Lecture Notes in Mathematics

Volume 2318

Editors-in-Chief
Jean-Michel Morel, CMLA, ENS, Cachan, France
Bernard Teissier, IMJ-PRG, Paris, France

Series Editors
Karin Baur, University of Leeds, Leeds, UK
Michel Brion, UGA, Grenoble, France
Annette Huber, Albert Ludwig University, Freiburg, Germany
Davar Khoshnevisan, The University of Utah, Salt Lake City, UT, USA
Ioannis Kontoyiannis, University of Cambridge, Cambridge, UK
Angela Kunoth, University of Cologne, Cologne, Germany
Ariane Mézard, IMJ-PRG, Paris, France
Mark Podolskij, University of Luxembourg, Esch-sur-Alzette, Luxembourg
Mark Policott, Mathematics Institute, University of Warwick, Coventry, UK
Sylvia Serfaty, NYU Courant, New York, NY, USA
László Székelyhidi , Institute of Mathematics, Leipzig University, Leipzig,
Germany
Gabriele Vezzosi, UniFI, Florence, Italy
Anna Wienhard, Ruprecht Karl University, Heidelberg, Germany
This series reports on new developments in all areas of mathematics and their
applications - quickly, informally and at a high level. Mathematical texts analysing
new developments in modelling and numerical simulation are welcome. The type of
material considered for publication includes:
1. Research monographs
2. Lectures on a new field or presentations of a new angle in a classical field
3. Summer schools and intensive courses on topics of current research.
Texts which are out of print but still in demand may also be considered if they fall
within these categories. The timeliness of a manuscript is sometimes more important
than its form, which may be preliminary or tentative.
Titles from this series are indexed by Scopus, Web of Science, Mathematical
Reviews, and zbMATH.
Jean Deteix • Thierno Diop • Michel Fortin

Numerical Methods
for Mixed Finite Element
Problems
Applications to Incompressible Materials
and Contact Problems
Jean Deteix Thierno Diop
GIREF, Département de Mathématiques et GIREF, Département de Mathématiques et
de Statistique de Statistique
Université Laval Université Laval
Québec, QC, Canada Québec, QC, Canada

Michel Fortin
GIREF, Département de Mathématiques et
de Statistique
Université Laval
Québec, QC, Canada

This work was supported by Agence Nationale de la Recherche (ANR-18-IDEX-0001).

ISSN 0075-8434 ISSN 1617-9692 (electronic)

Lecture Notes in Mathematics
ISBN 978-3-031-12615-4 ISBN 978-3-031-12616-1 (eBook)
https://doi.org/10.1007/978-3-031-12616-1

Mathematics Subject Classification: 74S05, 65N22, 65F10, 65F08, 74B20, 74M15

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
2 Mixed Problems .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
2.1 Some Reminders About Mixed Problems . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
2.1.1 The Saddle Point Formulation . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4
2.1.2 Existence of a Solution . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 4
2.1.3 Dual Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5
2.1.4 A More General Case: A Regular Perturbation .. . . . . . . . . . . . . . 6
2.1.5 The Case b(v, q) = (Bv, q)Q . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 6
2.2 The Discrete Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7
2.2.1 Error Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 8
2.2.2 The Matricial Form of the Discrete Problem . . . . . . . . . . . . . . . . . 9
2.2.3 The Discrete Dual Problem: The Schur Complement . . . . . . . . 11
2.3 Augmented Lagrangian . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 13
2.3.1 Augmented or Regularised Lagrangians . .. . . . . . . . . . . . . . . . . . . . 13
2.3.2 Discrete Augmented Lagrangian in Matrix Form . . . . . . . . . . . . 15
2.3.3 Augmented Lagrangian and the Condition Number
of the Dual Problem . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 15
2.3.4 Augmented Lagrangian: An Iterated Penalty .. . . . . . . . . . . . . . . . 17
3 Iterative Solvers for Mixed Problems . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 19
3.1 Classical Iterative Methods . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 19
3.1.1 Some General Points . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 20
3.1.2 The Preconditioned Conjugate Gradient Method . . . . . . . . . . . . . 22
3.1.3 Constrained Problems: Projected Gradient and
Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 24
3.1.4 Hierarchical Basis and Multigrid Preconditioning . . . . . . . . . . . 26
3.1.5 Conjugate Residuals, Minres, Gmres and the
Generalised Conjugate Residual Algorithm . . . . . . . . . . . . . . . . . . 27
3.2 Preconditioners for the Mixed Problem . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 32
3.2.1 Factorisation of the System . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 32

v
vi Contents

3.2.2 Approximate Solvers for the Schur Complement

and the Uzawa Algorithm . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
3.2.3 The General Preconditioned Algorithm . . .. . . . . . . . . . . . . . . . . . . . 38
3.2.4 Augmented Lagrangian as a Perturbed Problem . . . . . . . . . . . . . 41
4 Numerical Results: Cases Where Q = Q . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
4.1 Mixed Laplacian Problem . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
4.1.1 Formulation of the Problem . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
4.1.2 Discrete Problem and Classic Numerical Methods . . . . . . . . . . . 45
4.1.3 A Numerical Example . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 47
4.2 Application to Incompressible Elasticity . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 48
4.2.1 Nearly Incompressible Linear Elasticity . .. . . . . . . . . . . . . . . . . . . . 49
4.2.2 Neo-Hookean and Mooney-Rivlin Materials . . . . . . . . . . . . . . . . . 52
4.2.3 Numerical Results for the Linear Elasticity Problem . . . . . . . . 55
4.2.4 The Mixed-GMP-GCR Method . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 56
4.2.5 The Test Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 58
4.2.6 Large Deformation Problems . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 64
4.3 Navier-Stokes Equations . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 68
4.3.1 A Direct Iteration: Regularising the Problem . . . . . . . . . . . . . . . . . 71
4.3.2 A Toy Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 72
5 Contact Problems: A Case Where Q = Q . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 75
5.1 Imposing Dirichlet’s Condition Through a Multiplier .. . . . . . . . . . . . . . . 75
5.1.1 H00 (C ) and Its Dual H −1/2 (C ) . . . . . . .. . . . . . . . . . . . . . . . . . . .
1/2
77
5.1.2 A Steklov-Poincaré operator . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 78
5.1.3 Discrete Problems . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 79
5.1.4 A Discrete Steklov-Poincaré Operator . . . .. . . . . . . . . . . . . . . . . . . . 80
5.1.5 Computational Issues, Approximate Scalar Product . . . . . . . . . 81
5.1.6 The L2 (C ) Formulation . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 84
5.1.7 A Toy Model for the Contact Problem . . . .. . . . . . . . . . . . . . . . . . . . 84
5.2 Sliding Contact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 87
5.2.1 The Discrete Contact Problem . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 89
5.2.2 The Algorithm for Sliding Contact . . . . . . . .. . . . . . . . . . . . . . . . . . . . 92
5.2.3 A Numerical Example of Contact Problem . . . . . . . . . . . . . . . . . . . 93
6 Solving Problems with More Than One Constraint... . . . . . . . . . . . . . . . . . . . 97
6.1 A Model Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 97
6.2 Interlaced Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 98
6.3 Preconditioners Based on Factorisation . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 100
6.3.1 The Sequential Method . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 101
6.4 An Alternating Procedure . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 102
7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 105

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 107
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 113
Chapter 1
Introduction

Mixed Finite Element Methods are often discarded because they lead to indefinite
systems which are more difficult to solve than the nice positive definite problems of
standard methods. Indeed, solving indefinite systems is a challenge : direct methods
[4, 14, 78] might need renumbering (see [28]) of the equations and standard iterative
methods [22, 81, 86] are likely to stagnate or to diverge as proved in [86]. As an
example, consider the classical conjugate gradient method. Applied to a symmetric
indefinite problem it will generate a diverging sequence. As the conjugate gradient
method is (in exact arithmetic) a direct method, it will yield the exact solution if
the problem is small enough to avoid loosing orthogonality. Applying a minimum
residual method to the same problem will in most cases yield stagnation.
These two classical methods are the simplest in a list which grows constantly.
This monograph does not intend to introduce new iteration methods. We shall rely
mostly on existing packages, mostly Petsc from Argonne Laboratory [11].
Our concern is to solve algebraic systems associated to mixed discretisation.
Several approaches (see, for example, [8, 15, 43]) exist in the literature to solve
this type of problem but convergence is not always guaranteed. They are indefinite
systems but also structured systems associated with matrices of the form,
 
A Bt
(1.1)
B0

where A is often a positive definite matrix.

The key to obtain convergence is preconditioning. For general problem, a
vast number of preconditioners are available [15]. Our goal is to develop good
preconditioners for problems arising from (1.1).
We want to show in the present work that efficient iterative methods can be
developed for this class of problems and that they make possible the solution of
large problems with both accuracy and efficiency. We shall also insist on the fact that
these methods should be entirely automatic and free of user dependent parameters.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1

J. Deteix et al., Numerical Methods for Mixed Finite Element Problems,
Lecture Notes in Mathematics 2318, https://doi.org/10.1007/978-3-031-12616-1_1
2 1 Introduction

We also want to make clear that our numerical results should be taken as
examples and that we do not claim that they are optimal. Our hope is that they
could be a starting point for further research.
Here is therefore our plan.
• Chapter 1 rapidly recalls the classical theory of mixed problems, including
Augmented Lagrangian methods and their matricial form.
• Chapter 2 presents some classical iterative methods and describes the precondi-
tioner which will be central to our development. We come back to augmented
Lagrangian and a mixed form for penalty methods.
• Chapter 3 is devoted to numerical examples. The first one will be the approxima-
tion of a Dirichlet problem with Raviart-Thomas elements [20]. This is a simple
case which will however permit to consider the fundamental issues. We shall see
how an Augmented Lagrangian method enables us to circumvent the fact that we
do not have coercivity on the whole space.
We shall thereafter consider incompressible elasticity in solid mechanics, first
in the linear case and then for a non linear Mooney-Rivlin model.
In all those problems, the space of multipliers is L2 () and can therefore be
identified with its dual. We also present some ideas for the solution of the Navier-
Stokes equations. In those problems, with the discrete spaces that we employ, we
shall not be able to use a real augmented Lagrangian. However, we shall consider
a regularised formulation which will accelerate the convergence of our iterations.
• Chapter 4: We consider contact problems. In this case, the space of multipliers
is not identified to its dual. We shall present some ideas for which we do not
have numerical results but which we think could be research avenues for the
future. In particular, we present ideas about discrete Steklov-Poincaré operators.
Numerical results will be presented in the more classical formulation where the
duality product is approximated by the L2 scalar product.
• Chapter 5: Finally, we shall consider a case, arising from contact mechanics
between incompressible bodies, in which we have two different types of con-
straints. This will lead us to modify accordingly our preconditioners.
Chapter 2
Mixed Problems

2.1 Some Reminders About Mixed Problems

Basically, mixed problems arise from the simple problem of minimising a quadratic
functional under linear constraints. Let then V a function space and a bilinear form
defining an operator A from V into V  .

a(u, v) = Au, vV  ×V ∀ u, v ∈ V

This bilinear form should be continuous, that is

a(u, v) ≤ a uV vV (2.1)

where  · V is the norm of V and a is the norm of a(·, ·). In the same way,
consider another function space Q, a bilinear form on V × Q defining an operator
B from V into Q ,

b(v, q) = Bv, q ∀ v ∈ V , ∀ q ∈ Q. (2.2)

We also suppose this bilinear form to be continuous and thus

b(v.q) ≤ b vV qQ (2.3)

where b is the norm of b(·, ·). Defining the functional

1
F (v) := a(v, v) − f, v (2.4)
2

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 3

J. Deteix et al., Numerical Methods for Mixed Finite Element Problems,
Lecture Notes in Mathematics 2318, https://doi.org/10.1007/978-3-031-12616-1_2
4 2 Mixed Problems

we then want to solve the constrained problem,

inf F (v). (2.5)

Bv=g

2.1.1 The Saddle Point Formulation

Problem (2.9) can be classically transformed into the saddle-point problem,

1
inf sup a(v, v) − b(v, q) − f, v + g, q, (2.6)
v∈V q∈Q 2

for which the optimality system is


a(u, v) + b(v, p) = f, v ∀v ∈ V
(2.7)
b(u, q) = (g, q)Q ∀q ∈ Q,

2.1.2 Existence of a Solution

For problem (2.7) to have a solution, it is clearly necessary that there exists some ug
satisfying Bug = g. Moreover, the lifting from g to ug should be continuous. This
is classically equivalent [20] to the inf-sup condition

b(u, q)
inf sup ≥ k0 (2.8)
q∈Q v∈V uV qQ

a condition omnipresent in the following. Solving (2.5) is then equivalent to finding

u0 ∈ KerB such that,

a(u0 , v0 ) = (f, v0 ) − a(ug , v0 ) ∀v0 ∈ Ker B. (2.9)

We shall refer to (2.9) as the primal problem. We therefore suppose, as in [20] that
the bilinear form a(·, ·) is coercive on Ker B, that is there exists a constant α0 such
that,

a(v0 , v0 ) ≥ α0 v0 2V ∀v0 ∈ Ker B (2.10)

Remark 2.1 (Coercivity) Unless there is a simple way to build a basis of Ker B
or a simple projection operator on Ker B coercivity on Ker B is not suitable for
numerical computations. In our algorithms, we shall need, in most cases, coercivity
2.1 Some Reminders About Mixed Problems 5

on the whole of V .

a(v, v) ≥ αv2V ∀v ∈ V . (2.11)

It is always possible to get this by changing a(u, v) (see [43, 46, 79]) into


a (u, v) = a(u, v) + α(Bu, Bv)Q .

Such a change will have different consequences in the development of our algo-
rithms and will lead to augmented and regularised Lagrangians which will be
considered in detail in Sect. 2.3 

Remark 2.2 It should also be noted that (2.11) and (2.1) imply that a(v, v) is a norm
on v, equivalent to the standard norm. 

2.1.3 Dual Problem

Problem (2.6) has the general form,

inf sup L(v, q),

v∈V q∈Q

where L(v, q) is a convex-concave functional on V × Q. If one first eliminates q by

computing

J (v) = sup L(v, q),

q∈Q

one falls back on the original problem, the primal problem. Reversing the order of
operations, (this cannot always be done, but no problems arise in the examples we
present) and eliminating v from L(v, q) by defining

D(q) := inf L(v, q)

v∈V

leads to the dual problem

sup D(q).
q∈Q

The discrete form of the dual problem and the associated Schur’s complement will
have an important role in the algorithms which we shall introduce.
6 2 Mixed Problems

2.1.4 A More General Case: A Regular Perturbation

We will also have to consider a more general form of problem (2.7). Let us suppose
that we have a bilinear form c(p, q) on Q. We also suppose that it is coercive on Q,
that is

c(q, q) ≥ γ |q|2Q

we thus consider the problem


a(u, v) + b(v, p) = f, v ∀v ∈ V
(2.12)
b(u, q) − c(p, q) = (g, q)Q ∀q ∈ Q,

It is now elementary to show [20] that we have existence of a solution and

1 1
αu2V + γ p2Q ≤ f 2V  + g2Q (2.13)
α γ

In practice we shall consider the perturbed problem defined with

c(p, q) = (p, q)Q .

The bound (2.13) explodes for γ small which can be a problem. If we still have
the inf-sup condition and the coercivity on the kernel then we can have a bound on
the solution independent of  . We refer to [20] for a proof and the analysis of some
more general cases.
Remark 2.3 In this case, the perturbed problem can be seen as a penalty form for the
unperturbed problem. We shall come back to this in Sect. 2.3.4. It is then easy to see
that we have an O() bound for the difference between the solution of the penalised
problem and the standard one. The bound depends on the coercivity constant and
the inf-sup constant. 

2.1.5 The Case b(v, q) = (Bv, q)Q

The above presentation is abstract and general. We now consider a special case
which will be central to the examples that we shall present later. Indeed, in many
cases, it will more suitable to define the problem through an operator B from V into
Q. We suppose that on Q, we have a scalar product defining a Ritz operator R from
Q into Q

(p, q)Q = Rp, qQ ×Q = (Rp, Rq)Q

2.2 The Discrete Problem 7

We then define the bilinear form

b(v, q) = (Bv, q)Q = Bv, qQ ×Q

referring to (2.2) we then have,

B = RB (2.14)

and

(Bu, q)Q ×Q = (Bu, q)Q = Bu, RqQ×Q (2.15)

The constraint that we want to impose is then Bu = g with g ∈ Q or equivalently

Bu = g  = Rg. In the case Q = Q and R = I , the operators B and B coincide.
This will not be the case for the problems which we shall consider in Chap. 5. But
even in the case Q = Q , in the discrete formulation the distinction between B and
B will have to be taken into account.

2.2 The Discrete Problem

Our interest will be in solving discretisations of Problem (2.7). We thus suppose

that we have subspaces Vh ⊂ V and Qh ⊂ Q. We shall work in the framework of
Sect. 2.1.5 and we thus have,

b(vh , qh ) = (Bvh , qh )Q = (PQh Bvh , qh )Q (2.16)

and we can define

Bh = PQh B

The scalar product (·, ·)Q defines an operator Rh from Qh onto Qh and we can
introduce

Bh = Rh Bh

We want to solve

a(uh , vh ) + b(vh , ph ) = f, vh  ∀vh ∈ Vh
(2.17)
b(uh , qh ) = (gh , qh )Q , ∀qh ∈ Q

We denote gh the projection of g onto Qh . The second equation of (2.17) can be

read as Bh uh = gh or equivalently Bh u = Rh gh . Unless Buh ∈ Qh , this condition
8 2 Mixed Problems

is weaker than Buh = g which for almost all cases will lead to bad convergence or
even ‘locking ’, that is a null solution.
For example, we shall meet later the divergence operator div (which acts on a
space of vector valued functions which we shall then denote by u). When we take
for Qh a space of piecewise constant functions, divh uh is a local average of div uh .

2.2.1 Error Estimates

We recall here, in its simplest form the theory developed in [20]. This will be
sufficient for our needs. In the proof of existence of Sect. 2.1.2, we relied on two
conditions: the coercivity on the kernel (2.10) and the inf-sup condition (2.8). We
introduce their discrete counterpart. We thus suppose

∃ α0h > 0 such that a(v0h , v0h ) ≥ α0h v0h 2V ∀ v0h ∈ Ker Bh , (2.18)

or by the even simpler global ellipticity

∃ αh > 0 such that a(vh , vh ) ≥ αh vh 2V ∀ vh ∈ Vh .

We also write the discrete inf-sup condition,

b(vh , qh )
∃ βh > 0 such that sup ≥ βh qh Q . (2.19)
vh ∈Vh vh V

It will also be convenient, to define the approximation errors

Eu := inf u − vh V
vh ∈Vh

Ep := inf p − qh Q
qh ∈Qh

We recall that approximation errors depend on the regularity of the solution. We

refer to the classical work of [30] for an analysis of this dependence. Let a and
b be the norms of a(·, ·) and of b(·, ·) as defined in (2.1) and (2.3). The following
result is proved in [20].

Proposition 2.1 (The Basic Error Estimate) Assume that Vh and Qh verify (2.18)
and (2.19). Let f ∈ V  and g ∈ Q . Assume that the continuous problem (2.17) has
a solution (u, p) and let (uh , ph ) be the unique solution of the discretised problem
(2.17). If a(·, ·) is symmetric and positive semi-definite we have the estimates
 
2a 2a1/2b b
uh − uV ≤ + Eu + h Ep ,
α0h (α0h )1/2 βh α0
2.2 The Discrete Problem 9

 
2a3/2 a b 3a1/2 b
ph − pQ ≤ h 1/2
+ Eu + Ep
(α0 ) βh βh2
(α0h )1/2 βh

Remark 2.4 (Dependance on h) In the previous result, we allowed, in principle, the

constants βh and α1h (or α0h ) to depend on h. It is obvious (but still worth mentioning)
that if there exist constants β0 and α0 such that βh ≥ β0 and α1h ≥ α0 (or α0h ≥
α0 ) for all h, then the constants appearing in our estimates will be independent
of h. Considering constants depending on h will be useful for contact problems
(Sect. 5.2). 
We can also see from this estimate that the approximation of p is more sensitive
on the value of βh than the approximation of u. One also sees that, paradoxically,
improving Eu is better to counter this than improving Ep .

2.2.2 The Matricial Form of the Discrete Problem

To make explicit the numerical problem associated with (2.17) we need to introduce
j
basis vectors φhi for Vh and ψh for Qh and the vectors of coefficients u, q which
define uh and qh with respect to these bases.
  j
uh = ui φhi , qh = qj ψh . (2.20)
i j

Remark 2.5 (Notation) To avoid adding cumbersome notation, unless a real ambi-
guity would arise, we shall denote in the following by u and p either the unknowns
of the continuous problem or the unknowns of the numerical problems arising from
their discretisation.
We shall also denote by the same symbol the operators A and B of the continuous
problem and the matrices associated to the discrete problems. As they are used in
very different contexts, this should not induce confusion. 
Denoting ·, · the scalar product in Rn , we can now define the matrices associated
with the bases (2.20)

A u, v = a(uh , vh ).

We also have a matrix R associated with the scalar product in Qh which represents
the discrete Ritz operator

Rp, q = (ph , qh )Qh .

10 2 Mixed Problems

The scalar product in Qh will be associated with R −1 .

R −1 p, q = (ph , qh )Qh .

We then have

B u, q = b(uh , qh ) = (Bh uh , qh )Qh

and the operator Bh uh can be asociated with the matrix

B = R −1 Bu (2.21)

We summarise this in the following diagram.

Vh A Vh

B Bt
B Bt

R−1
Qh Qh
R

Remark 2.6 (Qh = Qh ) It is important to note that even in the case where Q =
Q and R = I , this is not the case in the discrete problem. The matrix R defined
above is not the identity matrix. 

Remark 2.7 (Change of Metric on Qh ) As we shall see in Sect. 2.2.3, it will

sometimes be useful in numerical applications to change the scalar product on Qh .
To distinguish this case we shall denote MQ the matrix defining the scalar product

(ph , qh )Qh = MQ p, q.

The choice MQ = R is frequent but not optimal in many cases [48]. In Sect. 2.2.3
we shall discuss the choice MQ = MS where MS is some approximation of the
discrete Schur complement. 
We now consider the finite dimensional matricial problems associated to our mixed
formulation, in fact the actual problem for which we want to build efficient solvers.
2.2 The Discrete Problem 11

We first write problem (2.17) in matrix form,

    
A Bt u f
= (2.22)
B0 p g

Although this block matrix is a non singular matrix the numerical solution of
(2.22) is not so simple. The main problem being that this matrix is indefinite. If one
wanted to employ a direct solver, we might have to introduce a renumbering of the
equations [28]. Moreover, our examples will come from the application of the mixed
finite element methods [20]. Indeed, we shall focus on methods suitable for large
systems arising from the discretisation of three-dimensional problems in mechanics
which lead to large problems of the form (2.22). Such large systems are not suitable
for direct solvers and require iterative methods. However, without preconditionning
the system (acting on its eigenvalues), iterative methods, such as Krylov methods,
are likely to diverge or stagnate on indefinite problems (see [86]). Therefore the key
to obtain convergence will be the construction of a good preconditioner.

2.2.3 The Discrete Dual Problem: The Schur Complement

If we eliminate u from (2.22) we obtain the discrete form of the dual problem,

− BA−1 B t p = g − BA−1 f. (2.23)

The matrix S = BA−1 B t is often called the Schur complement. The system (2.23)
is equivalent to the maximisation problem

1
sup {−A−1 B t q, B t q + g − BA−1 f, q} (2.24)
q 2

When building numerical methods, a crucial point will be the condition number
of S. Indeed the condition number is an important measure for the behaviour
(convergence) of iterative methods. Following Remark 2.7, we have on Qh a metric
defined by the matrix MQ . The standard choice would be to take MQ = R where
R is associated to the scalar product induced on Qh by the scalar product of Q. In
some cases, it will be convenient to change this scalar product, using a well chosen
matrix MS instead of R.
Remark 2.8 We write MS to emphasise that this matrix should be chosen to
approximate the Schur complement. 
We can change (2.23) into

MS−1 BA−1 B t p = MS−1 (g − BA−1 f ). (2.25)

12 2 Mixed Problems

In Sect. 3.1, this will correspond to a preconditioning. The idea is that MS should
be an approximation of BA−1 B t improving the condition number of the resulting
system.
To quantify this, we consider the eigenproblem

BA−1 B t φp = λMS φp (2.26)

Then the condition number for (2.25) can be defined as

λmax
Kh =
λmin

We have written Kh to emphasise that we have a discretised problem. A nice

property would be to have Kh independent of h. This would be the case if λmax ≤ C
and λmin ≥ c with C and c independent of h. We then say that MS and S are
spectrally equivalent [9]. The ideal case would be MS = S. Finding a good MS
means finding an easily computable approximation of S −1 .
Remark 2.9 (MS Depends on A) We shall come back to the choice of MS in
Sect. 4.2 in particular for linear elasticity (or the Stokes problem). It is worth already
remarking that MS should be an approximation of BA− B t and therefore depends
on A. In the simplest case, if A is changed into cA, one should change MS into
c MS . 
1

Although the question of spectral equivalence depends on the problem at hand, we

have a general case where it holds. We first recall that the minimum and maximum
eigenvalues in (2.26) are the minimum and maximum values of the Rayleigh
quotient,

A−1 B t q, B t q
RQ(q) =
MS p, p

Proposition 2.2 We suppose that the matrix A defines a norm on Vh , equivalent to

the standard norm.

αvh 2 ≤ Av, v ≤ A vh 2 .

We also suppose that we have the inf-sup condition (2.19). Then taking MS = R,
we have in (2.26)

βh2 B2
λmin ≥ , λmax ≤ . (2.27)
A α
2.3 Augmented Lagrangian 13

We recall that the dual norm associated to the norm defined by A is defined by A−1 .
We can therefore write the Rayleigh quotient, as

A−1 B t q, B t q v, B t 2
= sup
Rp, p v Av, vRq, q

The inf-sup condition yields the lower bound and the upper bound is direct.
This clearly also holds for another choice of MS if we have

c0 Rp, q ≤ MS p, q ≤ c1 Rp, q (2.28)

The bounds of (2.27) show that we have a condition number independent of the
mesh size if βh ≥ β0 > 0 . This is an important property if we want to solve large-
scale problems. However, if we consider MS as a preconditioner, we shall also want
to build it as a good approximation of S.

2.3 Augmented Lagrangian

Augmented Lagrangian is a popular method for the solution of mixed problems. We

shall present it in some details, trying to show its potential but also its shortcomings.

2.3.1 Augmented or Regularised Lagrangians

If we employ the regularised bilinear form 

a (u, v) defined in (2.11) we can define
an Augmented Lagrangian method.

a(u, v) + α(Bu, Bv)Q + b(v, p) = f, v + α(g  , Bv)Q ∀v ∈ V

(2.29)
b(u, q) = g , qQ ×Q , ∀q ∈ Q.

The extra term can be written differently using (2.15). We would then have,

a(u, v) + α(Bu, Bv)Q + b(v, p) = f, v + α(g, Bv)Q ∀v ∈ V
b(u, q) = (g, q)Q ∀q ∈ Q

with g = R −1 g  . This does not change the solution of the problem. In the discrete
problems, using an augmented Lagrangian is a classical way of accelerating some
algorithms and cancelling errors associated with penalty methods. We shall have
two different ways of implementing this idea.
14 2 Mixed Problems

The direct way to discretise (2.29) would be,

⎧
⎪ a(uh , vh ) + α(Buh , Bvh )Q + b(vh , ph )
⎪
⎨
= f, vh  + α(g, Bvh )Q ∀vh ∈ Vh (2.30)
⎪
⎪
⎩
b(uh , qh ) = (g, qh )Q , ∀qh ∈ Qh .

This corresponds, for α large, to a penalty method imposing exactly Buh = g,

in general a bad idea for many discrete problems. This will however be possible
in some cases, such as the example of Sect. 4.1. In general, for small values of α,
this formulation will be useful to regularise and improve some numerical solution
algorithms. This will be the case in Sect. 4.2. We shall call this a regularised
formulation.
To define a real discrete augmented Lagrangian, we suppose that Qh has a scalar
product defined by a metric MQ . As in (2.21) we take Bh = PQh B which depends
on MQ . A real discrete Augmented Lagrangian would then be written as
⎧
⎪
⎪ a(uh , vh ) + α(Bh uh , Bh vh )Qh + b(vh , ph )
⎨
= f, vh  + α(gh , Bh vh )Qh ∀vh ∈ Vh
⎪
⎪
⎩
b(uh , qh ) = (g, qh )Qh , ∀qh ∈ Qh ,
(2.31)

which we can also write as

⎧
⎪ a(uh , vh ) + α(Bh uh , Bh vh )Qh + b(vh , ph )
⎪
⎨
= f, vh  + α(gh , Bh vh )Qh ×Qh ∀vh ∈ Vh
⎪
⎪
⎩
b(uh , qh ) = (g, qh )Qh , ∀qh ∈ Qh .
(2.32)

This formulation depends on the choice of the metric MQ on Qh . The simplest

case being when this scalar product is the scalar product in Q, that is MQ = R.
We could also consider a scalar product induced by a matrix MS which we suppose
equivalent to the usual scalar product. As in Sect. 2.2.3 MS would be associated to
a good approximation of the Schur complement.
The solution is unchanged by the extra terms and everything looks perfect. The
question will rather be the actual implementation which will depend on the choice
of finite element spaces and also on the choice of the scalar product on Qh .
2.3 Augmented Lagrangian 15

2.3.2 Discrete Augmented Lagrangian in Matrix Form

We now consider the matricial form of the augmented Lagrangian formulations. In

the regularised form (2.30) we shall denote

Cu, v = (Buh , Bvh )Q

• For the regularised formulation (2.30) we have

    
A + αC B t u f + αB t g
= (2.33)
B 0 p g

• From (2.32) we have

    
A + αB t MS−1 B B t u f + αB t MS−1 g
= (2.34)
B 0 p g

Remark 2.10 (The Choice of MS ) The standard implementation is evidently to take

MS = R. If this is not the case, one should remark that in order to have (2.31) and
(2.32) to be equivalent, we have to write Bh = MS−1 B and not R −1 B as in (2.21). 

Remark 2.11 (MS−1 a Full Matrix?) Another important point is the presence of
MS−1 which is in general a full matrix. This makes (2.34) hardly usable unless MS
is diagonal or block diagonal and leads us to employ the regularised formulation. 

Remark 2.12 One could also write (2.34) writing the penalty term in mixed form,
that is
⎛ ⎞⎛ ⎞ ⎛ ⎞
A Bt Bt u f
⎝ B −MS 0 ⎠ ⎝ p
⎠ = ⎝ g ⎠
B 0 0 p g

The solution is of course p

 = 0. It is not clear if this form can be of some use. One
might perhaps consider it in the context of Chap. 6. 

2.3.3 Augmented Lagrangian and the Condition Number

of the Dual Problem

Using an augmented Lagrangian can be seen as a form of preconditioning for the

dual problem. We have already considered in Sect. 2.2.3 the condition number of
the dual problem (2.26), where MS defines a scalar product on Qh .
16 2 Mixed Problems

We had considered the eigenvalue problem,

BA−1 B t φp = λMS φp (2.35)

The eigenvalues of (2.35) are also the non zero eigenvalues of

A−1 B t MS−1 Bφu = λφu

with φu = A−1 B t φp which we can also write

1 t −1
Aφu = B MS Bφu
λ

If instead of A we use A + αB t MS−1 B this becomes

1 t −1
(A + αB t MS−1 B)φu = B MS Bφu .
λ
Denoting λα the corresponding eigenvalues, one easily sees that

λ
λα =
1+λα

Denoting λM and λm the largest and the smallest eigenvalues, the condition number
of the system is thus

λM (1 + α λm )
Kα =
λm (1 + α λM )

which converges to 1 when α increases. One sees that this holds for α large even
if the initial condition number is bad. One also sees that improving K = λM /λm
also improves Kα for a given α. Augmented Lagrangian therefore seems to be the
perfect solver. However, things are not so simple.
• The first problem is that the condition number of A + α B t MS−1 B worsens when
α increases. As solving systems with this matrix is central to the algorithms that
we shall introduce, we loose on our right hand what we gain with the left one. In
practice, this means finding the correct balance between the conflicting effects.
• The other point is the computability of B t MS−1 B. The matrix MS−1 could for
example be a full matrix. Even if we approximate it by a diagonal matrix, the
structure of the resulting matrix could be less manageable. This will be the case
in the examples of Sect. 4.2.
• When the real augmented Lagrangian cannot be employed, the regularised for-
mulation (2.33) might have a positive effect. However, the solution is perturbed
and only small values of α will be admissible.
2.3 Augmented Lagrangian 17

2.3.4 Augmented Lagrangian: An Iterated Penalty

Augmented Lagrangian formulations are often used as an iterative process, using a

penalty method to correct a previous value of p. Given pn , one solves

(A + αB t MS−1 B)un+1 + B t pn = f + αB t MS−1 g

pn+1 = pn + αMS−1 Bun

This could also be written as, writing  = 1/α

      
A Bt δu f − A un − B t p n run
= = (2.36)
B −MS δp g − Bun rpn

and

pn+1 = pn + δp, un+1 = un + δu.

System (2.36) is in fact a penalty method.

Remark 2.13 Problem (2.36) is of the perturbed form discussed in Sect. 2.1.4. Its
solution will rely on the same iterative procedures as the unperturbed problem which
we discuss in the next chapter. The question is whether there is some advantage in
doing so.
• It permits to use an augmented Lagrangian without handling the matrix
B t Ms−1 B.
• The price is that it is an iterative process.
We must note that the Schur complement now becomes

S = B A−1 B t + MS .

This makes the condition number of the perturbed problem better than that of the
standard one which makes one hope of a better convergence for the dual variable.
One also sees that if MS is an approximation of S one should rather use now,

MS = (1 + )MS

This means that there will be an optimum value of  = 1/α. Taking a larger
 for a better convergence enters in conflict with the paradigm of the augmented
Lagrangian which would need α large and thus  small. We shall present a numerical
example in Remark 4.11.

18 2 Mixed Problems

Remark 2.14 (Correcting the Regularised Form) One could use a similar idea to
eliminate the perturbation introduced by formulation (2.33), writing
      
A + αC B t δu f − A un − B t p n run
= = (2.37)
B 0 δp g − Bun rpn

again to price of an iteration. 

Chapter 3
Iterative Solvers for Mixed Problems

We now come to our main issue, the numerical solution of problems (2.22), (2.33)
and (2.34) which are indefinite problems, although with a well defined structure.
 
A Bt
A= . (3.1)
B0

Contact problems will also bring us to consider a more general non symmetric
system
 
A B1t
A= . (3.2)
B2 0

We intend to solve large problems and iterative methods will be essential. We shall
thus first recall some classical iterative methods and discuss their adequacy to the
problems that we consider. From there we introduce a general procedure to obtain
preconditioners using a factorisation of matrices (3.1) or (3.2).

3.1 Classical Iterative Methods

Iterative methods is a topic in itself and has been the subject of many books
and research articles. For the basic notions, one may consult [47, 50, 86] but this
is clearly not exhaustive. Moreover the field is evolving and new ideas appear
constantly. Our presentation will then be necessarily sketchy and will be restricted
to the points directly relevant with our needs.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 19

J. Deteix et al., Numerical Methods for Mixed Finite Element Problems,
Lecture Notes in Mathematics 2318, https://doi.org/10.1007/978-3-031-12616-1_3
20 3 Iterative Solvers for Mixed Problems

3.1.1 Some General Points

Linear Algebra and Optimisation

When considering iterative method, one may view things from at least two different
points of view:
• linear algebra methods,
• optimisation methods.
These are of course intersecting and each one can bring useful information.
When dealing with systems associated to matrix (3.1), from the linear algebra
perspective, we have an indefinite problem and from the optimisation perspective,
we have a saddle point problem.
• We have positive eigenvalues associated with the problem in u and the matrix A
is often symmetric positive definite defining a minimisation problem.
• On the other hand, we have negative eigenvalues associated to the problem in p,
that is the dual problem (2.24) which is a maximisation problem.
This induces a challenge for iterative methods which have to deal with conflicting
goals.

Norms

The linear systems which we consider arise from the discretisation of partial
differential equations, and are therefore special. It is also useful to see if the iteration
considered would make sense in the infinite dimensional case. Ideally, such an
iterative method would have convergence properties independent of the mesh size.
Considering system (3.1), we have a very special block structure and variables u
and p which represent functions uh ∈ Vh and ph ∈ Qh that have norms which are
not the standard norm of Rn .
• This is an important point: we have a problem in two variables which belong to
spaces with different norms.
In Remark 2.7 we introduced a matrix MQ associated with a norm in Qh , by the
same reasoning we associate a matrix MV for a norm in Vh . We thus have matrices
MV and MQ associated with these norms

MV u, v = (uh , vh )Vh , MQ p, q = (ph , qh )Qh .

This also means that residuals must be read in a space with the dual norms MV−1
−1
and MQ and this will have an incidence on the construction of iterative methods.
3.1 Classical Iterative Methods 21

Krylov Subspace

One should recall that classical iterative methods are based on the Krylov subspace,

Kr(A, b) = span(b, Ab, A2b, . . . , Ar−1 b),

looking for an approximate solution in this space. This is made possible by building
an orthogonal basis. If the matrix is symmetric one uses the Lanczos method [60]
to build orthogonal vectors. The important point is that symmetry allows to store a
small and fixed number of vectors. This is the case in the conjugate gradient method
and in the Minres [51, 74, 75, 83] algorithm.
• When A is symmetric positive definite (SPD) it defines a norm and one can look
for a solution in Kr(A, b) minimising x − A−1 b2A = Ax − b2A−1 . This is the
conjugate gradient method.
• When A is not positive definite, it does not define a norm. One then must choose
a norm M and minimise Ax − b2M −1 . This is the Minres algorithm.
When the matrix is not symmetric, the Arnoldi process [7] can be used to build an
orthogonal basis. This yields the GMRES algorithm [81] and related methods.

Preconditioning

We want to solve the problem Ku = b and we suppose that we have a preconditioner

P in the sense that P −1 is an approximation of K −1 (see [71, 88] for more details
on the preconditioner techniques).
• For K = A we shall consider a class of preconditioner in Sect. 3.2.
• For the subproblems in u that is (K = A), we shall rely on standard techniques
such as those available in the Petsc package of Argonne Laboratory [11]. For
example, we could employ as a preconditioner an Algebraic Multigrid method,
a SSOR method or an incomplete factorisation. But this is a choice and using
another technique is clearly permitted, the only criterion would be efficiency.
We shall also describe a special multigrid method in Sect. 3.1.4. The important
point is that we want to solve large scale three-dimensional problems and that
preconditioning will be a key issue.
We refer to [45] for a presentation and comparison of iterative methods and
references.
Remark 3.1 (Preconditioning and Acceleration) A preconditioner should ideally
be a converging iterative method in itself. We can then consider that the precondi-
tioned iterative method accelerates the preconditioner as well as the preconditioner
accelerating the iterative method. In our development, we shall indeed try to
build preconditioners which are convergent by themselves which may be hopefully
accelerated by some other classical iteration. 
22 3 Iterative Solvers for Mixed Problems

Given a preconditioner P and some iterative method IM we shall denote (P -IM)

the method IM preconditioned by P.
Before considering the indefinite systems (3.1) or (3.2), we shall consider the
subproblem associated with the matrix A which in many applications will be
symmetric and positive definite. Here we are on safe ground.

3.1.2 The Preconditioned Conjugate Gradient Method

We first consider the case where A is a positive definite matrix. Solving A u =

b is equivalent to minimising a functional of the form (2.4). This is a convex
functional and to minimise it, the standard procedure is a descent method. A simple
preconditioned gradient method would be

ui+1 = ui + αi P −1 (A ui − b)

The residual r = b − A u is then modified by

ri+1 = ri − αAP −1 ri .

One then minimises in α

ri+1 2A−1

which yields, denoting zi = P −1 ri

P −1 ri , ri  zi , ri 
α= −1 −1
= .
P ri , AP ri  zi , Azi 

We must remark that we have two different notions.

• A norm on the residual, here defined by A−1 .
• A preconditioner P , that is an approximate solver.
This can be improved by the classical preconditioned conjugate gradient method
(CG) [52, 81]. To use this we must have a symmetric positive definite P =
P 1/2 P 1/2 . In fact, the method is equivalent to applying the standard conjugate
gradient method to the system

P −1/2 AP −1/2 y = P −1/2 b. (3.3)

which is equivalent to the minimisation problem

1
inf (Av, v)P −1 − (b, v)P −1 (3.4)
v 2
3.1 Classical Iterative Methods 23

with u = P −1/2 y. The preconditioning can thus be seen as a change of metric on

the Euclidean space.
Remark 3.2 (Norm and Preconditioner) If we come back to the idea that the
residual is measured in the norm defined by A−1 one would now have this replaced
by P 1/2 A−1 P 1/2 . With respect to the case where the standard Rn norm would
be employed, we thus have two changes, one induced by A−1 and one by the
preconditioner. This is special to the symmetric positive definite case and will not
be directly transposed to the indefinite case where the two notions will have to be
considered in two different ways. 

Remark 3.3 (Convergence) The principle of a conjugate gradient method is to keep

the residual at every step orthogonal to all the previous directions in the norm
defined by A−1 . This is done implicitly and does not require extra computation.
It can therefore be seen as a direct method as it will provide the solution of a
problem in Rn in n steps. This would hold in exact arithmetic. Another nice property
is superlinear convergence [87]. 
To be complete, we present explicitly the preconditioned conjugate gradient method
Algorithm 3.1 P-CG algorithm
1: Initialization

• Let u0 the initial value.

• r0 = b − Au0
• z0 = P −1 r0
• p0 = z0
• i=0
2: while criterion > tolerance do
ri · zi
• αi =
pi · Api
• ui+1 = ui + αi pi
• ri+1 = ri − αi Api
• zi+1 = P −1 ri+1
zi+1 · ri+1
• βi =
zi · ri
• pi+1 = zi+1 + βi pi
• i =i+1
end while 
24 3 Iterative Solvers for Mixed Problems

3.1.3 Constrained Problems: Projected Gradient and Variants

When considering contact problems in Chap. 5 we shall have to solve a minimisation

problem with a positivity constraint. The method that we shall use is a special case
of a more general method.

Equality Constraints: The Projected Gradient Method

The projected gradient method is a very classical method for constrained optimi-
sation [18]. In the simpler case, let us suppose that the solution of a minimisation
problem must satisfy a linear constraint,

Bu = g (3.5)

If we have an initial value ug satisfying (3.5) and if we have a simple way of

projecting on Ker B, we can solve the problem,

1
inf A v0 − Aug , v0  − f, v0 .
v0 ∈KerB 2

This is in fact what we did to prove existence of the mixed problem. We shall
meet this procedure in Sect. 6.2. We can then apply the conjugate gradient method
provided the gradient is projected on Ker B. We shall also consider this method in
Sect. 4.3.

Inequality Constraints

We now consider a set of inequality constraints,

Bi u ≤ gi 1≤i≤m

Let u be the solution, we then split the constraints in two sets.

• Active constraints for which we have Bi u = gi .
• Inactive constraints for which Bi u < gi .
For more details on this strategy, we refer to [16, 17, 56, 57, 72]. If we know
which constraints are active, the problem reduces to the case of equality constraints.
In practice, this will be done iteratively.
• Given an initial guess for the active set, use an iterative method to solve the
equality problem.
• Monitor the possible changes of active constraints. If there is a change, restart
the iteration with the new set.
3.1 Classical Iterative Methods 25

The monitoring implies a change of active constraints if one has one of two
conditions.
• The solution is modified such that an inactive constraint is violated, one then
projects the solution on the constraint and restarts the iteration, now making this
constraint active.
• On an active constraint, the iteration creates a descent direction that would bring
the solution to the unconstrained region. This constraint is then made inactive.
This method is specially attractive in the following case.

Positivity Constraints

An important special case is when the solution must satisfy uj ≥ 0. The constraint
is active if uj = 0 and inactive if uj > 0. The projection on the active set is readily
computed by putting inactive values to zero (see [3, 54, 55]). The gradient (or more
generally the descent direction) is also easily projected. We shall meet this case in
contact problems (Sect. 5.2.2).

Convex Constraints

To complete this section, we give a hint on how this can be extended to a convex
constraint [18]. We consider as in Fig. 3.1 a point u0 on the boundary of the convex
set C. If the constraint is active, the gradient (or some descent direction) z is pointing
to the exterior of C. We can then project the gradient on the tangent to C (in red in
Fig. 3.1), search for an optimal point u∗ on the tangent as we now have a linear
constraint, and then project the result on C to obtain u1 . This will converge if the
curvature of the boundary of C is not too large.

Fig. 3.1 Projected gradient

for convex constraint
26 3 Iterative Solvers for Mixed Problems

3.1.4 Hierarchical Basis and Multigrid Preconditioning

There is a whole literature on preconditioning for standard finite element problems.

It would be out of scope to present it here. Let us just say, for example, that
incomplete factorisation is often very efficient. A classical combination is also to
use a SSOR method to precondition a CG method.
Multigrid methods [27, 70, 81] yield powerful preconditioning. On structured
meshes, they can be built very efficiently. On a general mesh, algebraic multigrid
methods (AMG) are also available. In our computational examples, we shall often
use a multigrid method based on a hierarchical basis. This was described in [1, 38].
The principle is simple and we use it in the case of a second degree finite element
approximation in which we represent these elements by a hierarchical basis.
• The basis function associated to vertices is the standard first degree basis
function.
• The basis function on edges is associated, not to a nodal value but to a correction
to the piecewise linear approximation.
When this is done, the matrix A can be subdivided in submatrices
 
A11 A12
A=
A21 A22

The preconditioning consists in solving with a block Gauss-Seidel iteration for the
piecewise linear part (A11) and the piecewise quadratic correction on the edges
(A22).
• It must be noted that for a typical three dimensional mesh, matrix A11 is about
eight times smaller than the global matrix A. This makes it possible to use a
direct method to solve the associated problem.
• This being a multigrid method, it can also be completed by applying an Algebraic
Multigrid (AMG) method to the problem in A11 .
• The matrix A22 has a small condition number independent of the mesh size [89]
and a SSOR method is quite appropriate.
In our examples, this preconditioner will be applied to the rigidity matrix of
elasticity problems. As we shall see, it is quite efficient for large problems.
The idea of hierarchical basis could be employed for other discretisations. In
some finite element approximations, degrees of freedom are added on the faces to
better satisfy a constraint.
Internal degrees of freedom are also amenable to this technique in a variant of
the classical ‘static condensation’.
3.1 Classical Iterative Methods 27

3.1.5 Conjugate Residuals, Minres, Gmres and the Generalised

Conjugate Residual Algorithm

We now consider the more general case of a system

Kx = b

having in mind K = A as in (3.1) or more generally (3.2).

When the matrix K is symmetric but not positive definite, the idea of keeping a
residual orthogonal to all the previous ones as in the conjugate gradient method is
still possible. However we must now choose a norm defining this orthogonality as
K −1 can no longer be used. This chosen norm will also be the norm in which we
shall minimise the residual. The simplest thing is to use the euclidean norm of Rn .
This is the Conjugate Residuals (CR) method for symmetric systems presented in
[65].
Considering Remark 3.1.1 one should rather introduce some norm which is
better adapted to our problem. Contrarily to the case of the Conjugate gradient
method, making the residuals orthogonal in some norm M −1 requires some extra
computation. Some examples can be found in [67] and also in [64] where a general
theory is developed. If r = b − Kx, z = M −1 r and T = Kr the basic idea is thus
to minimize

r − αT 2M −1

which yields

r, M −1 T 
α=
T , M −1 T 

This can be written in many equivalent ways. We refer for example to [41, 49] for
a complete presentation. In the classical Minres algorithm, the Lanczos method is
completed by Givens rotation to achieve orthogonality of residuals.
Remark 3.4 Although Minres is a well studied and popular method, we did not use
it for three reasons.
• Preconditioning: the change of metric is often said to be a SPD preconditioning.
We shall present in Sect. 3.2 preconditioners which are symmetric but not positive
definite.
• Contact problems are constrained problems and one must have access to the
part of the residual with respect to the multiplier. This was not possible to our
knowledge with the standard Minres implementation or indeed with GMRES
(see however [51]).
• Frictional contact leads to non symmetric problems.

28 3 Iterative Solvers for Mixed Problems

Our choice was rather to consider a method allowing both a general preconditioning
as a change of metric and non symmetrical matrix.

The Generalised Conjugate Residual Method

If the conjugate residual method introduced by Luenberger [65] seemed a natural

starting point, our interest in solving non symmetrical problems arising from
frictional contact led us to the use of the generalised conjugate residual (GCR)
method ([36, 39], which can be seen as a variant of the flexible GMRES method
[80]). Evidently more standard methods such as the Conjugate Gradient (CG)
method can be preferred when their use is possible.
Since the GCR method does not rely on symmetry, which means that non
symmetrical preconditioners can be employed. The price to pay is the storage of a
stack of vectors. When using the GCR method, we have two possibilities to consider
preconditioning.
• Left preconditioning: the system Kx = b is replaced by the equivalent system
P −1 Kx = P −1 b.
• Right preconditioner: the system is replaced by the system KP −1 y = b with
x = P −1 y.

Remark 3.5 (Change of Metric) We can introduce a metric M different from the
standard euclidean metric on Rn in the left and right preconditioned GCR algorithm.
The metric M is applied in the space of solutions while the dual metric M −1 is
applied to the residuals. 
We thus have two possibilities while using the preconditioner P to accelerate a
GCR method. We give here both left and right generic P -GCR using an arbitrary
metric M.

The Left Preconditioning

We first present the left preconditioner. This is the standard procedure as presented
to precondition the GMRES method in [81].
Algorithm 3.2 Left preconditioned P -GCR algorithm with an arbitrary metric M
1: Initialization

• i=0
• Let x0 the initial value.
• r0 = b − K x0
• r0P = P −1 (r0 )
3.1 Classical Iterative Methods 29

2: while criterion > tolerance do

• zi = riP
• Ti = Kzi
• zi = P −1 Ti
• From zi , using the modified Gram-Schmidt (MGS), compute zi⊥ orthonor-
mal in the M-norm to [z0⊥ , · · · , zi−1
⊥ ]. Using the same transformation on T
i
⊥ ⊥
compute Ti based on [T0 , · · · , Ti−1⊥ ].

• β = (riP , Mzi⊥ )
• Update
⎧
⎪
⎪ ri+1 = ri − βTi⊥
⎪
⎪
⎪
⎨ r P = r P − βz
i+1 i i
⎪
⎪ xi+1 = xi + βzi
⎪
⎪
⎪
⎩
i =i+1

endwhile 

The Right Preconditioning

This is the method which was used in most of our numerical results.
Algorithm 3.3 Right preconditioned P −GCR algorithm with an arbitrary metric
M −1
1: Initialization

• i=0
• Let x0 the initial value.
• r0 = b − K x0
2: while criterion > tolerance do

• zi = P −1 ri .
• Ti = K zi
• From Ti , using the MGS, compute Ti⊥ orthonormal in the M −1 -norm to
[T0⊥ , · · · , Ti−1
⊥ ]. Using the same transformation to z compute z⊥ based on
i i
⊥ ⊥
[z0 , · · · , zi−1 ].
• β = r, M −1 Ti 
30 3 Iterative Solvers for Mixed Problems

• Update
⎧
⎪
⎪ xi+1 = xi + βzi⊥
⎨
ri+1 = ri − βTi⊥
⎪
⎪
⎩
i =i+1

endwhile 

Remark 3.6 (Choice of Metric) If M −1 is the identity, that is the standard metric of
Rn , this reduce to the classical algorithm. This raise the question of the choice of
M. To fix ideas, let us consider two cases frequently met.
When K and the preconditioner P are both SPD. We would then want to obtain
a method equivalent to the preconditioned conjugate gradient method. As we have
seen in Remark 3.2 the norm should then be defined using P 1/2 K −1 P 1/2 . If the
preconditioner is good, this is close to identity. Thus M = I is not a bad choice
when we have a good precondioner.
For a symmetric system, if the preconditioner is SPD, we could take M = P .
One could then read this as a variant of Minres. 

The Gram-Schmidt Algorithm

In the GCR methods described above, we want to make a vector Ti orthogonal to

each vector in a stack in norm M or M −1 . We consider for example the case of M-
orthogonality. If the vectors in the stack are orthonormal in norm M, the modified
Gram-Schmidt method means computing,
Algorithm 3.4 Modified Gram-Schmidt orthonormalisation
1: Initialisation

• Ti given
2: for j = i − 1, . . . , 0 do

• sj = MTi , Tj⊥ ,
• Ti = Ti − sj Tj⊥
endfor
• Ti⊥ = Ti /Ti M

For the modified Gram-Schmidt method, one thus needs to compute M Ti at every
step as Ti is changed, which could be expensive. This can be avoided at the cost of
storing in the stack both Tj⊥ and MTj⊥ .
3.1 Classical Iterative Methods 31

GCR for Mixed Problems

We now come to our central concern: solving mixed problems. We now have an
indefinite system. In that case, iterative methods will often diverge or stagnate [45].
For example, without preconditioning a conjugate gradient method applied to an
indefinite system will diverge.
Remark 3.7 (We Have a Dream) As we have already stated,we are dealing with a
problem in two variables. We dream of a method which would really take this fact
into account. The preconditioner that we shall develop will. The GCR method is
then not really satisfying even if it provides good results.
Alas, nothing is perfect in our lower world. 
To fix ideas, let us consider a direct application of a right preconditioned GCR
algorithm. Let then A be as in (3.1) and b = (f, g) ∈ Rn × Rm . We are looking
for x = (u, p) and we use the right preconditioned GCR as in Algorithm 3.3. As
before we introduce an arbitrary metric. In this case the metric take into account the
mixed nature of the problem and is characterised by two different matrices Mu and
Mp giving a (Mu , Mp )-norm for (u, p).

(u, p)2(Mu ,Mp ) = u2Mu + p2Mp = (Mu u, u) + (Mp p, p)

Obviously, if both Mu and Mp are identities we have the usual right preconditionned
GCR algorithm.
Algorithm 3.5 Right preconditioned Mixed-P-GCR algorithm with arbitrary met-
ric (Mu , Mp )
1: Initialization

• i=0;
• Let x 0 = (u0 , p0 ) the initial value.
• ru = f − Au0 − B t p0 , rp = g − Bu0 and r = (ru , rp )
2: while criterion > tolerance do

• From r = (ru , rp ) the preconditionner yields zi = (ziu , zip ).

• Tiu = Aziu + B t zip , Tip = B ziu
• From T i = (Tiu , Tip ), using the MGS compute T ⊥ i orthonormal in
(Mu−1 , Mp−1 )-norm to [T ⊥ 0 , · · · , T ⊥ ]. Using the same transformation to
i−1
zi compute z⊥ i based on [z⊥ , · · · , z⊥ ].
0 i−1
• β = riu , Mu−1 Tiu
⊥  + r , M −1 T ⊥ 
ip p ip
32 3 Iterative Solvers for Mixed Problems

• Update
⎧
⎪
⎪ x = x i + βz⊥
⎨ i+1 i

r = r − βT ⊥
i
⎪
⎪
⎩
i =1+1

end while 

Remark 3.8 In the above GCR algorithm, we could decompose β in two component
⊥
β = βu + βp βu = ru , Mu Tiu , βp = rp , Mp Tip⊥ 

If the system is not properly preconditioned, βu or βp can become negative and

β = βu + βp can even become null, hence stagnation. This is the symptom of a
bad preconditioning. Another issue is that giving an equal weight to u and p in the
scalar product is not a priori the best choice, although the right choice is seldom
clear. This fact is in reality common to any iterative method applied to a system of
this form. 
The key is therefore in the construction of good preconditioners or in building
optimisation methods adapted to saddle-point problems. This will be our next point.
Remark 3.9 (The Case of Sect. 2.1.4: Perturbed Problems) Whenever one wants
to solve a problem of the form (2.12) one should make a few changes.
• In the initialisation phase, one should have rp = g − Bu − Rp
• After the preconditioner, one should have Tp = Bzu − Rzp .


3.2 Preconditioners for the Mixed Problem

3.2.1 Factorisation of the System

This section is centered on the development of an efficient preconditioner for the

system
   
u f
A = (3.6)
p g

with A an indefinite matrix of the form (3.1) or more generally for the non
symmetric case (3.2). Our first step toward a general solver for systems (3.6) will be
3.2 Preconditioners for the Mixed Problem 33

to consider a block factorisation LDU of the block matrix A

  
I 0 A 0 I A−1 B2t
A= := LDU (3.7)
B1 A−1 I 0 −S 0 I

where the Schur complement which we already considered in Sect. 2.2.3 S =

B1 A−1 B2 t is invertible, we then have,
  
−1 I −A−1 B2t A−1 0 I 0
(LDU) = (3.8)
0 I 0 −S −1 −1
−B1 A I

If A is symmetric and B1 = B2 we have U = LT and the full block decomposition

F LDLT is symmetric but indefinite. To simplify, we shall focus our presentation
on the symmetric case as the extension to the non symmetric case is straightforward.
Remark 3.10 (Regular Perturbation) We have considered in Sect. 2.1.4 the case of
a regular perturbation where Bu = g is replaced by Bu − Mp = g. In that case S
is replaced in the factorisation (3.7) by S + M 

Solving Using the Factorisation

Let (ru , rp ) be the residual of the system (3.6) for (u0 , p0 )

ru = f − Au0 − B t p0 rp = g − Bu0 .

Using the factorisation (3.8) to obtain (u, p) = (u0 + δu, p0 + δp) leads to three
subsystems, two with the matrix A and one with the matrix S.
Algorithm 3.6
⎧ ∗
⎨ δu = A−1 ru
δp = S −1 (Bδu∗ − rp ) (3.9)
⎩
δu = A−1 (ru − B t δp) = δu∗ − A−1 B t δp.


It must be noted that this is symmetric as it is based on (3.8). Particular attention
must be given to the solution

S δp = BA−1 B t δp = Bδu∗ − rp . (3.10)

With the exception of very special cases, S is a full matrix and it is not thinkable
of building it explicitly. Even computing matrix-vector product (needed by any
iterative solver) would require an exact solver for A which may become impossible
for very large problems. Therefore, except in rare cases where systems in A can
34 3 Iterative Solvers for Mixed Problems

be solved extremely efficiently Algorithm 3.6 is practically unusable. Thus in most

cases efficiency will lead us to iterative solvers for A which can avoid the need for
an exact solver for S.
Nevertheless (3.9) suggests a way to precondition A. To do so, we shall first
write an approximate form of the factorisation (3.7).
Assuming that we have an ‘easily invertible’ approximation A of A and inserting
it in (3.7) to replace A, we obtain
  
I 0 A 0 I A−1 B t
A= := LDU (3.11)
B A−1 I 0 −S 0 I

where

S = B A−1 B t (3.12)

If we now have an approximation S of S, we can introduce

Algorithm 3.7 A general mixed preconditioner (GMP)
⎧ ∗
⎨ zu = A−1 ru
z = S −1 (Bzu∗ − rp ) (3.8)
⎩ p
zu = A−1 (ru − B t zp ) = zu∗ − A−1 B t zp .


Using the factorisation (3.11), we may associate (3.8) with a matrix A

A Bt
A= (3.9)
B BA Bt − S
−1

This suggests many variants for a preconditioner of A. Their implementation

relies on two essential tools.
1. An easily computable and fast approximate solver A−1 . For this, we can rely
on classical iterative methods such as the Conjugate Gradient method when
the matrix A is symmetric and positive definite, the GMRES method or the
Generalised Conjugate Residual method. Many other methods are available on
general packages such as Petsc or Lapack. In all cases, a good preconditioner
will be essential. We shall often rely on the multigrid method of Sect. 3.1.4.
2. An easily computable and fast solver S to approximate S.
This contains in particular the case A = A and therefore S = S.
Remark 3.11 (A Solver in Itself) Although we shall use Algorithm 3.7 as a
preconditioner, it should be noted that it is also an iterative method in itself and that
it can be employed as a solver. Using it as a preconditioner in an iterative method
can be seen as a way of accelerating it (Remark 3.1).
3.2 Preconditioners for the Mixed Problem 35

3.2.2 Approximate Solvers for the Schur Complement and the

Uzawa Algorithm

In the following, we consider as in (3.12) approximations of the Schur complement

S or its approximate form S defined in (3.12).
When the solver in u is exact, we have S = S = BA−1 B t and the preconditioner
of Algorithm 3.7 reduces to using S instead of S. Therefore, in any cases, S is related
to a Schur complement matrix. We will not explicitly construct a matrix S. Instead
the equation S v = r in (3.8) should be interpreted as an approximate solution of
S v = r. We thus describe how we can employ an iterative method to obtain a
computable S. This iterative method can then be used with a more or less stringent
desired precision or a limited number of iterations. We also need a preconditioner
for this iteration, proceeding as follows
• We introduce an easily invertible approximation MS of S, following the discus-
sion Sect. 2.2.3
• We solve Szp = rp by a an iterative method using MS as a preconditioner. In
practice we restrict this step to a few (most often one) iteration.

Remark 3.12 (The Choice of MS ) This is an important point as it can make

the difference between an efficient method and a poor one. The choice evidently
depends on the problem. We have considered in Sect. 2.2.3 and in particular in
Proposition 2.2 how the choice for MS = R or MS satisfying (2.28) would be
adequate provided the inf-sup condition holds and Av, v is a norm on Vh , that are
the conditions to have a well posed problem.
• For problems of incompressible materials which we shall consider in Sect. 4.2,
the choice MS = M0 where we denote M0 the ‘mass matrix’ associated to the L2
scalar product in Q is a good choice. Indeed it was shown in [41] that this choice
leads to an iteration count independent of the mesh size (See Proposition 2.2.3).
In this case S is an operator of order 0.
• For contact problems, MS = M0 is less efficient as we do not have the inf-sup
condition in L2 and S is now of order 1. We shall discuss this in Sect. 5.1. We
shall also show how a discrete Steklov-Poincaré operator could be built.
• One could also think of a BFGS (Broyden, Fletcher, Goldfarb, and Shannon)
update [63] to improve MS during the iterative process.
• One should also recall (Sect. 2.3.3) that the Augmented Lagrangian Method
yields a preconditioning for the dual problem.

We now make explicit two iterative processes. As we noted above, they will be the
same for the solution of Sp = rp or Sp = rp . For the GCR method, we have the
following algorithm.
36 3 Iterative Solvers for Mixed Problems

Algorithm 3.8 A right preconditioned GCR iterative solver for S

1: Initialization

• i=0;
• Let zu and r0p be given values.
• zp = 0
2: while criterion > tolerance or maximum number of iterations do

• zzip = MS−1 rip

• zziu = −A−1 B t zzip
• Tip = Bzziu .
• Use the MGS procedure to obtain Tip⊥ orthonormal to the previous directions.
⊥ and zz⊥ .
Apply the same transformation to zzip and zziu to get zzip iu
• β = (Tip⊥ , rip )
• Update
⎧ ⊥
⎪
⎪ zp = zp − βzzip
⎨
⊥
⎪ zu = zu − βzziu
⎪
⎩
i =i+1

end while 

One could also use the more general version of Algorithm 3.3 with a change of
metric.
Remark 3.13 The computation of zu is optional. It avoids an additional use of A−1
when this is included in the preconditioner 3.7 to the price of some additional work
in the Gram-Schmidt process. 
If A and MS are symmetric, we can also use the simpler Conjugate Gradient
method, which takes the following form

Algorithm 3.9 A CG iterative solver for S

1: Initialisation

• i=0
• zu and rp , given values
2: while criterion > tolerance or maximum number of iterations do

• zzp = MS−1 rp
• If i > 0 α = (zzp , rp )/(zzp0 , rp0 )
• wp = zzp + α wp0
• zzu = −A−1 B t wp
• Tp = Bzzu .
3.2 Preconditioners for the Mixed Problem 37

• β = (wp , rp )/(Tp , wp )
• wp0 = wp , zzp0 = zzp , rp0 = rp
• Update
⎧
⎪
⎪ zp = zp + βzzp
⎪
⎪
⎪
⎨ zu = zu − βzzu
⎪
⎪ rp = rp − βTp .
⎪
⎪
⎪
⎩
i =i+1

end while 

The Uzawa Algorithm

When A = A we recall that S = S and for sufficiently strict convergence criteria the
last algorithm corresponds to a solver for S. Then when included in Algorithm 3.7
it yields solutions of (3.6) that is Algorithm 3.7 coincides with Algorithm 3.6. In
[43], this was called the Uzawa algorithm, which we can summarise as formed of
two steps illustrated in Fig. 3.2.
• Solve the unconstrained problem Au = f .
• Project this solution, in the norm defined by A, on the set Bu = g.
In [43] Uzawa’s algorithm was presented in its simplest form depending on an
arbitrary parameter β. The parameter β must then be chosen properly and depends
on the spectrum of BA−1 B t . This was studied in detail and convergence follows as
in the classical analysis of gradient methods (an acceleration by a conjugate gradient
method was also considered. Using this method implies that one has an efficient
solver for A.
Many variants of this method are proposed in the literature (see [15, 81]).
Several authors have studied numerically and in the theoretical framework of the
variants of the method. Elman and Golub [40] proposed the Uzawa method called

Fig. 3.2 Illustration of the

Uzawa algorithm
38 3 Iterative Solvers for Mixed Problems

inexact whose theoretical analysis is made in [23]. Bai et al. [10] presented the
so-called parametrised method. More recently, Ma and Zang in [66], dealt with
the so-called corrected method. They have shown that it converges faster than the
classical method and several of its variants under certain assumptions. However,
their approach come up against the question of determining the optimal parameter.

3.2.3 The General Preconditioned Algorithm

We now come to the use of Algorithm 3.7: the General Mixed Preconditioner.
• We must first chose A−1 . For this we rely on standard and well proven iterative
methods. Whenever possible we precondition these methods by a multigrid
procedure.
• We also need an approximate solver for the approximate Schur complement S.
• We also choose a norm N in which we minimise residuals. This will most of
times be the euclidean norm but better choices are possible
If for S we use Algorithm 3.9 or Algorithm 3.8 (i.e. we solve S) then we use a
limited number of iterations. Choosing S = S (i.e. fully converging Algorithm 3.9
or Algorithm 3.8) is a possibility, but in general it is not a good idea to solve too
well something that you will throw away at the next iteration. We shall thus develop
the case where only one iteration is done.
Remark 3.14 One should note that using a better A and more iterations for S is a
direct way to make things better.
• If A = A we have the Uzawa algorithms or close variants.
• If S is solved exactly, we have a form of the projected gradient method.

This being said we shall focus on a simple form of Algorithm 3.7.
Algorithm 3.10 A simple mixed preconditioner
1: Initialization

• ru , rp given
• zu = A−1 ru
• rp = Bzu − rp
2: Approximation of S −1

• zp = MS−1 rp
• zzu = A−1 B t zp
• Tp = Bzzu
(rp , Tp )
• β=
(Tp , Tp )
3.2 Preconditioners for the Mixed Problem 39

Fig. 3.3 Illustration of the

preconditioner

3: Final computation

• zp = βzp
• zu = zu − βzzu
4: End

This essentially amounts to obtaining a scaling by β of MS−1 rp . The computation of

β implies the computation of zzu , which implies a resolution with the approximate
solver A−1 . This computation would anyway be required by the last part of
Algorithm 3.7 so that there is in fact no extra cost. If we have symmetry of A and
MS , we can obtain β by one iteration of the gradient method instead of a minimum
residual.
Remark 3.15 (The Preconditioner Yields a Saddle-Point) It is interesting to see in
Fig. 3.3 that the preconditioner yields a saddle-point in a two-dimensional space.
One can also see that zzu is A-orthogonal (conjugate) to zu
This makes us think of the Partan method [82] where conjugation is obtained
after two descent steps (see Fig. 3.4).


Remark 3.16 (The Classical Method of Arrow-Hurwicz-Uzawa) We shall later use,

for the sake of comparison the classical Arrow-Hurwicz-Uzawa method, described
in [43], as a preconditioner.
Algorithm 3.11 A super-simple mixed preconditioner
1: Input ru , rp , Output zu , zp .
2: Compute zu and zp

• zu = A−1 ru
40 3 Iterative Solvers for Mixed Problems

Fig. 3.4 Illustration of

Partan’s method

• rp = rp − Bzu
• zp = βMS−1 rp

The problem here is that β has to be determined by the user, while in the previous
version, everything was automatic. The choice of parameters had been discussed in
[43]. Moreover, as we said above, the last part of Algorithm 3.7 requires an extra
resolution. 

Remark 3.17 (The Perturbed Problem) We have considered in Sect. 2.1.4, a

regular perturbation. We would then have to solve a problem of the form
    
A Bt u f
=
B −MQ p g

where MQ is the matrix defining the metric on Q. We have noted in Remark 2.13
that S then becomes S + MQ and that MS is changed into (1 + )MS if MQ = MS
.
The preconditioner of Algorithm 3.10 is easily adapted to the perturbed case. If
one uses MS = MQ one should change,
• rp = Bzu − MQ p − rp ,
1
• zp = M −1 rp
1+ Q
• Tp = Bzzu − MQ zp
When the preconditioner 3.10 is employed for the modified problem of Sect. 2.1.4,
one should also modify the computation of residuals in the associated CG or GCR
method, taking into account for example Remark 3.9. 
3.2 Preconditioners for the Mixed Problem 41

3.2.4 Augmented Lagrangian as a Perturbed Problem

We have considered in Sect. 2.3.4 a form of the augmented Lagrangian method in

which one has to solve a sequence of problems of the form (2.36), penalty problems
written in mixed form.
This could be considered when a classical Augmented Lagrangian as in (2.34)
would not be possible because R −1 is a full matrix. This could also be an alternative
to the regularised lagrangian.
Using this implies a two level iteration, which can be compared to a Newton
method. To fix ideas, we describe an implementation by the GCR algorithm.
• Initialise a GCR method with the residuals of the unperturbed problem.
• Solve for δu and δp with the penalty problem in mixed form using the
preconditioner as Remark 3.17 and the GCR iteration following Remark 3.9 but
without changing the initial residuals.
• Update u + δu and p + δp.
From there, one has many variants. The internal iteration may not be too exact.
The update could itself be accelerated by another GCR. When the problem is the
linearised part of a Newton iteration, one could in fact use this as an inexact Newton.
Chapter 4
Numerical Results: Cases Where Q = Q

We shall now illustrate the behaviour of these algorithms on some examples. In all
the cases considered here, the space of multipliers Q can be identified with its dual.
In fact we shall have in all cases Q = L2 ().
The first example will be the mixed approximation of a Poisson problem with
the simplest Raviart-Thomas element. This will allow us to consider a real discrete
Augmented Lagrangian method and its impact in iterative solvers. We shall then
consider incompressible problems in elasticity in both the linear and non linear
cases. Finally we shall introduce an application to the Navier-Stokes equations

4.1 Mixed Laplacian Problem

The first example that we consider is the mixed formulation of a Dirichlet problem
using Raviart-Thomas element. As an application, we can think of a potential fluid
flow problem in porous media as in the simulation of Darcy’s law in reservoir
simulation. This can be seen as the paradigm of mixed methods and the simplest
case of a whole family of problems. Higher order elements and applications to a
mixed formulation of elasticity problems are described in [20].

4.1.1 Formulation of the Problem

Let  a domain of Rd with d = 2 or 3. Its boundary  is divide into D and N on

which respectively Dirichlet and Neumann conditions will be imposed. We denote
n the normal to . Given f ∈ L2 () and g ∈ (L2 (N ))d , the problem consists in

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 43

J. Deteix et al., Numerical Methods for Mixed Finite Element Problems,
Lecture Notes in Mathematics 2318, https://doi.org/10.1007/978-3-031-12616-1_4
44 4 Numerical Results: Cases Where Q = Q

finding a function p ∈ H 1 () such that

⎧
⎪
⎨−p
⎪ = f in 
p = 0 on D (4.1)
⎪
⎪
⎩(grad p) · n = g on 
N

Taking u = grad p, Eq. (4.1) is equivalent to

⎧
⎪
⎪u − grad p =0 in 
⎪
⎪
⎨− div u =f in 
(4.2)
⎪
⎪p =0 on D
⎪
⎪
⎩
u·n =g on N

Let

V = H (div, ) = {u ∈ (L2 ())d , div u ∈ L2 ()}

and

Q = L2 ().

The variational formulation of (4.2) is then to find u ∈ V and p ∈ Q satisfying

⎧  
⎪
⎪ u · v dx + p div v dx = g · v ds ∀ v ∈ V,
⎪
⎨   N
⎪   (4.3)
⎪
⎪
⎩ div u q dx = − f q dx ∀q ∈ Q
 

In this context, the inf-sup condition [20] is verified and, the operator a(·, ·) defined
on V × V by

a(u , v) = u · v dx


is coercive on the kernel of the divergence operator but not on the whole space V .
The mixed formulation (4.3) is the optimality condition of the following inf-sup
problem
   
1 2
inf sup |v| dx − g · v ds + q div v dx + f q dx (4.4)
v∈V q∈Q 2  N  
4.1 Mixed Laplacian Problem 45

In order to obtain a discrete form of the problem, we introduce the Raviart-Thomas

elements. We suppose to have a partition Th of  into tetrahedra. If we denote by
Pk (K) the polynomials of degree k on a tetrahedron K, we define,

RTk () = {uh ∈ (Pk (K))d + xPk (K) ∀K ∈ Th } (4.5)

The properties of these spaces and related ones are well described in [20]. Indeed
they have been built to be applied to problem (4.3). If we take

Qh = {qh ∈ Pk (K) ∀K ∈ Th }

the matrix R corresponding to the scalar product on Qh is block diagonal and we

have the important property that div uh ∈ Qh , the inclusion of kernels property
which ensures that coercivity on the kernel is also valid for the discrete problem.
The discrete operator divh corresponding to the definition of (2.16) is indeed the
restriction of div to Vh .

4.1.2 Discrete Problem and Classic Numerical Methods

The discrete problem is clearly of the form (2.22) and indefinite. In [20, p. 427], this
was considered ‘a considerable source of trouble’. Let us consider things in some
detail.
• The matrix A is built from a scalar product in L2 ().
• The operator B is the standard divergence
• B t is a kind of finite volume gradient.
• The Schur complement BA−1 B t is a (strange) form of the Laplace operator
acting on piecewise constant.
It was shown in [12] how using a diagonalised matrix for the matrix A, one
indeed obtains a finite volume method. In the two-dimensional case, it is known
(see [68]) that the solution can be obtained from the non-conforming discretisation
of the Laplacian.
Another general approach (see [6] for example) is to impose the interface
continuity of the normal components in Vh by Lagrange multipliers to generate a
positive definite form.
We shall rather stick to the indefinite formulation an show how the methods that
we developed in Chap. 3 can be applied.

The Augmented Lagrangian Formulation

The first point is that we have a lack of coercivity on the whole space. The result of
Proposition 2.2 does not hold and the convergence of Uzawa’s method, for example
46 4 Numerical Results: Cases Where Q = Q

would not be independent of h. To obtain the coercivity on the whole of V we shall

consider an augmented Lagrangian method, which also means to use

a(u, v) + α(div u, div v).

Solving the equation with the augmented Lagrangian method gives us the inf-sup
problem
 
1 α
inf sup |v|2 dx+ | div v + f |2 dx
v q 2  2 
   (4.6)
− g · v ds + q div v dx + f q dx
N  

where α > 0 is a scalar representing the parameter of regularisation. It’s well known
that the equation (4.4) is equivalent to (4.6) for which optimality conditions are
⎧  
⎪
⎪
⎪
⎪ u · v dx + α div u div v dx + p div v dx
⎪
⎪
⎪
⎪
 
 

⎨
= g · v ds − α f div v dx ∀ v ∈ V, (4.7)
⎪
⎪ N 
⎪
⎪  
⎪
⎪
⎪
⎪
⎩ div u q dx = − f q dx ∀ q ∈ Q.
 

Here, for all α greater than zero, the coercivity is satisfied on V since

a(v, v) + α div u div v dx ≥ min(1, α 2 )v2V , ∀ v ∈ V,


with
  
v2V = |u|2 dx + | div v|2 dx
 

A consequence is that the two forms of the augmented Lagrangian (2.30) and
(2.32) are the same and that using them will not change the solution of the problem.
Obviously (2.32) corresponds to the linear system of equations associated to the
discrete version of (4.7). As we said earlier, for the Raviart-Thomas elements MS =
R is a block diagonal matrix associated to the L2 scalar product on Qh and we are
in the perfect situation (Remark 2.11) for the augmented Lagrangian.
    −1 
A + αB t R −1 B B t u g − αB t MD f
= .
B 0 p f
4.1 Mixed Laplacian Problem 47

4.1.3 A Numerical Example

For our numerical tests, we take k = 0 in (4.5). Qh is thus a space of piecewise

constants. For the numerical example, we consider the domain  as a cube of unit
edges. We take D empty so N = ∂. The functions f and g are given by:

f (x, y, z) = −3π 2 sin(πx) sin(πy)sin(πz)

g(x, y, z) = 0

We use a tolerance of 10−12 for the convergence of the algorithms.

The interest of this example is to see the effect of the augmented Lagrangian
formulation and the role of the parameter α. We would like to emphasise again
that we are in an ideal case where the augmented Lagrangian does not change the
solution.
• Small values of α act as regularisation parameter. They provide coercivity on the
whole of V and not only on the kernel of the divergence operator.
• Large values of α yield a penalty method which is then corrected by an iteration.
This is the classical augmented Lagrangian method.
To evaluate the effect of the augmented Lagrangian we first present results in which
we use a relatively crude mesh with 50 688 tetrahedra and 24 576 faces.
We first consider the case of a direct solver for the primal problem in u. This
is therefore the Uzawa algorithm of Sect. 3.2.2. The number of iterations in p
depends on the condition number of the dual problem (2.23) which improves when
α becomes large. This is observed in our experiences. Table 4.1 illustrate that a
larger α yields a better convergence as should be expected. In fact we see that the
largest improvement is to have coercivity on the whole space (taking α > 0) rather
then having coercivity only on the kernel (α = 0). Notice that increasing values of α
does not improve things as the condition number of the problem in u becomes bad.
We can compare this with an incomplete resolution in u. This is done by using
a simple iterative method: a fixed number of the conjugate gradient method with an
SSOR preconditioner, which we denote CG(n). The results presented in Table 4.2
were obtained for n = 5 and n = 10.
We observe once again a large difference between α = 0 and α > 0. We also
see that there is an optimal α and that for too large values the condition number

Table 4.1 Laplace problem α # iterations CPU time (s)

with coarse mesh and
complete (LU) resolution for 0 133 18.57
the primal problem: global 10−1 7 3.2995
number of iterations and 102 5 2.8836
computing time in seconds 105 3 2.5867
according to the value of α 107 3 2.5305
48 4 Numerical Results: Cases Where Q = Q

Table 4.2 Laplace problem CG(5)-SOR CG(10)-SOR

with coarse mesh with
α # it. CPU time(s) # it. CPU time(s)
CG(5)-SOR and
CG(10)-SOR for the primal 0 133 13.6105 133 14.8259
solver: number global of 1.0 × 10−4 80 7.0878 79 8.8246
iterations (# it.) and 5.0 × 10−4 45 4.2140 42 5.1697
computing time in seconds 1.0 × 10−3 38 3.7885 31 4.2228
according to the value of α
2.5 × 10−3 37 3.9470 25 3.7092
5.0 × 10−3 37 3.7123 24 3.6260
1.0 × 10−2 47 4.4617 22 3.4336

Table 4.3 Laplace problem LU CG(100) CG(50)

with fine mesh: total number
(α = 105 ) (α = 2.5 × 10−3 ) (α = 10−2 )
of iterations and computing
time in seconds according to # it. 3 15 25
the solver of primal problem CPU (s) 808.5 586.7 569.1

in u worsens the performance. If we improve the resolution, we have indeed better

results.
With a better solver, the optimal value of α increases and the number of iterations
in p decreases. However, the computing time does not as each iteration becomes
more expensive.
Since the goal is to solve big problems, we took a rather fine mesh (with
3,170,304 faces and 1,572,864 tetrahedra) to prove the interest of the algorithm with
an incomplete resolution. In this context, the direct resolution is very expensive in
time as can be seen in Table 4.3 whereas in iterative resolution, the computation
time is less.
We can see that an iterative method for u can be better than a direct solver.
This result could surely be much improved by a more clever choice of this iterative
method.
Remark 4.1 (This Is not Optimal) We would like to emphasise that the choice of
the solver in u is not optimal and that one could improve the results. Our main point
is that the indefinite form of mixed methods is not such a ‘source of trouble’! 

4.2 Application to Incompressible Elasticity

We consider a second class of examples, arising from the discretisation of incom-

pressible elasticity problems. We consider two cases : a standard linear elasticity
model and a Mooney-Rivlin model [21, 53, 84]. This relatively simple case will
allow us to compare the preconditioners introduced and to discuss strategies to
improve coercivity.
4.2 Application to Incompressible Elasticity 49

4.2.1 Nearly Incompressible Linear Elasticity

Given a body , we want to determine its displacement u under external forces.

Denoting  the boundary of , we define D the part of  on which Dirichlet
(displacements) conditions are imposed and N the part where we have Neumann
(forces) conditions. To avoid unnecessary developments, we restrict ourselves to the
case of null conditions on D . We thus define,

V = {v |v ∈ (H 1 ())d , v = 0 on D }

For v ∈ V , we define the linearised strain tensor

1
ε (v) = (∂i uj + ∂j ui ) (4.8)
ij 2
and its deviator
1
εD = ε − tr(ε)I
3
One then has,

1
|εD (v)|2 = |ε(v)|2 − tr(ε(v))2 (4.9)
3
To define our problem, we have to define some parameters. Elasticity problems are
usually described by the Young Modulus E and the Poisson ratio ν. We shall rather
employ the Lamé coefficients μ, λ

E Eν
μ= , λ= (4.10)
2(1 + ν) (1 + ν)(1 − 2ν)

and the bulk modulus k

E 2μ
k= =λ+ (4.11)
3(1 − 2ν) 3

We then consider the problem

  
k
inf μ |εD (v)|2 dx + | div v − g|2 dx − f · v dx (4.12)
v∈V  2  

or equivalently,
  
λ
inf μ |ε(v)|2 dx + | div v − g|2 dx − f · v dx. (4.13)
v∈V  2  
50 4 Numerical Results: Cases Where Q = Q

Problem (4.13), for example, leads to the variational form,

  
2μ ε(u) : ε(v) dx + λ (div u − g) div v dx = f · v ds ∀ v ∈ V.
  
(4.14)

It it is well known that a brute force use of (4.14) or (4.13) could lead to bad
results for large values of λ (or as the Poisson ratio ν nearing its maximal value of
1/2). In extreme cases, one gets a locking phenomenon that is an identically zero
solution.
The standard way to circumvent this locking phenomenon is to switch to a
mixed formulation with a suitable choice of elements. Essentially, we introduce the
variable

p = λ(div v − g)

and we consider the saddle-point problem,

 
1
inf sup μ |ε(v)| dx− 2
|q|2 dx
v q  2λ 
 
+ q (div v − g)dx − f · v ds
 

for which optimality conditions are and denoting (u, p) ∈ V × Q the saddle point,
⎧   
⎪
⎨ 2μ ε(u) : ε(v) dx +
⎪ p div v dx = f · v ds ∀ v ∈ V,
  
   (4.15)
⎪
⎪ 1
⎩ div u q dx − pq dx = g qdx ∀q ∈ Q
 λ  

In the limiting case of λ becoming infinite, the case that we want to consider, the
second equation of (4.15) becomes
 
div u q dx = g qdx ∀q ∈ Q
 

Problem (4.15) is well posed. Indeed we have,

• The bilinear form a(u, v) =  ε(u) : ε(v) dx is coercive on V. This is Korn’s
inequality : there exist a constant α such that

a(v, v) = μ(|ε(v)|2 ) ≥ αv2V ∀v ∈ V

4.2 Application to Incompressible Elasticity 51

• We have an Inf-sup condition [20]: there exists a constant β such that

b(v, q)
sup ≥βq ∀q ∈ Q.
v∈V v

Remark 4.2 If we use formulation (4.12) instead of (4.13) as the starting point, the
bilinear form

a D (u, v) = εD (u) : εD (v) dx


is coercive on the kernel of B. Indeed, by Korn’s inequality and (4.9) : there exist a
constant α such that
 
1 μ
μ|εD (v)|2 = μ |ε(v)|2 − tr(ε(v))2 ≥ αv2 − tr(ε(v))2 ) ∀v ∈ V
3 3

This makes it possible for the matrix AD defined by the bilinear form a D (·, ·) to be
singular, a situation which is not be acceptable in our algorithms. 

Remark 4.3 In the limiting case of infinite λ, that is of incompressible materials,

problems (4.12) and (4.13) are equivalent as we have div u = tr(ε(u)) = 0, the
solution is unchanged and for (4.13) we have coercvity on the whole of V and
not only on the kernel of B. However, for the discretised problem, the solution is
modified as in general one does not have div uh = tr(ε(uh )) = 0. 
In our numerical tests, we shall use the Augmented Lagrangian but as in (2.30) that
is a ‘regularised Lagrangian’
⎧  
⎪
⎪ 2μ ε(u) : ε(v) dx + λ̂ (div u − g) div v dx
⎪
⎪
⎪
⎪  
⎪
⎨  
+ p div v dx = f · v ds = 0 ∀ v ∈ V,
⎪
⎪
⎪
⎪  
 
⎪
⎪
⎪
⎩ div u q dx = g qdx ∀q ∈ Q
 
(4.16)

Where we define λ̂ using an artificial Poisson ratio ν̂ as in (4.10) which should in

no way be close to 1/2 unless the choice of elements allows really divergence-free
solution. The parameter λ̂ is arbitrary and is chosen to obtain a good convergence
without downgrading the condition number of the system. This will be discussed in
Sect. 4.2.5.
Remark 4.4 (The Stokes Problem) It is clear that all we say about this linear
elasticity problem is valid for the equivalent Stokes problem for creeping flow
52 4 Numerical Results: Cases Where Q = Q

problems. The numerical solution of the Stokes problem has been the object of a
huge number of articles and books. 

4.2.2 Neo-Hookean and Mooney-Rivlin Materials

We now consider non linear elasticity models described in a Lagrangian formula-

tion. We thus have a reference domain denoted  and a deformed domain ω while
their respective boundaries are denoted as  and γ . We denote X and x respectively
their coordinates. We write the deformation of a body as

x = u(X)

∂x
The deformation F of this transformation is given by F = = I + ∇X u and its
∂X
determinant detF is denoted J . Note that ∇X stands for the gradient with respect
to the variable X. The Cauchy-Green tensor is then defined as C = FT F and its
principal invariants I1 , I2 and I3 are given by :

1 2
I1 = C : I, I2 = (I − C : C), I3 = det(C) = J 2 .
2 1
As in the case of linear elasticity, the boundary  is composed of (at least) two
parts: D where a Dirichlet condition is given and N where a Neumann (pressure)
condition g is imposed.
• The Neo-Hookean model
Although there are many formulations of Neo-Hookean models for compress-
ible materials, they share a unique elastic potential energy function or strain
energy function W as named in [73]. Following [91], we define a particular
Neo-Hookean material where the potential energy function W is given by

μ λ λ
W= (I1 − 3) + (J 2 − 1) − ( + μ) ln J
2 4 2
 
μ − 13 1
= I I1 − 3 + κ(J − 1)2
2 3 2

where the material is characterized by μ and λ the Lamé coefficients as in the

linear model and κ the bulk modulus.
• The Mooney-Rivlin model
For a Mooney-Rivlin hyperelastic material, the energy functional is given by
 1   2 
−3 −3 1
W = μ10 I3 I1 − 3 + μ01 I3 I2 − 3 + κ(J − 1)2
2
4.2 Application to Incompressible Elasticity 53

where κ, the bulk modulus, μ10 and μ01 are parameters characterizing the
material.
The elasticity problem consists in minimizing the potential energy W under appro-
priate boundary conditions. The weak formulation on the reference configuration 
can be written as a nonlinear problem,
  
(F · S) : ∇X v dX = f · v dX + g · v dγ (4.17)
  N

for any v in a proper functional space and where S = 2∂W /∂C is the second Piola-
Kirchoff stress tensor. More details on the formulation can be found in [36].

Mixed Formulation for Mooney-Rivlin Materials

We are interested in the simulation of incompressible materials, more particularly

rubberlike materials. As in the case of linear materials, it is well known that
displacement-only formulations are inadequate. As the material becomes more
and more incompressible (corresponding to an increasing bulk modulus), the
conditioning of the linearized system grows. And the usual locking phenomena may
occur when the bulk modulus κ becomes large. As a consequence, direct solvers and
iterative solvers are ineffective (even with preconditionning).
It is then convenient to split the second Piola-Kirchoff tensor S into a volumic
part and an isochoric part and to use a mixed formulation in which p is explicitly an
unknown. We thus write

S = S − pJ C−1 .

Here p is the pressure which is defined by p = −κ(J − 1). For a Mooney-Rivlin

hyperelastic material (of which a Neo-Hookean material is a special case),
   
−1/3 1 −2/3 2
S = 2μ10I3 I − I1 C−1 + 2μ01I3 I1 I − C − I2 C−1
3 3

The mixed formulation can now be written as:

⎧ 
⎪
⎪ · 
: ∇ − pJ F−T : ∇X v dX
⎪
⎪ (F S ) X v dX
⎪
⎪
⎪ 
⎪ 


⎨
= g · v dS + f · v dX ∀v ∈ V (4.18)
⎪
⎪ N 
⎪
⎪  
⎪
⎪
⎪
⎪ 1
⎩ (J − 1)q dX + pq dX = 0 ∀q ∈ Q
 κ 
54 4 Numerical Results: Cases Where Q = Q

We are interested in the really incompressible case when the bulk modulus
becomes infinite. As we have seen in the linear case, this may lead to an ill
conditioned or even singular matrix in u. To get good results, we shall again
introduce a stabilisation parameter K̂. We then define Ŝ using this artificial small
bulk modulus. and solve
⎧ 
⎪
⎪
⎪
⎪ (F · Ŝ) : ∇X v dX − pJ F−T : ∇X v dX
⎪
⎪
⎪
⎪




⎨
g · v dS + f · v dX ∀v ∈ V (4.19)
⎪
⎪ N 
⎪
⎪ 
⎪
⎪
⎪
⎪
⎩ (J − 1)q dX = 0 ∀q ∈ Q


This is a non linear formulation, its resolution will be based on a Newton-like

method and will imply a linearisation of (4.18). To lighten the presentation (we
refer to [42] for the derivation of the linearised problem), let us rewrite the non
linear problem (4.18) as

R1 ((u, p), v) = 0,
R2 ((u, p), q) = 0.

At each Newton’s iterations, we consider a linearized version of this system with

respect to u and p. Linearizing around (un , pn ), we get the followings bilinears
operators

∂R1 ((un , pn ), v)
an (δu, v) = · δu
∂u

 T
= S (un ) : ∇X (δu) · ∇X v dX



+ C(un ) : FT (un ) · ∇X (δu) : FT (un ) · ∇X v dX,


∂R1 ((un , pn ), v)
bn (v, δp) = · δp
∂p

= − J δp F−T (un ) : ∇X v dX


∂R2 ((un , pn ), q)
cn (δp, q) = · δp
∂p

1
= − (δp)q dX.
 k
4.2 Application to Incompressible Elasticity 55

The linearised variational formulation is, knowing (un , pn ), the previous solution,
to find (δu, δp) such that
⎧
⎨ an (δu, v) + bn (v, δp) = −R1 (u , pn ), v , ∀ v ∈ V ,
n
(4.20)
⎩ bn (δu, q) − cn (δp, q) = −R2 (u , pn ), q , ∀ q ∈ Q.
n

Remark 4.5 An important point is that the linearised system (4.20) depends on
some initial value of the displacement. In general, we do not have the equivalent
of Korn’s inequality and the matrix can in fact be singular if one has chosen as
initial guess a bifurcation point. 

4.2.3 Numerical Results for the Linear Elasticity Problem

To illustrate the behaviour of our algorithms, we first consider the simple case of
linear elasticity on a three-dimensional problem. The results will also be applicable
to the Stokes problem.
To obtain an approximation of (4.16), we have to choose a finite element space
Vh ⊂ V and a space Qh ⊂ Q. In our numerical experiments, we consider a three-
dimensional problem. We thus had to make a choice of a suitable finite element
approximation. The catalogue of possibilities has been well studied [20] and we
made a choice which seemed appropriate with respect to the standard engineering
applications. We employ tetrahedral elements, a choice motivated by our eventual
interest in mesh adaptation and we want to respect the inf-sup condition without
having to use elements of too high degree. The popular Taylor-Hood element was
retained:
• A piecewise quadratic approximation for the displacement u and a piecewise
linear approximation of the pressure p.

Remark 4.6 This choice of element is good but there is a restriction for the
construction of element at the boundary: no element should have all its vertices
on the boundary. This might happen on an edge if no special care is taken when the
mesh is built. A bubble can be added at the displacement to avoid this restriction,
but to the price of more degrees of freedom. 

Remark 4.7 (Regularised Lagrangian) We therefore have a continuous approxi-

mation for pressure and this has some consequences on the choice of solvers, in
particular if one would like to use the augmented Lagrangian methods. It is almost
impossible to employ (2.32) as for all reasonable choices of MS , MS−1 is a full
matrix. However, as we shall see later the regularised formulation (2.30) term may
improve convergence.
56 4 Numerical Results: Cases Where Q = Q

In order to use the real augmented Lagrangian (2.32) , one would need a
discontinuous pressure element, which would make MS = R block diagonal. For
three dimensional problems such elements are of high polynomial degree [20] or
induce a loss in the order of convergence. 
We therefore consider the discrete regularised Lagrangian corresponding to
(4.16)
⎧  
⎪
⎪ 2μ ε(u ) : ε(v ) dx + λ̂ (div uh − g) div v h dx
⎪
⎪ h h
⎪
⎪  
⎪
⎨  
+ ph div v h dx = f · v h ds = 0 ∀ v h ∈ Vh , (4.21)
⎪
⎪
⎪
⎪



⎪
⎪
⎪
⎩ div uh qh dx = g qh dx ∀ qh ∈ Qh
 

We shall later discuss in detail the choice of λ̂.

4.2.4 The Mixed-GMP-GCR Method

In the following, we employ Algorithm 3.5 (Mixed-P-GCR) using the general

mixed precondiioner GMP of Algorithm 3.7. The approximation of S is the one
iteration version Algorithm 3.10. We shall call this combination Mixed-GMP-GCR.
To complete the description, we must define a solver in u for which we consider
different possibilities.

Approximate Solver in u

Our mixed solver relies on an approximate solver for the problem in u. We shall
explore various possibilities for this choice, corresponding to different choice for
−1 in (3.8).
A
• The direct solver is denoted LU.
• The conjugate gradient method CG, GMRES or GCR methods can also be used.
• We also consider the HP method of [37] already discussed in Sect. 3.1.4. In this
method, the quadratic approximation P2 is split into a linear P1 part defined on
the vertices and a complement P2 part defined on edges and the matrix A is split
into four submatrices
 
A11 A12
A=
A21 A22
4.2 Application to Incompressible Elasticity 57

From this splitting, we obtain a two-level algorithm in which one solves in

sequence the P1 part and the P2 part. From [89], we know that the P2 part is
well conditioned so that a simple iteration is suitable,
– For the P1 part, we use an Algebraic Multigrid Method (AMG) or a direct LU
factorisation.
– For the P2 part, we use a SSOR iteration.
HP-AMG will denote the use of the HP solver with AMG; HP-LU the use of
the HP solver with LU. We then use these solvers, as a preconditionner in the
following ways :
– as a preconditioner on its own, limited to one iteration, denoted as PREONLY,
– as a preconditioner for a GCR method or a GMRES method (see below).
With n the (limited) number of iterations of the GCR or GMRES, we shall denote
GCR(n)-HP-AMG, GCR(n)-HP-LU, GMRES(n)-HP-AMG or GMRES(n)-HP-
LU the approximate solver for A using one iteration of the HP-AMG or HP-LU
solver as a preconditioner.
These approximate solver of A are used in a general mixed preconditioner (GMP,
here we use Algorithm 3.10) and the global system is solved using a Mixed-GMP-
GCR solver (Algorithm 3.5).
Remark 4.8 For the approximation MS of the Schur’s complement in Algo-
rithm 3.10, we take the matrix associated with the L2 scalar product which we
denote M0 . Computing M0−1 can be solved by a direct or iterative solver. For the
numerical tests, we use the LU factorisation (which can be done once for all). 

Remark 4.9 (Variable Coefficients) Related problems for non Newtonian flows
lead to variable coefficients. In [48] one considers the choice of the approximation
MS (Sect. 2.2.3) to the Schur complement, which also defines a scalar product on
Qh .They show that if the bilinear form

2μ ε(u) : ε(v) dx


is changed into

2 μ(x)ε(u) : ε(v) dx (4.22)


one should take for MS


1
MS = ph qh dx.
μ(x)
58 4 Numerical Results: Cases Where Q = Q

From Sect. 2.3 one should change the regularised form (4.16) into
 
2 μ(x)ε(u) : ε(v) dx + λ̂ μ(x)(div u − g) div v dx
 
 
+ p div v dx = f · v ds = 0 ∀v ∈ V
 

4.2.5 The Test Case

We consider a simple academic example : a cube [0, 1]3 clamped at the bottom (the
plane z = 0) is submitted to a vertical displacement imposed on the top (plane
z = 1) (see Fig. 4.1).
We consider a linear incompressible material with a Young’s modulus
E = 102 MPa and, depending on the numerical experiment, an artificial
Poisson coefficient ν̂ varying between 0.0 and 0.4. Four meshes of sizes
h = 0.5, 0.25, 0.125, 0.0625 (respectively 2 187, 14 739, 107 811 and 823 875
degrees of freedom) will be considered. Exceptionally, for Fig. 4.3, a very coarse
mesh (h = 0.5) will also be used. Although this can be seen as a simple problem, it
must be noted that it is a true three-dimensional case.
Remark 4.10 We shall present some examples illustrating the use of our mixed
solvers. In these experiments, we have imposed a rather strict tolerance of 10−10
on the l 2 norm of the residual in p. 
We have introduced in (4.21) a ‘regularised formulation’ parametrised by λ̂
which we may associate with an artificial Poisson ratio ν̂. We emphasise that this
is not an Augmented Lagrangian: the penalty term is introduced for the continuous
divergence-free condition and not for the discrete one so that, for the discretisation
that we employ, the penalty parameter must be small in order not to perturbate the


u(x, y, 1) := [0, 0, −2]
z
u(x, y, 0) := [0, 0, 0]
y
x

Fig. 4.1 Geometry and boundary conditions

4.2 Application to Incompressible Elasticity 59

100
ν̂ =0 ν̂ =0
100
ν̂ = 0.1 ν̂ = 0.1
ν̂ ν̂
Residual in u

Residual in p
= 0.2 = 0.2
10−4
10−3
ν̂ = 0.4 ν̂ = 0.4

10−6 10−8

10−9
10− 12
0 10 20 30 0 10 20 30

Global iteration number Global iteration number

Fig. 4.2 Linear elasticity problem with GCR(3)-HP-AMG as primal solver: convergence in l 2 -
norm of the primal and dual residuals according to the artificial Poisson ratio ν̂

1.8
10−1
uν̂ − u0 L2

1.0

10−2 ν̂ = 0.1
ν̂ = 0.2
ν̂ = 0.3
10−3 ν̂ = 0.4

h/16 h/8 h/4 h/2 h

h
Fig. 4.3 Linear elasticity problem with GCR(3)-HP-AMG as primal solver: uν̂ − u0 L2 with
respect h according to the value of ν̂

solution. The global system is solved by Mixed-GMP-GCR as described above .

The problem in u by three iterations of GCR with a preconditioning by HP-AMG.
Figure 4.2 shows that increasing ν̂ yields to a better convergence in p but has
an adverse effect on the convergence in u when ν̂ is larger than 0.3. If we take the
solution for ν̂ = 0 (the non-regularised problem) as a reference, we may quantify
the effect of the regularising parameter ν̂.
Figure 4.3 illustrate this effect using the difference in l 2 -norm of u for various
values of ν̂ with respect to the mesh size h (from 0.2 to 0.0125). One sees in Fig. 4.3
that the regularised formulation changes the solution but still yields convergence to
the same solution when h goes to zero. It must be noted that the convergence is not
quadratic as expected for a smooth solution. In fact the solution of our test problem
is not regular [32] and the slope 1.8 corresponds to this lack of regularity.
60 4 Numerical Results: Cases Where Q = Q

Fig. 4.4 Linear elasticity 10−1

problem: convergence in =0
l 2 -norm of the dual residuals  = 0.01
according to the penalty value  = 0.05

Residual in p
 10−3
 = 0.1

10−5

10−7

0 5 10 15

Global iteration number

Remark 4.11 (Iterated Exact Augmented Lagrangian) In Sect. 2.3.4 we presented

a way of using an exact augmented Lagrangian method, instead of a regularised one,
avoiding the matrix B t MS−1 B which cannot be handled as MS−1 is a full matrix with
our choice of elements. This implies solving the problem
    
A Bt δu ru
= . (4.23)
B −MS δp rp

We then consider the same test case and following Sect. 3.17, we solve by the same
algorithm which we used for the regularised problem.
As we had discussed in Sect. 2.3.4, one sees in Fig. 4.4 that increasing 
accelerates the convergence in p. In this test the acceleration does not justify the
extra iteration in δu, δp. we conclude that, at least with the solver employed, this
method would be useful only if the acceleration of the convergence of p is very
important. 

Number of Iterations and Mesh Size

When the solver in u is LU, the algorithm becomes the standard Uzawa method
(Sect. 3.2.2). For the problem that we consider, the number of iterations in p is then
independent of the mesh size as is the condition number of the dual problem (see
[90]). It is interesting that this property holds even with our iterative solver in u as
can be seen in Table 4.4
4.2 Application to Incompressible Elasticity 61

Table 4.4 Linear elasticity problem with GCR(3)-HP-AMG as primal solver: global number of
iterations with ν = 0.4 according to the size of the mesh
Number of degrees of freedom 2187 14,739 107,811 823,875
Number of iterations 21 22 22 21

Table 4.5 Elasticity problem: Algorithm 3.7 method. Performance with optimal (automated)
parameter β
Value of n (CG(n)) 1 2 3 4 5
# iterations 662 188 134 130 117
CPU time (s) 128 53 44 52 61

Comparison of the Preconditioners of Sect. 3.2

In the previous computations, we employed a combination of methods which were

indeed quite efficient. One could wonder if this choice is the best. Obviously we
will not definitively answer to this question, but we can certainly illustrate the merit
of specific mixed methods built from Algorithm 3.5.
We first try to establish whether
• the use of Algorithm 3.7 as a solver in itself can be considered and if the GCR
really accelerates.
• the cost of computing the optimal parameter β in Algorithm 3.10 is worth the
effort of a second resolution in u.
We thus still consider Algorithm 3.7 but we use three variants.
• Algorithm 3.7 as an iterative solver with Algorithm 3.10 for 
S,
• Algorithm 3.7as an iterative solver with Algorithm 3.11 for 
S,
• Algorithm 3.7 using Algorithm 3.10 as a preconditioner for a GCR method.
We want to compare these three variants with respect to numerical behaviour and
performance on a single mesh (here we chose a mesh of size h = 0.0125).
We first present the result for Algorithm 3.7 with Algorithm 3.10 which implies
computing β, to the price of an additional solution in u.
We shall use the conjugate gradient CG(n) method, preconditioned by SSOR for
u. For Algorithm 3.10 we use for MS = M0 and we illustrate its dependency to
the solver in u by using different number of iterations in CG(n). We take the same
tolerance as above.
In this case CG(3) yields the best performance (Table 4.5).
We now want to know if computing β is worth the cost. To do so, we take CG(3)
as the solver in u which was best in the previous test and we replace Algorithm 3.10
by Algorithm 3.11. This corresponds to the classical Arrow-Hurwicz method
as described in [43]. Since β is arbitrary, we considered various values. Based
on Table 4.6 the optimal value for a fixed β is between 8 and 9 and the corresponding
number of iterations is about 400.
62 4 Numerical Results: Cases Where Q = Q

Table 4.6 Elasticity problem: Algorithm 3.7, number of iterations and CPU time according to the
value of β
β 1 5 7 8 9
# iterations Max iter 649 471 414 diverged
CPU time (s) 283 188 127 109 –

Table 4.7 Elasticity problem: performance of GCR’s method preconditioned with Algorithm 3.7

Value of n (CG(n)) 1 2 3 4 5
# iterations 117 91 82 98 81
CPU time (s) 29 30 33 47 51

One sees that the optimal β is slightly higher that 8 and that the computing time is
the double than with a computed β. Moreover, the optimal values has to be guessed.
In the last comparison we use a GCR to accelerate the precedent solver, once
again different number of iteration are taken for the CG(n).
In Table 4.7 the CPU time is increasing with n. However choosing a value of n
between 1 and 3 achieves a good reduction of CPU time with respect to the optimal
value of Table 4.5.

Effect of the Solver in u

The next point that we will address is related to the use of HP solver. We are
interested in the effect of HP-AMG and HP-LU when employed in the solver in
u.
In the next numerical tests we use a Mixed-GMP-GCR with different variants
of Algorithm 3.7 with Algorithm 3.10 using solvers in u based on the HP-AMG or
HP-LU.
We present in Fig. 4.5 the convergence of the residuals on the finer mesh for
different solvers in u based on the HP-AMG.
One sees that the better the solver in u, the better the convergence. Furthermore,
the gain becomes negligible if this solver is good enough while using HP in
PREONLY mode seems to be rather poor. However the picture is quite different
if one considers computing time. This is what we present in Table 4.8.
For all three meshes, we present the computing time for the different versions of
the mixed-GMP-GCR solver. Each row corresponds to the use of a specific version
of the solver in u. For coarse grids, the direct solver is much more efficient than the
iterative ones but this advantage rapidly disappears when the mesh gets finer. The
solvers using HP-AMG are clearly the best for large meshes. The good news is that
there is little difference as soon as a good enough solver is employed.
Remark 4.12 (GCR or GMRES?) In the previous results, we have chosen GCR(n)
for the solution in u. One might wonder why this choice. Although mathematically
4.2 Application to Incompressible Elasticity 63

PREONLY GCR(1) PREONLY GCR(1)

102
10−1

Residual norm of u
Residual norm of p

GCR(2) GCR(3) GCR(2) GCR(3)

GCR(4) GCR(5) GCR(4) GCR(5)

10−5 10−2

10−9 10−6

10−13 10−10
0 20 40 60 80 0 20 40 60 80

Number of global iterations Number of global iterations

Fig. 4.5 Convergence of the residuals in p and u using a PREONLY(HP-AMG) or GCR(n)-HP-

AMG with n = 1, .., 5

Table 4.8 Linear elasticity No. of degress of freedom

problem: CPU time in
14,739 107,811 823,875
seconds according to the
solver used for the primal LU 7.09 70.41 1903.65
problem and the size of the PREONLY HP-AMG 11.30 49.43 253.11
mesh. On top, as a reference, HP-LU 10.21 39.37 325.87
the CPU times using a direct GCR(1) HP-AMG 13.35 58.57 250.38
solver (LU) for u at each
iteration HP-LU 11.39 50.92 350.62
GCR(2) HP-AMG 10.10 46.92 179.11
HP-LU 8.92 41.70 306.84
GCR(3) HP-AMG 9.23 43.41 164.3
HP-LU 7.92 39.93 313.00
GCR(4) HP-AMG 8.44 38.08 159.78
HP-LU 8.02 40.71 320.47
GCR(5) HP-AMG 8.86 37.72 158.69
HP-LU 8.20 42.41 353.74

GCR is equivalent to GMRES, this could indeed be discussed. To evaluate our

choice, we have made a comparison with the GMRES(n). As we can see in Fig. 4.6,
and as predicted, the behaviour is essentially the same, however for such saddle
point system the GCR seems to benefit from orthogonality of the residuals as we
observe a slight difference in the number of iterations. As to computing time, we
still have a slight difference in favor of GCR as we see in Table 4.9 which should be
compared with the last column of Table 4.8. 
All this shows that it is possible to solve correctly and rapidly incompressible
linear elasticity problems, with a method which is totally automatic and does not
rely on arbitrary parameters. We now consider the non linear case.
64 4 Numerical Results: Cases Where Q = Q

Fig. 4.6 Iterations number

for the Mix-GMP-GCR GCR(n)
preconditioned with a GMRES(n)

Iterations number
GCR(n)-HP or
60
GMRES(n)-HP
preconditioner

40

20

1 2 3 4 5

Table 4.9 Elasticity problem with a fine mesh (823 875 dof): CPU time according to the
preconditionner of GMRES(n)-HP of the primal problem
GMRES(n)
1 2 3 4 5
HP-AMG 32,163 22,051 21,581 21,381 19,980
HP-LU 50,314 39,616 48,920 33,563 50,829

4.2.6 Large Deformation Problems

In this section, we consider the solution of non linear elasticity problems for
incompressible materials. We thus consider a Newton method which brings us to
solve a sequence of linearised problems. We take as examples the neo-hookean
and Money-Rivlin models. As we know, for non-linear elasticity, the coercivity of
the linearised system (4.17) is not guaranteed. It might indeed fail near bifurcation
points and this would require techniques [61] which are beyond the scope of our
presentation. We shall focus on two points.
• The algorithms presented for the linear case are directly amenable to the
linearised problems which we now consider.
• The stabilising terms are important.
In our numerical tests, we apply the Mixed-GMP-GCR method to the linearised
problem. The solver in u is GCR(3) prconditioned by HP-AMG.

Neo-Hookean Material

The parameter of the considered material is E = 102 with a Poisson ratio ν = 0

which gives μ = 50. We regularize the problem with an artificial bulk modulus
4.2 Application to Incompressible Elasticity 65

102
displ. (u)
pressure (p)
Residual

10−5

10−12
0 100 200 300 400 500 600 700

Iteration number
102 displ. (u)
pressure (p)
Residual

10−5

10−12
0 20 40 60 80 100 120 140 160

Iteration number

Fig. 4.7 Non linear elasticity problem (Neo-Hookean): convergence in l 2 -norm of the residuals
for Poisson ratio ν̂ = 0 (left) and ν̂ = 0.4 (right)

k̂ corresponding to an artificial Poisson ratio ν̂. We can see in Fig. 4.7, how the
stabilizing term accelerates the convergence of the problem.
In this result, GCR(3)-HP-AMG is used as the solver in u. Each decreasing curve
corresponds to a Newton iteration. Stabilisation does provide a better non linear
behaviour. All the parameters of the algorithm are computed automatically.

Mooney-Rivlin Material

The material is rubber and the associated parameters are c1 = 0.7348 et c2 =

0.08164. With this material, the problem in u might not be coercive and as we solve
at each Newton step a linearised problem the condition number will depend on the
solution. In the formulation of (4.19), we introduced a stabilisation term K̂. This
stabilisation is indeed necessary.
The behaviour of the material for K̂ = 0 is somewhat paradoxical. When we
solve in displacement only, we obtain the result on the left of Fig. 4.8. This kind of
behaviour was also observed in the technical report of ADINA [2].
When we solve the incompressible problem with Mixed-GMP-GCR, we find
a good solution, presented at the right of Fig. 4.8. However, we got a very slow
convergence of the algorithm as we see in Fig. 4.9. For the first Newton’s iteration
all seems well but convergence is slow for the second one.
66 4 Numerical Results: Cases Where Q = Q

Fig. 4.8 Non linear elasticity problem (Mooney-Rivlin): solution of the problem for K̂ = 0. On
the left using a displacement only formulation and on the right using a mixed formulation

Displ. (u)
101
Pression(p)
Residual

10−1

10−3

10−5
0 500 1,000

Iteration number
Fig. 4.9 Non linear elasticity problem in mixed formulation: convergence in l 2 -norm of the
residuals for K̂ = 0 for the two first Newton steps

When we stabilise with K̂ > 0 we have a better convergence (see Fig. 4.10). In
this case an optimal value on K̂ is 5 and is independent of the size of the problem. As
we could expect, taking K̂ too large has an adverse effect on the condition number
of the problem in u and the algorithm slows down.
Finally we would like to confirm that the computed parameter β of Algo-
rithm 3.10 is still correct for this non linear problem. In order to do so, we consider
the same Mooney-Rivlin case as above using K̂ = 10 for stabilisation and tested
with Algorithm 3.11 using some fixed values of β. Figure 4.11 allows us to compare
the convergence for these fixed values of β with the behaviour when using βopt (that
is using Algorithm 3.10).
Obviously the choice of fixed values of β was not totally random, ‘educated
guess’ were involved giving reasonable numerical behaviour. We can see again that
the computed value of β is nearing the optimal value for this parameter. This, once
again, justifies the use of Algorithm 3.10 and it also shows that the preconditioner
is independent of the problem and avoids more or less justified guesses.
4.2 Application to Incompressible Elasticity 67

100 K̂ = 5
Residual in u

K̂ = 7
−5
10 K̂ = 10
K̂ = 20
10−10 K̂ = 50

0 20 40 60 80 100 120 140 160 180 200 220

Iteration number

100
K̂ = 5
Residual in p

K̂ = 7
K̂ = 10
10−5
K̂ = 20
K̂ = 50
10−10

0 20 40 60 80 100 120 140 160 180 200 220

Iteration number
Fig. 4.10 Non linear elasticity problem (Mooney-Rivlin): convergence in l 2 -norm of the residuals
according to different artificial bulk modulus K̂

101
β = 20
Residual in u

β = 25
β = 30
10−5 β = 40
βopt

10−11
0 20 40 60 80 100 120 140

Number of iterations
101
β = 20
Residual in p

β = 25
β = 30
10−6 β = 40
βopt

10−13
0 20 40 60 80 100 120 140

Number of iterations

Fig. 4.11 Non linear elasticity problem (Mooney-Rivlin): convergence in l 2 -norm of the residuals
according to the value of β
68 4 Numerical Results: Cases Where Q = Q

4.3 Navier-Stokes Equations

We now rapidly consider Navier-Stokes equations for incompressible flow prob-

lems. This will evidently be very sketchy as numerical methods for Navier-Stokes
equations is the subject of an immense literature. We consider the subject only as it
is related to the previous Sect. 4.2 but also as it shares some points with Sect. 4.1 for
the importance of the coercivity condition. This will lead us to present the classical
projection method in a new setting. This should be seen as exploratory. Projection
methods have been a major development in the solution methods for the Navier-
Stokes equations, [29, 85]. We refer to [76] for a discussion of its variants and their
relation to the type of boundary conditions imposed on the system.
We now present a simple form of the problem. Let  be a Lipschitzian domain
in Rd , d = 2 or 3. Let ∂ = D ∪ N the boundary of  with N = ∅, n is the
exterior normal vector and t the open set  × (0, T ), where T > 0 is the final
time.
We consider an unsteady flow of an incompressible fluid in which the density
variations are neglected and we restrict ourselves to a Newtonian fluid, that is with
a constitutive equations of the form

σ = −pI + 2με(u),

where ε(u) is defined as in (4.8). We consider homogeneous Dirichlet boundary

conditions on D and a homogeneous Neumann condition on N ,

u(t, x) = 0 on D σ (u(t, x), p(t, x)) = 0 on N .

We thus define,

V = {v ∈ (H 1 ())d | v = 0 on D }, Q = L2 (),


a(u, v) = ε(u) : ε(v) dx,



c(u, v, w) = u · grad v · w dx.


We also denote

(u, v) = u · v dx

4.3 Navier-Stokes Equations 69

and we write the Navier-Stokes equations for the fluid velocity u and pressure p as

(∂u/∂t, v) + c(u, u, v) − 2μa(u, v) − (p, div v) = (f , v) ∀v ∈ V

(4.24)
(div u, q) = 0 ∀q ∈ Q.

Here f represents eventual external volumic forces. System (4.24) is completed

with the following initial data:

u(0, x) = u0 (x) ∈ L2 ()d with div u0 = 0

Here, we choose a uniform time step δt and a backward (also called implicit)
Euler time discretisation. For the spatial discretisation we choose a finite element
approximation (Vh , Qh ) for the velocity and pressure. At time t k = kδt < T ,
knowing (uhk−1 , phk−1 ) we consider the system
⎧
⎪
⎪ (ukh , v h ) + δt c(ukh , ukh , v h ) − 2μδt a(ukh , v h )
⎪
⎨
− δt (phk . div v h ) = (uk−1
h , v h ) + δt (f , v h ) ∀v h ∈ Vh
⎪
⎪
⎪
⎩
(div ukh , qh ) = 0 ∀qh ∈ Qh .
(4.25)

Remark 4.13 We present this simple implicit time discretisation to fix ideas. Our
development is in no way restricted to this example. In (4.25) we have a non linear
problem for ukh for which we can consider a Newton linearisation. One could also
consider a semi implicit formulation with c(uhk−1 , ukh , v h ) instead of c(ukh , ukh , v h ).
 = δt p. We can write (4.25) in the form,
Let us denote p
⎧
⎪
⎪ (ukh , v h ) − (
phk . div v h ) + δt c(ukh , ukh , v h ) − 2μ δt a(ukh , v h )
⎪
⎨
= (uk−1
h , v h ) + δt (f , v h ) ∀v h ∈ Vh
⎪
⎪
⎪
⎩
(div ukh , qh ) = 0 ∀qh ∈ Qh .

We now apply he classical projection method. In its simplest form, we first

compute a predictor 
uh for ukh solution of,

uh , v h ) + δt c(
( uh , v h ) − 2μ δt a(
uh , uh , v h )
(4.26)
= (uk−1 k−1
h , v h ) + (p̃h , div v h ) + δt (f , v h ) ∀v h ∈ Vh
70 4 Numerical Results: Cases Where Q = Q

then project 
uh on the divergence-free subspace by solving formally,
⎧
⎪
⎪ − δ p̃ = div
uh
⎪
⎪
⎨
∂δ p̃
= 0, on D (4.27)
⎪
⎪ ∂n
⎪
⎪
⎩
δ p̃ = 0, on N

and finally updating the velocity and pressure

ukh = 
uh − grad δ p̃
p̃k = p̃k−1 + δ p̃

The basic flaw in this approach is that this projection takes place in H (div, ) and
not in H 1 (). Tangential boundary values are lost. Many possibilities have been
explores to cure this and we shall propose one below.
In a finite element formulation, the exact meaning of  uh − grad δp must also be
precised. We consider a non standard way of doing this.
Referring to our mixed formulation (4.3) and defining δuh = ukh −  uh

V0h = {vh | v h · n = 0 on γD }. (4.28)

The Poisson problem (4.27) can be written as finding δuh ∈ V0h , solution of

(δuh , v h ) + (δ p̃h . div v h ) = 0 ∀v h ∈ V0h

(4.29)
(div δuh , qh ) = (div
uh , q h ) ∀qh ∈ Qh .

We have an incompressibility condition and we have to choose a discrete

formulation satisfying the inf-sup condition. To fix ideas, we may assume the same
choice as we used earlier for incompressible elasticity (cf. Sect. 4.2.3), that is the
Taylor-Hood element which is also classical for flow problems. It is not, however,
suitable for the formulation (4.29). The trouble is with the lack of coercivity on the
kernel. This has been studied in [25] and the cure has already been introduced :
we add in (4.29) a stabilising term to the first equation and we now have to find
δuh ∈ V0h solution of
⎧
⎪
⎪ (δuh , v h ) + α(div δuh − div uh , div v h )
⎨
+ (δ p̃h . div v h ) = 0 ∀v h ∈ V0h (4.30)
⎪
⎪
⎩
(div δuh , qh ) = (divuh , q h ) ∀qh ∈ Qh .

With the discretisation considered, this will not yield a fully consistent aug-
mented Lagrangian and the stabilising parameter will have to be kept small. We are
4.3 Navier-Stokes Equations 71

fortunate: our experiments of Sect. 4.1 show that even a small value will be enough
to get a good convergence.
Remark 4.14 (Why This Strange Mixed Form) Using the mixed form (4.30) makes

uh and grad ph to be in the same finite element space, which is a nice property.
We recall that the normal condition δuh · n = 0 must be imposed explicitly on
D in either (4.29) or (4.30). If no condition is imposed on δuh · n, one then imposes
δ p̃h = 0 in a weak form.

Since we have no control on the tangential part of grad ph on D this method
potentially leaves us with a tangential boundary conditions which is not satisfied.
The simplest remedy would be an iteration on the projection. To fix ideas, we shall
use the semi-implicit problem
⎧
⎪
⎪ (ukh , v h ) + δt c(uk−1h , uh , v h ) − 2μδt a(uh , v h )
k k
⎪
⎨
− (p̃hk . div v h ) = (uk−1
h , v h ) + δt (f , v h ) ∀v h ∈ Vh (4.31)
⎪
⎪
⎪
⎩
(div ukh , qh ) = 0 ∀qh ∈ Qh .

One could then, given an initial p̃hk ,

1. Solve the first equation of (4.31)
2. Project the solution by (4.30).
3. Update p̃hk + δ p̃h .
4. Repeat until convergence.
From this, one can think of many possibilities using approximate solutions and
imbedding the procedure, for example, in a GCR method. In all cases, solving (4.30)
will require a mixed iteration. and this leads us to another approach.

4.3.1 A Direct Iteration: Regularising the Problem

Finally we consider the direct use of the same Algorithms 3.7 and Algorithm 3.10
that we used for incompressible elasticity. To get a better coercivity we had the
regularised penalty terms and we change the first equation of (4.25) into

(ukh , v h )+α(div ukh , div v h ) + δt c(ukh , ukh , v h ) − 2μδt a(ukh , v h )

− δt (phk . div v h ) = (uk−1

h , v h ) + δt (f , v h ) ∀v h ∈ Vh
(4.32)

This is a non linear problem which should be linearised. We can then apply the
Mixed-GMP-GCR method to the resulting linearised form.
72 4 Numerical Results: Cases Where Q = Q

4.3.2 A Toy Problem

In order to show that this technique is feasible, we consider a very simple example.
We consider  =]0, 1[×]0, 1[ and an artificial solution,

u(t, x, y) = (y 2 (1 + t), x 2 (1 + t))

with a viscosity’s value equal to 10−2 , the source force associated is

f (t, x, y) = 0.98(1 + t) + y 2 + 2x 2(1 + t)2 y,


− 1.02(1 + t) + x 2 + 2y 2 (1 + t)2 x

Results for the method (4.32) are presented in Table 4.10, which represents the
global number of iterations and CPU time in seconds to reach T = 1. This is done
for different values of the time step δt and the regularity coefficient α. The table
shows the interest of the stabilisation parameter α, which can reduce the number
of global iterations and CPU time. The optimal value of the stabilisation parameter
is also stable with respect to the time step. We used an iterative solver (precisely
CG(10) ) for solving the problem in u, a choice which could clearly be improved.
The mesh is composed of 3200 triangles and 6561 degrees of freedom.
The following figures show the convergence for the problem with α = 0.1 and
δt = 10−1 . Figure 4.12 presents the convergence of both primal and dual residuals
for Newton iterations of a single arbitrary time step (here the fifth one corresponding
to t = 0.5).

Table 4.10 Navier-Stokes problem: iteration’s number and CPU time in second according to the
step time (δt) and the regularity coefficient (α)
δt = 0.05 δt = 0.1 δt = 0.25
α # it. CPU (s) # it. CPU (s) # it. CPU (s)
0 13,593 945 6701 475 3542 254
1 × 10−2 10,567 739 5167 366 2765 178
1 × 10−1 5709 412 3203 233 2231 161
2.5 × 10−1 5440 396 3756 267 3150 221
1 8315 580 6767 468 6033 441
4.3 Navier-Stokes Equations 73

10−3 primal
dual
Residual

10−6

10−9

10−12
1,250 1,300 1,350 1,400 1,450 1,500

Number of cumulative iterations

Fig. 4.12 Navier-Stokes problem: convergence curves for the primal and dual residuals for t = 0.5

The same behaviour repeats at every time step.

This is a very simple case. Nevertheless, we may think that this direct approach
is feasible. Clearly, a better solver in u is needed and it is thinkable, following [26]
to include the projection method in the preconditioner.
Chapter 5
Contact Problems: A Case
Where Q = Q

This chapter presents solution methods for the sliding contact. We shall first develop
some issues related to functional spaces: indeed, in contact problems, we have a case
where the space of multipliers is not identified to its dual. To address this, we first
consider the case where a Dirichlet boundary condition is imposed by a Lagrange
multiplier and present the classical obstacle problem as a simplified model for the
contact problem. We shall then give a description of the contact problem and its
discretisation with a numerical example.

5.1 Imposing Dirichlet’s Condition Through a Multiplier

We temporarily put aside contact problem to consider a simple Dirichlet problem

which will nevertheless enable us to address some issues relative to contact
problems.
We thus consider a domain  of boundary . We divide  into two parts 0
where null Dirichlet are imposed and C where we have a non homogeneous
Dirichlet condition which will provide a simplified model of the contact region.
We then want to solve
⎧
⎪
⎨−u = f in
⎪
u = 0 on0 (5.1)
⎪
⎪
⎩u = g on C

We define the space V = {v |v ∈ H 1 (), v|0 = 0}. We suppose that f ∈ V  and

1/2 1/2
g ∈ H00 (D ). Following Sect. 2.1.5, we take  = H00 (C ) and the operator B
1/2
is the trace of V in  on C . One should note that we take g ∈ H00 (C ) to make
our presentation simpler but that this is not essential.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 75

J. Deteix et al., Numerical Methods for Mixed Finite Element Problems,
Lecture Notes in Mathematics 2318, https://doi.org/10.1007/978-3-031-12616-1_5
76 5 Contact Problems: A Case Where Q = Q

Remark 5.1 (Change of Notation) In the following, we denote  instead of Q the

space of multipliers. 

1/2
Remark 5.2 (Sobolev Spaces) The space H00 (C ) was introduced in [62]. The
elements of this space are in a weak sense null at the boundary of C and thus match
with the zero boundary conditions on 0 . The dual of H00 (C ) is H −1/2(C ). The
1/2
1/2
scalar product on H00 (C ) is usually defined by an interpolation norm which also
defines a Ritz operator R from H00 (C ) onto H −1/2 (C ). We shall consider later
1/2

other possibilities to define the scalar product. 

1/2
Let then Bv be the trace of v in H00 (C ). To write our problem in mixed form, we
define

a(u, v) = grad u · grad v dx (5.2)


b(v, μ) = (Bv, μ)H 1/2 ( = Bv, RλH 1/2 ( −1/2 (

00 C) 00 C )×H C)

1/2
To simplify the notation we write (λ, μ)1/2 the scalar product in H00 . We want to
find u ∈ V , λ ∈  solution of

a(u, v) + b(v, λ) = (f, v) ∀v ∈ V ,
(5.3)
b(u, μ) = (g, μ)1/2 ∀μ ∈ .

This problem is well posed. Indeed the bilinear form a(u, v) is coercive and we
have an inf-sup condition. To show this, we use the fact that there exists a continuous
1/2
lifting L from H00 (C ) into V . Denoting vλ = Lλ, we have

|vλ |V ≤ C|λ|

b(v, λ) b(vλ , λ) 1
sup ≥ ≥ |λ|λ .
v |v|V |vλ |V C

1/2
Remark 5.3 (This May Seem a Non Standard Formulation!) Taking λ ∈ H00 (C )
may seem a little strange. Indeed, a more standard formulation would define

b(v, μ ) = Bv, μ H 1/2 ( −1/2 (

00 C )×H C)

In fact the two formulations are equivalent as we have

Bu, μ H 1/2 ( −1/2 ( = (Bu, R−1 μ )1/2 .

00 C )×H C)
5.1 Imposing Dirichlet’s Condition Through a Multiplier 77

where R is the Ritz operator on . We therefore have the choice of working with λ
or with λ , and this choice will be dictated by numerical considerations. 
We may then introduce the operator B = RB from V onto  and we have

Bu, λ × = b(v, λ)

Remark 5.4 We have considered a case where the choice of  and  is simple. In
more realistic situations  is a subspace of H 1/2(C ) corresponding to the trace of
the elements of a subspace of H 1 (). The space  is then the dual of  which in
general will contain boundary terms. 
As a simple numerical procedure to solve (5.3) we could use Algorithm 3.7. A
central point of this algorithm would be to compute,

zλ = 
S −1 rλ = 
S −1 (Bu − g).

It will thus be important to have a good approximation of the operator S −1 . To better

understand the issue, we have to analyse more closely the spaces considered here.

5.1.1 H00 (C ) and its dual H −1/2 (C )

1/2

We rapidly present some results that should help to understand the spaces with
which we have to work. In the present case, R is the Ritz operator from  into
 , that is from H00 (D ) onto H −1/2 (D ). This corresponds to the Dirichlet-
1/2

Neumann Steklov-Poncaré operator which we shall consider in more detail below.

1/2
The classical way to define H00 (C ) is through an interpolation between L2
1
and H0 . There has been in the last years many developments to define fractional
order Sobolev spaces in term of fractional order derivatives. We refer to [31] for a
general presentation. We show informally how this could bring some insight about
our problem. If we suppose to simplify that C lies in a plane, the elements of
1/2 1/2
H00 (C ) can be prolonged by zero as elements of H00 (Rn−1 ) and we can then
introduce on C a fractional order gradient of order 1/2 denoted grad 1/2 . It must
be noted that fractional derivatives are not local operators.
We can define for  ⊂ Rn and C ⊂ Rn−1
1/2
H00 (C ) = {v |v ∈ L2 (C ), grad 1/2 v ∈ (L2 (C ))n−1 }

1/2
We can also define on H00 (C ) a norm
 
|v|21/2 = |v|2 dx + |grad 1/2 v|2 dx (5.4)
C C

This norm depends on a fractional tangential derivative.

78 5 Contact Problems: A Case Where Q = Q

Remark 5.5 (Discrete Interpolation Norms) This should be made in relation to

discrete interpolation norms [5], which involve taking the square root of the matrix
associated to a discrete Laplace-Beltrami operator on the boundary. This is an
operation which might become quite costly for large problems. One should also
recall the work of [44]. 
1/2
We can then identify the dual of H00 with L2 (C ) × (L2 (C ))n−1 and the
duality product as
 
λ, v = λ0 vdx + λ1 · grad 1/2 v dx
C C

with λ0 ∈ L2 (C ) and λ1 ∈ (L2 (C ))n−1 . As regular functions are dense in
H00 (C ), we could say that H −1/2(C ) is the sum
1/2

λ0 + div1/2 λ1

where div1/2 is a fractional order divergence operator.

5.1.2 A Steklov-Poincaré operator

Another important component of our problem is the Dirichlet to Neumann Steklov-

Poincaré operator. Essentially, the idea is to associate to a Dirichlet condition the
corresponding Neumann condition. This is a wide subject and we present the simple
case associated with our model problem. We refer to [77] for the discussion of
domain decomposition methods.
1/2 1/2
Given r ∈ H00 (C ), we first use the continuous lifting from H00 (C ) into V
and find ur such that Bur = r on C . We then solve for u0 ∈ H01 ()

a(u0 , v0 ) = −a(ur , v0 ), ∀v0 ∈ H01 ().

and setting φr = u0 + ur , we can then define λ ∈ H −1/2 (C ) by

λ , Bv × = a(φr , v) ∀v ∈ H 1 ()

Remark 5.6 Taking v = u, λ , Bu × is a norm on H00 (C ).

1/2

5.1 Imposing Dirichlet’s Condition Through a Multiplier 79

Using This as a Solver

To solve the Dirichlet problem we can now build a Neumann problem using the SP
operator. Assuming we have some initial guess λ0 we shall first solve the Neumann
problem

a(uN , v) = f, v + v, λ0 × ∀v ∈ V .

We then set r = g − BuN on C , and get λ = λ0 + SPr.

Solving the Dirichlet problem (5.1) is then equivalent to solving the Neumann
problem

a(u, v) + Bv, λ × = f, v ∀v ∈ V . (5.5)

The condition Bu = g on C can be written as (Bu − g, Bv) = 0 which can also

be written as

Bu − g, SP v× = Bu − g, μ × = 0 ∀μ ∈  . (5.6)

Remark 5.7 We have thus written our problem in the form (5.3). One could argue
that we did nothing as the Steklov-Poincaré operator implies the solution of a
Dirichlet’s problem. This will become useful whenever the λ has an importance
by itself. This will be the case in a similar formulation of the contact problem where
λ is the physically important contact pressure. 

5.1.3 Discrete Problems

The direct way of discretising problem (5.3) would be to consider a space Vh ⊂ V

and to take for h a subspace of the space of the traces of Vh on C . We then have
Bh vn = Ph Bvh .We define

b(vh , λh ) = (Bh vh , λh )1/2,h ,

where (λh , μh )1/2,h is some discrete norm in h . We look for uh and λh solution of

a(uh , vh ) + b(vh , λh ) = (f, vh ) ∀vh ∈ Vh ,
(5.7)
b(uh , μh ) = (gh , μh )1/2,h ∀μh ∈ h ,

where we may take gh the interpolate of g in h . More generally gh might be the

projection of g on h in (·, ·)1/2,h norm.
80 5 Contact Problems: A Case Where Q = Q

The Matrix Form and the Discrete Schur Complement

The scalar product defines a discrete Ritz operator Rh from h onto h . We have,

(Bh vh , μh )1/2,h = Rh Bh vh , μh h ×h = Bh vh , μh h ×h

In the framework of Sect. 2.2.2 and Remark 2.5 we can associate to uh and λh their
coordinates u and λ on given bases.
R will the matrix associated to Rh that is

Rλ, μ = (λh , μh )1/2,h

We then define the matrix B

B u, λ = (Bh uh , λh )1/2,h

• We emphasise that these matrices depend on the discrete scalar product.

The Schur complement BA−1 B t operates from h onto λh and can be read as
1. Solve the Neumann problem a(uh , vh ) = (λh , Bh vh )1/2,h .
2. Take the trace Bh uh and compute λh = Rh Bh uh . 
This is usable if the bilinear form b(vh , λh ), that is the discrete scalar product in
1/2
H00 is easily computable.
Remark 5.8 We can also write the discrete problem using the discrete form of (5.5)
and (5.6). We then look for uh ∈ Vh and λh in h solution of

a(uh , vh ) + vh , λh h ×h = (f, vh ) ∀vh ∈ Vh ,
Bh uh − gh , μh h ×h =0 ∀μh ∈ h .

As in the continuous case, we can write the problem using h ∈ h or λh ∈ h .
The two formulations are equivalent for the equality condition u = g but this will
not be the case for inequality condition u ≥ g. 
We shall first rely, on a discrete Steklov-Poncaré operator which will enable us
to define a ‘perfect’ discrete scalar product in h .

5.1.4 A Discrete Steklov-Poincaré Operator

We thus take as h a subspace of the traces of functions of Vh on C and we write

Bh vh the projection of Bvh on h .
5.1 Imposing Dirichlet’s Condition Through a Multiplier 81

We shall also denote the matrix B associated to Bh ., the same notation as the
continuous operator !B as they are used in a different context.
The goal is to associate to an element rh ∈ h an element λh in h . To do so,
we first build φhr a function of Vh such that Bh φhr = rh and we solve

a(φh0 , vh0 ) = −a(φhr , vh0 ) ∀vh0

where vh0 = 0 on C and the bilinear form is as in (5.2). Let φh = φh0 + φhr . We
now define SPh rh = λh , an element of h by

Bh vh , λh h ×h = a(φh , vh ) ∀vh ∈ Vh . (5.8)

This would enable us to solve the problem (5.7) just as we had done for the
continuous case:
• Given λ0h solve the Neumann problem

a(ûh , vh ) = (f, vh ) + vh , λ0h  ∀vh ∈ Vh .

• Take rh = Buh − gh and compute λh = λ0h + SPh rh

• Solve

a(uh , vh ) = (f, vh ) + λh , vh  ∀vh ∈ Vh . (5.9)

What we have done here is to define the discrete scalar

(λh , μh )1/2,h = SPh λh , μh h ×h .

Again, this is somehow tautological: we solve a Dirichlet problem by solving a

Dirichlet problem. The gain is that we also have λh which has in contact problems a
physical significance. If we write u and λ the vectors asociated with uh and λh , we
can introduce a matrix SP and write (5.1.4) as

(λh , μh )1/2,h = SP λ, μ. (5.10)

We shall consider later simplified formulations. It is worth, however to give a

look at the computation of the discrete Steklov-Poincaré operator.

5.1.5 Computational Issues, Approximate Scalar Product

We first note that to compute SPh rh as in (5.8), we need only to compute a(φh , vih )
for all vih associated to a node i on C as in Fig. 5.1.
82 5 Contact Problems: A Case Where Q = Q

Fig. 5.1 Node of the

boundary

It is also interesting to see what this result looks like. To fix ideas, let us consider
a piecewise linear approximation for Vh .
Referring to Fig. 5.1, we obtain at node i

a(φ, vih ) = hx hy [(grady φ (−1/ hy ) + gradx φ gradx vih ]. (5.11)

The first term can be read as


∂φ
hx grady φ ≈ v ds.
C ∂n

But this is not all: the second term depends on the tangential derivative of φr near
the boundary, and could be thought as a half order derivative of the boundary value
r. We have
(ri+1 − 2ri + ri−1 )
hy
hx

It is interesting to compare this to (5.4). This suggests to define a discrete scalar

1/2
product to approximate the H00 (C ) scalar product
 
(λh , μh )h = λh μh ds + h grad λh grad μh ds. (5.12)
C C

Formulas more or less similar to this are often advocated. 

Remark 5.9 (A Matrix Representation of SPh )
In Sect. 5.1.4 we have introduced a matrix representing SPh We could also define
this matrix by computing φih for every uih on C , computing

SPij = a(φih , φj h )

As noted above this reduces to a computation near the boundary.

There is a cost to this: one must solve a Dirichlet problem for every nodal value.
Whether this is worth the effort would depend on each problem and the number of
5.1 Imposing Dirichlet’s Condition Through a Multiplier 83

cases where this matrix form would be used and the cleverness with which we solve
the Dirichlet problems. 

Simplified Forms of the SPh Operator and Preconditioning

We shall ultimately solve our discrete problem (5.7) by Algorithm 3.7 with a suitable
preconditioner. In Algorithm 3.10, our standard preconditioner, we compute an
approximation of the Schur complement B A −1 B t and approximate its inverse by
MS . In this case, the iteration is don in λ.
One can see SPh as a representation of the Schur complement. To employ it in
Algorithm 3.7, one could rely on an approximation SP h
• One could think of building SP h on a subdomain around C and not on the
whole .
• The computation of SP h could also be done using a simpler problem, for
instance using a Laplace operator instead of an elasticity operator.

• The Dirichlet problem defining SP h could be solved only approximately with A.
One should now modify Algorithm 3.10 as we are now iterating in λ and not in
λ.We must also note that we have an approximation SP of the Schur complement
S.

Algorithm 5.1 A preconditioner using SP h

1: Initialization

• ru , rλ given ,
• zu = Ã−1 ru
• r̃u = ru + Bzu
• zλ = SP r̃u

• zzu = A−1 = B t (λ )

• Tu = Bzzu
z , Tu 
• β= λ
Tu , Tu 
2: Final computation

• zλ = βzλ
• zu = zu − βzzu
3: End 

The real efficiency of using a SPh operator is to be tested. We leave this as an

open point.
84 5 Contact Problems: A Case Where Q = Q

5.1.6 The L2 (C ) Formulation

A very common choice is to take as a scalar product


(λh , μh )1/2,h = λh μh ds (5.13)
C

In this formulation, although λh ∈ L2 (C ), it is in fact a representation of λh . We

must expect a weak convergence in H −1/2 (C ).
Formally, this also means that, using (5.13) we shall have an inf-sup condition
(2.19) in L2 with βh = O(h1/2).

The Choice of h

We have considered the case where h is the trace of Vh . We could also take a
subspace of the traces. A simple and important example would be to have piecewise
quadratic elements for Vh and a piecewise linear subspace for h . We refer to [24]
for an analysis of the inf-sup condition and the choice of spaces. Why would we
do this? Essentially because of β in the inf-sup condition: A larger h means a
smaller βh . This is a standard point in the analysis of mixed methods, augmenting
the space of multipliers makes the inf-sup condition harder to satisfy. We then have
two consequences.
• As λh is a representation of λh which converges in H −1/2, a richer h will
produce an oscillatory looking λh .
• A smaller β will mean a slower convergence of the solver.
If we take a reduced h it must be noted that the solution is changed as Bh uh =
gh means

Ph (Buh − gh ) = 0

which is weaker than Buh = gh .

5.1.7 A Toy Model for the Contact Problem

Introducing a multiplier to manage a Dirichlet condition was a first step in exploring

some of the technicalities related to contact problems. Another issue will arise when
considering contact problems. We shall have to introduce positivity constraints. To
illustrate this, we consider an obstacle problem in which the Dirichlet condition

Bu = g on C
5.1 Imposing Dirichlet’s Condition Through a Multiplier 85

is replaced by an inequation Bu ≥ g on C . For the continuous problem, we use

1/2
 = H00 (C ) and

b(v, μ) = (Bv, μ)1/2 = RBv, μ ×λ = Bv, μ ×λ .

We can then consider in , the cone + of almost everywhere positive μ+ and on



+ = {μ+ | μ+ , μ+  ×λ ≥ 0 ∀μ+ ∈ + } (5.14)

We thus solve for u ∈ V , λ+ ∈ + solution of


a(u, v) + b(v, λ+ ) = (f, v) ∀v ∈ V ,
b(u, μ+ − λ+ ) ≥ (g, μ+ − λ+ )1/2 ∀μ+ ∈ + .

The solution satisfies the Kuhn-Tucker conditions.

λ+ ∈ + , (Bu − g, μ+ )1/2 ≥ 0 ∀μ+ ∈ + , (Bu − g, λ+ )1/2 = 0.

The second condition can be read as

R(Bu − g), μ+  ×λ ≥ 0 ∀μ+ (5.15)

We have thus imposed the weak condition

(Bu − Rg) ≥ 0 in  . (5.16)

The Discrete Formulation

To get a discrete version, we have to choose Vh and h and we must give a sense to
λh ≥ 0.
For the choice of Vh , we take a standard approximation of H 1 (). For h we
take the traces of Vh or more generally a subspace. For example, if Vh is made of
quadratic functions, we can take h the piecewise quadratic traces or a piecewise
linear subspace.
To define +h the obvious choice is to ask for nodal values to be positive. As
we can see in Fig. 5.2 this works well for piecewise linear approximation. For
a piecewise quadratic approximation one sees that positivity at the nodes does
not yield positivity everywhere. Piecewise linear approximations are thus more
attractive even if this is not mandatory.
86 5 Contact Problems: A Case Where Q = Q

Fig. 5.2 Positivity of λ

As in the continuous case, the definition of positivity on h also defines positivity

on h

λh ≥ 0 ⇐⇒ λh , μh h ×h ≥ 0 ∀μh ≥ 0

Let (λh , μh )1/2,h be a scalar product on h . We then define b(vh , μh ) =

(Bh vh , μh )1/2,h and we solve,

a(uh , vh ) + b(vh , λh ) = (f, vh ) ∀vh ∈ Vh ,
b(uh , μh − λh ) ≥ (gh , μh − λh )1/2,h ∀μh ∈ h .

If we denote μ+h ≥ 0 the elements of h positive at the nodes and write rh =

Bh uh − gh the discrete Kuhn-Tucker conditions are
⎧ + +
⎨ λh ∈ h
(r , μ+ ) ≥ 0 ∀μ+
⎩ h +h 1/2,h h
(rh , λh )1/2,h = 0

Referring to Sect. 5.1.3, this can be written in terms of nodal values,

Br, μ+  ≥ 0 ∀μ+ Br, λ+  = 0.

This means that r is positive in some average sense.

The Active Set Strategy

In Sect. 5.2 we shall us the active set strategy [3, 54, 55] which we already discussed
in Sect. 3.1.3
This is an iterative procedure, determining the zone where the equality condition
u = g must be imposed. We define the contact status dividing C in two parts.
• Active zone : λh > 0 or λh = 0 and Brh < 0
• Inactive zone : λh = 0 and Brh ≥ 0
5.2 Sliding Contact 87

Remark 5.10 (A Sequence of Unconstrained Problems) Once the contact status is

determined, one solves on the active zone equality constraints Brh = 0 using
algorithms Algorithm 3.7 with Algorithm 3.10. One could also, as in Sect. 5.1.4,
take an approximate Steklov-Poincaré operator on the active zone in Algorithm 5.1.
We shall discuss this in more detail in Remark 5.13.
The active zone is checked during the iteration and if it changes, the iterative process
is reinitialised.
We shall illustrate below this procedure for contact problems.

5.2 Sliding Contact

We consider, for example as in Fig. 5.3, an elastic body in contact with a rigid
surface. We shall restrict ourselves to the case of frictionless contact as the frictional
case would need a much more complex development. We refer to [34] for a more
general presentation. We thus look for a displacement u minimising some elasticity
potential J (v) under suitable boundary conditions. In the case of linear elasticity,
we would have
 
J (v) = μ |ε(v)|2 dx − f · v dx.
 

Here we take v ∈ V ⊂ (H 1 ())3 , the space of displacements. In the following

we denote,
• C is a part of the boundary where contact is possible.
• B is the trace operator of the elements of V in  ⊂ (H 1/2(C ))3 .

ΓD

d(u(x)) > 0 d(u) = 0

Ω

• x ΓC
n(x)
rigid plane

Fig. 5.3 Contact with a horizontal rigid plane. An elastic body  submitted to a displacement u.
Illustration of the oriented distance computed on the potential contact surface C
88 5 Contact Problems: A Case Where Q = Q

• R is the Ritz operator from  onto  .

• B = RB is an operator from V into  .
The basic quantity for the contact problem is the oriented distance, computed by
projecting the elastic body on a target surface. The oriented distance to this target
is defined as negative in case of penetration and this projection implicitly define the
normal vector n. Since u is a displacement of the body , the oriented distance is in
fact a function of u, we will express this dependency by denoting it d(u).
The distance d(u) is a function on the boundary which belongs to a functional
space which we assume to be a subspace  of H 1/2(C ). If the boundary is not
smooth (e.g. Lipschitzian), we have to introduce a space defined on each smooth
part. We place ourselves in the setting of Sect. 2.1.5. In this context Bv would
be d(v) but in general d(v) is non linear so that we will need a linearisation to
apply our algorithms. The first (non linear) condition to be respected is then the non
penetration, that is

d(u) ≥ 0 on C

and we can write the first step of our model as,

inf J (v).
d(v)≥0

Contact Pressure
The next step is to introduce a Lagrange multiplier λ ∈  for the constraint: the
contact pressure [35, 59]. We thus transform our minimisation problem into the
saddle-point problem,

inf sup J (v) + (λn , d(v)) .

v λn ≥0

The optimality conditions are then


A(u) u − f , v + (λn , v · n) = 0 ∀v
(d(u), μn − λn ) ≥ 0 ∀μn ≥ 0.

where the operator A(u) represent the constitutive law of the material. From this
system we deduce the Kuhn-Tucker conditions,
⎧
⎪
⎪ λn ≥ 0,
⎨
(d(u), μ) ≥ 0 ∀μ ≥ 0,
⎪
⎪
⎩
(d(u), λn ) = 0.
5.2 Sliding Contact 89

Newton’s Method, Sequential Quadratic Programming

It must be noted that even in the case of linear elasticity, the problem is non linear.
To solve this problem, we apply the Newton method or in this context the sequential
quadratic programming (SQP) method [58]. Let u0 be some initial value of u and
gn0 = d(u0 ) the corresponding initial gap. We recall [33] that the derivative of the
distance function is given by

d  (u0 ) · δu = δu · n

The linearised problem is then,

⎧
⎨ A (u0 )δu, v + (λn , v · n) = f − A(u0 ), v ∀v ∈ V
(5.17)
⎩ (φ − λ , g 0 − δu · n) ≥ 0 ∀φn ≥ 0
n n n 

Here δu is the correction of the initial value u0 , and we have linearised both the
constitutive law of the material represented by the operator A and the distance
function. The Kuhn Tucker conditions then become
⎧
⎪ λ ≥ 0,
⎪ n
⎨
(gn0 − δu · n, μn ) ≥ 0 ∀μn ≥ 0,
⎪
⎪
⎩ 0
(gn − δu · n, λn ) = 0.

Remark 5.11 (The Choice of Multipliers) In the above formulation, we have

worked with the bilinear form

b(v, λn ) = (v · n, λn )

that is with the scalar product in  ⊂ H 1/2(C ). As we have seen in Remark 5.3
we can also write (5.17) with

λ , v · n = Rλ, v · n

We recall that this is the same formulation written in two different ways. 

5.2.1 The Discrete Contact Problem

In the previous Sect. 5.2 we developed a formulation of sliding contact using a

Lagrange multiplier. We have presented a simpler problem in Sect. 5.1. We now
come to the contact problem. We suppose that we have chosen a space Vh to
90 5 Contact Problems: A Case Where Q = Q

discretise displacements and a space h ⊂ . The choice of these spaces must

of course be done to satisfy some requirements.
• The material will often be incompressible : we shall consider this case in
Chap. 6. This imposes some restrictions on the choice of Vh and of the associated
pressure space Qh . We refer to [20] for details. In practice, for three-dimensional
problems, we use as in Sect. 4.2 the Taylor-Hood approximation, that is piecewise
quadratic elements for Vh and piecewise linear for Qh , but this is one of many
possibilities.
• The problem (5.17) is a (constrained) saddle-point problem. This means that the
choice of h requires an inf-sup condition to be satisfied, [13]. In the following,
we use piecewise linear elements for h .

Remark 5.12 (Scalar Product on h ) We also have to define a scalar product on

h . Ideally this should be the scalar product in . As we have discussed earlier the
scalar product in H 1/2 is not easily computed and we shall have to employ some
approximation. In general we consider as in Section 5.1.5 a discrete H 1/2 norm
defined by a discrete scalar product (λh , μh )1/2,h .
We thus have an operator Rh from h onto h

Rh λh , μh h ×h = (λh , μh )1/2.h .

and the associated matrix R. In our computations we shall rely on Algorithm 3.10.
An important issue is the choice of MS and its inverse MS−1 . the normal choice is
here MS = R. From Proposition 2.2, to obtain a convergence independent of the
mesh size, R should define on h a scalar product coherent with the scalar product
of H 1/2 in order to have an inf-sup condition independent of h.
We thus have to make a compromise: a better R yields a better convergence but
may be harder to compute.
In the results presented below, we use the simple approximation by the L2 scalar
product in h and the matrix R becomes M0 defined by,

M0 λ, μ = λh μh ds. (5.18)
C

We could also employ a diagonal approximation MD to M0 . This can be obtained by

replacing (5.18) by a Newton-Cotes quadrature using only the nodes of h . In the
case of piecewise linear elements, this is equivalent to using a ‘lumped’(diagonal)
matrix obtained by summing the rows of M. 
The geometrical nodes of the mesh which support degrees of freedom for λh
and u are not identical. To emphasise this difference we denote by x i the nodes
supporting the values of λh ∈ h and y j the nodes for uh . As in Remark 2.5, we
now denote u and λ the vectors of nodal values, whenever no ambiguity arises. We
5.2 Sliding Contact 91

thus have

u = {uh (y j ), 1 ≤ j ≤ NV }

λ = {λh (x i ), 1 ≤ i ≤ N }

Denoting by ·, · the scalar product in RNV or RN , we thus have the matrices
associated with operators Rh and Bh

Rλ, μ = (λh , μh )1/2,h , R ∈ RNA ×NA

Bu, μ = (Bh uh , μh )1/2,h , B ∈ RNA ×NV

and then R −1 B is the matrix associated with the projection of Buh on h .

• These matrices depend on the choice of the discrete scalar product.
We also suppose that we have a normal ni defined at x i and we define for λh its
normal component such that at all nodes,

λnh (x i ) = λh (x i ) · ni

which is a scalar. We also use its vectorial version

λnh (x i ) = λn (x i )ni (5.19)

We denote by nh the subset of h of normal vectors of the form (5.19) and

+
nh = {λnh ∈ nh | λn (x i ) ≥ 0}

We want to have as in Sect. 5.1.7 a discrete form of the condition u · n ≥ 0. Recall

that Bu is a vector corresponding to an element of h . We consider at node x i its
normal component

Bn u = {(Bu)i · ni }

We then have the condition

Bn u, μ+
n
 ≥ 0 ∀μ+
n
.

which is a weak form of u · n ≥ 0 .

We can now introduce our discrete contact problem. We first consider the
unconstrained case, Bn u = g where we do not have to consider a positivity
condition. Indeed our final algorithm will solve a sequence of such unconstrained
92 5 Contact Problems: A Case Where Q = Q

problems, we thus look for the solution of

1
inf sup Av, v + Rλn , (R −1 Bv − g) − f , v (5.20)
v λn 2

that is also
1
inf sup Av, v + λn , (Bn v − Rg) − f , v. (5.21)
v λn 2

This problem can clearly be solved by the algorithms of Chap. 3. We must however
introduce a way to handle the inequality constraint. To do this we first need the
notion of contact status.

Contact Status

Let rn = Bn u − g the normal residual. A basic tool in the algorithms which follow,
will be the contact status. It will be represented point wise by the operator P (λ, rn )
defined by,
⎧
⎪
⎪ if λn = 0,
⎪
⎪
⎪
⎪
⎪
⎪ (1) if rn ≤ 0 then P (λ, rn ) = 0
⎨
(2) if rn > 0 then P (λ, rn ) = rn
⎪
⎪
⎪
⎪
⎪
⎪ if λn > 0,
⎪
⎪
⎩
(3) P (λ, rn ) = rn

The Kuhn-Tucker optimality conditions can then written as P (λ, rn ) = 0. We say

that in case (1), the constraints are inactive and that on (2) and (3), they are active.We
denote CA subset of C where the constraints are active.

5.2.2 The Algorithm for Sliding Contact

We can now present our solution strategy for the sliding contact problem.

A Newton Method

We have a nonlinear problem. As we have seen in Sect. 5.2, we rely on a Newton

method in which the distance is linearised.
5.2 Sliding Contact 93

The Active Set Strategy

The algorithm proposed here is based on the active set strategy

• Given (u0 , λ0n ), compute P (r 0n , λ0n ), that is the contact status.
• Solve (on the active set) an equality constraint problem.

A (u0 )δu, v + λn , v = f , v ∀v
(5.22)
δu · n − gn0 , φn  = 0 ∀φn

• Check the contact status after a few iterations.

– If the status has changed: project λ on the admissible set h +; recompute the
active and inactive sets; restart.
– If the contact status does not change, we continue the iteration.
• Iterate till those sets are stabilised.
The contact status is thus checked during the iterations to avoid continuing to solve
a problem which is no longer relevant.
Remark 5.13 (Ms and the Active Zone) The active set strategy solves a sequence
of problems with equality constraints on a subset CA of C where the constraints
are actually imposed. This will mean in Algorithm 3.10 to restrict MS to CA .
Let us first consider the case where we would take MS = I . One would then
have a correction zλ on λn given by zλ = rn . If we do not want to modify λn on
the inactive zone we should make rn = 0 on inactive nodes. This is analogue to
the classical conjugate projected gradient method [69]. The same is true if one has
MS = MD where MD is the diagonal matrix introduced in Remark 5.12.
In general we can proceed by,
• Make rn = 0 outside the active zone CA . Solve MS z = rn imposing z = 0
outside of CA .
If one uses the diagonal matrix MD , this is automatically taken into account to
the price of a slower convergence.
A similar procedure should be used if one introduces a discrete Steklov-Poincaré
operator. 

5.2.3 A Numerical Example of Contact Problem

To illustrate the behaviour of the algorithm we present the case of an elastic cube
being forced on a rigid sphere (see Fig. 5.4). In this case a displacement is imposed
on the upper surface of the cube. We want to illustrate the behaviour of the algorithm
as the mesh gets refined (see Table 5.1).
94 5 Contact Problems: A Case Where Q = Q

Fig. 5.4 Contact pressure

over the potential contact area

Table 5.1 Iteration’s number and CPU time according to the total number of degrees of freedom
(dof) of u and the solving method for the primal system
14,739 dof 107,811 dof 823,875 dof
# it. CPU (s) # it. CPU (s ) # it. CPU (s)
LU 32 6.85 39 208.46 41 13099
GCR(10,HP-LU) 40 8.96 42 59.06 53 477.36

We consider a linear elastic material with a Poisson coefficient ν = 0.3 and

an elasticity modulus E = 102. The primal problem will be solved by either a
direct or an iterative method and we present at Table 5.1 the two usual performance
indicators i.s.e the total number of iterations and the computational time in seconds.
Here we take R = M0 (Remark 5.12) for which we do not expect convergence to
be independent of h. There is room for improvement but even then the results show
that the method is suitable for large problems.
In Table 5.1 the iterations numbers show that even when using the L2 scalar
product for b(v h , λh ), convergence of the dual problem is only slightly dependent of
the discretization. The large differences in computation time illustrate the important
gain in efficiency of the iterative solver for large problems.
It is also interesting to consider convergence curves of the normal residual and
the importance of the solver in u. In Fig. 5.5 one compares the convergence for three
mesh sizes. Red dots indicate a change in contact status in the active set strategy.
The changes occur at the beginning of the process and the iteration is then restarted
on the new active region.. Most of the computational efforts consist in solving a
5.2 Sliding Contact 95

101
LU
GCR-HP-LU

Normal residual
Contact status
Newton
10−4

10−9

10−14
0 10 20 30 40 50

Number of cumulative iterations

100
LU 100 LU
GCR-HP-LU GCR-HP-LU
Normal residual

Normal residual
10−2 Contact status Contact status
Newton 10−2 Newton

10−4

10−4
10−6

10−8
10−6
0 10 20 30 40 50 0 10 20 30 40 50

Number of cumulative iterations Number of cumulative iterations

Fig. 5.5 Convergence for rn . On top a coarse mesh (14,739 dof for u). Bottom left a mesh with
107,811 dof, on the right a mesh with 823,875 dof

single equality constraint problem at each Newton step. If we ignore computational

time obviously the direct solver in u perform better since it produces less iterations
compared to the iterative solver. Finally, for smaller mesh (top of Fig. 5.5) the direct
solver seems too good: time is lost with excessive tolerance at each Newton step.
Chapter 6
Solving Problems with More Than One
Constraint

In this chapter, we shall consider problems of the form

⎛ ⎞⎛ ⎞ ⎛ ⎞
A Bt Ct u f
⎝B 0 0 ⎠ ⎝p⎠ = ⎝gp ⎠ . (6.1)
C 0 0 λ gλ
 
A

In theory, this is not fundamentally different from the problem (2.22). From the
numerical point of view, things are not so simple as the presence of two independent
constraints brings new difficulties in the building of algorithms and preconditioners.
We present some ideas which indeed yield more questions than answers.

6.1 A Model Problem

To fix ideas, we consider a case where two simultaneous constraints are to be

satisfied. We have already considered in Sect. 4.2 problems of incompressible
elasticity and in Sect. 5 problems of sliding contact. We now want to solve problems
where both these features are involved.
Let  be a domain of R3 ,  its boundary and C a part of  where sliding contact
conditions are imposed by Lagrange multipliers. We suppose that suitable Dirichlet
conditions are imposed on a part of the boundary to make the problem well posed.
We shall use the same procedure as in Sect. 5.2, the only difference will be
that we now denote C the operator which was then denoted B, this operator now
being associated with the discrete divergence: we now want the material to be
incompressible. Proceeding again to linearise the problem, and using the active set

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 97

J. Deteix et al., Numerical Methods for Mixed Finite Element Problems,
Lecture Notes in Mathematics 2318, https://doi.org/10.1007/978-3-031-12616-1_6
98 6 Solving Problems with More Than One Constraint

technique, we are led to solve a sequence of saddle-point problem of the form

1
inf sup Av, v + λn , (Cv − rλ ) + p, (Bv − rp ) − r u , v.
v p,λn 2

This is indeed a problem similar to (6.1). We shall first consider a naive solution
technique, the interlaced method and then reconsider the use of approximate
factorisations.

6.2 Interlaced Method

It would be natural to rewrite the system (6.1) as a problem with a single constraint.
To do so we introduce block matrices

A Ct u r
= u
C 0 λ rλ

where the blocks are defined by

A Bt u r rλ
A= , C= C0 , u= , ru = u and rλ = .
B 0 p rp 0

It is then possible to handle this problem as a contact problem where each

solution of the elasticity problem is in fact the solution of an incompressible
elasticity problem which can be solved (with more or less precision) by the methods
presented in Sect. 4.2. We call this the interlaced method.
Remark 6.1 The choice of the ‘inner’ problem, here the incompressible elasticity
problem is arbitrary. We could also have chosen the contact problem. Our choice is
directed by the fact that the incompressible elasticity problem is linear. 
Starting from this rewriting of the system, we can use directly the technique of
Sect. 3.2.1, in particular the preconditioner of Algorithm 3.10 in which we replace
the approximate solver for A by an approximate solver for A. This is simple but
carries some difficulties.
• The matrix A is not definite positive. This essentially renders invalid the classical
proofs of convergence, for example of Uzawa’s method.
• We have to solve an internal problem which will make the method expensive.
The downside of this approach is thus the significant cost of the iterations in λ. In
fact, at each iteration, two solutions to the problem in (u, p) are asked. These latter
being made iteratively they also require at least two displacement only solutions.
6.2 Interlaced Method 99

Remark 6.2 (Projected Gradient) If we refer ourselves to Sect. 3.1.3 one sees that
what we are doing is iterating in the divergence-free subspace provided the problems
in (u, p) are solved precisely. 
To illustrate the behaviour of the algorithm, we first consider the case of an accurate
solution in (u, p). This will be done as in Sect. 4.2.4. For the solve in u, we take
either a direct LU solver or some iterations of GCR preconditioned by HP-lU. This
is clearly an expensive procedure only used for comprehension.
We present the convergence in λ for the intermediate mesh of Sect. 5.2.3 with
107811 degrees of freedom.
As can be expected the results are comparable for the two solver in u as we use
essentially the same information to update λ. Indeed if we have an accurate solution
in (u, p) it should not be dependent on the way it is computed (Fig. 6.1).
When an incomplete solution in (u, p) is considered by limiting the number of
iterations permitted in the Mixed-GMP-GCR method, the iterative solution in λn is
loosing effectiveness (Fig. 6.2).

Fig. 6.1 Interlaced method

using complete solver for A: LU
convergence for the normal GCR
Normal residual

10−2
residual according to the Contact status
primal solver with the Newton
intermediate mesh
10−5

10−8

0 20 40 60

Number of cumulative iterations

Fig. 6.2 Interlaced method
using incomplete solver for LU
A: convergence according to GCR
10−2
Normal residual

the primal solving method Contact status

using intermediate mesh Newton
10−5

10−8

10−11
0 20 40 60 80 100

Number of cumulative iterations

100 6 Solving Problems with More Than One Constraint

Table 6.1 Interlaced method 107 811 dof

using incomplete solver for
CP U (s )
#it.
A: number of iterations and
CPU time in seconds LU 93 464.64
according to the solver in u GCR(10,HP) 108 4928.12

Although it works, it is clearly an expensive method (Table 6.1) and we did not
push the test to finer meshes. We now consider another approach which looked more
promising.

6.3 Preconditioners Based on Factorisation

In Sect. 3.2.1 we presented a general preconditioner based on the factorisation of the

matrix associated to our problem. This can be done in the present case but will also
imply to develop variants. It is easily seen that the matrix of equation (6.1) can be
factorised in the form,
⎡ ⎤⎡ ⎤⎡ ⎤
I 00 A 0 0 I A−1 B t A−1 C t
A = ⎣ BA−1 I 0 ⎦ ⎣ 0 −SBB −SCB ⎦ ⎣ 0 I 0 ⎦
−1
CA 0 I 0 −SCB −SCC 0 0 I

where

SBB = BA−1 B t , SBC = BA−1 C t , SCB = CA−1 B t , SCC = BA−1 C t .

We shall denote

SBB SBC
S=
SCB SCC

We now see what the Algorithm 3.7 would now become,

Algorithm 6.1 General mixed preconditioner for two constraints Assuming that
we have A −1 and S−1 approximate solvers for A and S, from the residual ru , rp
and rλ , we compute the vectors zu , zp and zλ .
• Initialise

−1 ru
zu∗ = A
rp = Bzu∗ − rp ,
rλ − Czu∗ − rλ . (6.1)
6.3 Preconditioners Based on Factorisation 101

• Using S−1 , solve approximately for zp and zλ the coupled problem,

SBB SBC zp rp
= (6.2)
SCB SCC zλ rλ

• Compute

−1 B t zp − A
zu = zu∗ − A −1 C t zλ (6.3)

The key is thus to solve (6.2). There is clearly no general way of doing this.
Starting from the point of view that we have approximate solvers for SBB and SCC ,
we may use a Gauss-Seidel iteration for zp and zλ .
In the simplest case we do one iteration for each problem.

6.3.1 The Sequential Method

In this case we use

• One iteration in p and one in λ
• The result is used in a GCR global method.
In the following Table 6.2, we present the number of iterations and the solution
time, for different ways of solver for the matrix A of the primal problem.
For small meshes, LU factorisation is clearly the best way and this also
the true in all cases with respect to iteration number. A GCR method without
preconditioning rapidly deteriorates. The multigrid HP-AMG method becomes the
best for computing time for the fine mesh.
Looking to convergence curves, we see that even this last method could be
improved as the convergence seems to slow down when the residual becomes
small. For a still finer mesh one could need a more accurate solver for A to avoid

Table 6.2 Number of iterations and CPU time (s) according to the size of the mesh and the solving
method for the primal system
14,739 dof 107,811 dof 823,875 dof
#it. CPU (s) #it . CPU (s ) # it. CPU (s)
LU 55 8.36 88 120.94 124 3432.87
GCR(10) 129 92.97 141 724.54 209 7940.04
CG(10,GAMG) 129 62.98 99 332.34 162 4459.59
GCR(5,HP-AMG) 174 36.66 159 191.78 159 1533.71
102 6 Solving Problems with More Than One Constraint

101
LU LU
HP 10−3 HP
Normal residual

Contact status

Residual in p
10−3 Newton
10−6

10−7
10−9

10−11 10−12
0 50 100 150 0 50 100 150

Number of cumulative iterations Number of cumulative iterations

Fig. 6.3 Convergence curves for the normal residual (left) and hydrostatic pressure p (right)
according to the primal problem solving method when the middle mesh is considered

101
LU LU
GCR 10−3 HP
Normal residual

Contact status
Residual in p

10−3 Newton 10−6

10−7 10−9

10−12
10−11
0 50 100 150 0 50 100 150

Number of cumulative iterations Number of cumulative iterations

Fig. 6.4 Convergence curves for the normal residual (left) and hydrostatic pressure p (right)
according to the primal problem solving method when the finest mesh is considered

stagnation, keeping in mind that LU becomes more and more inefficient (Figs. 6.3
and 6.4).

6.4 An Alternating Procedure

Finally, to end this presentation, we present some results for an ‘Alternating

Direction’ procedure. The idea is simple: solve alternatingly the contact problem
for (u, λn ) fixing the hydrostatic pressure p and then solve for (u, p) fixing the
contact pressure λn . Another issue is that for the contact part, a restart procedure
has to be considered when a change occurs in contact status (Fig. 6.5).
6.4 An Alternating Procedure 103

Fig. 6.5 Illustration of the

alternating procedure

Table 6.3 Alternating method: iteration’s number/CPU(s) according to the size of the mesh and
the solving method for the primal system
14,739 dof 107,811 dof 823,875 dof
# it. CPU (s) # it. CPU (s ) # it. CPU (s)
LU 66 7 110 93 120 2585
GCR(10) 110 35 120 294 150 2966
GCR(5,HP-LU) 120 17 150 139 160 1168

Remark 6.3 (A Projection Method) This could be seen as a Projection method.

After a few iterations in , the solution is projected on the divergence-free
subspace. 
As we did in the previous section with the sequential method, some computations
are presented with this alternating method for the three different meshes. In the
Table 6.3, we present the number of global iteration on λ and the computing time
for different solvers for the primal problem. The conclusion is the same as for all
previous examples: when the mesh size gets large, the iterative method (in particular
when using the HP-LU method of Sect. 3.1.4) becomes better in computing time.
For the convergence, we can see in Figs. 6.6 and 6.7 respectively the middle and
finest mesh, the curves for the hydrostatic and contact pressure according to the
solver of primal problem. We see a good convergence of thoses curves particulary
the contact one.
104 6 Solving Problems with More Than One Constraint

LU LU
10−3
GCR GCR
Normal residual

GCR-HP GCR-HP

Residual in p
10−2

10−7

10−6
10−11

10−10 10−15
0 50 100 150 0 20 40 60 80

Number of cumulative iterations Number of cumulative iterations

Fig. 6.6 Alternating method: convergence curves for the normal residual (left) and hydrostatic
pressure (p) according to the primal problem solving method when the middle mesh is considered

LU,GCR,GCR-HP xticklabel style LU

10−3
100
GCR
Normal residual

GCR-HP
Residual p

10−3 10−7

10−6 10−11

10−9
10−15

0 50 100 150 0 20 40 60 80 100

Number of cumulative iterations Number of cumulative iterations

Fig. 6.7 Alternating method: convergence curves for the normal residual (left) and p the
hydrostatic pressure (right) according to the primal problem solving method when the finest mesh
is considered

This is a crude first testing and things could clearly be ameliorated. In particular,
one could think of marrying the alternating idea into the sequential technique of the
previous Sect. 6.3.1. There is room for new ideas...
Chapter 7
Conclusion

We hope to have shown that solving mixed problems can be accomplished effi-
ciently. This work is clearly not exhaustive and we have indeed tried to open the
way for future research. We have relied as building bricks on rather classical iterative
methods. However, we think that we have assembled these bricks in some new ways.
We have also insisted in developing methods as free as possible of user depending
parameters.
We have also considered Augmented Lagrangians, either in an exact or a
regularised version. This was done in the mind that direct methods should be
avoided for large scale computations and that penalty terms destroy the condition
number of the penalised system.
• For mixed formulations based on elements satisfying the equilibrium conditions,
we have shown that Augmented Lagrangian is efficient. The problem presented
was very simple but we think that the results could be extended to more realistic
situations. In the case of mixed elasticity in which a symmetry condition has to
be imposed [19] one would have to deal with two constraints as in Chap. 6. The
situation would be better than in the example presented there as the equilibrium
constraint is amenable to a real augmented Lagrangian.
• Problems involving incompressible elasticity are of central importance in many
applications. Unfortunately, they are often solved with poor methods using
penalty and low order elements. We have shown that continuous pressure
elements, which are essential for accurate three-dimensional computations at
reasonable cost, are manageable and can be accelerated by a stabilisation term.
• For contact problems, we have considered some possible alternative avenues
to the standard approximations where the constraint is treated in L2 instead
of the correct H 1/2. This is still an open area. We have shown that using the
more classical formulation, one can obtain results for large meshes with good
efficiency.
• For problems involving two constraints, we have explored some possibilities and
many variants are possible.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 105
J. Deteix et al., Numerical Methods for Mixed Finite Element Problems,
Lecture Notes in Mathematics 2318, https://doi.org/10.1007/978-3-031-12616-1_7
106 7 Conclusion

The methods that we discussed can in most cases be employed for parallel
computing. We did not adventure ourselves in this direction which would need a
research work by itself. We must nevertheless emphasise that the Petsc package
which we used is intrinsically built for parallel computing.
Bibliography

1. B. Achchab, J.F. Maître, Estimate of the constant in two strengthened C.B.S. inequalities
for F.E.M. systems of 2D slasticity: application to multilevel methods and a posteriori error
estimators. Numer. Linear Algebra Appl. 3(2), 147–159 (1996)
2. ADINA Inc., Instability of two-term mooney-rivlin model. https://www.adina.com/newsgH48.
shtml
3. P. Alart, A. Curnier, A mixed formulation for frictional contact problems prone to Newton like
solution methods. Comput. Methods Appl. Mech. Eng. 92(3), 353–375 (1991)
4. G. Allaire, S.M. Kaber, Numerical Linear Algebra. Texts in Applied Mathematics, vol. 55
(Springer, New York, 2008)
5. M. Arioli, D. Kourounis, D. Loghin, Discrete fractional Sobolev norms for domain decompo-
sition preconditioning. IMA J. Numer. Anal. 33(1), 318–342 (2013)
6. D.N. Arnold, F. Brezzi, Mixed and nonconforming finite element methods: implementation,
postprocessing and error estimates. ESAIM: Math. Model. Numer. Anal. 19(1), 7–32 (1985)
7. W.E. Arnoldi, The principle of minimized iterations in the solution of the matrix eigenvalue
problem. Quart. Appl. Math. 9(1):17–29 (1951)
8. K.J. Arrow, L. Hurwicz, H. Uzawa, Studies in Linear and Non-linear Programming. Stanford
Mathematical Studies in the Social Sciences, vol. 2 (Stanford University Press, Stanford, 1958)
9. O. Axelsson, G. Lindskog, On the rate of convergence of the preconditioned conjugate gradient
method. Numer. Math. 48(5), 499–523 (1986)
10. Z.Z. Bai, B.N. Parlett, Z.Q. Wang, On generalized successive overrelaxation methods for
augmented linear systems. Numer. Math. 102(1), 1–38 (2005)
11. S. Balay, S. Abhyankar, M. Adams, J. Brown, P. Brune, K. Buschelman, L. Dalcin, A. Dener,
V. Eijkhout, W. Gropp, D. Karpeyev, D. Kaushik, M. Knepley, D. May, L. Curfman McInnes,
R. Mills, T. Munson, K. Rupp, P. Sanan, B. Smith, S. Zampini, H. Zhang, H. Zhang, PETSc
Users Manual: Revision 3.10. (Argonne National Lab. (ANL), Argonne, 2018)
12. J. Baranger, J.F. Maitre, F. Oudin, Connection between finite volume and mixed finite element
methods. ESAIM Math. Model. Numer. Anal. 30(4), 445–465 (1996)
13. K.J. Bathe, F. Brezzi, Stability of finite element mixed interpolations for contact problems.
Atti della Accademia Nazionale dei Lincei. Classe di Scienze Fisiche, Matematiche e Naturali.
Rendiconti Lincei. Matematica e Applicazioni 12(3), 167–183 (2001)
14. L. Beilina, E. Karchevskii, M. Karchevskii, Solving systems of linear equations, in Numerical
Linear Algebra: Theory and Applications (Springer, Cham, 2017), pp. 249–289
15. M. Benzi, G.H. Golub, J. Liesen, Numerical solution of saddle point problems. Acta Numer.
14, 1–137 (2005)

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 107
J. Deteix et al., Numerical Methods for Mixed Finite Element Problems,
Lecture Notes in Mathematics 2318, https://doi.org/10.1007/978-3-031-12616-1
108 Bibliography

16. M. Bergounioux, K. Kunisch, Primal-dual strategy for state-constrained optimal control

problems. Comput. Optim. Appl. 22(2), 193–224 (2002)
17. M. Bergounioux, M. Haddou, M. Hintermüller, K. Kunisch, A comparison of a Moreau–Yosida
based active set strategy and interior point methods for constrained optimal control problems.
SIAM J. Optim. 11(2), 495–521 (2000)
18. D.P. Bertsekas, Nonlinear Programming, 2nd edn. (Athena Scientific, Belmont, 1999)
19. D. Boffi, F. Brezzi, M. Fortin, Reduced symmetry elements in linear elasticity. Commun. Pure
Appl. Anal. 8(1), 95–121 (2009)
20. D. Boffi, F. Brezzi, M. Fortin, Mixed Finite Element Methods and Applications. Springer Series
in Computational Mathematics, vol. 44 (Springer, Berlin 2013)
21. J. Bonet, R.D. Wood, Nonlinear Continuum Mechanics for Finite Element Analysis, 2nd edn.
(Cambridge University Press, Cambridge, 2008)
22. D. Braess, Finite elements: theory, fast solvers and applications in solid mechanics. Meas. Sci.
Technol. 13(9), 365 (2007)
23. J.H. Bramble, J.E. Pasciak, A.T. Vassilev, Analysis of the inexact Uzawa algorithm for saddle
point problems. SIAM J. Numer. Anal. 34(3), 1072–1092 (1997)
24. F. Brezzi, K.J. Bathe, A discourse on the stability conditions for mixed finite element
formulations. Comput. Methods Appl. Mech. Eng. 82(1), 27–57 (1990)
25. F. Brezzi, M. Fortin, L.D. Marini, Mixed finite element methods with continuous stresses.
Math. Models Methods Appl. Sci. 03(02), 275–287 (1993)
26. J. Cahouet, J.P. Chabard, Some fast 3D finite element solvers for the generalized Stokes
problem. Int. J. Numer. Methods Fluids 8, 869–895 (1988)
27. M. Chanaud, Conception d’un solveur haute performance de systèmes linéaires creux couplant
des méthodes multigrilles et directes pour la résolution des équations de Maxwell 3D en régime
harmonique discrétisées par éléments finis. Ph.D. Thesis, Université de Bordeaux 1, France
(2011). https://www.theses.fr/2011BOR14324/abes
28. S.H. Cheng, Symmetric indefinite matrices: linear system solvers and
modified inertia problems. Ph.D. Thesis, University of Manchester (1998).
https://www.maths.manchester.ac.uk/~higham/links/theses/cheng98.pdf
29. A.J. Chorin, Numerical solution of the Navier–Stokes equations. Math. Comput. 22(104), 745–
762 (1968)
30. P.G. Ciarlet, The Finite Element Method for Elliptic Problems. Studies in Mathematics and Its
Applications (North-Holland, Amsterdam, 1978)
31. G.E. Comi, G. Stefani, A distributional approach to fractional Sobolev spaces and fractional
variation: existence of blow-up. J. Function. Anal. 277(10), 3373–3435 (2019)
32. M. Dauge, Elliptic Boundary Value Problems on Corner Domains. Lecture Notes in Mathe-
matics, vol. 1341 (Springer, Berlin, 1988)
33. M.C. Delfour, J.P. Zolésio, Shapes and Geometries: Analysis, Differential Calculus and Opti-
mization. Advances in Design and Control (Society for Industrial and Applied Mathematics,
Philadelphia, 2001)
34. T. Diop, Résolution itérative de problèmes de contact frottant de grande taille. Ph.D. Thesis,
Université Laval, Canada (2019). https://corpus.ulaval.ca/jspui/handle/20.500.11794/34968
35. G. Duvaut, J.L. Lions, Inequalities in Mechanics and Physics. Grundlehren der mathematis-
chen Wissenschaften, vol. 219 (Springer, Berlin, 1976)
36. S.C. Eisenstat, H.C. Elman, M.H. Schultz, Variational iterative methods for nonsymmetric
systems of linear equations. SIAM J. Numer. Anal. 20(2), 345–357 (1983)
37. A. El maliki, R. Guénette, M. Fortin, An efficient hierarchical preconditioner for quadratic
discretizations of finite element problems. Numer. Linear Algebra Appl. 18(5), 789–803 (2011)
38. A. El maliki, M. Fortin, J. Deteix, A. Fortin, Preconditioned iteration for saddle-point systems
with bound constraints arising in contact problems. Comput. Methods Appl. Mech. Eng. 254,
114–125 (2013)
39. H.C. Elman, Iterative methods for large, sparse, nonsymmetric sys-
tems of linear equations. Ph.D. Thesis, Yale University (1982).
http://www.cvc.yale.edu/publications/techreports/tr229.pdf.
Bibliography 109

40. H.C. Elman, G.H. Golub, Inexact and preconditioned Uzawa algorithms for saddle point
problems. SIAM J. Numer. Anal. 31(6), 1645–1661 (1994)
41. H.C. Elman, D.J. Silvester, A.J. Wathen, Finite Elements and Fast Iterative Solvers: With
Applications in Incompressible Fluid Dynamics. Numerical Mathematics and Scientific Com-
putation (Oxford University Press, Oxford, 2005)
42. A. Fortin, A. Garon, Les Élements Finis de La Théorie à La Pratique. GIREF, Université Laval
(2018)
43. M. Fortin, R. Glowinski, Augmented Lagrangian Methods: Applications to the Numerical
Solution of Boundary-Value Problems. Studies in Mathematics and its Applications, vol. 15
(North-Holland, Amsterdam, 1983)
44. M.J. Gander, Optimized Schwarz methods. SIAM J. Numer. Anal. 44(2), 699–731 (2006)
45. A. Ghai, C. Lu, X. Jiao, A comparison of preconditioned Krylov subspace methods for large-
scale nonsymmetric linear systems. Numer. Linear Algebra Appl. 26(1), e2215 (2019)
46. G.H. Golub, C. Greif, On solving block-structured indefinite linear systems. SIAM J. Sci.
Comput. 24(Part 6), 2076–2092 (2003)
47. G.H. Golub, C.F. Van Loan, Matrix Computations, Johns Hopkins Studies in the Mathematical
Sciences, 3rd edn. (Johns Hopkins University Press, Baltimore, 1996)
48. P.P. Grinevich, M.A. Olshanskii, An iterative method for the stokes-type problem with variable
viscosity. SIAM J. Sci. Comput. 31(5), 3959–3978 (2010)
49. A. Günnel, R. Herzog, E. Sachs, A note on preconditioners and scalar products in Krylov
subspace methods for self-adjoint problems in Hilbert space. Electron. Trans. Numer. Anal.
41, 13–20 (2014)
50. W. Hackbusch, Iterative Solution of Large Sparse Systems of Equations. Applied Mathematical
Sciences, vol. 95, 2nd edn. (Springer, Amsterdam, 2016)
51. R. Herzog, K.M. Soodhalter, A modified implementation of Minres to monitor residual
subvector norms for block systems. SIAM J. Sci. Comput. 39(6), A2645–A2663 (2017)
52. M.R. Hestenes, E. Stiefel, Methods of conjugate gradients for solving linear systems. J. Res.
Natl. Bureau Standards 49(6), 409–436 (1952)
53. G.A. Holzapfel, Nonlinear Solid Mechanics : A Continuum Approach for Engineering (Wiley,
New York, 2000)
54. S. Hüeber, B.I. Wohlmuth, A primal-dual active set strategy for non-linear multibody contact
problems. Comput. Methods Appl. Mech. Eng. 194(27), 3147–3166 (2005)
55. S. Hüeber, G. Stadler, B.I. Wohlmuth, A primal-dual active set algorithm for three-dimensional
contact problems with coulomb friction. SIAM J. Sci. Comput. 30(2), 572–596 (2009)
56. K. Ito, K. Kunisch, Augmented Lagrangian methods for nonsmooth, convex optimization in
Hilbert spaces. Nonlinear Anal. 41(5), 591–616 (2000)
57. K. Ito, K. Kunisch, Optimal control of elliptic variational inequalities. Appl. Math. Optim. Int.
J. Appl. Stoch. 41(3), 343–364 (2000)
58. E.G. Johnson, A.O. Nier, Angular aberrations in sector shaped electromagnetic lenses for
focusing beams of charged particles. Phys. Rev. 91(1), 10–17 (1953)
59. N. Kikuchi, J.T. Oden, Contact Problems in Elasticity: A Study of Variational Inequalities and
Finite Element Methods. SIAM Studies in Applied Mathematics, vol. 8 (SIAM, Philadelphia,
1988)
60. C. Lanczos, An iteration method for the solution of the eigenvalue problem of linear differential
and integral operators. J. Res. Natl. Bureau Standards 45, 255–282 (1950)
61. S. Léger, Méthode lagrangienne actualisée pour des problèmes hyperélastiques
en très grandes déformations. Ph.D. Thesis, Université Laval, Canada (2014).
https://corpus.ulaval.ca/jspui/handle/20.500.11794/25402
62. J.L. Lions, E. Magenes, Non-Homogeneous Boundary Value Problems and Applications.
Grundlehren der mathematischen wissenschaften; bd. 181, vol. 1 (Springer, Berlin, 1972)
63. D.C. Liu, J. Nocedal, On the limited memory BFGS method for large scale optimization. Math.
Program. 45(1–3), 503–528 (1989)
64. D. Loghin, A.J. Wathen, Analysis of preconditioners for saddle-point problems. SIAM J. Sci.
Comput. 25(6), 2029–2049 (2004)
110 Bibliography

65. D.G. Luenberger, The conjugate residual method for constrained minimization problems.
SIAM J. Numer. Anal. 7(3), 390–398 (1970)
66. C.F. Ma, Q.Q. Zheng, The corrected Uzawa method for solving saddle point problems. Numer.
Linear Algebra Appl. 22(4), 717–730 (2015)
67. K.-A. Mardal, R. Winther, Preconditioning discretizations of systems of partial differential
equations. Numer. Linear Algebra Appl. 18(1), 1–40 (2011)
68. L.D. Marini, An inexpensive method for the evaluation of the solution of the lowest order
Raviart-Thomas mixed method. SIAM J. Numer. Anal. 22(3), 493–496 (1985)
69. H.O. MaY, The conjugate gradient method for unilateral problems. Comput. Struct. 22(4),
595–598 (1986)
70. S.F. McCormick, Multigrid Methods. Frontiers in applied mathematics, vol. 3 (Society for
Industrial and Applied Mathematics, Philadelphia, 1987)
71. G.A. Meurant, Computer Solution of Large Linear Systems. Studies in Mathematics and Its
Applications, vol. 28 (North-Holland, Amsterdam, 1999)
72. J. Nocedal, S. Wright, Numerical Optimization. Springer Series in Operations Research and
Financial Engineering, 2nd edn. (Springer, New York, 2006)
73. R.W. Ogden, Non-linear Elastic Deformations. Ellis Horwood Series in Mathematics and Its
Applications (Ellis Horwood, Chichester, 1984)
74. C.C. Paige, M.A. Saunders, Solution of sparse indefinite systems of linear equations. SIAM J.
Numer. Anal. 12(4), 617–629 (1975)
75. J. Pestana, A.J. Wathen, Natural preconditioning and iterative methods for saddle point
systems. SIAM Rev. 57(1), 71–91 (2015)
76. L. Plasman, J. Deteix, D. Yakoubi, A projection scheme for Navier-Stokes with variable
viscosity and natural boundary condition. Int. J. Numer. Methods Fluids 92(12), 1845–1865
(2020)
77. A. Quarteroni, A. Valli, Theory and application of Steklov-Poincaré operators for boundary-
value problems, in Applied and Industrial Mathematics: Venice - 1, 1989, ed. by R. Spigler,
Mathematics and Its Applications (Springer, Dordrecht, 1991)
78. A. Quarteroni, R. Sacco, F. Saleri, Méthodes Numériques: Algorithmes, Analyse et Applica-
tions (Springer, Milano, 2007)
79. R.T. Rockafellar, The multiplier method of Hestenes and Powell applied to convex program-
ming. J. Optim. Theory Appl. 12(6), 555–562 (1973)
80. Y. Saad, A flexible inner-outer preconditioned GMRES algorithm. SIAM J. Sci. Comput. 14(2),
461–469 (1993)
81. Y. Saad, Iterative Methods for Sparse Linear Systems, 2nd edn. (Society for Industrial and
Applied Mathematics, Philadelphia, 2003)
82. B.V. Shah, R.J. Buehler, O. Kempthorne, Some algorithms for minimizing a function of several
variables. J. Soc. Ind. Appl. Math. 12(1), 74–92 (1964)
83. D.J. Silvester, V. Simoncini, An optimal iterative solver for symmetric indefinite systems
stemming from mixed approximation. ACM Trans. Math. Softw. 37(4), (2011)
84. J.C. Simo, T.J.R. Hughes, Computational Inelasticity. Interdisciplinary Applied Mathematics,
vol. 7. (Springer, New York, 1998)
85. R. Temam, Navier–Stokes Equations Theory and Numerical Analysis. Studies in Mathematics
and Its Applications (North-Holland, Amsterdam, 1977)
86. L.N. Trefethen, D. Bau, Numerical Linear Algebra (Society for Industrial and Applied
Mathematics, Philadelphia, 1997)
87. A. van der Sluis, H.A. van der Vorst, The rate of convergence of conjugate gradients. Numer.
Math. 48(5), 543–560 (1986)
88. H.A. van der Vorst, Iterative Krylov Methods for large Linear Systems. Cambridge Monographs
on Applied and Computational Mathematics, vol. 13 (Cambridge University Press, New York,
2003)
89. R. Verfürth, A posteriori error estimation and adaptive mesh-refinement techniques. J. Comput.
Appl. Math. 50(1), 67–83 (1994)
Bibliography 111

90. A.J. Wathen, Realistic eigenvalue bounds for the Galerkin mass matrix. IMA J. Numer. Anal.
7(4), 449–457 (1987)
91. P. Wriggers, Computational Contact Mechanics, 2nd edn. (Springer, Berlin, 2006)
Index

MS , 11 choice of elements, 56
Mooney Rivlin model, 52
neo-hookean model, 52
Active constraints, 24 Ellipticity
Active set stategy, 86 global, 8
Active set strategy, 93 on the kernel, 8
Arrow-Hurwicz-Uzawa, 39, 61 Error estimates, 8
Augmented Lagrangian, 5, 13, 41, 45 Existence, 4
dual problem, 15
iterated penalty, 17
discrete, 15 Factorisation, 33
Fractional order derivatives, 77

Choice of MS , 35
variable coefficients, 57 GCR, 28
Coercivity, 46 GCR solver for the Schur complement, 36
on the kernel, 4, 6, 51 General mixed preconditioner, 34, 38, 56
Condition number, 12 GMRES, 62
Conjugate gradient, 22, 23
contact pressure, 88
Contact problems, 75 Hierarchical basis, 26
Contact status, 92
Convergence
independence of mesh size, 60 Incompressible elasticity, 48
Convex constraints, 25 inequality, 24
Inequality constraints, 24
inf-sup condition, 4
Dirichlet condition, 75
Discrete dual problem, 11
Discrete mixed problem, 7 Kuhn-Tucker conditions, 85, 92
Discrete norm, 82
Discrete scalar product, 90
Dual problem, 5 Linear elasticity, 49

Elasticity Matricial form, 9

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 113
J. Deteix et al., Numerical Methods for Mixed Finite Element Problems,
Lecture Notes in Mathematics 2318, https://doi.org/10.1007/978-3-031-12616-1
114 Index

Matrix form, 80 Reduced h , 84

Metric on Qh , 10 Regularised Lagrangian, 51, 58
Mixed Laplacian, 43 Regular perturbation, 6
Mixed problem, 3 Right preconditioned GCR, 29
Model for discrete contact, 79
Mooney-Rivlin
numerical results, 65 Saddle-Point, 4
Multigrid, 26, 56, 62 Schur complement, 5, 11
approximate solver, 35
Sequential quadratic programming, 89
Navier-Stokes, 68 Simple mixed preconditioner, 38
projection, 69 Sliding contact, 87
Neo-Hookean discretisation, 89
numerical results, 64 incompressible material, 97
Non linear elasticity, 52 numerical example, 93
Sobolev spaces of fractional order, 76
Solver, 33
Obstacle problem, 84 Spectrally equivalent, 12
Operator B, 6 Steklov-Poincaré, 78
Oriented distance, 88 discrete, 80
computation, 82
matrix, 82
Penalty method, 17 Stokes problem, 51
Perturbed problem, 32, 40
Positivity constraints, 84, 91
Preconditioner for the mixed problem, 32 Two constraints, 97
Primal problem, 4 factorisation, 100
Projected gradient, 24, 99 projection method, 102

Raviart-Thomas elements, 45 Uzawa, 37

Rayleigh quotient, 12
LECTURE NOTES IN MATHEMATICS 123
Editors in Chief: J.-M. Morel, B. Teissier;

Editorial Policy
1. Lecture Notes aim to report new developments in all areas of mathematics and their
applications – quickly, informally and at a high level. Mathematical texts analysing new
developments in modelling and numerical simulation are welcome.
Manuscripts should be reasonably self-contained and rounded off. Thus they may, and
often will, present not only results of the author but also related work by other people.
They may be based on specialised lecture courses. Furthermore, the manuscripts should
provide sufficient motivation, examples and applications. This clearly distinguishes
Lecture Notes from journal articles or technical reports which normally are very concise.
Articles intended for a journal but too long to be accepted by most journals, usually do not
have this “lecture notes” character. For similar reasons it is unusual for doctoral theses to
be accepted for the Lecture Notes series, though habilitation theses may be appropriate.

2. Besides monographs, multi-author manuscripts resulting from SUMMER SCHOOLS or

similar INTENSIVE COURSES are welcome, provided their objective was held to present
an active mathematical topic to an audience at the beginning or intermediate graduate level
(a list of participants should be provided).
The resulting manuscript should not be just a collection of course notes, but should require
advance planning and coordination among the main lecturers. The subject matter should
dictate the structure of the book. This structure should be motivated and explained in
a scientific introduction, and the notation, references, index and formulation of results
should be, if possible, unified by the editors. Each contribution should have an abstract
and an introduction referring to the other contributions. In other words, more preparatory
work must go into a multi-authored volume than simply assembling a disparate collection
of papers, communicated at the event.

3. Manuscripts should be submitted either online at www.editorialmanager.com/lnm to

Springer’s mathematics editorial in Heidelberg, or electronically to one of the series edi-
tors. Authors should be aware that incomplete or insufficiently close-to-final manuscripts
almost always result in longer refereeing times and nevertheless unclear referees’ rec-
ommendations, making further refereeing of a final draft necessary. The strict minimum
amount of material that will be considered should include a detailed outline describing
the planned contents of each chapter, a bibliography and several sample chapters. Parallel
submission of a manuscript to another publisher while under consideration for LNM is not
acceptable and can lead to rejection.

4. In general, monographs will be sent out to at least 2 external referees for evaluation.
A final decision to publish can be made only on the basis of the complete manuscript,
however a refereeing process leading to a preliminary decision can be based on a pre-final
or incomplete manuscript.
Volume Editors of multi-author works are expected to arrange for the refereeing, to the
usual scientific standards, of the individual contributions. If the resulting reports can be
 forwarded to the LNM Editorial Board, this is very helpful. If no reports are forwarded
or if other questions remain unclear in respect of homogeneity etc, the series editors may
wish to consult external referees for an overall evaluation of the volume.

5. Manuscripts should in general be submitted in English. Final manuscripts should contain

at least 100 pages of mathematical text and should always include
– a table of contents;
– an informative introduction, with adequate motivation and perhaps some historical
remarks: it should be accessible to a reader not intimately familiar with the topic
treated;
– a subject index: as a rule this is genuinely helpful for the reader.
– For evaluation purposes, manuscripts should be submitted as pdf files.

6. Careful preparation of the manuscripts will help keep production time short besides
ensuring satisfactory appearance of the finished book in print and online. After
acceptance of the manuscript authors will be asked to prepare the final LaTeX source
files (see LaTeX templates online: https://www.springer.com/gb/authors-editors/book-
authors-editors/manuscriptpreparation/5636) plus the corresponding pdf- or zipped ps-
file. The LaTeX source files are essential for producing the full-text online version of
the book, see http://link.springer.com/bookseries/304 for the existing online volumes
of LNM). The technical production of a Lecture Notes volume takes approximately 12
weeks. Additional instructions, if necessary, are available on request from lnm@springer.
com.

7. Authors receive a total of 30 free copies of their volume and free access to their book on
SpringerLink, but no royalties. They are entitled to a discount of 33.3 % on the price of
Springer books purchased for their personal use, if ordering directly from Springer.

8. Commitment to publish is made by a Publishing Agreement; contributing authors of

multiauthor books are requested to sign a Consent to Publish form. Springer-Verlag
registers the copyright for each volume. Authors are free to reuse material contained in
their LNM volumes in later publications: a brief written (or e-mail) request for formal
permission is sufficient.

Addresses:
Professor Jean-Michel Morel, CMLA, École Normale Supérieure de Cachan, France
E-mail: [email protected]

Professor Bernard Teissier, Equipe Géométrie et Dynamique,

Institut de Mathématiques de Jussieu – Paris Rive Gauche, Paris, France
E-mail: [email protected]

Springer: Ute McCrory, Mathematics, Heidelberg, Germany,

E-mail: [email protected]