inertia of the disc. right). SOME TYPICAL PROBLEMS 17 Ω O γ Figure 1. In this second example (see Fig.5.4.4: Flow around an obstacle conditions : as the motion of the piston is not compatible with the no-slip condition on the lateral boundary. J = 0. and by m its mass. t·σ·n = 0 on Γℓ u = ẋ ex on Γm supplemented by Newton’s law for the piston: Z d2 x m = −k(x − x0 ) − σ · n. No-slip condition on the boundary of the body writes u= U+ω∧r on γ. the global problem may be written   ∂u ρ + (u · ∇) u − µ∆u + ∇p = 0 in Ω(t) ∂t ∇ · u = 0 in Ω(t) u = 0 on Γ0 σ·n = 0 on Γin u·n=0. we shall assume that the fluid may slip freely along this part of the boundary. the fluid problem can be written in the form of incompressible Navier-Stokes equations in the moving domain Ω(t) (with right-hand side ρg. Assuming the fluid sticks to the piston. we consider a rigid disc flowing freely in a viscous fluid. . 1. Kinematic coupling. Denoting by ρs the density of the solid. supplemented by two types of coupling conditions: 1. Newton’s law for the particle translational and angular velocity write Z dU dω m = − σ · n + mg .1. Dynamic coupling. 2.

but it can be extended straightforwardly to the many-body situation (with 3 degrees of freedom per particle. MODELS FOR INCOMPRESSIBLE FLUIDS Γℓ Ω Ω γ U Γ0 O k Γin Γm ω Γ0 Γℓ Figure 1. .18 CHAPTER 1.5: Interaction with a rigid structure The problem is written here for a single particle. together with a three dimensional coupling between each particle and the fluid).

1). We hereafter discuss those three aspects of the problem and detail the natural links between them as far as the discretization by the finite element method is concerned. To this aim. we consider only Stokes equations. 2. such Zthat for all (v.1 The incompressibility constraint. the viscosity has been taken equal to unity :  −∆u + ∇p = f . one faces three problems that seem quite uncoupled and follow different (though complementary) strategies: • The incompressibility constraint div u = 0 and the related pressure gradient . The equation is non stationary. p) ∈ V × M .Chapter 2 Navier-Stokes equations. Indeed. there are typically two variational formulations associated to (2. numerics Navier-Stokes equations possess several terms that need to be handled adequately in view of a full numerical scheme. q) ∈ V × M . namely the discretization of the constraint div u = 0 . q)Z ∈ V × M . namely Find (u. (2.3)  q div u dx = 0 .2)  q div u dx = 0 . Ω 19 . • The nonlinearity present in the convective term .  Z ∇u : ∇vdx − pdiv v dx = f · v dx .  ZΩ Ω Ω (2. depending on the Neumann boundary conditions that one con- siders. such that forZ all (v. where. We have already seen that.1) div u = 0 .  ZΩ Ω Ω (2. • The time dependence.  Z (∇u + t ∇u) : ∇vdx − p div v dx = f · v dx . and a suitable time marching algorithm needs to be written. p) ∈ V × MZ . in order to focus only on the mathematical aspects of the problem. Ω or Find (u. Stokes equations Let us start with the first problem.

there might be only one function satisfying all those which is uh = 0. ph ) ∈ Vh × M Z (vh .6) h>0 qh ∈Mh vh ∈Vh ||qh ||M ||vh ||V 1 Of course.5) Ω One sees that if the constraints expressed by (2.20 CHAPTER 2.4)  qh div uh dx = 0 .2) for instance (the corresponding holds for (2. To finish for the time being on this question. suitable boundary conditions need to be considered. qh ) ∈ Vh × Mh . (2. NAVIER-STOKES EQUATIONS. Ω where Vh ⊂ V and Mh ⊂ M are two finite dimensional subspaces of V and M respectively. at this level of the discussion to discuss these problems. it turns out that there is an additional problem due to the discretization of the Hilbert spaces V and M . Some choices are known to be unstable (they do not lead to a convergent discrete solution as the space step h goes to 0) like the natural couples Vh = P 1 .5). Korn’s inequality is needed to prove existence and uniqueness of the second formulation. etc.3)) reads as follows Find (uh .5) are too numerous. Therefore the finite element spaces Vh and Mh need not be chosen separately. . ∀qh ∈ Mh . This might for instance be the case if unfortunately dim(Mh ) > dim(Vh ). It is not the aim of the authors. P k stands for the classical polynomial approximation of degree k on a triangula- tion. the most known being Vh = P 2 and Mh = P 1 . When the discretization by the finite element method is concerned. then there might be a convergence problem. the discrete variational formulation associated with (2. They however lead to different Neumann boundary conditions when one integrates by part. The preceding problem has a clear energetic formulation. Namely uh can be seen as minimizing Z Z 1 2 E(uh ) = |∇uh | dx − f · uh dx 2 Ω Ω under the constraint(s) Z qh div uh dx = 0 . It is unclear why the discrete solution found uh should be close to the continuous one u. Mh = P 0 . Indeed. extra work is needed here. If on the other hand. Mh = P 1 or Vh = P 1 . Z h . (2. Lax-Milgram lemma does not apply directly because the bilinear form is not coercive. Other choices lead to both convergent and well posed discrete formulations. R)/R. there are not enough constraints in (2. such that for all  Z ∇uh : ∇vh dx − ph div vh dx = f · vh dx . R3 ) and M = L2 (Ω. It can be shown that both formulations leads to well- posed problems1 provided V ⊂ H 1 (Ω. also known as the Lagrange finite element of degree k. In the former. let us say that stable and convergent choices for the finite element spaces Vh and Mh are those which satisfy the following Inf-Sup condition Z qh div vh dx inf inf sup Ω ≥ β > 0.  ZΩ Ω Ω (2. NUMERICS Both variational formulation lead to the same Stokes equations inside the domain due to the fact that  div t ∇u = ∇(div u) = 0 since the vector field must be incompressible.

Of course. Namely. P 1 ). in the sense that forgetting the nonlinearity for the time being. ∂t time discretization strategies can be described straighforwardly. The implicit version writes un+1 − un − ∆un+1 = f.7) Z Z Z Z n+1 n n u v dx = u v dx − δt ∇u ∇v dx + δt f v dx . (2. we have for the explicit Euler scheme (2.8) δt 1 where θ = 0 corresponds to the explicit scheme.2. 1] writes un+1 − un  − (1 − θ)∆un + θ∆un+1 = f. At first sight.2 Time discretization 2. There is again a stability issue given by the following proposition. calling Vh the finite element space in which the solution is sought. typically (P 2 . all these schemes must be thought in weak form. All these schemes enter in the class of the so-called θ−scheme which for a parameter θ ∈ [0.2.7) δt where δt is the time step. δt Extensions of those two approaches can be built to obtain a better accuracy in the time reso- lution. the remaining terms of the equations are parabolic. Proposition 3 After a finite element space discretization. δt 2 is more accurate (second order in time). (2.2.1 Generalities Let us focus now on the time discretization. the Crank-Nicholson scheme (which can be described as half explicit - half implicit) un+1 − un 1  − ∆un + ∆un+1 = f. TIME DISCRETIZATION 21 Temporary conclusion: Due to stability problems the velocity/pressure need to be computed using a stable finite element. Ω Ω Ω Ω Similar expressions can be derived for the implicit or Crank-Nicholson schemes. For scalar and linear problems like the heat equation ∂u − ∆u = f. Navier-Stokes equations are very similar to the scalar heat equation. 2. whereas it preserves the stability properties of the implicit scheme. θ = 1 the implicit scheme and θ = 2 is the Crank-Nicholson scheme. As an example. The simplest time-discretization approach is the explicit Euler scheme un+1 − un − ∆un = f. . the θ−scheme is uniformly and un- conditionally stable (and convergent) provided θ ≥ 21 .

22 CHAPTER 2. when the right-hand side f vanishes the two following du estimates abstained respectively by multiplying the equation by u and : dt Z Z 1 d u2 dx + |∇u|2 dx = 0 . The solution to the heat equation.) A recipe to prove the stability is to look for a discrete energy bound. verifies. NUMERICS Proof : (Hint. NAVIER-STOKES EQUATIONS. 2 dt Ω Ω and Z .

.

2 Z .

du .

.

.

dx + 1 d |∇u|2 dx = 0 . Ω dt 2 dt Ω .

.

Both equations give after integration in time a bound of a positive quantity. multiplying now (2. noticing that uθ = + (2θ − 1) leads to the estimate 2 2 Z Z Z Z (un+1 )2 (un+1 − un )2 2 (un )2 dx + (2θ − 1) dx + |∇uθ | dx = dx . one can use the second estimate.7) by δt gives Z  n+1 2 Z  n+1  u − un u − un dx + ∇uθ · ∇ dx = 0 . 0 Ω dt The idea is to mimic those results in the discrete case. Ω 2δt Ω 2δt Ω Ω 2δt 1 When θ ≥ 2 this gives the uniform bound (for all n. Ω δt Ω un + un+1 un+1 − un which. writing again uθ = + (2θ − 1) leads now to the estimate 2 2 Z . multiplying the scheme (2.8) by uθ := (1 − θ)un + θun+1 (with f = 0) and integrating by parts leads to Z n+1 Z u − un uθ + |∇uθ |2 dx = 0 . For the L2 stability. h) for the L2 norm Z Z (u ) dx ≤ (u0 )2 dx . Namely. n 2 Ω Ω un+1 − un Similarly. Ω δt Ω δt un + un+1 un+1 − un which. namely L2 and the 10 norms respectively Z TZ 2 ||u(T )||L2 + |∇u|2 dx = ||u(0)||2L2 0 Ω and Z Z  2 T du ||∇u(T )||2L2 + dx = ||∇u(0)||2L2 . δt.

.

n+1 .

2 Z  n+1 2 Z

 n+1 

2 Z ∇u u − un 1 .

u − un .

.

|∇un |2 dx + dx + (θ − )δt .

.

∇ dx = dx . Ω 2δt Ω δt 2 Ω δt .

Ω 2δt Again. h) Z Z n 2 . δt. when θ ≥ 12 . we obtain the uniform estimate (for all n.

0 .

2 |∇u | dx ≤ .

∇u .

Ω Ω . dx .

2. ρ n where α = ρ/δt. there is still the problem of the non linear term that can be treated explicitly or implicitly. δt with ∇ · un+1 = 0. we deduce that the Laplace term is more stable when discretized implicitly. and the new right-hand side f̃ = f − ρ (un · ∇) un + δt u accounts for inertial terms. Fully explicit scheme. condition of the type uδt ≤ Ch. where U = ρun is known. TIME DISCRETIZATION 23 Exercise 2 Extend those results for the case of a non vanshing external force f . The problem is still linear. i. 2. As for Burger’s equation. it is defined by Z Z Z a(u. where C is a dimensionless constant smaller than 1. Note that complete space discretization may necessitate special quadrature formulae to compute (or at least estimate) the term Z (un · ∇) un · v.2 Going back to Navier-Stokes equations From the preceding section. The semi explicit scheme is also very popular.2. Temporary conclusion: Implicitizing the Laplace term in the heat equation increases the stability of the underlying finite element scheme. it consists in writing the nonlinear term as an advection of the unknown velocity by the known one: un+1 − un ρ + ρ (un · ∇) un+1 − µ∆un+1 + ∇pn+1 = f . Note that it leads to a so called generalized Stokes problem  αu − µ∆u + ∇p = f̃ . However. Ω Ω Ω . δt with ∇ · un+1 = 0 which leads to a problem of the type  αu − µ∆u + (U · ∇)u + ∇p = f̃ . it obviously requires a so called CFL condition. In particular. This problem has exactly the same mathematical structure as the standard Stokes problem. The fully explicit scheme is based on an explicit evaluation of the non linear term in the Navier Stokes equations un+1 − un ρ + ρ (un · ∇) un − µ∆un+1 + ∇pn+1 = f . Semi-explicit scheme. or the even simpler transport equations. and can be handled numerically as such.e. in the variational formulation. ∇ · u = 0. ∇ · u = 0. but the associated bilinear form is no longer symmetric. this approach is expected to raise stability issues.2. It is typically unconditional between the Crank-Nicolson scheme (θ = 12 ) and the fully implicit scheme (θ = 1). v) = α u · v + µ ∇u : ∇v + v · (U · ∇)u.

homogeneous Dirichlet boundary conditions). Exercise 3 Notice that if u and v are two vector fields with div u = 0 (and suitable. e. Consider t 7−→ χ(t) the trajectory of a fluid particle located at x at time τ For any function Φ (scalar or vector).24 CHAPTER 2. the total derivative of Φ at (x. Ω Then apply the energy method to prove unconditional stability of the semi explicit and implicit schemes. The fully implicit scheme is based on an implicit evaluation of the non linear term in the Navier Stokes equations un+1 − un  ρ + ρ un+1 · ∇ un+1 − µ∆un+1 + ∇pn+1 = f .g. Note that this approach requires to solve a nonlinear problem at each time step. NAVIER-STOKES EQUATIONS. one has Z (u · ∇)v · v dx = 0 .3 The method of characteristics The method of characteristics can be seen as a direct discretization of the total derivative. τ ) is . NUMERICS Fully implicit scheme.2. 2. Compared to the previous methods. it is also a way of handling the non linear term explicitly but which enjoys better stability properties than the previous explicit scheme. δt with ∇ · un+1 = 0.

.

DΦ .

.

d .

t). = Φ(χ(t).

.

Dt (x.τ ) .

τ ) − Φ(χ(τ − δt). for any point of the domain. Using the preceding formula. τ ) = lim ε→0 ε Φ(χ(τ ). δt Now consider that the solution un at time nδt is known. τ − δt) ∼ . τ ) − Φ(χ(τ − ε). we see that . τ − ε) = lim ε→0 ε Φ(χ(τ ). τ + ε) − Φ(χ(τ ). dt t=τ Φ(χ(τ + ε).

Du .

.

un+1 (x) − un (y) ∼ Dt .

(x. Calling X : Ω −→ Ω x 7→ y .τ ) δt where y is the position at time nδt of the characteristic which ends at x at time (n + 1)δt (y is the so-called “foot” of the characteristic). Notice that y is obtained by integrating backward along the vector field un which is known.

(2. +int2d(Th)(-convect([ux. without the divergence constraint (notice that the pressure has be frozen to its previous value). the term containing X goes to the right hand side. We shall present the approach coupled with the method of characteristics. The variables (tux and tuy) stand for the test functions in the weak formulation.-dt. 2. As for the explicit scheme.2. ∇·u = 0 where f̃ carries the initial forcing term supplemented by the inertial term αun ◦ X. PROJECTION METHODS 25 the scheme is usually written as αun+1 − αun ◦ X − µ∆un+1 + ∇pn+1 = f .uy)*tuy) In the preceding command. This method can be implemented in Freefem++ using the command convect as follows.9) ∇ · un+1 = 0 . which yields −∇ · ∇r n+1 = −∆r n+1 = −α∇ · ũn+1 . This step is nothing that an implicit heat flow time step. It is clear that the incompressibility constraint has been lost during this step.uy]. Notice that the integration of the characteristic flow is done backward.-dt.uy]) of the convect command is the vector field along which the quantity needs to be advected. (Prediction) The velocity is predicted by solving αũn+1 − αun ◦ X − µ∆ũn+1 + ∇pn = f . . one writes  αun+1 − αũn+1 + ∇r n+1 = 0 .3. as they split the saddle point problem onto two Poisson like problems. 2. and " 1 un+1 = ũn+1 − ∇r n+1 α pn+1 = pn + r n+1 . which yields a generalized Stokes problem: ( αu − µ∆u + ∇p = f̃ .ux)*tux-convect([ux. The approach described here does not rely to the way the inertial term is handled. More precisely. but the other approaches where it is treated explicitly or implicitly could be used as well: 1.3 Projection methods Projection methods are very popular. the first argument ([ux. the second argument (-dt) is the time of the advection and the third argument (ux or uy) what has to be advected.uy]. The velocity is now corrected by adding a pressure gradient correction term ∇r n+1 to ũn+1 in order that un+1 be divergence free.

which is the Helmholtz’s decomposition of ũn+1 as ũn+1 = ∇φ + curl ψ . Notice that (2. this splitting algorithm is consistent with the original Navier-Stokes equations since one has αun+1 − αun ◦ X − µ∆ũn+1 + ∇(pn + r n+1 ) = f .9) is nothing than the projection of ũn+1 on divergence free vector fields. NAVIER-STOKES EQUATIONS. NUMERICS This discretization approach therefore to Poisson like problems only : a vector problem for the momentum equation.10) ∇ · un+1 = 0 . Indeed. (2.26 CHAPTER 2. one can rewrite (2. and a scalar one for the incompressibility constraint.9) as  n+1  u + ∇ α1 r n+1 = ũn+1 . . Notice also that. although there are two seemingly uncoupled steps.

5) Navier-Stokes equations. situation.4) 0 The domain at time t can be written as Ω(t) = {L(x0 . We consider that the domain moves according to Navier-Stokes equations.3)  L(x0 . t). . All surfaces and functions are supposed to be sufficiently regular. We shall consider here a 2D free surface flow. the kinematical approach is applied to a particular. and somewhat idealized. It can be written in integral form Z t L(x0 .Chapter 3 Moving domains 3. u0 (x0 . s) = u(L(x . t) . (3. (3. s) ds. (3. t) = x0 + u (L(x0 s). The fact that the domain Ω(t) is “moved” by the physical velocity can be expressed the following way (Lagrangian approach): one can build mappings L( . with no inlet nor outlet. We shall disregard here the equations to which the velocity u(x. x0 ∈ Ω(0)}. t) : Ω(0) −→ Ω(t) x0 7−→ xt = L(x0 . nor any contact with any wall. and focus on the kinematical aspects. t) is a solution.1) where Γ(t) is the free surface. 0) = x0 . Let us denote by u0 the Lagrangian velocity with respect to the initial configuration.e. The domain Ω(t) is delimited by ∂Ω(t) = Γ(t).1 Lagrangian standpoint For the sake of clarity. t) (3. s) 0 0 ∂s (3. s) is the characteristic curve from x0 to xt in R2 :   ∂ L(x . s). i. t). t) = u(L(x0 . 27 .2) where s 7−→ L(x0 .

The next triangulation is obtained by performing T n+1 = T n + δtun . for x0 fixed. is actually trivial in terms of computation. un+1 n is an array of values associated n n to the vertices of the triangulation T of Ω . because in practice the underlining mesh is likely to degenerate. . The equation defining un+1 from un+1n simply consists in doing nothing to the array of velocities1 . In the discrete setting. t) = (u(L(x0 .e.. it is restricted to very few situations. t). ∂t Dt As a consequence. (3. the partial time derivative of u0 is the total derivative.6) ∂t with ∇ · u0 = 0. when a field is defined on a certain mesh. It is necessary to use auxiliary arrays to bypass this automatic interpolation (see Section 3. Indeed. t). This linearization of Navier-Stokes equations is only valid instantaneously. t)) = u + ( L(x0 . Nevertheless. in a nonlinear way. t) + (u · ∇)u = . when the initial field is called. the term (u · ∇)u is dropped). Indeed. Indeed. and is based on solutions of linear systems (the nonlinear term has been dropped).4). MOVING DOMAINS Since the space variable is Lagrangian. a special attention to this step has to be paid.7) δt The domain Ωn+1 is now defined the as the push forward of Ωn by δtun : Ωn+1 = {x + δtun (x) . with the command movemesh).6) is valid at the first order in a neighborhood of t = 0. for piecewise affine functions. If one aims at expressing it at a time t different from that of the reference domain. x ∈ Ωn } . Note that the latter operation. the eulerian coordinate at time t (according to which the Laplacian et gradient operators are defined) and the Lagrangian one x0 are related by x0 7−→ x = L(x0 . t) · ∇)u dt dt ∂t ∂t ∂ Du = u(L(x0 . The method that has been presented is quite simple to implement.g. the partial time derivative will still express the total derivative (i. if the mesh is moved (e. The time stepping approach can be described in the following way. 1 With some softwares like Freefem++. it is possible to design a numerical method of the Lagrangian type. d d ∂ d u0 (x0 . Freefem++ favors the function represented by a discrete field before the array of values that represented. assuming that (3. (3. an interpolation between the two meshes is automatically performed. In a way.g. It can be computed as the solution to un+1 − un ρ n − µ∆un+1 n + ∇pn+1 n = f. The velocity un+1 is then defined by un+1 (x + δtun (x)) = un+1 n (x). up to creation of reversed triangles. We denote by un+1 n the next velocity. by using the Lagrangian coordinates that corresponds to the current configuration. but the terms that correspond to space derivatives will be deeply affected. the Navier-Stokes equation expressed in the Lagrangian variable at t = 0 write ∂u0 ρ − µ∆u0 + ∇p0 = f . Yet. Denote by Ωn and un the domain and velocity at time tn . that will stop the computation. as tricky as it may look. e. defined in the lagrangian coordinates associated with Ωn .28 CHAPTER 3. t).

(3. for any Ω. t∈[0.5)). ARBITRARY LAGRANGIAN EULERIAN (ALE) POINT OF VIEW 29 3. the 3D field (c(x. Let t = (tx . ∀x ∈ Γ.9) where n(x) is the normal vector to Γ = ∂Ω at x. Ω) 7−→ c = C(u.10) 0 1 tx uy − ty ux so that. The approach is based on a definition of the domain velocity that is partially decoupled from the fluid velocity: it has to satisfy some consistency relations so that the computational domain will follow the actual fluid domain (globally Lagrangian behavior). and the associated domain velocity field c(x. Proposition 4 The characteristics associated to the space–time field (c(x. we investigate the possibility to build mappings between the domains (similar to (3.2. We consider a consistent ALE mapping C. Ω) is said to be consistent if. (3. The definition of the domain velocity is based on a mapping (u. on the lateral boundary.       cx ty c · N =  cy  ·  −tx  = cx ty − cy tx + tx uy − ty ux (3. continuous case) The ALE mapping u 7→ c = C(u. (3. t).12) = 0 (3.9) ensures that this field is tangent to the “lateral” boundary of S (boundary of S excepting Γ(0)× {0} and Γ(T )× {T }). 1) can be integrated in the physical space-time domain S corresponding to the time interval [0. The fact that the second step can be performed arbitrarily explains the name of the approach: Arbitrary Lagrangian Eulerian. t).13) . ty ) be the vector tangent to Γ(t) at x. t) can be expressed       tx ux ty N =  ty  ×  uy  =  −tx . t). The normal vector to the space–time domain S at (x. but it will defined within the domain in order to minimize the domain motion (Eulerian tendency: move the vertices as little as possible). Ω).3.8) where Ω is a smooth domain.2 Arbitrary Lagrangian Eulerian (ALE) point of view In what follows. u and c are regular fields defined in Ω (domains and fields are supposed at least C 1 ). T ]. Proof :Let us show how equation (3. 1) remain in the space–time domain [ S= Ω(t) × {t}. Definition 6 (Consistent ALE mapping. with a field which is not the fluid velocity.T ] In other words. u.11) 1 1 tx uy − ty ux = u·n−c·n (3. u(x) · n(x) = c(x) · n(x) .

(3. t1 . t) is the characteristic curve from (x1 . τ . t) = u(γ(x. t) (x ∈ Ω(τ ).1: Space time domain as soon as u and c verify the consistency relation (3. t2 ) : Ω(t1 ) −→ Ω(t2 ) x1 ∈ Ω(t1 ) 7−→ x2 = γ(x1 .15) Partial derivation with respect to t of (3. t1 .9). t) ∈ Ω(t)) (3. t2 ) where (γ(x1 . t) (x ∈ Ω(t)).30 CHAPTER 3. t). t. t1 . t1 . t).15) at t = τ gives .  Proposition 4 makes it possible to define applications between the Ω(t): γ(·. t) = u(x. t2 ) in S: ( d (γ(x1 . t1 ) For each τ .1) y Particle pathline ALE trajectory c u x Figure 3. t1 ) = (x1 . τ . t) = [c(γ(x1 . τ ).14) which is equivalent to uτ (γ(x. t).1) (u. t). t1 . t). t1 ) to (x2 . 1] dt γ(x1 . t1 . γ(x. the ALE velocity is then defined by uτ (x. MOVING DOMAINS t (c.

∂u ∂uτ ∂γ .

.

= + ∇uτ · ∂t ∂t ∂t .

16) ∂t Note that identity (3. . t) defined in the space time domain (for x ∈ Ω(t)).15) can be used to define the ALE version of any scalar or vector field Φ(x.t=τ ∂uτ = − (cτ · ∇) uτ (3.

2: Fields u. We shall restrict ourself here to a simple way which consists in prescribing the direction R of the boundary velocity. For a given class of domains.2.18) R·n See Fig. this space–time S is non longer cylindrical. We chose to solve a Laplace equation for each component of the domain velocity c. and it is such that R · n > 0 on Γ. (3. Definition of the domain velocity within the domain plays little role in the present approach. which can . Remark 4 Evolution problems in a fixed domain lead to a cylindrical space–time domain (set product of the space domain by the time interval). then the consistency relation (3. Ω) be a given family of vector fields : for any Ω. R. 3.3.9) imposes the consistent ALE mapping (u.17) to be defined uniquely (on the boundary) as u·n c= R.2. If we prescribe R to be the direction of the boundary velocity. When the domain moves. ARBITRARY LAGRANGIAN EULERIAN (ALE) POINT OF VIEW 31 u R c Figure 3. Ω) (3. and c. Ω) 7−→ c = C(u. R is defined on Γ = ∂Ω. let R(·.1 Actual computation of the domain velocity The proper way to define the velocity c on the boundary is highly dependent on the type of flow which is considered. 2 for an example of field c. The approach which has been carried out can be seen as a way to straighten S locally by introducing the ALE velocity: the partial time derivative of the new field follows the lateral boundary of S.

this operation can be performed at zero cost. verify the ALE Navier–Stokes equations at t = τ . which are defined on Ω(τ ) and related to the Eulerian velocity by (3. The next domain is then obtained as Ωn+1 = Ωn + δtcn . and σ is the total stress tensor σ = ν(∇u +t ∇u) − p 1.e. since the motion of the domain is partially decoupled from the actual motion of fluid particles. with ∇ · un+1 n = 0. κ is the surface tension coefficient (between the fluid and the external gas). t) and ALE pressure. In the time discretized setting.2. the nonlinear term disappears.23) ∇ · uτ = 0 The ALE Navier-Stokes equations can be discretized in time. The difference with the full eulerian case is that the velocity must be advected by the velocity that is relative to the domain velocity. a modified nonlinear term remains in the equations. pe the external pressure (which we shall take equal to 0 in the following). . and it has to be handled in some way. i.22) ALE velocity uτ (x.14).19) with boundary conditions given by (3. and X is the foot of the characteristic. as explained at the end of the previous section. 3. (3. δt in the current domain Ωn . for every τ ∈]0. and then solving un+1 n − un ◦ X ρ − µ∆un+1 n + ∇pn+1 n = g. MOVING DOMAINS be expressed in vector form △c = 0 (3. As a consequence. T ]: ( ∂uτ ρ + ρ ((uτ − cτ ) · ∇) uτ − µ△uτ + ∇pτ = g ∂t (3.18). The method of characteristics can be used to handle it.20) ∇·u = 0 with conditions on the free boundary n σ · n = −pe n − κ (3.32 CHAPTER 3. u − c. it does not in the present context. the scheme consists in computing the domain velocity cn . Whereas in the Lagragian approach. which is obtained by upstream integration of un − cn during δt. The last step consists in transferring ALE quantities to the new domain but.2 Navier–Stokes equations in the ALE context We consider now a fluid motion governed by the incompressible Navier–Stokes equations ( ∂u + u · ∇u − ν△u + ∇p = g ∂t (3.21) R where R is the radius of curvature of Γ(t).

3. SOME APPLICATIONS OF THE ALE APPROACH 33 σ·n=0 u·n=0 u=0 Figure 3. .24) on Γ3 . (3.3.4. with an open top (free surface). i. c · n = u · n =⇒ c = . It may not be so straightforward in practice: for any mesh vertex inside the domain. The fluid is subject to gravity forces. and relation (3. 3.3. left). In order to avoid diffusion in the horizontal direction. ∂c/∂n = 0 on Γ2 ∪ Γ4 . right). 3. 3. Since its normal component must equal that of the fluid velocity.3. for the vertical component of the domain velocity. This task is not trivial for a general mesh.3 Some applications of the ALE approach 3. An alternative way consists in solving a Laplace problem over the whole domain. t · σ · n = 0 on Γ2 (t) ∪ Γ4 (t)    u = 0 on Γ1 . it is natural to define a vertical velocity of the domain. in contact with a gas considered as perfect.24) ey · n A natural way to define the interior domain velocity is to interpolate between the upper boundary and 0. one has to determine the velocity of the corresponding boundary point.2 Falling water The next situation corresponds to water falling out of a tap (see Fig. and to slip with no friction on the lateral walls:  ∂u    ρ + ρ (u · ∇) uτ − µ△u + ∇p = g in Ω(t)    ∂t     ∇ · u = 0 in Ω(t)  σ · n = 0 on Γ3 (t)       u · n = 0 .e. with c = 0 on Γ1 .3: Fluid in a container 3. In this context. It is assumed to stick to the bottom boundary. the vertical component is determined: u·n c = c ey .4.1 Water waves We consider the situation illustrated by Fig.3. A viscous fluid is located in a container. to solve −∇ · D∇c = 0. and falling down a corner (see Fig.3. it is possible to solve use an anistropic diffusion tensor.

the following trick can be used: tmp=ux[]. To keep the same nodal values for ux and pux. If the mesh is moved according to Th = movemesh(Th. pux[]=tmp . . and if one updates pux with the instruction pix = ux.4: Falling water In the first situation. In the second case.4 Implementation issues As previously pointed out. the position of the vertex will be located in the previous configuration.[x+dt*cx. Freefem++ will interpolate the values between the two meshes. pux=0. 3. to compute the nodal value of pux at a vertex i of the new mesh.y+dt*cy]). 3.n).4. the domain velocity can be taken horizontal. and another field pux that is defined in the space P1 associated to the same mesh. MOVING DOMAINS Figure 3. Consider a field like ux (first component of the velocity) that has been computed as a piecewise affine function on a mesh Th. where tmp has been defined previously as an auxiliary array : real[int] tmp(ux[].34 CHAPTER 3. the direction cannot be taken constant: it can be prescribed in the way indicated by the dotted lines in Fig. and the corresponding value of the field will be computed by interpolation of ux in the previous configuration. More precisely. Freefem considers a discrete field as a function more than an array of values. right.

Navier-Stokes equations which describe a continuum.Chapter 4 Fluid-structure interactions Many problems involve the interaction between a fluid and a solid. Classical mechanics. On the right hand sides. the fluid and the solid is not obvious since. or pieces immersed in the fluid or at its surface like for instance a boat on a lake or the sand in concrete. This is a discrete system. but could be as well magnetic or electric forces if the body is magnetized or charged for instance. The fluid domain will be denoted by ΩF while the total domain (fluid+solid) is Ω = ΩS ∪ ΩF .1)  dL  = T.  Z∂ΩS ΩS Z (4. etc. m stands for the mass of the solid. The solid can be either a deformable boundary for the fluid like for the blood flow in the arteries and veins.g. 4. the fluid must also fulfill the no-slip boundary condition which imposes that the velocity of the fluid equates a rigid 1 Those external forces are commonly the gravity. so that assuming that the fluid obeys Navier-Stokes equations  Z Z   F = − σ · n dσ(x) + fext dx . given by Newton’s law describes the motion of the solid by the system of differential equations   m dV = F . in which only a few degrees of freedom need to be computed. here due to the fluid and external forces density1 fext .2)   T=−  (x − G(t)) × σ · n dσ(x) + (x − G(t)) × fext dx . particles in the air.  dt (4. A second difficulty stands for the fact that at the boundary of the solid. one has to write Newton’s laws for the solid where the external forces contains the ones given by the fluid. 35 . which depends on time.1 A non deformable solid in a fluid We focus for the time being on the case of a non deformable solid ΩS in the fluid. while the fluid is modeled through e. at first sight the physics involved is different. F and T are respectively the total force and torque induced by external forces. ∂ΩS ΩS where G(t) stands for the position of the center of mass of the solid ΩS . The coupling between both elements. Indeed. V the velocity of its center of mass and L its angular momentum. for a non deformable solid immersed in a fluid. dt In the former equations. and the normal n points from the fluid to the solid.

since u is a rigid body motion in the solid. but this is not mandatory. namely those of the solid. (4. (4. Putting everything together. ΩS where ρS is the density of the solid2 . However this rigid motion is unknown and precisely obtained through the coupling.5) Z mG̈ = − σ · n dσ + Fext . The total momentum of the solid is then given by Z mV = ρS u dx . we extend the velocity field inside the body by setting u(x) = Ġ + ω × (x − G) for all x in ΩS . (4. and V = Ġ is the velocity of the center of mass G of the solid. and Text = (x − G) × fext dx ΩS ΩS respectively stand for the total force and torque on the solid due to external forces.8) Notice that the continuity of the velocity at the boundary ∂ΩS is naturally enforced.3) ∂t div u = 0 in ΩF . (4.7) dt ∂ΩS where we have denoted by ρF the density of the fluid.9) dt ΩS ∂ΩS Z Z d d L= ρS (x − G) × u dx = − (x − G) × σ · n dσ + Text . we therefore consider the system ∂u ρF + ρF (u · ∇)u − µ∆u + ∇p = f in ΩF . and by ω the rotational vector of the solid. Moreover. (4.6) ∂ΩS Z d (Jω) = − (x − G) × σ · n dσ + Text . (4. and Z Z Fext = fext dx . |ΩS | . Indeed. 4.36 CHAPTER 4.2 Rigid body motion The coupling of the PDE and the ODE in the preceding problem needs very specific numerical development. we can rewrite the balances of momentum and angular momentum respectively as Z Z d ρS u dx = − σ · n dσ + Fext (4. and the numerical methods for the PDE and the ODE as well. The idea we aim at explaining hereafter is to consider the solid as a very viscous fluid. the angular momentum of the solid is given by Z L= ρS (x − G) × u dx . ΩS Doing so. Similarly. one has ∇u + t ∇u = 0 in ΩS . FLUID-STRUCTURE INTERACTIONS motion.10) dt dt ΩS ∂ΩS 2 m One has typically ρS = if the solid is homogeneous. Therefore. the mathematical features of the equations are different.4) u = Ġ + ω × (x − G) on ∂ΩS . (4.

2. In view of this. We now consider a test function φ (smooth or at least in H 1 (Ω. RIGID BODY MOTION 37 On the left hand side. R3 )) defined on all Ω such that φ is a rigid body motion in ΩS φ(x) = φ̄ + φω × (x − G) in ΩS . both expression can be written as Z Z   d ∂u ρS u dx = ρS + (u · ∇)u dx (4.4. φ and u being rigid body motions in ΩS we have ∇u + t ∇u = ∇φ + t ∇φ = 0 in ΩS .12).11) and (4. Ω ΩF Indeed. (4. where both φ̄ and φω are constant vectors in R3 .12) dt ΩS ΩS ∂t and we recognize terms that are present in Navier-Stokes equations. ρ(x) = ρS if x ∈ ΩS . and eventually Z Z Z − f · φ dx = − f · φ dx − fext · φ dx Ω ΩF ΩS Z = − f · φ dx − Fext · φ̄ − Text · φω . ΩF Summing up all contributions. we define the density in all Ω by setting  ρF if x ∈ ΩF . we also extend the pressure inside the body by setting p = 0 in ΩS and the force f that we take to be equal to fext inside ΩS . On the other hand Z Z t   µ ∇u + ∇u : ∇φ dx = µ ∇u + t ∇u : ∇φ dx . Similarly Z Z − p div φ dx = − p div φ dx Ω ΩF since div φ = 0 in ΩS . ∂ΩF ∂ΩF . We now compute Z   Z   ∂u ∂u d dL ρ + ρ(u · ∇)u · φ dx = ρF + ρF (u · ∇)u · φ dx + φ̄ · (mV) + φω · Ω ∂t ΩF ∂t dt dt because of (4.11) dt ΩS ΩS ∂t Z Z   d ∂u ρS (x − G) × u dx = ρS (x − G) × + (u · ∇)u dx . though with a different density (ρS instead of ρF ). and noticing that Z Z Z t  µ ∇u + ∇u : ∇φ dx − p div φ dx = (−µ∆u + ∇p) · φ dx ΩF ΩF ΩF Z Z +φ̄ · σ · n dσ + φω · (x − G) × σ · n dσ .

3 Enforcing the constraint Nevertheless. 4. q)  Z  n+1  Z u − un n n n+1 t n+1     ρ + (u · ∇)u · φ dx + µ ε ∇u + ∇u : ∇φ dx   Ω  δt ZΩ Z − pn+1 div φ dx − f · φ dx = 0 . ε > 0 is meant to be a small parameter. Ω . FLUID-STRUCTURE INTERACTIONS we obtain Z   Z Z Z ∂u t  ρ + ρ(u · ∇)u · φ dx + µ ∇u + ∇u : ∇φ dx − p div φ dx − f · φ dx Ω ∂t Ω Ω Ω Z    Z  ∂u d = ρF + ρF (u · ∇)u − µ∆u + ∇p − f · φ dx + φ̄ · (mV) − σ · n dσ − Fext ΩF ∂t dt ∂ΩF  Z  dL +φω · − (x − G) × σ · n dσ − Text dt ∂ΩF = 0. pn+1 ) such that for all (v. The time discretization can then be obtained using one of the strategies that were described in chapter 2. ZΩ ∂t Ω Ω Ω    q div u = 0 . µε (x) = 1 (4. What is the easiest way to take the constraint into account? Should we use special finite element spaces? How to handle such a constraint inside a FreeFem++ code? A possible way that we explain below consists in penalizing the rigid body motion constraint. this variational formulation discretizes the continuous variational formulation  Z   Z Z Z ∂u  µε ∇u + t ∇u : ∇φ dx −    ρ + (u · ∇)u · φ dx + p div φ dx − f · φ dx = 0 .3–4. a typical discrete variational formulation at time nδt for the whole problem will be given by Find (un+1 .7) in a whole is given by testing with functions that are rigid motions inside the solid. Ω Formally. the idea consists in changing the variational formulation in order that the constraint will be naturally approximately satisfied. Namely. Therefore the variational formulation. The situation is very symmetric since at a time t the solution u also belongs to the space of velocity fields that are rigid body motions inside the solid.   Z Ω Ω     q div un+1 = 0 .38 CHAPTER 4.13) otherwise . which encodes the system (4. hence. the question of implementation of the constraint of being a rigid motion inside the solid remains entire. ε In the preceding definition. In order to do so we modify the viscous term in the variational formulation as follows: Z Z   µ ∇u + t ∇u : ∇φ dx −→ µε ∇u + t ∇u : ∇φ dx Ω Ω where the viscosity has been extended to all Ω by ( µ if x ∈ ΩF .

4.4. YET OTHER NON DEFORMABLE CONSTRAINTS 39 for which one has the energy inequality (test formally with φ = u) Z Z Z 1 d .

.

2 ρ|u|2 dx + µε .

∇u + t ∇u.

(4. dx = f · u.14) 2 dt Ω Ω Ω from which we deduce a uniform L2 bound for u and the estimation Z tZ .

.

2 ∃C > 0 . such that ∀t > 0 . µε .

∇u + t ∇u.

15) 0 Ω This leads to Z tZ . dx ≤ C . (4.

.

.

∇u + t ∇u.

This field is still active at the research level. At the limit ε → 0 it seems reasonable to think that the solution converges towards the solution of the previous variational formulation3 4. Therefore. we can use the previous approach. u is approximately a rigid motion. in the solid. This can be also enforced by penalty by replacing again the same term by Z 1 u · φ dx .17) 0 ΩS which ensures that u is close to a constant inside the solid (though this constant may depend on time). and it is expected that the fluid with very high viscosity only deforms very little. if the solid can not rotate but only translate. both fluids never mix. As we know. ε ΩS by Z 1 ∇u : ∇φ dx . 3 This is true although not obvious to prove. ε ΩS Similarly to the former case. the solution u satisfies an estimate Z tZ |∇u|2 dx ≤ Cε (4.4 Yet other non deformable constraints The method described above can be easily generalized to other type of constraints. in the penalized setting by replacing the term Z 1 (∇u + t ∇u) : ∇φ dx .2 dx ≤ Cε . Eventually. proving the very same statement but for Stokes equations is much easier and left as an interesting exercise to the curious reader. . the method above amounts to saying that the solid is a fluid with very high viscosity (actually 1/ε). However. in the continuous setting by restricting both the unknown and the test functions to be constant in the solid ΩS . Physically speaking. a fixed rigid body can be modeled by u = 0 inside the solid.16) 0 ΩS which means that ∇u + t ∇u tends to 0 as ε tends to 0 inside the solid. For instance. due to the incompressibility constraint and the conservation of mass. ε ΩS Proofs of convergence as ε tends to 0 of the corresponding formulation to the constrained Navier-Stokes problem are possible although not obvious. (4.

40 CHAPTER 4. . and a very viscous one which becomes solid when its viscosity tends to infinity.5 Conclusion The aim of this chapter was to design slight changes that can be made in the classical variational formulation of Navier-Stokes equations in order to handle different material. FLUID-STRUCTURE INTERACTIONS 4. We have focused our attention to the case of a non deformable solid in a viscous (incompressible) fluid. and we have seen that it can be seen as a mixture of two incompressible fluids: the original one.

The attachment can be modeled by prescribing that the mean velocity in a ball contained in the rigid part. by prescribing that the strain tensor identically vanishes. Constrained rigid body: consider the typical case of an helix in a fluid flow. and centered at the fixed point. We present here a general approach based on a dual expression of the constraints. In some situations. it may be more relevant to prescribe the global flow rather than the whole velocity profile. is large. 1 The pressure has been introduced in the first chapter according to physical considerations. 3. In particular the pressure will be interpreted as a Lagrange multiplier of the incompressibility constraint. in some cases. Firstly. and necessitates a deep change in the assembling procedure. 1. presents some drawbacks. 3 It means that the condition number. it modifies the underlying opera- tor (like the Laplace operator) by addition of a term in 1/ε. it can be of great interest to prescribe the value of some average quantity. that is the ratio between extremal eigenvalues. 2 This is actually the way it is handled in Freefemm++. Incompressibility: this condition can be seen as a constraint on the fluid velocity1 . and we shall see that it circumvents those problems. 5. 4.Chapter 5 Duality methods This chapter describes an general approach to deal with constraints in fluid mechanics. We shall see in this chapter that it can be formalized from a constrained minimization point of view. Some of these problems can be addressed by the penalty approach presented in the previous chapter. Obstacle: a fixed body in a fluid domain can be considered as a subdomain where the velocity is zero. it can be considered at each instant as a subdomain where the velocity field is that of a rigid motion (translation + rotation). But it can be also considered as a constraint2 . It can be modeled as a rigid body attached at one of its point.g. Secondly. Rigid body: when a rigid object moves in a fluid. No slip conditions: this type of boundary condition has been considered in an essential manner. For example. Global constraints. Yet. the penalty approach significantly affects the structure of the matrix. this approach. by integrating them to the space where the unknown velocity is defined. e. 6. 2. Discretization of the corresponding problem will lead to matrices that are ill-conditionned3 . on the outlet of a pipe flow. 41 . Some typical situations have already been presented. therefore the systems will be harder to solve.

but it can be rigorously proven that the approach is efficient in terms of approximation of the primal part u. for an explanation of this term). µ) = 0 ∀µ ∈ Λ. the corresponding formulation can be discretized in space.1) might be ruled out. In the case where im B ⋆ is closed (which is equivalent to saying that im B is closed). the existence of a solution to (5. existence of a saddle point (u. In general (Λ is infinite dimensional). 2. the minimizer of J over K Interpretation of the Lagrange multipliers in a mechanical context. λ) = hϕ . .e.. i. v) − hϕ . 1 to 2. vi ∀v ∈ V (5. Λ). existence of a saddle-point (u. from which we deduce that (∇J(u). 2 2 i=1 .e. . the saddle-point formulation is still valid. 1 J(v) = (v. . i. 1. It holds that J(u + tv) ≥ J(u) ∀v ∈ K . defined as the kernel of a continuous linear operator B ∈ L(V. λ) is guaranteed.1 General presentation of the approach Consider a smooth functional J in a Hilbert space V . B ⋆ is one-to-one. The latter system is called a saddle point formulation of the minimization problem (see Proposition 10.) by springs with stiffness k and length at rest 0 . since the limit does not exist in Λ. we have (ker B)⊥ = im B ⋆ . The question of the behavior of the discrete Lagrange multiplier when the mesh size goes to 0 is delicate. Note also that if B is onto. in the context of PDE’s. n. . etc. so that there exists λ ∈ Λ such that  ∇J(u) + B ⋆ λ = 0 Bu = 0. λ) implies that u minimizes the quadratic functional J over K (see Proposition 10.1) (Bu. When the image of B is not closed.e. DUALITY METHODS 5. i. (ker B)⊥ = im B ⋆ no longer holds. .42 CHAPTER 5. It can also be written in a variational way:  a(u.. v) = 0 ∀v ∈ K. x). In the case where J is a quadratic functional. linked to each other (0 to 1. Suppose that such a minimizer u exists. thus the Lagrange multiplier is unique. page 68). which leads to a finite dimensional problem for which existence of a saddle point is guaranteed. We want to find a minimizer of this functional on a subspace K. The positions are represented by the vector (x0 . v) + (Bv. page 68. Consider a horizontal system of n + 1 masses 0. where Λ is also a Hilbert space. xn ) ∈ Rn+1 . ∀t ∈ R. but we have (ker B)⊥ = im B ⋆ . . On the other hand. If the dimension of Λ is finite. The potential energy writes n 1 X 1 J(x) = k |xi − xi−1 |2 = k(Ax. Yet. vi. . . x1 . 2 it takes the form  Au + B ⋆ λ = ϕ Bu = 0. ∇J(u) ∈ (ker B)⊥ .

. according to the following procedure 1. The two end lines of this system are k(x0 − x1 ) + λ0 = 0 k(−xn−1 + xn ) + λ1 = 0.. then the sequence un converges to the solution u to the constrained minimization problem. by the wall at 1 upon the mass n). Any configuration (x. 0 0 B= . and the corresponding saddle point problem reads simply  Au + B ⋆ λ = ϕ Bu = z. x0 = 0 . . xn = L} = X + ker B . 0 0 . and the second one writes λn+1 = λn + ρ(Bun+1 − z). λ) (not necessarily unique). 2.1. and if ρ > 0 is sufficiently small (ρ < 2α/ kBk. v) + (Bv. The matrix B writes   1 0 . .. Given a parameter ρ > 0.. xn ) ∈ R2 . We now want to minimize J over the affine space E = {x . Non-homogeneous constraint.1). . The first step of the Uzawa algorithm is unchanged. λ1 ) ∈ R2 such that Ax + Bλ = 0. Uzawa algorithm. and the n − th mass to the point xn = L > 0. up to a multiplicative constant. In the car where the constraint reads Bu = z 6= 0. −λ1 ) as the force exert by the wall at 0 upon the mass 0 (resp. and an initial guess λ0 . Assuming that λn is known. vi∀v ∈ V. . x.5. This algorithm can be shown to converge in the following sense: if the saddle-point problem admits a solution (u. . . where α is the ellipticity constraint of A). it consists in in computing λ1 . a similar approach can be carried out. . These relations express the balance of end masses. avec B : x ∈ Rn+1 7−→ (x0 . Update the Lagrange multiplier λn+1 = λn + ρBun+1 . We shall end this introductory section by describing a practical way to solve the saddle-point problem. λ2 . x) de Rn+1 obviously minimizes this energy Let us now consider the situation where 0 is fixed at x0 = 0. 0 1 There exists λ = (λ0 . The Uzawa algorithm is an iterative procedure to solve Problem (5. and allows to interpret −λ0 (resp. GENERAL PRESENTATION OF THE APPROACH 43 where A is. the discrete Laplacian with Neuman boundary conditions. λn ) = hϕ . compute un+1 as the solution to a(un+1 . .

For instance. 2 Ω Ω over K = ker B .2 The divergence-free constraint for velocity fields This constraint plays a special role in the context of contained minimization. At the continuous level. λ = 1). Ω Ω Thus. and homogeneous Dirichlet boundary conditions. B : v ∈ H01 (Ω)d 7−→ −∇ · v ∈ L2 (Ω). but it is very delicate in terms of implementation. From the mathematical standpoint. Consider the Stokes problem: −µ∆u + ∇p = f . the problem which consists in finding a function in H01 . It can be done in practice. 2. and it is rarely done in practice. The pressure. Yet. with ∇ · u = 0. As the discrete level. has disappeared from the formu- lation. The fact that the constrained is distributed over the whole fluid domain requires to pay a special attention to the well-posedness of the saddle-point formulation. existence of a pressure as a Lagrange multiplier of incompressibility constraint is a consequence of the surjective character of the divergence operator from H 1 (Ω)d to L2 (Ω). directly fits in Lax-Milgram framework. This so called inf-sup (or Babuska-Brezzi) condition expresses in some way that the discrete divergence operator asymptotically behaves like a surjection between two infinite dimensional Hilbert spaces. u1 ) the solution to the primal problem associated to λ = 0 (resp. DUALITY METHODS Small number of constraints. to obtain Z Z µ ∇u : ∇v = f · v. the exact solution can be written u = (1 − θ)u0 + θu1 . When the number of scalar constraints is small. The simplest way to handle it consists in taking divergence free test-functions. denoting by u0 (resp. this approach presents two main drawbacks: 1. . The value of θ is obtained by prescribing that the constrained is met: Bu0 θ= . Bu0 − Bu1 5. good approximation properties required special conditions between the two discretization spaces (in velocity and pressure). Numerical implementation of this approach necessitates to account for the divergence free constraint in the discretization space for the velocity. that has a meaning in terms of modeling. that is divergence free. the Lagrange multiplier can be directly computed. in the case of a unique linear constraint. the pressure can be recovered by considering that the problem consists in minimizing the functional Z Z µ J(v) = |v|2 − f · v.44 CHAPTER 5.

without the use of an iterative algorithm). Ω Z µ · u = 0 ∀µ ∈ L2 (ω). Choose λ0 ∈ L2 (ω).1 Obstacle problem We consider here the flow around a fixed cylinder. Yet.5. since B maps a field in H 1 to its restriction in ω as a function of L2 .3 Other types of constraints 5. Compute un+1 (together with pn+1 ) as the solution to  Z Z Z n+1 n+1  µ ∇u   : ∇v − p ∇·v + λ · v = hϕ . In that case. the Uzawa algorithm on λ writes: 1. . ω ω and the Uzawa algorithm can be written accordingly. and it can even be proved that the discretized velocity field uh converges to the exact solution. The (double) saddle-point formulation can be written Z Z Z Z µ ∇u : ∇v − p∇ · v + λ·v = f ·v ∀v ∀v ∈ H01 (Ω).3. the approach can be shown to be work in practice. We denote by ω the subdomain occupied by the cylinder. Ω 3. Note that the continuous saddle-point formulation is not well-posed. Ω Ω ω ω Z q∇ · u = 0 ∀q ∈ L2 (Ω). Another approach would consist here in defining B as the restriction operator to the space H 1 (ω)d . vi ∀v ∈ H01 (Ω) Ω Ω ω Z   − q∇ · un+1  = 0 ∀q ∈ L2 (Ω). 2.3. Compute λn+1 as λn+1 = λn + ρun+1 |ω . the saddle-point problem writes Z Z Z Z Z µ ∇u : ∇v − p∇ · v + λ·v+ ∇λ · ∇v = f ·v ∀v ∀v ∈ H01 (Ω).e. OTHER TYPES OF CONSTRAINTS 45 5. Consider the Stokes problem in a domain Ω \ ω around the cylinder. and homogeneous Dirichlet boundary condition on the external boundary ∂Ω. Ω Z Z µ·u+ ∇µ · ∇u = 0 ∀µ ∈ H 1 (ω). Ω Ω ω Z q∇ · u = 0 ∀q ∈ L2 (Ω). ω If one considers that the Stokes problem can be directly solved (i.

Prescribing a zero velocity at the center is not sound in the present context. where the point lies within a zone that is considered rigid. Consider the case where the disc is attached at its center. Z u = 0. It would lead to an extra term in the Stokes (or Navier-Stokes) variational formulation of the type Z  Z  1 u v . u1 ) the velocity field corresponding to λ = 0 (resp. It is more appropriate to prescribe a zero mean velocity over the disc.3 Prescribed mean velocity Consider a Stokes (or Navier-Stokes) flow in a domain Ω. which makes it tricky to formalize properly.2 Prescribed flux Consider the flow in a pipe-like domain. We aim at prescribing the flux through Γout . and that verify Z v · n = Q. say. it would actually make sense. Indeed. and is free rotate around it. it is not necessary to run the Uzawa algorithm. since pointwise values are not4 defined in H 1 .e. with two end boundaries Γin and Γout .3. Ω Ω Γout Z q∇ · u = 0 ∀q ∈ L2 (Ω).46 CHAPTER 5.e. a disc (subdomain ω ⊂ Ω). . DUALITY METHODS 5. the exact solution u can be written u = (1 − λ)u0 + λu1 . it consists in minimizing the functional Z µ J(v) = |∇v| 2 2 Ω among all those fields v that are divergence free.3. denoting by u0 (resp. Γout where Q is prescribed. Γout Γout u1 · n − Γout u0 · n 5. ω The could be done by penalizing this quantity. λ = 1). without prescribing the whole velocity profile. i. But it would involve linear functionals that are not defined over the unconstrained space. For the Stokes problem. coupled with the motion of a rigid object. ε ω ω 4 In the present case. Ω Z µu · n = Qµ ∀µ ∈ L2 (Γout ). The (double) saddle-point formulation can be written Z Z Z Z µ ∇u : ∇v − p∇ · v + λv · n = f · v ∀v ∈ H01 (Ω). Z R Q − Γout u0 · n µu · n = Q =⇒ λ = R R . Γout Note that. The value of λ is then obtained by requiring that the constraint is verified. i. in this situation.

Such a mesh can be build in the following way . we will handle the rigid body constraint by following a penalty approach. 0 0 1 and by u00 . as in the previous case. Finally. The penalty method does not require a conforming mesh. In order to illustrate the possibility to couple different approaches.5. λ01 = . By linearity of the Stokes problem. ε ω ω Z q∇ · u = 0 ∀q ∈ L2 (Ω). Yet. This motivates the use of a duality method to account for this constraint. it cannot be straightforwardly implemented in a software like freefem++. −hΨε . so that a direct strategy can be followed: Denote by       0 1 0 λ00 = . vi is the power that the forces exerted on the fluid to maintain a quasi rigid motion would develop. ε ω For any test function considered as a virtual velocity field. λ = (λ1 . we obtain the variational formulation (for the Stokes problem) Z Z 1 µ (∇u + t ∇u) : (∇v + t ∇v) − p∇ · v+ 2 Ω Ω Z Z Z 1 t t (∇u + ∇u) : (∇v + ∇v) + λ · v = f · v ∀v ∈ H01 (Ω)2 . i. verifies the (two) constraints. obtained by writing that u (expressed as above). consider the linear form Z 1 Ψε : v 7−→ hΨε . λ2 )T can be determined as a solution to a 2 × 2 linear system. Implementation issues. using a conforming mesh (i.e. the approach presented above can be discretized over a conforming mesh. the variational term in λ corresponds to an extra term in the right-hand side of the momentum equation: · · · = −λ1ω . In case of homogeneous Dirichlet boundary conditions on the outer boundary. the exact velocity writes u = u00 + λ1 (u10 − u00 ) + λ2 (u01 − u00 ) . u10 . The corresponding λ accounts for the force exerted by the fluid on the particle. More precisely. ω Note that. OTHER TYPES OF CONSTRAINTS 47 Note that this term will couple all the degrees of freedom associated to vertices within the particle. Interpretation as forces. the penalty term can also be interpreted as a force. Indeed. a mesh that covers the whole computational domain (including ω). In this spirit.3. the space of Lagrange multiplier has low dimension (2 in the present case). vi = (∇u + t ∇u) : (∇v + t ∇v). Ω Z µ· u = 0 ∀µ ∈ R2 . so that |ω| λ is the resultant force exerted by the obstacle on the fluid. Following the principle of virtual power. u01 the corresponding velocity fields (solutions of the Stokes problem with the correspond choice of λ). without any element crossed by the boundary) can improve space accuracy. It follows immediately that the corresponding force field has a zero resultant and a zero torque. but that contains a piecewise approximation of the disc boundary.e. λ10 = . Besides.

y=yB+rB*sin(t).} mesh Th=buildmesh(Gs(L*Np)+Ge(Np)+Gn(L*Np)+Gw(Np)+Gin(NB)). DUALITY METHODS int Np = 20.label=5. Note that Gin(NB) is written without a minus sign. .y=1.y=0.} border Ge(t=0.L){x=t. NB = 2*Np . To compute integrals over the discretized disc.1 .5 .L){x=L-t.2*pi){x=xB+rB*cos(t). border Gs(t=0.label=1.y=1-t. so that a mesh will be created within the particle.} border Gw(t=0. rB = 0. yB = 0.label=2.label=3.48 CHAPTER 5.1){x=0. a robust way consists in defining the characteristic function of the disc as a piecewise constant function. and to multiply the integrand by this function when the integration is to be limited to the disc.} border Gn(t=0.label=4.5 . real L = 2.} border Gin (t=0.1){x=L.y=t. xB = 0.

vi. v) + b(v.Chapter 6 Elements of numerical analysis for saddle-point problems 6. The formulation reads  a(u. p) can also be written hB ⋆ p . vi 2 49 . Note that b(v.2 consists in minimizing 1 J(v) = a(v. µ) ∈ V × Λ 7−→ b(v. µ) . µ) = b(v. and the discrete constrained space is denoted by Kh = Vh ∩ ker Bh = {vh . 2 defined in a Hilbert space V . approximation spaces for V and Λ. Λh ) by (Bh v. ph ) = hϕ . We define the operator Bh ∈ L(V. µh ) ∀µh ∈ Λh . vh i ∀vh ∈ Vh (6. µh ) = 0 ∀µh ∈ Λh } The problem 6. respectively. One denotes by B ∈ L(V. when the minimizers is subject to belong to a linear space that is written as the V -kernel of a bilinear form: (v. p) = hϕ .1) b(u. v) − hϕ .1 Mixed finite element formulations We consider here the general saddle-point formulation of the problem which consists in mini- mizing a quadratic functional 1 J(v) = a(v. v) − hφ . vi.2) b(uh . Λ) the operator defined by (Bv. µ). K = {v ∈ V . vh ) + b(vh . µh ) = (Bv. vi ∀v ∈ V (6. b(vh . µh ) = 0 ∀µh ∈ Λh . µ) = 0 ∀µ ∈ Λ. µ) = 0 ∀µ ∈ Λ} . It is natural to introduce Vh and Λh . b(v. and the associated discretized problem  a(uh .

Λ).3) (Bh uh . Remark 5 In the case of Stokes problem.1. so that the saddle-point problem may be ill-posed in terms of existence of a Lagrange multiplier. Of course. vi 2 . it would be very natural to approximate velocities with continuous. in case it is uniquely defined. Extension of the theory developed for standard (unconstrained) variational problems is in fact not straightforward. Concerning approximation spaces. but not necessarily for the Lagrange multiplier ph . The question is now: assuming that (Vh ) and (Λh ) are suitable sequences of approximation spaces. vi 2 over K is well-posed by Lax Milgram theorem. but Λh is any finite dimensional space. Although this estimate will not be used as such in the context of divergence free constraint. Λh ) both go to zero as the discretization parameter h goes to 0. i. we shall start the numerical analysis by a very general estimation of the sole primal quantity. 6. e. As we shall see. v) − hφ . vh i ∀vh ∈ Vh (6. so that the solution uh to the discrete constrained minimization problem might converge (if ever) outside K.e. The problem which consists in minimizing 1 J(v) = a(v.2 is obviously well-posed. ph ) = hϕ . qh ) = 0 ∀qh ∈ Λh .e. piecewise P 1 functions (as for the poisson problem). and we denote by u its solution. we simply assume that Vh ⊂ V . Uniqueness holds for the primal variable uh . for reasons that we shall detail below. while pressures would be approximated by P 0 functions. v) − hφ . the problem might be underconstrained. for the fictitious domain method.50 CHAPTER 6. We assume here that a(·. We do not make extra assumptions on B. as the velocity is H 1 and the pressure L2 . On the other way. but it is not necessary for the time being. Vh ) and dist(q. vh ) + (Bh vh . Both approximations have obviously to be balanced in some way. this straightforward approach does not work. so that Kh does not approximate V properly (it might even be reduced to {0}).g. Considering finite dimensional approximation spaces Vh and Λh . Kh has finite di- mension. can uh be expected to converge to the solution u of the continuous constrained minimization problem ? As for the dual part ph . it will be necessary to make extra assumptions on Λh and Bh to obtain useful estimates. symmetric and coercive bilinear form over V × V . if Λh is much “richer” than Vh in terms of approximation. which will be expressed by the discrete inf-sup (or BBL) condition.1 A general estimate without inf-sup condition To enlight how this inf-sup condition allows a proper error estimation of primal and dual quanti- ties (i. if Λh is too poor. at least in terms of existence. and Bh ∈ L(V. so that K = ker B is closed. |u − uh | and |p − ph |). it proves powerful in other contexts. one may also wonder whether it converges to p. Λh ). This problems consists in minimizing 1 J(v) = a(v. the discrete problem might be overconstrained. STOKES EQUATIONS over Kh . therefore it is closed. so that Problem 6. dist(v. and B ∈ L(V. ·) is a continuous. We consider the (abstract) discretized saddle-point problem:  a(uh . Informally. This balance necessitates a fine tuning between the two discretization spaces. under very loose conditions (in particular without the inf-sup condi- tion).

ph ) = 0. We denote by u the unique minimizer of J over K = ker B. Let us start by a very simple remark on the continuous problem: there exists a unique ξ ∈ V ′ (more precisely in K ⊥ ) such that Au + ξ = φ. vh i. vh i. Λh ). vh ) = hφ . any qh ∈ Λh . vh ) + hξ − Bh⋆ µh . symmetric.6. vh ) = hφ . Let Vh ⊂ V and Λh be finite-dimensional spaces. vh i. The basis of the following estimate will be to prove that. vh i = hφ .4) wh ∈Kh qh ∈Λh h Proof : For any vh ∈ Kh . Let (uh . (Bh vh . As u is uniquely defined. Then   ⋆ |u − uh | ≤ C inf |wh − u| + inf kξ − Bh qh kV ′ (6. some convergence of uh toward u can be expected as soon as ξ can be approximated by Bh⋆ µh . where wh is in Kh . and uh is uniquely defined as the minimizer of J over Kh . so that this quantity can be substracted to the right-hand side. B ∈ L(V. But if Λh is too rich. ph ). which ends the proof. the saddle-point problem is well-posed in the following sense: it admits a solution (uh .1. coercive. It comes a(vh . so that a(uh . As vh is in Kh . vh i = (Bh vh . As Λh is finite dimensional. vh ) = a(u − wh . µh ) = 0 for any µh ∈ Λh . φ ∈ V ′ . Λ). vh i − a(wh . ·) be a continuous. MIXED FINITE ELEMENT FORMULATIONS 51 over Kh = ker Bh . h for any wh ∈ Kh .4) expresses the required balance between both approximation spaces. hBh⋆ qh . so that a(vh . Proposition 5 Let a(·. vh ). and Kh = ker Bh ∩ Vh . The core of the proof consists in taking vh in the form uh − wh . it may constrain excessively the problem. Λh has to be rich enough (and Bh has to approximate B in some sense) for the second term to be small. a(vh . so that Kh may not be a good approximation space for K.2). vh ) = a(u − wh . Firstly. preventing the first term to go to 0. vh ) + hξ . ξ is uniquely defined as φ − Au. bilinear form over V × V . vh ) + hξ . One obtains α |vh |2 ≤ kak |u − wh | |vh | + kBh⋆ qh − ξk |vh | ⇒ α |uh − wh | ≤ kak |u − wh | + kBh⋆ qh − ξk and therefore   |uh − u| ≤ C |u − wh | + kBh⋆ qh − ξkV ′ . The exact solution u verifies a(u. ph ) be a solution of the saddle point-problem (6. . even if ξ cannot be written B ⋆ p. Bh ∈ L(V. Estimate (6. vh i.

β yh |yh | β 1 As always in such contexts. we will show that the estimate on |u − uh | induces an estimate on |p − ph |. we build an approximation wh = vh + zh ∈ Kh with the same approximation properties (up to a multiplicative constant which does not depend on h). p). i.1. qh ) inf sup ≥ β > 0. we implicitly consider sequences of approximation spaces (Vh ) and (Λh ). We consider here a saddle point formulation which is well-posed at the continuous level.2 Estimates with the inf-sup condition We may now introduce the framework which we will favor for Stokes equation.5) wh ∈Vh qh ∈Λh Proof : The proof is based on estimate (6. As preliminary step.e. Step 1. wh = vh + zh is the projection of vh onto Kh ). qh ∈Λh vh ∈Vh |vh | |qh | where β is independent1 of h. Bh vh ) ∀µh ∈ Λh . consists in showing that the first infimum in (6.4) can be replaced by an infimum over the unsconstrained approximation space Vh . indexed by a parameter h (which will represent the mesh diameter in actual Finite Element discretization. . We also assume that both approximations are conforming. we proceed as follows: consider an approximation vh ∈ Vh of u. B is surjective. We furthermore assume here that B ∈ L(V. qh ) = (Bv.e.52 CHAPTER 6. yh ) + (ηh .4). qh ∈Λh qh ∈Λh The first step of the proof. which goes to 0 as the dimension of the spaces goes to infinity. Let vh ∈ Vh be given. As for the first term in estimate (6. Λ) is surjective. In a second step. This zh is the primal part of the solution to the saddle point problem (zh . let us first note that the infimum over Λh of kBh⋆ qh − ξkV ′ takes the form (as ξ = B ⋆ p) h inf |B ⋆ (qh − p)| ≤ inf kB ⋆ k |qh − p| . that Bh is defined by (Bh v. We denote by zh the element of Vh which verifies Bh zh = −Bh vh and which minimizes the norm (i. Proposition 6 Notations and assumptions are those of Proposition 5. Bh yh ) = 0 ∀yh ∈ Vh (µh . yh i 1 |ηh | ≤ sup h = |zh | . qh ) = b(v. Thanks to the inf-sup condition. Bh zh ) = −(µh . and that it verifies the discrete inf-sup condition: b(vh . so that the continuous saddle point problem admits a unique solution (u. i. which is essential to use approximation properties of Vh . we have 1 hB ⋆ ηh . Vh ⊂ V and Λh ⊂ Λ. qh ).4).e. Then we have the following error estimate   |u − uh | + |p − ph | ≤ C inf |wh − u| + inf |qh − p| (6. STOKES EQUATIONS 6.

Bvh ) ∀qh ∈ Λh . By the inf-sup condition. it comes |zh |2 ≤ |(ηh . Step 2. with |Πh v| ≤ C |v| . and by taking wh = vh in the first line. Λ) verifies the inf-sup condition. The new approximant wh = vh + zh is then such that |u − wh | ≤ |u − vh | + |zh | ≤ C |u − vh | . we have 1 |a(u − uh . vh ) + (p − qh . qh ) = 0 ∀(v. 2 As always. The saddle point formulation of the problem writes (vh . this assertion refers to a family (Πh ) of operators which is uniformly bounded with respect to h. Bh zh )| ≤ |ηh | |Bh vh | = |ηh | |Bh (u − vh )| ≤ C |zh | |Bh (u − vh )| . Bvh ) = a(u − uh .6. Proof : Assume that b(·. We define Πh v as vh (orthogonal projection onto B ⋆ (Λh ). Bv) ∀qh ∈ Λh .1. Bvh )| |ph − qh | ≤ sup β vh ∈Vh |vh | 1 ≤ (kak |u − uh | + kBk |p − qh |) . β so that. so that. To estimate |p − ph |. finally. wh ) + (ph . By the inf-sup condition. with Πh ∈ L(V. taking yh = zh . vh ) + (p. MIXED FINITE ELEMENT FORMULATIONS 53 On the other hand. which is not easily verified in its native form. we obtain |vh | ≤ C |v|. with a bounded norm2 . we substract the discrete variational formulation from the con- tinuous one. qh ) ∈ V × Λh . such that b (Πh v − v. Vh ). Bwh ) = 0 ∀wh ∈ Vh (qh . ·) verifies the inf-sup condition. we build vh which mini- mizes the norm among all those wh that verify Bwh = Bv (as in the beginning of Proposition 6). For any v ∈ V . Then the sequence (Vh . Λh ) verifies the discrete inf-sup condition iff there exists a constant C > 0 and a family (Πh )h . More precisely. We get (ph . β β qh ∈Λh The discrete inf-sup condition. Bvh ) = (qh .   kak kBk |p − ph | ≤ |u − uh | + 1 + inf |p − qh | . for any qh ∈ Λh (ph − qh . is equivalent to the existence of an operator from V to Vh which preserves the discrete constraint. we have Proposition 7 (Fortin’s criterium) We assume that B ∈ L(V. one has |ph | ≤ |vh | /β. Bvh ) = a(u − uh . Bvh ). . vh ) + (λ − µh .

For a given v in H01 (Ω)d . and we shall keep the same notation Πh to denote the operator which acts on scalar functions. 6.3 Some stable finite elements for Stokes equations A finite element (Vh . Λh ) defined previously is stable. The simplest stable element is the so called mini element. the second sum goes over all simplices of Ωh . consider qh ∈ Λh ⊂ Λ. such that the mean value of each of its component equals the mean value of the corresponding component of v. The family (Vh . We have b(vh . 7): we give en explicit construction of Πh . Consider . neither (P 1 . It is actually built component by component. we shall have Z Z b(Πh v − v. denote by v its reciprocal image by B which minimizes the norm. Let us now detail how Πh can be built. Let us denote by wK the bubble function associated to element K. The approximation space for pressure is simply ( ) X Λh = qh = qi wi i Proposition 8 We consider a polygonal domain. STOKES EQUATIONS On the other way. This norm is controlled by that of qh because the range of B is closed. its gradient is constant in each element K. Vh is defined as the set of all those vector fields which can be written d ! X X X uh = uαi wi eα + α βK wK eα α=1 i K where the first sum goes over “free” vertices (vertices of Ωh which do not belong to a boundary on which Dirichlet boundary condition is prescribed). It is obtained from the unstable family (P 1 . Proof : The proof is based on the Fortin’s criterium (see Prop. qh ) b(Πh v. nor (P 1 . qh ) 1 sup ≥ ≥ |qh | . For a triangulation Ωh . the idea consists in building Πh v with a norm controlled by the norm of v (uniformly with respect to h). and a regular family (Ωh ) of triangulations. Among natural candidates. P 1 ) by adding a degree of freedom in the center of each element. and wi is the hat function associated to vertex i. Bubble element. Ω Ω As qh is piecewise affine. P 1 ) (piecewise affine functions for the velocity and the pressure). qh ) = qh ∇ · (v − Πh v) = − (v − Πh v) · ∇qh . P 0 ) (piecewise constant functions for the pressure) meet this require- ment. Λh ) for the mixed formulation of Stokes problem is said to be stable when- ever it verifies the discrete inf-sup condition uniformly with respect to the discretization param- eter h. which is the product of the barycentric coordinates.54 CHAPTER 6. so that the latter quantity is 0 by construction. It is usually performed by considering the so-called bubble function associated to each element. vh |v h | |Π h v| C which ends the proof.1. If we are able to build such an operator. or bubbleelement.

where wh is a bubble function. To overcome this difficulty. K K Let us denote by wh the mean value of wh over K. MIXED FINITE ELEMENT FORMULATIONS 55 v ∈ H01 (Ω). The difficulty is that standard pointwise interpolation is impossible. One has R . we consider instead Rh v. More precisely.K ≤ ChK |v|1.1. for any element K. The approach consists then in correcting Rh v within each element (using the bubble degree of freedom) to respect the mean value. chosen in such a way that Z Z Πh v = v. where ∆K is the set of elements in contact with K. we write Πh v = Rh v + wh .6. The first step consists in associating a piecewise affine function vh which approxi- mates v. which is such that |v − Rh v|0.∆K . where Rh is the so-called Clément operator (the existence of which we admit here). as pointwise values of v are not defined (as soon as d ≥ 2).

Z Z .

K wh .

−1 .

.

wh = = |K| .

v − Rh v .

.

e. which means that B is surjective. It requires to restrict pressure fields to scalar fields with zero mean value.Ω ≤ |v|1.Ω ≤ C |v| 1.K ≤ C |K|−1/2 hK |v|1. or P 2 − P 1 . the set of discrete pressures is that of piecewise affine fields with zero mean value over Ω. ≤ |K|−1/2 |v − Rh v|0. Taylor-Hood element. which ends the proof Remark 6 We assumed that B verifies the continuous inf-sup condition.Ω . one commonly considers a modified saddle-point problem of the form Au + B ⋆ p = φ. As the semi-norm in the right-hand side involves a neighborhood of K. we have |wh |1. An higher order element is the so-called Taylor-Hood.K ≤ Ch−1 K w h |K| . in the case where Neuman conditions are prescribed on some part of the boundary. In practice. element. the H 1 semi-norm of wh verifies 1/2 |wh |1. as hat functions do not verify this constraint. We finally have |wh |1. whereas the pressure is piecewise affine. As suggested .K ≤ C |v|1. hK /ρK is uniformly bounded). We therefore have |Πh v| 1. Λh ) verifies directly the inf-sup condition (without any additional requirement on discrete pressure fields). This constraint is a priori difficult to handle. . for elements with a controlled aspect ratio (i. Note that.∆K . so that the couple (Vh . Bu + εp = 0. then the pressure is no longer defined up to a constant.∆K |K| K K Besides. It is based on a quadratic description of the velocity in each element. where ε > 0 is a small parameter.Ω . As a consequence.

56 CHAPTER 6. STOKES EQUATIONS .

5.Chapter 7 Complements 7.5) (radius 0. Γ2 . 57 . Γ3 Γ4 Γ2 Γ1 Figure 7. Γ3 .1 Estimation of interaction forces We consider the situation illustrated by figure 7. and free outlet boundary conditions on the rest of the boundary µ∇u · n − pn = 0 on Γ1 . according to Navier-Stokes equations   ∂u ρ + (u · ∇) u − µ∆u + ∇p = 0 in Ω ∂t ∇ · u = 0 in Ω Boundary conditions will be taken as follows: Dirichlet boundary conditions on the inlet bound- ary Γ4 (uniform horizontal velocity). A newtonian fluid flows around a fixed cylinder centered at (0. The time discretized problem leads to the generalized Stokes problem ρ ρ u − µ∆u + ∇p = ũ ◦ X in Ω (7.1. No slip conditions (u = 0) will be prescribed on the boundary of the disc.1: Flow around a cylinder We propose here to use the method of characteristics to handle the advection term. 0.1). the domain is a 2 × 1 rectangle.1) δt δt ∇ · u = 0 in Ω.

therefore a relevant energy balance can be expected. we consider a smooth vector field v that is ex on γ and 0 on other parts of the fluid domain boundary. In the expression of the horizontal (drag) force Z D = (µ∇u · n − pn) · ex . Homogeneous Dirichlet boundary conditions.1). δt δt Ω Ω The form we obtain is exactly the variational formulation of the force balance equation. COMPLEMENTS where ũ is the velocity at the previous time step. The simplest situation is that of a Navier- Stokes fluid in a closed container Ω. 7. This section describes the way such an energy blanches can be established in some typical situation.58 CHAPTER 7. δt δt so that D can be expressed Z Z ρ ρ D = u − ũ ◦ X + ∇u : ∇v − p∇ · v. by re-using the implemented variational formulation. The first term can be written Z d ρ 2 |u| .which yields Z Z Z Z ∆u · v + ∇u : ∇v − ∇p · v − p∇ · v. checking that a “numerical” energy balance can be obtained is a efficient way to check the physical relevancy of the computed solution. γ ex can be replaced by v. dt Ω 2 . so that effective computation of the force is straightforward under freefem++. Multiplying the momentum equations by the velocity itself and integrating over the domain yields Z Z Z Z Z ρ ∂t u · u + ρ ((u · ∇)u) · u − µ ∆u · u + Ωu · ∇p = f · u. Ω Ω Ω Ω Thanks to Eq. we have that ρ ρ ∆u − ∇p = u − ũ ◦ X. We aim at estimating the forces exerted by the cylinder on the fluid. This standpoint can be fruitful in the context of numerical simulation. Ω Ω Ω Ω The right-hand side is the power of external forces. Since an analytical solution is rarely known.2 Energy balance The models proposed in these lecture notes express standard physical laws (mainly mass con- servation and Newton’s law). with no-slip conditions conditions on the wall. In can be integrated by parts. Freefem++ makes it possible to directly estimate the boundary integral Z (µ∇u · n − pn) . To that purpose. γ but it is more stable to use a volume integral. (7.

with two main differences: since the domain is moving according to the fluid velocity u . at constant pressure P . thanks to (7. We consider the free fall a drop of viscous fluid:  ∂u    ρ + ρ (u · ∇) u − µ△u + ∇p = g in Ω(t)  ∂t  ∇·u = 0 in Ω(t)    σ · n = µ(∇u + t ∇u) · n − pn = −P n on Γ(t) We multiply again by the velocity u. integrate over Ω. Ω Ω Ω Ω Γ(t) =0 .2). Free boundary problem.2) dt Ω(t) Ω(t) Γ(t) Besides. the quantity Z d ρ 2 |u| . and its boundary is assumed to separate the viscous fluid in consideration from an outside fluid that is assumed to be perfect. we can use the formula 1 (u · ∇)u = ∇ |u|2 − u × (∇×)u. ENERGY BALANCE 59 it is the time derivative of the kinetic energy. (7. We finally obtain that the time derivative of kinetic energy equals the power of external forces (rate of energy injected into the system) minus the rate of energy lost as heat by viscous effects. t)u · n. Ω Ω 2 Γ 2 that vanishes since u is zero on the boundary. the time derivative of an integral quantity writes Z Z Z d Ψ(x. t) = ∂t Ψ(x. Ω Ω it corresponds to viscous dissipation. Putting things together. dt Ω 2 that is the time derivative of the kinetic energy. t) + Ψ(x.7. the term Z Z Z 1 ρ 2 ρ ((u · ∇)u) · u = ρ ∇ |u|2 u = |u| u · n Ω Ω 2 Γ 2 is no longer 0. The viscous term is Z Z −µ ∆u · u = µ |u|2 .2. Let us consider a richer situation: the fluid domain moves. we obtain Z   Z ρ  Z ρ ∂u ρ + ρ (u · ∇) u · u = ∂t |u|2 + |u|2 u · n Ω ∂t Ω 2 Γ 2 that is exactly. The power of external forces writes Z Z Z Z Z g · u = − gey · u = − g∇y · u = gy ∇ · u | {z } − yu · n. 2 which leads to Z Z Z 1 ρ 2 ρ ((u · ∇)u) · u = ρ ∇ |u|2 u = |u| u · n. For the second one.

60 CHAPTER 7. Γ(t) dt Ω(t) We finally obtain that the derivative of the total energy (kinetic + potential) balances the viscous losses: ! Z Z Z d ρ 2 µ .2)) Z Z d − yu · n = − gy. COMPLEMENTS The latter quantity can be written (using again (7.

.

.

2 |u| + gy = − ∇u + t ∇u.

The corresponding boundary condition can be written σ · n = −κC n. Another approach consists in solving a variational problem discretized in the space of piecewise affine functions. short range forces between the fluid induce a non zero resulting force whenever the interface is not flat. This effect is delicate to be accounted for numerically. R ds . This can be done under Freefem++ by directly computing the partial derivatives: om = dy(ux)-dx(uy) Since the discrete velocities are not C 1 over the domain. An alternative way consists in representing the vorticity. and on the chemical state of the interface itself. The effect can be modeled by a pressure drop through the interface. The approach is based on elementary calculus in differential geometry. This coefficient depends of the properties of both fluids. Consider the case of a (two-dimensional) viscous fluid flowing freely in a perfect fluid. the inverse of the radius of curvature.e. Let us show here that there is a natural way to integrate it in the finite element method. Defining it as a P 1 function will force freefem++ to perform some interpolation to define nodal values. This curvature will be considered positive in case of a convex body. i. the interface is the whole boundary of the domain. so that the force points toward the center of the radius of curvature. i. namely the formula n d Cn = = . . this function is a priori discontinuous. It can be defined in a natural way as a P 0 function.e. We shall consider the case of a boundary with no ends. where C is the curvature. that is scalar in 2d: ω = ∂uy /∂x − ∂ux /∂y.4 Surface tension When two immiscible fluids are separated by an interface. which we will consider at pressure zero. 7. This pressure drop is the product of the mean curvature of the interface and a coefficient κ. which rules out the possibility to draw isolines.3 Visualization The graphic representation of the velocity field is often not sufficient to give a clear idea of the fluid flow structures. dt Ω 2 Ω(t) 2 7. since it involves a quantity that involves the second derivative of the boundary.

and s the curvilinear abscissa. . SURFACE TENSION 61 where is the tangent vector.7.4. Now consider the term that corre- sponds to the surface tension effect in the variational formulation: Z Cn · v. Γ where v is a vector test function.

62 CHAPTER 7. COMPLEMENTS .

Check your Freefem++ implementation by comparing your results in terms of drag and lift forces.u-psud. in a specific geometry.1 Which mesh for which Reynolds number ? Given a problem to solve. At the end. penalty. Investigate the roles of the time step. This is intentional since we expect the students to bring the chosen subject into personal directions. and to accurately compute relevant quantities.1 General 8.math. and perpendicular to this direction (lift). we present subjects that are open to students for personal projects.g.Chapter 8 Projects In the following. the mesh diameter.BenchmarkTurek. 8. It is expected that the students choose one of them and spend 10 to 15 hours of personal time working on them.2 Drag and lift of an obstacle Describe a method to compute the two components of the instantaneous force exerted by a Navier-Stokes fluid on an obstacle: in the direction of the flow (drag). saddle-point formulation).fr/~maury/paps/NUM_NS. e.pdf for the 2d flow around a cylinder. By “right” we mean the cheapest one that allows to properly describe the features of the flow. Both professors are available for discussion upon request during the total duration of the projects. and possibly the chosen approach (direct approach. 8. 63 .1 Benchmark A benchmark is proposed in http://www.2. propose an approach (possibly a heuristic/experimental one) to determine the “right” mesh to use in order to provide relevant simulations.1. All the subjects are described very shortly. 8. the flow around a cylinder at a certain Reynolds number (say Re=1000). they are asked to present their own work in a short 15 mn defense.

3 oscillating cylinder Propose a numerical approach to compute the problem presented at page 25 in http://pastel. 8. when the disc is assumed to be connected to a fixed point by a spring of a certain stiffness and a zero length.4 Free surface 8. .fr/docs/00/54/59/37/PDF/Duclercq-Marion_these.4. by considering the tin as a perfect fluid. We aim here at modeling the motion of the glass sheet over the tin bath.3. Investigate the interaction between the fluid flow and the moving object. with a very small viscosity).pdf 8. 8.3 Moving objects 8.2 Density waves Propose a way to simulate the motion of two layers of immiscible fluids.e. PROJECTS 8. Investigate the effect of bumps in the supporting surface upon the shape of the free boundary. The case of an elongated body (ellipse or rectangle) is of particular interest. and pulled mechanically at the other end of the bath. investigate the dependence of the mean lift force upon the angle of attack.3. 8.2 Fluid structure interaction Propose a way to compute the motion of a rigid disc in a Navier-Stokes flow. in the case where the Reynolds number is sufficiently high to induce a oscillations downstream the cylinder.1 Free fall Propose and implement a numerical strategy to compute (in 2d) the motion of a rigid object (different from a disc) falling freely in a Navier-Stokes fluid. with a pressure that is considered hydrostatic.1 Glass forming process The production of flat glass is based on the so called float process: glass at high temperature (i.4. one on the top of the other.64 CHAPTER 8.2. and a given regime (choice of a Reynolds number between 20 and 500).3 Inclined plane Situation: a viscous and inertial fluid flows down an inclined plane.2. 8.3 Airfoil For a given shape of airfoil. 8. and investigate the propagation of waves on the interface between the two fluids.archives-ouvertes. 8. a light one and a heavier one. liquid glass) is poured on a glass of tin (liquid metal.3.2 Rotating cylinder We consider that the obstacle is a rotating cylinder (disc in 2d).4. Investigate numerically the lift induced by the rotation of the cylinder around its center.

due to the tendency of warm fluid to go up. 8.4 Waves on the beach Investigate the effect of a depth change upon a traveling water wave. whereas its top is maintained at 0 (no-flux conditions on the lateral boundaries). Investigate numerically the so-called Rayleigh Bénard instabilities.5 Rayleigh-Taylor instability A container is filled with two fluids of different densities. and raises up to another height in the right hand part of the domain. 8. If the free surface between them were flat.8.5 Free convection We consider the following situation: The bottom of a container filled with a viscous fluid is maintained at temperature 1.4. FREE CONVECTION 65 8. the heavier being on top of the lighter.4. the situation would be stationary (though highly unstable). . At time t = 0. Give a method which can compute this evolving situation. but the surface between both fluids is sinusoidal. we thus consider that the velocity vanishes. Consider for example the case of a bottom profile that is flat on the left hand of the domain.5.

PROJECTS .66 CHAPTER 8.

there exists a positive constants M < +∞ and α > 0 |a(u. minimization under constraint We collect here some basic properties of Hilbert spaces. vi ∀v ∈ V. ·) a bilinear form which is continuous and elliptic. v) = hϕ . We consider the following set of assumptions . there exists a unique u such that a(u. v)| ≤ M |u| |v| . and we refer e. vi ∀v ∈ V. and a(·. Theorem 2 (Riesz. (A.Fréchet ) Let V be a Hilbert space and ϕ ∈ V ′ a continuous linear functional over V . In this case. v) = hϕ . to [?] for a more detailed presentation.Appendix A Appendix A.1 Hilbert spaces. u) ≥ α |u|2 . a(u.e. continuity and ellipticity ensure that the corresponding scalar product defines a norm which is equivalent to the native one. the solution u can be characterized as the unique minimizer of the func- tional 1 v 7−→ J(v) = a(v.g. i.1) If a(·. 2 Note that. There exists a unique u such that (u. v) − hϕ . vi. the Lax-Milgram theorem is a straightforward consequence of Riesz-Fréchet theorem. ·) is symmetric. in case the bilinear form is symmetric. For any ϕ ∈ V ′ . Theorem 3 (Lax Milgram) Let V be a Hilbert space.

 .

V and Λ Hilbert spaces.  .

  .

 .

elliptic functional in V × V . symmetric. a(·. ·) bilinear.    .

  .

 .

Λ). B ∈ L(V.  . ϕ ∈ V ′ .

.

2) . (A.

Bu = 0}   . K = ker B = {u ∈ V .

  .

  .

1   .

J(v) = a(v.   . vi. v) − hϕ .

2  67 .

∇J ∈ (ker B)⊥ . We shall restrict this section to the simplest case.g. i.68 APPENDIX A. ·) : V × Λ 7→ R. (A.e. we simply have im B ⋆ ⊂ (ker B)⊥ .e. (u. we have (ker B)⊥ = im B ⋆ . (u. In the finite dimension setting. Proposition 12 In the previous proposition. if B is furthermore assumed to be surjective. µ) = J(v) + (µ. where A ∈ L(V. v) ∀v ∈ V. λ) is a solution to (A. λ) is a saddle-point for L (defined by (A. On the other hand. µ) ∈ V × Λ 7−→ L(v. λ) be a saddle point for L (or. λ) ≤ L(v. APPENDIX and the constrained optimization problem: Find u ∈ K such that J(u) = inf J(v).5) Bu = 0. 7] for a general presentation of the approach. vi = a(u. namely the first order Lagrange Finite Element. Bv). to [17. Proposition 10 Let (u.4)) if and only if  Au + B ⋆ λ = ϕ (A. and to [16] for its application to fluid flows. We define the Lagrangian of the problem as L : (v.e.4) Proposition 9 Under the set of assumptions (A. J(u + th) ≥ J(u) for all t ∈ R and h ∈ K.5)). ∀µ ∈ Λ. Let Ω be a domain of Rd . i.5) is tightly related to the original problem. Then u minimizes J over K. But in general (infinite dimensional setting).2). so that existence of λ such that ∇J + B ⋆ λ is straightforward.2 Elliptic problems and Finite Element Method We give here some basic facts on the discretization of an elliptic problem by the Finite Element approach. so that full equivalence between the saddle point formulation (A.2). Proposition 11 Under the set of assumptions (A. The saddle-point problem (A. equivalently. w) = 0 for all w ∈ ker B. λ) is a saddle point for the Lagrangian L. A. (u. λ) is a solution to (A. µ) ≤ L(u. then the Lagrange multiplier λ is unique. we consider the Poisson problem in Ω: ( −∆u = f in Ω u = 0 on Γ = ∂Ω . V ′ ) is defined by hAu . λ) ∀v ∈ V .5).5) and the initial minimization problem does not hold in general. if L(u. expressing that u minimizes J over K. if B has a closed range. there exists λ ∈ Λ such that (u. we obtain (∇J. λ) is said to be a saddle point to L(·. We refer e. i.3) v∈K Definition 7 (Saddle point) A couple (u. which consists in minimizing J over K. (A.

We now define Vh as the set of all those functions of V such that their restriction to any element of Th is affine. This behavior is ensured by the following lemma. space of all those functions in L2 (Ω) with a gradient in L2 (Ω)d . the intersection K ∩ K is either empty. The discrete problem is now   Find uh ∈ Vh such that  Z Z (A. Note that. ELLIPTIC PROBLEMS AND FINITE ELEMENT METHOD 69 The variational formulation of the problem can be formally obtained by multiplying the equation by a test function that vanishes on Γ. for any two simplices K and K ′ in Th . or (in the case d = 3).1).6)   ∇u · ∇v = f v ∀v ∈ V Ω Ω which can also be written in an abstract way as (A. Such a space can be built as follows: We first consider a triangulation of Ω in the sense of the following definition. The initial phase which consists in obtaining the varia- tional formulation does not guarantee any rigorous link between the initial problem and its new formulation. tetrahedra for d = 3). a rigorous proof has to be provided after the solution to the variational problem has been identified and characterized. and all operation (like integration by parts) are performed formally. a common vertex. in which the variational formulation above makes sense. they are continuous over Ω. Assume that f is in L2 (Ω).7)   ∇uh · ∇vh = fh ∀vh ∈ Vh Ω Ω If one considers a sequence of such spaces Vh which converges in some sense to the infinite dimensional space V . Definition 8 (Conforming triangulation) A conforming triangulation Th is a set of non degenerated simplices (triangles in the two- dimensional setting. If one aims at establishing properly that the solution to the variational problem is a solution to the initial problem in some sense. so it does not make sense to make precise assumptions on the regularity of the test func- tions. The process allows to obtain a new formulation of the problem. the problem can be written   Find u ∈ V such that  Z Z (A. and integrating by part: Z Z ∇u · ∇v = f v. K∈Th ′ such that. a common edge. which fits into a precise mathematical framework as soon as functional spaces are defined. which can be seen as the abstract core of the Finite Element Method: . as they are required to belong to V . Ω Ω Remark 7 The elaboration of the variational formulation is not covered by strict mathematical rules. a common face.A. The Finite Element Method is based on the introduction of a finite dimensional approxima- tion space Vh ∈ V . with [ Ω= K. Considering the Sobolev space V = H01 (Ω). it is natural to expect a convergence of uh to u. the problem fits into the framework of the Lax-Milgram theorem.2.

We define u and uh as solutions to the problems a(u. The maximum of hK over all elements of a given triangulation Th is called the diameter of the triangulation.3. respectively. It remains to establish that the latter infimum converges to 0 as h goes to 0. and it is simply1 denoted by h. 2 For the physical dimensions d = 1. . which maps any function in H 2 to the unique element in Vh which takes the same values2 at the vertices of the triangulation Th . Consider Problem (A.  a(uh .e. 1 Committing here a common abuse of notations. vi. Considering regular families of triangulation consists in ruling out the possibility that simplices may degenerated (i. vh i ∀vh ∈ Vh . v) − hϕ . The discrete counterpart of Problem (A. 2. Finite element method for constrained problems. We denote by Ih the interpolation operator. functions of the Sobolev space H 2 (Ω) can be identified with continuous functions. Handling constraints in the finite element framework is a delicate matter. vh ) + (Bvh . and Vh ⊂ V a subspace. vh i ∀vh ∈ Vh . (A. vi ∀v ∈ V . APPENDIX Lemma 1 (Céa’s) Let V be a Hilbert space. which belong to Kh = {vh ∈ Vh . A sequence (Th )h is said to be regular if the ratio hK /ρK is bounded uniformly with respect to all the triangulations of the sequence. µh ) = 0 ∀µh ∈ Λh } . It holds |u − Ih u| ≤ Ch |u|2 . we denote by h both the index (i.8) (Buh .e. It is natural to introduce discretization spaces Vh ⊂ V and Λh ⊂ Λ of the primal and dual spaces.2). vh ) = hϕ . in a variational form. Proposition 13 Let u ∈ H 2 (Ω) ∩ H 1 (Ω) be given. Definition 9 (Regular sequence of triangulations) For a simplex K. It holds |u − uh | ≤ C inf |u − vh | . Convergence of uh toward u (for the H 1 norm) is now a direct consequence of Proposition 13 and Lemma 1. This property can be established in the present situation thanks to the H 2 regularity of the solution.e. so that pointwise values are well-defined. under assumptions (A. the label) and the diameter (which is a positive number). 2 over the the subspace of Vh of all elements that satisfy the discrete counterpart of the constraint. a(uh . or 3. we denote by hK its diameter. (Bvh . become flat). λh ) = hϕ . v) = hϕ .5) writes. and by ρK the diameter of the largest ball included in K. and to compute the minimum of the same functional 1 v 7−→ J(v) = a(v. Let a(·. ·) a bilinear form over V × V which is continuous and elliptic.70 APPENDIX A. µh ) = 0 ∀µh ∈ Λh . i. vh ∈Vh where C > 0 does not depend on h.

e.3.5). Even if one is interested in the primal part uh only. convergence of primal and dual components of the solution can be proven Proposition 14 Under assumptions (A. Its divergence is  ∂ !  ∂x1  u1 ∂u1 ∂u2  ∂ · ∇·u=  = + . a first consequence can be the non uniqueness of the discrete Lagrange multiplier λh . there exists C > 0 such that   |u − uh | + |λ − λh | ≤ C inf |u − vh | + inf |λ − µh | . if its dimension is too small compared to that of Vh ). we assume that B is surjective. and we denote by (u. reducing the discrete constrained space Kh to {0} in some situations.A. if Λh is too rich.8. µ∈Λ v∈V |v| |µ| Inf-sup condition (A. λh ) the solution to Problem A. it might harm the convergence of λh toward the exact Lagrange multiplier λ. DIFFERENTIAL CALCULUS. Λh )h is said to verify the inf-sup condition (also called Ladyzenskaya-Babuska-Brezzi. Let q be a scalar field.3 Differential calculus. Besides. INTEGRATION BY PARTS 71 The choice of Λh is obviously related to Vh . (A. µ) inf sup > 0. If Λh it is too poor (i. and (uh . Considering now a sequence (Vh . On the other hand. u2 )T be a vector field. µh ) inf sup ≥C ∀h.9) can therefore be seen as a discrete counterpart of this characterization. Integration by parts Notations. Its gradient is the vector   ∂q  ∂x1  ∇q =   ∂q   ∂x2 Let u = (u1 .2). λ) the solution to Problem (A. Λ) is equivalent to (Bv. When such a condition is met.9) µh ∈Λh vh ∈Vh |vh | |µh | Note that the surjective character of an operator B ∈ L(V. u2 ∂x1 ∂x2 ∂x2 . the discrete problem can become overconstrained. Definition 10 (Inf-sup condition) The sequence of discretization spaces (Vh . The perfect balance is met whenever the so called inf-sup condition is verified.9). and the solution uh might converge to something which does not verify the constraint. the problem might be underconstrained. Λh ) which verifies inf-sup condition (A. or LBB condition) there exists C > 0 such that (Bvh . vh ∈Vh µh ∈Λh A.

each component of which is the divergence (in the usual sense) of the corresponding matrix line :   ∂σ11 ∂σ12 σ11 σ12 !  ∂x1 + ∂x2  ∇·σ =∇· =   ∂σ21 ∂σ22  σ21 σ22 + ∂x1 ∂x2 Let u = (u1 . The matrix u ⊗ u is the matrix (ui uj )i.j . APPENDIX Let u = (u1 .j Note that |A| = (A : B)1/2 is a Euclidean norm on the space of matrices (called Frobenius norm). . Its gradient is the matrix   ∂u1 ∂u1  ∂x1 ∂x2  ∇u =   ∂u2  ∂u2  ∂x1 ∂x2 For any vector n. Its divergence is a vector. we write ∂u ∇u · n = . ∂n Let u be a vector field. ∂u2 ∂u2  ∂v2 ∂v2  ∂xj ∂xj i=1 j=1 ∂x1 ∂x2 ∂x1 ∂x2 The notation |∇u|2 is used to denote ∇u : ∇u.72 APPENDIX A. ∆u2 For A = (aij ) and B = (bij ) two square matrices. i. u2 )T be a vector field. Its Laplacian ∆u is the vector ! ∆u1 ∆u = . we have     ∂u1 ∂u1 ∂u1 ∂u1  ∂x1 n1 + n2  !  ∂x1 ∂x2  n1 ∂x2 ∇u · n =   ∂u2  = . so that u(x + εn) = u(x) + ε∇u · n + o(ε). For u and v two vector fields     ∂u1 ∂u1 ∂v1 ∂v1 2 X 2  ∂x1 ∂x2   ∂x1 ∂x2  X ∂ui ∂vi ∇u : ∇v =   :   = . ∂u2  n2  ∂u2 ∂u2  n1 + n2 ∂x1 ∂x2 ∂x1 ∂x2 which is the derivative of u in the n direction. Let σ be a matrix field. A : B denotes the scalar X A:B= aij bij . u2 )T be a vector field. If n a unit vector.

Z Z t t  0+ ∇u : ∇v = v· ∇u · n (A. (A.3. It holds (Green’s identity for scalar functions) Z Z Z ∂u v∆u = ∇u · ∇v + v . As a consequence. if ∇ · u = 0. INTEGRATION BY PARTS 73 If ∇ · u = 0. and q a scalar field.12) Ω Γ Let v be a vector field.13) Ω Ω Γ Let u and v be scalar fields. if ∇ · u = 0. Let u and v be vector fields.14) Ω Ω Γ ∂n where n is the outward normal vector. 2 If ∇ · u = 0. then ∇ · t ∇u = 0. (A.11) Ω Γ Let q be a scalar field.15) Ω Ω Γ ∂n And. Integration by parts.A. It holds (Divergence Theorem for a matrix) Z Z ∇·σ = σ·n (A. ! |u|2 (∇ · (u ⊗ u)) · u = ((u · ∇) u) · u = ∇ · u .10) Ω Γ Let σ be a matrix field.16) Ω Γ . (A. It holds (Divergence Theorem) Z Z ∇·v= v·n (A. Let v be a vector field. It holds (Green’s identity) Z Z Z ∂u ∆u · v + ∇u : ∇v = v· . then   ∂u1 ∂u1  u1 ∂x1 + u2 ∂x2  ∇ · (u ⊗ u) = (u · ∇) u =    ∂u2 ∂u2  u1 + u2 ∂x1 ∂x2 Still under the condition that ∇ · u = 0. DIFFERENTIAL CALCULUS. (A. It holds Z Z Z q∇ · u + u · ∇q = qu · n. It holds Z Z ∇q = qn. then  ∇ · ∇u + t ∇u = ∇ · ∇u = ∆u.

i j i j j i Le second terme ci-dessus est nul car u est à divergence nulle. and F (x. t) a scalar function. (A. On a alors Z Z t  0=− ∇u : ∇v + v · t ∇u · n ω ∂ω Proof :On écrit Z Z Z t  v· ∇u · n = n · (∇u · v) = ∇ · (∇u · v) ∂ω ∂ω ω X X XX X X = ∂i vj ∂j ui = ∂i vj ∂j ui + vj ∂j ∂i ui . t). (A. if ∇ · u = 0. On a Z Z Z Z ∇u : t ∇v = (∇ · u) (∇ · v) + (∇ · u)v · n − (∇u · v) · n Ω Ω Γ Γ Proof : On a Z Z (∇u · v) · n = ∇ · (∇u · v) Γ ZΩ X X = ∂i vj ∂j ui Ω i j Z XX = ((∂i ∂jui )vj + ∂j ui ∂i vj ) Ω i j Z Z = v (∇∇ · u) + ∇u : t ∇v ZΩ Ω Z Z = (∇ · u) v · n − (∇ · u) (∇ · v) + ∇u : t ∇v Ω Ω Ω . For any two vector fields u and v.74 APPENDIX A. APPENDIX As a consequence. then Z Z Z t  ∆u · v + ∇u : (∇v + ∇v) = v · ∇u + t ∇u · n. d’où l’on déduit l’identité annoncée.18) Ω Ω so that Z Z 1 t ∇u : (∇v + ∇v) = (∇u + t ∇u) : (∇v + t ∇v) (A. Proposition 16 Soient u et : v deux champs rguliers sur Ω. since (c̃ · σ) · n = c̃ · (σ · n) as soon as σ is symmetric. On suppose que u est à divergence nulle.17) Ω Ω Γ (note that extra parentheses are not necessary in the boundary integral above. (A.20) dt ω(t) ω(t) ∂t ∂ω(t) Proposition 15 Soient u et v deux champs de vecteurs réguliers définis sur Ω. t)u · n. we have Z Z t t ∇u : ∇v = ∇u : ∇v. It holds Z Z Z d ∂F F (x.19) Ω 2 Ω Let ω be a material system advected by the velocity field u(x. t) = − F (x. which is the case here).

Valli. Springer. Kluwer ed. [11] J. Numerical Approximation of Partial Differential Equations. Mathematical Methods for the Magnetohydrodynam- ics of Liquid Metals. Sur le principe de Dirichlet. 34 (1906). 159 of Applied Mathematical Series. [3] D. Raviart. Paris. [14] J.-L. A. Brenner.C. Springer. Bull. Acta Math. p.L. C. Hadamard. Duvaut and J. Springer-Verlag 1986. fast solvers ans applications in solid mechanics. de la SMF. Quarteroni. Springer Series in Computational Mathematics. Leray Essai sur le mouvement d’un liquide visqueux emplissant l’espace. Low Reynolds number hydrodynamics: with special applications to particulate media. Finite Element Methods for Navier-Stokes Equations. [4] S. Scott. 63 (1933). Les inquations en mcanique et en physique. Modélisation mathématique et mécanique des milieux conti- nus. [8] FreeFem++ reference manual [9] J. Boyer and P. Lions. Analyse fonctionnelle. 193?248. pp 135–138. Theory and Algorithms. Braess. [13] J. Brezis. Lions. [10] V. [16] A. [5] H. Temam. Springer 2007. third edition.. Eds de l’Ecole Polytechnique. The Mathematical Theory of Finite Element Methods (Texts in Applied Mathematics). vol. 1972. Theory. Quelques mthodes de rsolution des problmes aux limites non linaires. [2] F. 2004. [6] G. 75 . Brenner. Ern and J. Masson Eds.-L. Miranville and R. Eléments d’analyse pour l’étude de quelques modèles d’écoulements de fluides visqueux incompressibles. Finite elements. [7] A. Dunod. 2nd edition 2008. 1969. [12] J.A. vol. Scopos.Bibliography [1] Grégoire Allaire. [15] A. Springer. 2006. Cam- bridge University Press. Springer. Berlin-Heidelberg-New York-Tokyo.F.T. Fabrie. 52. Mathématiques et Applications. Theory and Practice of Finite Elements. Gerbeau. Happel and H. Oxford Science Publications. Analyse numérique et optimisation. Le Bris. New York. R. Girault. P. Lelievre. Gauthier- Villars ed. Guermond.

. Théorie des distributions. M. Thomas. J. Schwartz. [18] L. Introduction l’analyse numrique des quations aux drives partielles. [19] R. American Mathemat- ical Society. Dunod 2004. Raviart. Navier-Stokes equations. Temam. 3rd edition 2000.76 BIBLIOGRAPHY [17] P. Hermann 1997. A. Theory and numerical analysis.